Functional

Statistics
Outliers Detection and
Quality Control
Edited by
Javier Martínez Torres
Printed Edition of the Special Issue Published in Mathematics

www.mdpi.com/journal/mathematics
Functional Statistics
Functional Statistics
Outliers Detection and Quality Control

Editor
Javier Martı́nez Torres

MDPI • Basel • Beijing • Wuhan • Barcelona • Belgrade • Manchester • Tokyo • Cluj • Tianjin
Editor
Javier Martı́nez Torres
Escuela Superior de Ingenierı́a y Tecnologı́a,
Universidad Internacional de La Rioja
Spain

Editorial Office
MDPI
St. Alban-Anlage 66
4052 Basel, Switzerland

This is a reprint of articles from the Special Issue published online in the open access journal
Mathematics (ISSN 2227-7390) (available at: https://www.mdpi.com/journal/mathematics/special
issues/Functional Statistics Outliers Detection Quality Control).

For citation purposes, cite each article independently as indicated on the article page online and as
indicated below:

LastName, A.A.; LastName, B.B.; LastName, C.C. Article Title. Journal Name Year, Volume Number,
Page Range.

ISBN 978-3-03943-963-8 (Hbk)

ISBN 978-3-03943-964-5 (PDF)


c 2020 by the authors. Articles in this book are Open Access and distributed under the Creative
Commons Attribution (CC BY) license, which allows users to download, copy and build upon
published articles, as long as the author and publisher are properly credited, which ensures maximum
dissemination and a wider impact of our publications.
The book as a whole is distributed by MDPI under the terms and conditions of the Creative Commons
license CC BY-NC-ND.
Contents

About the Editor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii

Preface to ”Functional Statistics” . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix

Manuel Oviedo-de La Fuente, Celestino Ordóñez and Javier Roca-Pardiñas

Functional Location-Scale Model to Forecast Bivariate Pollution Episodes
Reprinted from: Mathematics 2020, 8, 941, doi:10.3390/math8060941 . . . . . . . . . . . . . . . . . 1

Ofélia Anjos, Miguel Martı́nez Comesaña, Ilda Caldeira, Soraia Inês Pedro, Pablo Eguı́a
Oller and Sara Canas
Application of Functional Data Analysis and FTIR-ATR Spectroscopy to Discriminate Wine
Spirits Ageing Technologies
Reprinted from: Mathematics 2020, 8, 896, doi:10.3390/math8060896 . . . . . . . . . . . . . . . . . 13

Ishaq Adeyanju Raji, Muhammad Hisyam Lee, Muhammad Riaz, Mu’azu Ramat Abujiya
and Nasir Abbas
Outliers Detection Models in Shewhart Control Charts; An Application in Photolithography:
A Semiconductor Manufacturing Industry
Reprinted from: Mathematics 2020, 8, 857, doi:10.3390/math8050857 . . . . . . . . . . . . . . . . . 35

Miguel Martı́nez Comesaña, Sandra Martı́nez Mariño, Pablo Eguı́a Oller, Enrique Granada
Álvarez and Aitor Erkoreka González
A Functional Data Analysis for Assessing the Impact of a Retrofitting in the Energy Performance
of a Building
Reprinted from: Mathematics 2020, 8, 547, doi:10.3390/math8040547 . . . . . . . . . . . . . . . . . 53

Elı́as Berriochoa, Alicia Cachafeiro, Alberto Castejón and José Manuel Garcı́a-Amor
Classical Lagrange Interpolation Based on General Nodal Systems at Perturbed Roots of Unity
Reprinted from: Mathematics 2020, 8, 498, doi:10.3390/math8040498 . . . . . . . . . . . . . . . . . 73

Javier Martı́nez Torres, Jorge Pastor Pérez, Joaquı́n Sancho Val, Aonghus McNabola, Miguel
Martı́nez Comesaña and John Gallagher
A Functional Data Analysis Approach for the Detection of Air Pollution Episodes and Outliers:
A Case Study in Dublin, Ireland
Reprinted from: Mathematics 2020, 8, 225, doi:10.3390/math8020225 . . . . . . . . . . . . . . . . . 91

Miguel Flores, Salvador Naya, Rubén Fernández-Casal, Sonia Zaragoza, Paula Raña, Javier
Tarrı́o-Saavedra
Constructing a Control Chart Using Functional Data
Reprinted from: Mathematics 2020, 8, 58, doi:10.3390/math8010058 . . . . . . . . . . . . . . . . . . 111

v
About the Editor
Javier Martı́nez Torres has a PhD in Mathematics and Engineering from the University of Vigo.
He is currently an Assistant Professor at the University of Vigo and has participated in more than
20 research projects as a principal investigator. He has published more than 50 papers in JCR indexed
journals and participated in more than 25 international conferences.

vii
Preface to ”Functional Statistics”
Why use statistics and functional analysis? This is one of the most frequently asked questions
with the rise of these types of techniques, especially when they are applied in areas other than
mathematics, engineering, economics, etc. The reason is very straightforward: in this era of large
amounts of data, most of the data available can be seen from a functional perspective, and thus,
problems can be addressed through functional analysis and/or functional statistics. This book aims
to build bridges between these two disciplines, which are sometimes far apart but are undoubtedly
much closer than one imagines.

Javier Martı́nez Torres

Editor

ix
 mathematics
Article
Functional Location-Scale Model to Forecast Bivariate
Pollution Episodes
Manuel Oviedo-de La Fuente 1 , Celestino Ordóñez 2,∗ and Javier Roca-Pardiñas 3
1 Department of Statistics, Mathematical Analysis and Optimization, Universidad de Santiago de Compostela,
15782 Santiago de Compostela, Spain; [email protected]
2 Department of Mining Exploitation and Propsecting, Universidad de Oviedo, Escuela Politécnica de Mieres,
33600 Mieres, Spain
3 Department of Statistics and Operation Research, Universidad de Vigo, 36310 Vigo, Spain; [email protected]
* Correspondence: [email protected]

Received: 7 May 2020; Accepted: 29 May 2020; Published: 8 June 2020

Abstract: Predicting anomalous emission of pollutants into the atmosphere well in advance is crucial
for industries emitting such elements, since it allows them to take corrective measures aimed to avoid
such emissions and their consequences. In this work, we propose a functional location-scale model to
predict in advance pollution episodes where two pollutants are involved. Functional generalized
additive models (FGAMs) are used to estimate the means and variances of the model, as well as
the correlation between both pollutants. The method not only forecasts the concentrations of both
pollutants, it also estimates an uncertainty region where the concentrations of both pollutants should
be located, given a specific level of uncertainty. The performance of the model was evaluated using
real data of SO2 and NOx emissions from a coal-fired power station, obtaining good results.

Keywords: pollution episodes; functional data; bivariate analysis; uncertainty region; generalized
additive models

1. Introduction
Forecasting air quality and concentrations of pollutants in the atmosphere by means of statistical
methods is an active area of research given the transcendence of the problem and the difficulty to
find optimal solutions using deterministic mathematical models. Among the different methods that
can be found in the literature to tackle this problem, models for time series analysis such as the
integrated autoregressive moving average—ARIMA [1–3], multivariate regression [4–7], generalized
linear or additive models (GAM) [8–11] and artificial neural networks (ANN) [12–19] are the most
extended. Due to the increased access to continuous data over time, functional data analysis [20,21]
was also proposed for air quality forecasting and outlier detection [22–24]. Parametric [25,26]
and nonparametric [27–29] functional regression methods were tested. A functional framework
allows considering the inherent correlation between observations, instead of considering them as
independent realizations of an underlying stochastic process. Some functional approaches add related
meteorological variables to the models [30–34], which can improve the result of the predictions and
help to understand the process underlying the evolution of the pollutants.
Most of the documents in the literature propose solutions to predict the concentration of each
pollutant individually, being much scarcer those focused on predicting more than one pollutant at
a time. Vector autoregressive moving average (VARMA) [35,36] and vector autoregressive integrated
moving average (VARIMA) [37] models were applied to reach this objective. In this work, we proposed
a method for the simultaneous forecasting of pollution episodes when two pollutants, i.e., SO2 and
NOx , are involved. Apart from transport, one of the main sources of these pollutants is public electricity
and heating. Their negative effects on human health are well known, and goes for mild (i.e., eyes

Mathematics 2020, 8, 941; doi:10.3390/math8060941 1 www.mdpi.com/journal/mathematics

Mathematics 2020, 8, 941

irritated, nose or headache) to severe (i.e., lung damage or reduced oxygenation of tissues). They also
have negative effects on animals and plants, as well as in other substances, such as water and soils. In
addition, NOx is a precursor of the tropospheric ozone. High levels of ozone contributes to climate
change, cause adverse impacts on health and can damage vegetation.
Pollution episodes (incidents) are abnormally large emissions of one or more pollutants in short
periods of time. Although the improvement of the chemical processes and particle filter systems have
significantly reduced the amount and intensity of the pollution episodes, they are still of particular
interest for the industries, as they may be subject to sanctions, or for other reasons, such as public
health deterioration or industry discredit. Therefore, pollution industries, such as coal-fired power
plants, are very interested in determining in advance when these episodes of excessive contamination
might occur. Specifically, this is the purpose of our work: forecasting pollution episodes of SO2 and
NOx early enough to allow corrective measures to be taken. Our approach uses a location-scale
model [11,38,39] that treats the predictors, the concentrations of both pollutants over time, as functions,
while the response is a scalar, the concentration of the pollutants some time in advance. The novelty of
our approach is the combination of a biviariate location-scale model with functional additive models.
This method combines the simplicity of the location and scale models with the capacity of functional
data analysis to deal with data in the form of functions.
The document is structured as follows: In Section 2 we show the mathematical model proposed to
solve the problem under analysis and the algorithm used to estimate a solution from the data. Section 3
is devoted to test the validity of the model using real data. Finally, a discussion of the results and the
main conclusions of our work are exposed in Section 4.

2. Methodology

2.1. Mathematical Model

 
Let {Xi , Yi }in=1 be a set of observations of a stochastic process, X = X 1 (t), . . . , X p (t) ,
where X j (t) ∈ L2 [0, T ] , j = 1, . . . , p, are predictor covariates and Y = (Y1 , Y2 ), with Yj ∈ R, a response
variable. In this context, the following bivariate location-scale model [40,41] is assumed
     
Y1 μ1 (X) ε1
= + Σ1/2 (X) (1)
Y2 μ2 (X) ε2

where Σ1/2 (X) represents the Cholesky decomposition of the variance-covariance matrix Σ(X)
 
σ12 (X) σ12 (X)
Σ (X) = (2)
σ12 (X) σ22 (X)
 T
so that Var(Y|X) = Σ(X) = Σ1/2 (X) Σ1/2 (X) . To guarantee the model identification in (1), the
bivariate residuals (ε 1 , ε 2 ) are assumed to be independent of the covariates, with zero mean, unit
variance, and zero correlation. Despite we do not assume any distribution for the error term, within
the framework of functional data analysis this work might be addressed under the assumption of other
structures for error distribution: generalized Gauss-Laplace distribution that relax the constrictive
assumption of the normal distribution errors [42], generalized linear mixed models (GLMMs) [38]
to estimate random effects and dependent (temporal or spatial) errors, and generalized additive
models for location, scale and shape (GAMLSS) [43] to model the dynamically variable distribution,
considering skewness and kurtosis.
We define the unconditionally probabilistic region for the errors (ε 1 , ε 2 ) as

ε τ (k) = {(ε 1 , ε 2 ) ∈ R2 | f (ε 1 , ε 2 ) ≥ k}

2
Mathematics 2020, 8, 941

f being the density function of the bivariate residuals (ε 1 , ε 2 ) and k the τ −quantile of f (ε 1 , ε 2 ).
Then, for a given X, we define a conditional τ th - uncertainty region for (Y1 , Y2 ) containing τ% of
the observations as:
 
μ1 (X)
R τ (X) = + Σ1/2 (X)ε τ (3)
μ2 (X)

2.2. Estimation Algorithm

To implement an algorithm that allows applying the mathematical model exposed in the
previous section, we propose using a functional additive models to estimate the
 means, variances
 and
p
covariances in (1). Given a sample of size n, {Xi , (Yi1 , Yi2 )}in=1 , where Xi = Xi1 (t), . . . , Xi (t) , the
steps of the proposed estimation algorithm are the following:
Step 1: Perform a decomposition of each covariate X j (t) in basis functions of the form X j (t) ≈
j
∑k=1 ξ k φk (t), where φk (k = 1, . . . , K) are K basis functions (i.e., B-splines, wavelets), and ξ il are either
K

the coefficients of an expansion in fixed basis or the principal component scores of the Karhunen-Loève
expansion [44,45]. As a result, we obtain the transformed covariates
     
p p
X̃i = 1
ξ i1 1
, . . . , ξ iK 2
; ξ i1 2
, . . . , ξ iK ; . . . ; ξ i1 , . . . , ξ iK i = 1, . . . , n

Step 2: For r = 1, 2, fit an additive model to the sample {X̃i , Yi1 , Yi2 }in=1 and obtain an estimation
of the means
p K
∑∑
j j
μ̂r (Xi ) = αr + fˆrk (ξ ik ) (4)
j =1 k =1

and then estimate σr2 (X) from the sample {X̃i , (Yir − μ̂r (Xi ))2 }in=1 as

p K
∑ ∑ ĝrk (ξ ik )
j j
σ̂r2 (Xi ) = exp β̂ r + (5)
j =1 k =1

Then, compute the correlation ρ(X), which is related to the covariance by σ12 (X) =
σ1 (X)σ2 (X)ρ(X), using the sample {Xi , δ̂i }in=1 , as follows:

p K
∑ ∑ m̂k (ξ ik )
j j
ρ̂(Xi ) = tanh γ̂ +
j =1 k =1

being   
Yi1 − μ̂1 (Xi ) Yi2 − μ̂2 (Xi )
δ̂i =
σ̂1 (Xi )σ̂2 (Xi )
j j j
where f rk , grk and mk are smooth and unknown functions, αr , β r and γ are coefficients, p the number of
predictors (covariates), and K the number of basis. Please note that the link functions Hσ (·) = exp(·)
and Hρ (·) = tanh(·) used in the variance and correlation structures, respectively, ensure that the
restrictions on the parameter spaces (σr2 (X) ≥ 0 and 0 ≤ ρ(X) ≤ 1) are maintained. Moreover,
in order to guarantee the identification of the model, we assume that all the means of functions f j , g j
and m j are zero.
Step 3: Compute the standardized residuals
   
ε̂ i1 Yi1 − μ̂1 (Xi )
= Σ̂−1/2 (Xi ) i = 1, . . . , n (6)
ε̂ i2 Yi2 − μ̂2 (Xi )

where

3
Mathematics 2020, 8, 941

 
σ̂12 (Xi ) σ̂12 (Xi )
Σ̂(Xi ) = (7)
σ̂12 (Xi ) σ̂22 (Xi )

with σ̂12 (Xi ) = σ̂1 (Xi )σ̂2 (Xi )ρ̂(Xi ).

Step 4: Obtain the kernel estimation of the bivariate density fˆ(ε 1 , ε 2 ) given by
n  
1 ε 1 − ε̂ i1
fˆ((ε 1 , ε 2 ), H) =
n ∑ KH ε 2 − ε̂ i2
(8)
i =1

where K (·) is the kernel which is a symmetric probability density function and H is a 2 × 2 positive
definite matrix. Then, obtain the τ th unconditional bivariate uncertainty region on the residual scale as

ε̂ τ = {(ε 1 , ε 2 )) ∈ R2 | fˆ(ε 1 , ε 2 )) ≥ k̂} (9)

k̂ being the empirical τ quantile of the values fˆ(ε 11 , ε 12 ), . . . , fˆ(ε n1 , ε n2 ). Finally, for a given X,
the conditional bivariate uncertainty region Rτ (X) is estimated according to (3)
 
μ̂1 (X) 1/2
R̂τ (X) = + Σ̂ (X)ε̂ τ (10)
μ̂2 (X)

3. Case Study: Joint Forecasting of (SO2 , NOx ) Pollution Episodes

The mathematical model exposed in the previous section was applied to the forecasting of
pollution episodes registered at a coal-fired power station located in the northwest of Spain. SO2 and
NOx are two of the main air pollutants generated by combustion processes, and both have harmful
effects on human health. Moreover, it was proven that both pollutants are correlated [46], which is
consistent with the model in (1). Fortunately, pollution episodes are not very frequent and the trend is
that they will become scarcer as technology advances.
Let t0 be the present time measured each five minutes, and SO(t0 ) and NO(t0 ) the concentrations
obtained respectively by the series of bi-hourly SO2 and NOx means at instant t0 . Being th the
prediction horizon time, the interest is to predict

(Y1 , Y2 ) = (SO(t0 + th ), NO(t0 + th ))

and provide an uncertainty region for these estimations given a specific value of τ, using the
predictive covariates
   
X = X 1 (t), X 2 (t), X 3 (t), X 4 (t) = SO(t), NO(t), SO (t), NO (t) with t ∈ [t0 − tlag , t0 ]

where (NO (t), SO (t)) represents the first derivatives of the functions that approximate the
concentrations of both pollutants. These derivatives are obtained from the functional representation of
the discrete data, according to Step 1 of the estimation algorithm. Please note that tlag represents the
lagged time used in the predictors. In particular„ we are interested in predicting an hour in advance,
according to the requirements of current Spanish legislation and, therefore, we will consider th = 12
(60 min) from now on.
Most of the time, these concentrations times series are low, close to zero, and in order to obtain
a reasonably large number of pollution incidents, we took as our sample a historical matrix {(Xi , Yi )}iN=1
with pollution data of approximately 12 years, which includes a considerable number of pollution
episodes (see [9] for a detailed description of the historical matrix construction). In summary, in
the historical matrix not all the data are used, but only part of them, following a quantile-weighted
criterion. This means that the larger the concentration, the greater the number of observations of
that concentration in the sample. Figure 1 shows a sample of the historical matrix, on top are the
curves of both pollutants (NO(t), SO(t)) measured in tlag = 20 discretization points and evaluated in

4
Mathematics 2020, 8, 941

5 B-spline basis functions of order p = 4. On the bottom, the first derivative of the B-spline curves
(NO (t), SO (t)) with order p − 1 are represented.

Figure 1. Five curves of the historical matrix for pollutants (NO(t), SO(t)) and their corresponding
first derivatives (NO (t), SO (t)), observed in a period of time tlag = 20.

In this paper, we tested four models using as predictors different combinations of the covariates
 
X 1 (t), X 2 (t), X 3 (t), X 4 (t) that include the concentrations of both pollutants and their first derivatives.
In particular, we will consider models given by:
      
SO(t0 + 12) μ1 ( X1 ) σ12 (X1 ) σ12 (X1 , X2 ) ε1
= + (11)
NO(t0 + 12) μ2 ( X2 ) σ12 (X1 , X2 ) σ22 (X2 ) ε2

The four considered models, M1 , M2 , M3 and M4 , are configured in Table 1 where the cross X
indicates the covariates included in each model.

Table 1. Selected models from equation in (11) . Cross X indicates the covariates included in each of
the four considered models. The derivatives of the functions are indicated with a single quote.

X1 X2
Model SO(t ) NO(t ) SO (t ) NO (t ) SO(t ) NO(t ) SO (t ) NO (t )
M1 X X
M2 X X X X
M3 X X X X
M4 X X X X X X X

To validate and compare the four proposed models, we randomly select from the full historical
ntrain N
matrix a training set M I = (XiI , YiI ) and a test set M I I = (XiI I , YiI I ) .
i =1 i =ntrain +1
The estimates μ̂1 , μ̂2 , Σ̂ were obtained from the samples in the first matrix M tI . The bivariate
uncertainty regions for the values of the covariates on the second matrix M I I were obtained using (3).
The estimated coverage τ̂ is given by

N
1
τ̂ =
ntest ∑ I {YiI I ∈ R̂τ (XiI I )}; ntest = N − ntrain (12)
i =ntrain +1

The performance of the proposed predictors was evaluated in two pollution incidents. A bivariate
representation of these episodes is shown in Figure 2. The orientation of the points shows a clear
correlation between both pollutants although the range of concentrations is quite different.

5
Mathematics 2020, 8, 941

Figure 2. Observed and forecasted concentrations of SO2 and NOx for two pollution episodes.

The nominal and the estimated coverages for different time lags and training sample sizes are
shown in Table 2. The coverages correspond to the bivariate solution, and were obtained for ntest
consecutive observations that might or might not correspond to pollution incidents.

Table 2. Nominal τ and estimated τ̂ coverages for each of the four models under study. Two time
lags, tlag = 10, tlag = 20, two sizes of the training sample ntrain = 10, 000 and ntrain = 4900, and two
numbers of principal components, K = 3 and K = 5, were considered. Results correspond to the test sample.

τ̂
τ tlag n•train K M1 M2 M3 M4
3 0.43 0.47 0.45 0.51
10 20
5 0.42 0.48 0.47 0.49
3 0.51 0.52 0.52 0.52
10 49
5 0.51 0.50 0.50 0.50
0.50
3 0.45 0.49 0.44 0.49
20 20
5 0.48 0.46 0.43 0.46
3 0.50 0.54 0.51 0.50
20 49
5 0.49 0.49 0.49 0.48
3 0.70 0.73 0.72 0.75
10 20
5 0.69 0.74 0.73 0.74
3 0.76 0.78 0.78 0.78
10 49
5 0.77 0.76 0.76 0.76
0.75
3 0.70 0.72 0.70 0.72
20 20
5 0.70 0.71 0.69 0.69
3 0.77 0.78 0.75 0.73
20 49
5 0.75 0.74 0.72 0.72
3 0.88 0.87 0.87 0.89
10 20
5 0.86 0.88 0.87 0.88
3 0.91 0.90 0.90 0.90
10 49
5 0.90 0.90 0.90 0.90
0.90
3 0.87 0.87 0.85 0.86
20 20
5 0.87 0.86 0.86 0.84
3 0.90 0.89 0.90 0.86
20 49
5 0.87 0.87 0.88 0.85
3 0.93 0.93 0.93 0.93
10 20
5 0.93 0.93 0.93 0.93
3 0.96 0.93 0.93 0.93
10 49
5 0.95 0.94 0.94 0.94
0.95
3 0.93 0.93 0.92 0.92
20 20
5 0.92 0.92 0.93 0.90
3 0.95 0.94 0.95 0.92
20 49
5 0.92 0.92 0.93 0.90

6
Mathematics 2020, 8, 941

RMSE values for each model are shown in Table 3 , considering an expansion of the functions in
three or five principal components, and lags tlag = 10 and tlag = 20. Please note that this table makes
reference to the marginal distributions, and that the range of concentrations for each pollutant is very
different, therefore the RMSE values are also different.

Table 3. RMSE values for two pollution episodes and the four models tested, considering curves with
two different time lags, size of the training samples and number of principal components.

Episode 1 Episode 2
Repsonse tlag n•train K M1 M2 M3 M4 M1 M2 M3 M4
3 20.7 25.2 20.0 23.0 1.9 1.7 1.4 1.4
20
5 19.6 18.6 20.6 16.4 0.9 0.8 0.8 0.6
10
3 20.9 24.7 18.2 23.8 1.8 1.9 1.4 1.4
49
NOx 5 19.2 17.5 19.3 16.7 0.9 0.9 0.8 0.7
3 34.0 24.7 19.9 19.6 3.3 5.5 1.7 3.1
100
5 40.9 23.3 46.5 29.2 1.5 2.5 1.0 1.8
20
3 30.2 18.2 18.7 20.4 3.4 5.3 1.7 2.8
49
5 36.2 27.6 39.8 35.6 1.5 2.5 1.0 1.8
3 505.0 841.0 407.5 837.3 544.8 531.9 419.8 419.1
100
5 914.9 868.6 669.4 516.3 215.6 230.3 184.5 199.1
10
3 686.0 685.3 515.8 518.3 481.4 484.4 338.0 361.1
49
SO2 5 991.1 925.5 846.3 682.9 199.9 214.7 170.6 179.9
3 1463.4 2172.7 825.4 1199.9 1154.5 1133.1 709.5 659.0
100
5 1470.6 2531.2 1002.9 1428.5 525.5 482.1 352.0 341.3
20
3 1.458.7 2485.4 768.4 698.4 1162.6 1125.8 644.8 628.7
49
5 1787.6 2811.0 1111.2 951.5 548.1 492.0 352.7 359.3
Note: ntrain = 100 · n•train .

For the two episodes analyzed, Figure 3 shows the observed and the predicted values as well as
the quantile for τ̂ = 0.95, calculated for the test sample. The results correspond to curves observed in
ten points (tlag = 10) and represented in a basis expansion in three functional principal components.
These univariate confidence intervals were respectively constructed from (11) as μ1 (X1 ) + σ1 (X1 )ε0.95
1
and μ2 (X2 ) + σ2 (X2 )ε0.95
2 , ε1
0.95 and ε0.95 being the 0.95 quantile of the distributions of errors ε and
1 1
ε 2 , respectively.

Figure 3. Observations (solid black line), mean (solid gray line) and 0.95th quantile estimations
(discontinuous line) for both pollutants, SO2 and NOx and for two pollution episodes.

7
Mathematics 2020, 8, 941

Table 4 shows the maximum consumed memory and the runtime (in seconds) for the four models
tested and two different dimensions of the submatrices are executed in a Intel Core i7-2600K with 16
GB of RAM.

Table 4. Maximum memory consumption (MB) and computation time (seconds) for the four models
tested, Mi , following different strategies concerning the time lag, tlag , the size of the training sample,
ntrain , and the number of basis functions, K.

Memory (MB) Runtime (seconds)

tlag ntrain K M1 M2 M3 M4 M1 M2 M3 M4
3 652.70 1083.03 1319.57 2266.34 17.98 29.66 35.99 61.62
20
5 1090.78 1855.39 2299.14 4075.88 35.55 55.29 63.15 124.7
10
3 329.94 548.58 669.42 1153.38 10.53 17.42 21.38 37.01
49
5 552.74 942.68 1171.67 2089.53 19.79 31.28 46.62 88.20
3 653.34 1084.15 1320.66 2268.16 18.11 29.63 35.97 61.21
20
5 1091.90 1857.26 2300.99 4078.95 32.14 49.51 72.45 124.81
20
3 330.10 548.84 669.69 1153.82 10.51 17.56 21.49 37.82
49
5 553.01 943.13 1172.20 2090.36 17.80 36.44 39.03 76.11
Note: ntrain = 100 · n•train .

4. Discussion
We begin the discussion of the results analyzing Table 2 that show the estimated coverage for
the bivariate prediction depending on the time lag, the size of the training sample, the number of
principal components and the model used. It can be appreciated that the estimated coverages are
generally lower than the theoretical coverages, although very close. Therefore, the mathematical
models proposed show a good performance although there is a trend to underestimate the observed
values. This effect can also be appreciated in Figure 3, where the mean tends to be under the observed
values. Then, in order to be on the safe wide, it would be preferable to use the quantile τ = 0.95,
which provides greater guarantee of predicting the highest values of the pollution episodes. Regarding
the rest of the parameters, it is not possible to establish a combination of them that provides the best
results. However, in general they were obtained for the lowest training size, ntrain = 49, 000, and for
models that includes one or two derivatives (models M3 and M4 ).
The prediction errors are shown in Table 1, where the best results (minimum RMSE) are marked
in bold. As can be seen, they correspond to model M4 for NOx and model M3 for SO2 . In both cases,
these models incorporate the derivatives of the original functions. Accordingly, we conclude that the
derivatives contribute positively to improve the results, which reinforces the role of the functional
approach. However, there is an asymmetry between both pollutants: using the concentrations of
SO2 and their derivatives improves the results for NOx , but using the concentrations of NOx and
their derivatives is not an advantage in the estimation of SO2 . When SO2 and NOx concentrations of
both episodes are plotted against time (Figure 4), a slight advance can be seen on the first pollutant
compared to the second, which would explain this asymmetry.
With respect to the time lag, the minimum RMSE values were obtained for the shorter period
of time tlag = 10, so it seems that using 20 observations to predict one our in advance introduced
noise into the model instead of adding useful information. This result is in agreement with those
obtained for the same data in previous studies of some of the authors, which indicated that only a few
observations close to the time of prediction contribute to that prediction. Talking about the size of
the training sample, simplifying the original data by removing small values of the concentrations
improves the results in most of the cases, so this would be the advisable option.
When the effect of the number of principal components used as basis functions is analyzed, using
K = 5 is always favorable for episode 2, for both SO2 and NOx , but not for episode 1, for which the
trend is opposite.

8
Mathematics 2020, 8, 941

Figure 4. Example of a pollution incident showing SO2 and NOx concentrations versus time. Notice
that there is an advance in the first pollutant compared to the second.

Although they are not shown in the article, so as not to overstretch it, a comparison of the
estimated coverages using 3 or 5 principal components, or 5 B-splines basis functions, tell us that
there are not substantial differences among them, so it seems that one or other base functions can be
used interchangeably.
Finally, regarding memory consumption and runtime for model training, it is evident, from Table 4
that more complex models consume more resources and requires more computing time. For fixed
values of the time lag (tlag ) and the size of the training sample (ntrain ), model M4 is between 3 and
7 times more expensive than model M1 in terms of memory consumption and runtime. Using time
lags tlag = 10 or tlag = 50 has no effect in terms of computation requirements; and employing 5
principal components instead of 3 principal components implies an approximately double memory
consumption and runtime.
To conclude, it is possible to establish that the functional location-scale model proposed were quite
a good approach (in terms of coverage and prediction error) to forecast bivariate pollution episodes
one hour in advance, as it is required by the Spanish legislation. The best results were obtained when
the derivatives of functions adjusted to the observed data are included in the model, when the raw data
are filtered and when the shorter period of time is used for the prediction. The size of the training data
and the type and number of basis functions are, instead, parameters on which definitive conclusions
could not be drawn.

Author Contributions: Conceptualization, J.R.-P. and C.O.; methodology, J.R.-P. and M.O.-d.L.F.; software, J.R.-P.
and M.O.-d.L.F.; validation, J.R.-P., C.O. and M.O.-d.L.F.; writing—original draft preparation, J.R.-P. and C.O.;
writing—review and editing, J.R.-P., C.O. and M.O.-d.L.F.; supervision, C.O. All authors have read and agreed to
the published version of the manuscript.

9
Mathematics 2020, 8, 941

Acknowledgments: The authors acknowledge financial support from: (1) UO-Proyecto Uni-Ovi
(PAPI-18-GR-2014-0014), (2) Project MTM2016-76969-P from Ministerio de Economía y Competitividad—Agencia
Estatal de Investigación and European Regional Development Fund (ERDF) and IAP network StUDyS from
Belgian Science Policy, (3) Nuevos avances metodológicos y computacionales en estadística no-paramétrica y
semiparamétrica—Ministerio de Ciencia e Investigación (MTM2017-89422-P).
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Siew, L.Y.; Ching, L.Y.; Wee, P.M.J. ARIMA and integrated ARFIMA models for forecasting air pollution
index in Shah Alam, Selangor. Malay. J. Anal. Sci. 2008, 12, 257–263.
2. Ibrahim, M.Z.; Roziah, Z.; Marzuki, I.; Muhd S.L. Forecasting and Time Series Analysis of Air Pollutants in
Several Area of Malaysia. Am. J. Enverion. Sci. 2009, 5, 625–632. [CrossRef]
3. Abhilash, M.S.K.; Thakur, A.; Gupta, D.; Sreevidya, B. Time Series Analysis of Air Pollution in Bengaluru
Using ARIMA Model. In Ambient Communications and Computer Systems; Advances in Intelligent Systems
and Computing; Perez, G., Tiwari, S., Trivedi, M., Mishra, K., Eds.; Springer: Singapore, 2018.
4. Liu, P.W.G. Simulation of the daily average PM10 concentrations at Ta-Liao with Box-Jenkins time series
models and multivariate analysis. Atmos. Environ. 2009, 43, 2104–2113. [CrossRef]
5. Nazif, A.; Mohammed, N.I.; Malakahmad, A.; Abualqumboz, M.S. Regression and multivariate models for
predicting particulate matter concentration level. Environ. Sci. Pollut. Res. Int. 2018, 25, 283–289. [CrossRef]
6. Zhao, R.; Gu, X.; Xue, B.; Zhang, J.; Ren, W. Short period PM2.5 prediction based on multivariate linear
regression model. PLoS ONE 2018, 13, e0201011. [CrossRef]
7. Ng, K.Y.; Awang, N. Multiple linear regression and regression with time series error models in forecasting
PM10 concentrations in Peninsular Malaysia. Environ. Monit. Assess. 2018, 190, 63. [CrossRef]
8. Roca-Pardiñas, J.; Gonzàlez Manteiga, W.; Febrero-Bande, M.; Prada-Sànchez, J.M.; Cadarso-Suàrez, C.
Predicting binary time series of SO2 using generalized additive models with unknown link function.
Environmetrics 2004, 15, 729–742. [CrossRef]
9. Martínez-Silva, I.; Roca-Pardiñas, J.; Ordóñez, C. Forecasting SO2 pollution incidents by means of quantile
curves based on additive models. Environmetrics 2016, 27, 147–157. [CrossRef]
10. Garcia, J.M.; Teodoro F.; Cerdeira, R.; Coelho, L.M.R.; Prashant, K.; Carvalho, M.G. Developing
a methodology to predict PM10 concentrations in urban areas using generalized linear models.
Environ. Technol. 2016, 37, 2316–2325. [CrossRef]
11. Roca-Pardiñas, J.; Ordóñez, C. Predicting pollution incidents through semiparametric quantile regression
models. Stoch. Environ. Res. Risk Assess. 2019, 33, 673–685. [CrossRef]
12. Azid, I.A.; Ripin, Z.M.; Aris, M.S.; Ahmad, A.L.; Seetharamu, K.N.; Yusoff, R.M. Predicting combined-cycle
natural gas power plant emissions by using artificial neural networks. In Proceedings of the 2000
TENCON Proceedings, Intelligent Systems and Technologies for the New Millennium (Cat. No.00CH37119),
Kuala Lumpur, Malaysia, 24–27 September 2000; Volume 3, pp. 512–517.
13. Perez, P.; Trier, A.; Reyes, J. Prediction of PM2.5 Concentrations Several Hours in Advance Using Neural
Networks in Santiago, Chile. Atmos. Environ. 2000, 34, 1189–1196. [CrossRef]
14. Ferretti, G.; Piroddi, L. Estimation of NOx Emissions in Thermal Power Plants Using Neural Networks.
J. Eng. Gas Turbines Power 2001, 132, 465–471. [CrossRef]
15. Siwek, K.; Osowski, S. Improving the accuracy of prediction of PM10 pollution by the wavelet transformation
and an ensemble of neural predictors. Eng. Appl. Artif. Intell. 2012, 25, 1246–1258. [CrossRef]
16. Muñoz, E.; Martín, M.L.; Turias, I.J.; Jimenez-Come, M.J.; Trujillo, F.J. Prediction of PM10 and SO2
exceedances to control air pollution in the Bay of Algeciras, Spain. Stoch. Environ. Res. Risk Assess.
2014, 28, 1409–1420. [CrossRef]
17. He, H.D.; Lu, W.Z.; Xue, Y. Prediction of particulate matters at urban intersection by using multilayer
perceptron model based on principal components. Stoch Environ. Res. Risk. Assess. 2015, 29, 2107–2114.
[CrossRef]
18. Antanasijević, D.; Pocajt, V.; Perić-Grujić, A.; Ristić, M. Multiple-input–multiple-output general regression
neural networks model for the simultaneous estimation of traffic-related air pollutants. Atmos. Pollut. Res.
2018, 9, 388–397. [CrossRef]

10
Mathematics 2020, 8, 941

19. Gilson, M.; Dahmen, D.; Moreno-Bote, R.; Insabato, A.; Helias M. The covariance perceptron: A new
paradigm for classification and processing of time series in recurrent neuronal networks. BioRxiv 2019.
[CrossRef]
20. Ramsay, J.O.; Silverman, B.W. Applied Functional Data Analysis: Methods and Case Studies; Springer: New York,
USA, 2002.
21. Ferraty, F.; Vieu, P. Nonparametric Functional Data Analysis; Springer: New York, USA, 2006.
22. Febrero-Bande, M.; Galeano, P.; González-Manteiga, W. Outlier detection in functional data by depth
measures with application to identify abnormal NOx levels. Environmetrics 2008, 19, 331–345. [CrossRef]
23. Martinez, J.; Saavedra, Á.; García-Nieto, P.J.; Piñeiro, J.I.; Iglesias, C.; Taboada, J.; Sanchoa, J.; Pastor, J.
Air quality parameters outliers detection using functional data analysis in the Langreo urban area
(Northern Spain). Appl. Math. Comput. 2014, 241, 1–10. [CrossRef]
24. Shaadan, N.; Jemain, A.A.; Latif, M.T. Anomaly detection and assessment of PM10, functional data at several
locations in the Klang Valley, Malaysia. Atmos. Pollut. Res. 2015, 6, 365–375. [CrossRef]
25. Ignaccolo, R.; Mateu, J.; Giraldo, R. Kriging with external drift for functional data for air quality monitoring.
Stoch. Environ. Res. Risk Assess. 2014, 28, 1171–1186. [CrossRef]
26. Wang, D.; Zhong, Z.; Kaixu, B.; Lingyun, H. Spatial and Temporal Variabilities of PM2.5 Concentrations in
China Using Functional Data Analysis. Sustainability 2019, 11, 1620. [CrossRef]
27. Aneiros-Pérez, G.; Cardot, H.; Estévez-Pérez, G.; Vieu, P.H. Maximum ozone concentration forecasting by
functional non-parametric approaches. Environmetrics 2004, 15, 675–685. [CrossRef]
28. Fernández de Castro, B.M.; González-Manteiga, W.; Guillas, S. Functional samples and bootstrap for
predicting sulfur dioxide levels. Technometrics 2005, 47, 212–222. [CrossRef]
29. Quintela-del-Río, A.; Francisco-Fernández, M. Nonparametric functional data estimation applied to ozone
data: Prediction and extreme value analysis. Chemosphere 2001, 82, 800–808. [CrossRef]
30. Besse, P.C.; Cardot, H.; Stephenson, D.B. Autoregressive forecasting of some functional climatic variations.
Scand. J. Stat. 2000, 27, 673–687. [CrossRef]
31. Damon, J.; Guillas, S. The inclusion of exogenous variables in functional autoregressive ozone forecasting.
Environmetrics 2002, 13, 759–774. [CrossRef]
32. Ruiz-Medina, M.D.; Espejo R.M. Spatial autoregressive functional plug-in prediction of ocean surface
temperature. Stoch. Environ. Res. Risk. Assess. 2012, 26, 335–344. [CrossRef]
33. Ruiz-Medina, M.D.; Espejo, R.M.; Ugarte, M.D.; Militino A.F. Functional time series analysis of
spatio-temporal epidemiological data. Stoch. Environ. Res. Risk Assess. 2014, 28, 943–954. [CrossRef]
34. Alvarez-Liebana, J.; Ruiz Medina, M.D. Prediction of air pollutants PM10 by ARBX(1) processes. Stoch.
Environ. Res. Risk Assess. 2019, 33, 1721–1736. [CrossRef]
35. Hsu K.J. Time series analysis of the interdependence among air pollutants. Atm. Environ. Part B Urban Atmos.
1992, 26, 491–503. [CrossRef]
36. Kadiyala, A.; Kumar, A. Vector time series models for prediction of air quality inside a public transportation
bus using available software. Environ. Prog. Sustain. 2014, 33, 337–341. [CrossRef]
37. García-Nieto, P.J.; Sánchez-Lasheras, F.; García-Gonzalo, E.; de Cos Juez F.J. Estimation of PM10 concentration
from air quality data in the vicinity of a major steelworks site in the metropolitan area of Avilés (Northern
Spain) using machine learning techniques. Stoch Environ. Res. Risk Assess. 2018, 32, 3287–3298. [CrossRef]
38. Hedeker, D.; Mermelstein, R.J.; Demirtas, H. An Application of a Mixed-Effects Location Scale Model for
Analysis of Ecological Momentary Assessment (EMA) Data. Biometrics 2008, 64, 627–634. [CrossRef]
39. Taylor, J.; Verbyla, A. Joint modelling of location and scale parameters of the t distribution. Stat. Model.
2004, 4, 91–112. [CrossRef]
40. Pugach, O.; Hedeker, D.; Mermelstein, R. A Bivariate Mixed-Effects Location-Scale Model with application
to Ecological Momentary Assessment (EMA) data. Health Serv. Outcomes Res. Methodol. 2014, 14, 194–212.
[CrossRef]
41. He, W.; Lawless, J.F. Bivariate location-scale models for regression analysis, with applications to lifetime
data. J. R. Statist. Soc. B 2005, 67 Pt 1, 63–78. [CrossRef]
42. Jäntschi, L.; Bálint, D.; Bolboaca, S.D. Multiple Linear Regressions by Maximizing the Likelihood under
Assumption of Generalized Gauss-Laplace Distribution of the Error. Comput. Math. Methods Med.
2016, 2016, 8578156. [CrossRef]

11
Mathematics 2020, 8, 941

43. Rigby, R.A.; Stasinopoulos, D.M. Generalized additive models for location, scale and shape. J. R. Stat. Soc.
Ser. C 2005, 54, 507–554. [CrossRef]
44. Karhunen, K. Zur Spektraltheorie Stochastischer Prozesse. Annales Academiae Scientiarum Fennicae Series A1
Mathematica-Physica 1946, 54, 1–7. Available online: https://katalog.ub.uni-heidelberg.de/cgi-bin/titel.cgi?
katkey=67295489 (accessed on 24 February 2020).
45. Febrero-Bande, M.; Oviedo de la Fuente, M. Statistical Computing in Functional Data Analysis: The R
Package fda.usc. J. Stat. Softw. 2012, 51, 12. [CrossRef]
46. Dogruparmak, S.C.; Özbay, B. Investigating Correlations and Variations of Air Pollutant Concentrations
under Conditions of Rapid Industrialization–Kocaeli (1987–2009). Clean-Soil Air Water 2011, 39, 597–604.
[CrossRef]

c 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).

12
 mathematics
Article
Application of Functional Data Analysis and
FTIR-ATR Spectroscopy to Discriminate Wine Spirits
Ageing Technologies
Ofélia Anjos 1,2,3, *, Miguel Martínez Comesaña 4 , Ilda Caldeira 5,6 , Soraia Inês Pedro 1 ,
Pablo Eguía Oller 4 and Sara Canas 5,6
1 Instituto Politécnico de Castelo Branco, Escola Superior Agrária, 6001-909 Castelo Branco, Portugal;
[email protected]
2 CEF, Instituto Superior de Agronomia, Universidade de Lisboa, 1349-017 Lisboa, Portugal
3 CBPBI, Centro de Biotecnologia de Plantas da Beira Interior, 6001-909 Castelo Branco, Portugal
4 Department of Mechanical Engineering, Heat Engines and Fluid Mechanics, Industrial Engineering School,
University of Vigo, Maxwell s/n, 36310 Vigo, Spain; [email protected] (M.M.C.);
[email protected] (P.E.O.)
5 INIAV, INIAV-Dois Portos, Quinta da Almoínha, 2565-191 Dois Portos, Portugal; [email protected] (I.C.);
[email protected] (S.C.)
6 MED—MediterraneanInstitute for Agriculture, Environment and Development, Instituto de formação
avançada, Universidade de Évora, Pólo da Mitra, Ap. 94, 7006-554 Évora, Portugal
* Correspondence: [email protected]; Tel.: +351-272-339-900

Received: 28 April 2020; Accepted: 27 May 2020; Published: 2 June 2020

Abstract: Fourier transform infrared spectroscopy (FTIR) with Attenuated Total Reflection (ATR)
combined with functional data analysis (FDA) was applied to differentiate aged wine spirits
according to the ageing technology (traditional using 250 L wooden barrels versus alternative
using micro-oxygenation and wood staves applied in 1000 L stainless steel tanks), the wood species
used (chestnut and oak), and the ageing time (6, 12, and 18 months). For this purpose, several features
of the wine spirits were examined: chromatic characteristics resulting from the CIELab method,
total phenolic index, concentrations of furfural, ellagic acid, vanillin, and coniferaldehyde, and total
content of low molecular weight phenolic compounds determined by HPLC. FDA applied to spectral
data highlighted the differentiation between all groups of samples, confirming the differentiation
observed with the analytical parameters measured. All samples in the test set were differentiated and
correctly assigned to the aged wine spirits by FDA. The FTIR-ATR spectroscopy combined with FDA
is a powerful methodology to discriminate wine spirits resulting from different ageing technologies.

Keywords: FTIR-ATR; FDA; vector analysis; wine spirit; ageing technology; micro-oxygenation;
wood; oak; chestnut; ageing time

1. Introduction
The contact of wine distillate with wood is recognised as a pivotal step of the aged wine spirit
production, during which its quality increases and sensory fullness can be reached. Scientific evidence
exists on the key role of several physicochemical phenomena, particularly the extraction and oxidation
reactions involving the wood-derived compounds of low molecular weight, on the chemical changes
(quantitative and qualitative aspects of the beverage’s volatile and non-volatile composition) and
sensory changes (colour, aroma, and taste) occurred [1–3]. Besides, they mainly depend on the ageing
technology, the kind of wood used (oak and chestnut), and the length of the ageing process [1,4,5].
Despite the high-quality spirits attained by the traditional ageing technology, using wooden
barrels, this expensive and lengthy process led to the search and study of sustainable alternative

Mathematics 2020, 8, 896; doi:10.3390/math8060896 13 www.mdpi.com/journal/mathematics

Mathematics 2020, 8, 896

technologies. Research has been focused on a technology based on adding wood pieces to the distillate
kept in stainless steel tanks, as for other alcoholic beverages [6,7]. Recently, the micro-oxygenation
technique, reproducing the oxygen transfer that occurs in the wooden barrel, was applied to optimise
this ageing technology. Promising outcomes based on the phenolic composition and colour of the wine
spirits in the beginning of ageing were attained [8]. Nevertheless, a comprehensive approach is needed
towards the full/robust characterisation of this novel technology, exploring the data acquired over
the ageing period through different analytical and statistical methodologies.
Spectroscopic techniques are very useful for food and beverages quality evaluation because
they require minimal or no sample preparation (absence of extraction, reactions with some other
chemical species, treatment with a chelating agent, masking, sub-sampling, or other manipulation),
they are rapid and non-destructive (causing no physicochemical changes during the analysis), and
they can be used to simultaneously assess several parameters of a sample. In recent years, significant
developments related to the applicability of vibrational spectroscopy combined with multivariate data
analysis has been made to give the rapid quantification of several compounds in different matrices or
to discriminate different groups of samples. Concerning the alcoholic beverages, the studies were more
centred on other drinks than wine spirits. FTIR-ATR (Fourier transform infrared spectroscopy with
Attenuated Total Reflection) was used to evaluate different parameter in alcoholic beverages, such
as the determination of important quality parameters of beers [9], determination of ethanol content
in liquors [10], quantitative analysis of methanol (an adulterant in alcoholic beverages) [11], analysis
of ethanol and methanol content in traditional fruit spirits [12], and the authentication of whisky
according to its botanical origin [13]. Actually, the studies on wine spirits with FTIR-ATR are scarce.
One of them was carried out by Anjos et al. [14], applying the FTIR-ATR methodology to predict
the alcoholic strength, methanol, acetaldehyde, and fusel alcohols contents of grape-derived spirits.
In this paper, a new application of functional data analysis (FDA) in the field of quality evaluation
using spectroscopic techniques is presented. In recent years, FDA has been used in numerous
investigations to analyse processes in continuous time; some examples are energy efficiency [15], medical
research [16], econometrics [17], optimisation problems [18], industrial processes [19], environmental
research [20,21], and food science [22]. In all these works, FDA showed its usefulness for the study of
functions, defined in a specific interval, without missing the correlation between the observations.
FDA allows the analysis of the entire curves, which represent individual observations of the sample
under study, detecting different behaviours throughout the curves [15,23]. Ordoñez et al. [23], in
a similar context that this paper, showed significant reasons to analyse the sample with curves instead
of individual observations. In addition, the contrast of similarity has been carried out from a vectorial
approach, but summarising the curves with a single value; in this case, the mean. This is necessary
because the curves, although representing individual data, are formed by a set of observations
correlated to each other. Martínez et al. [20] explained how this correlation is not taken into account
from a vectorial approach. Furthermore, from this point of view, to contrast the similarity between
the samples, a classical analysis of variance (ANOVA) [24,25] and anon-parametric Kruskal’s test [26]
have been performed. Although this is a simpler approach, it contributes to highlight the strengths
of FDA. On the one hand, Martínez et al. [15] presented how different conclusions can be reached
for each approach because of the biased sample used in the vector analysis (mean of the curves) or
different detection of outliers. On the other hand, among others, Sancho et al. [19] demonstrated that
FDA presents more realistic and accurate results, showing how and at which specific part of the curve
the groups have different spectrometric behaviours.
The aim of this research is to contrast the similarity between the samples obtained with FTIR-ATR
spectrometry for wine spirits aged by different ageing technologies (traditional and alternative), with
different kinds of wood (chestnut and oak) and over the ageing period (6, 12, and 18 months), using
functional data analysis of variance (FANOVA).
Thus, the second section presents, on the one hand, the data used in this study and, on the other
hand, the specific methodology applied to obtain the results. The third section presents the results of

14
Mathematics 2020, 8, 896

the comparison and analysis of the differences between the wine spirit samples. Then, in the fourth
section, a discussion is carried out about the results and the information obtained with them. Finally,
in Section 5, the conclusions of this work are presented.

2. Materials and Methods

2.1. Samples
A total of 30 samples of wine spirit (produced at industrial scale) aged with Limousinoak (Quercus
robur L.) and Portuguese chestnut (Castanea sativa Mill.) for 6, 12, and 18 months of ageing time were
analysed (Figure 1), covering the following categories:

• TL—aged in 250 L new oak wooden barrels;

• TC—aged in 250 L new chestnut wooden barrels;
• AL—aged in 1000 L stainless steel tanks with oak wood staves and micro-oxygenation;
• AC—aged in 1000 L stainless steel tanks with chestnut wood staves and micro-oxygenation.

Figure 1. Scheme of the essay. 1, 2, and 3 indicate the essay replicates.

From each sample, four independent samples were taken for spectral analysis, totalising 120
spectra analysed.
The wooden barrels and the stainless-steel tanks were placed in the cellar in similar environmental
conditions. The same wine distillate (alcohol strength, 77.4 v/v; pH, 5.44; total acidity, as acetic acid,
0.13 g/hL of absolute ethanol; volatile acidity, as acetic acid, 0.11 g/hL of absolute ethanol) produced
by the Adega Cooperativa da Lourinhã, Portugal, was used. The wood pieces and barrels were
manufactured by J.M. Gonçalves cooperage (Palaçoulo, Portugal) with the medium plus toasting level,
as described by Canas et al. [8].
To characterise the wine spirits aged by different technologies, with different wood species,
through the ageing time, several analytical determinations were performed for all the samples: CIELAB
colour parameters (lightness, saturation, and chromaticity coordinates), total phenolic index, and
low molecular weight compounds contents, according to the methodologies described below. All
the analyses were done in duplicate.

15
Mathematics 2020, 8, 896

2.2. Analytical Procedures

2.2.1. Analysis of Chromatic Characteristics

The chromatic characteristics of the wine spirits—lightness (L *), varying between 100% (fully
transparent) and 0% (fully opaque); saturation (C *); chromaticity coordinates (a * and b *), of which
the coordinate a * varies between green (a * < 0) and red hues (a * > 0), and the coordinate b * varies
between blue (b * < 0) and yellow hues (b * > 0)—were analysed according to the CIELab method.
The analysis was performed in a Varian Cary 100 Bio spectrophotometer (Santa Clara, CA, USA) with
a 10 mm glass cell, considering a D65 illuminant and a 10◦ standard observer. The transmittance
measurement was made every 10 nm from 380 to 770 nm. The analysis was performed in duplicate.

2.2.2. Determination of the Total Phenolic Index

The total phenolic index (TPI) of the wine spirits was analysed as described by Cetó et al. [27]:
dilution of the samples with ethanol/water 77:23 v/v; absorbance measurement at 280 nm, using a Varian
Cary 100 Bio spectrophotometer (Santa Clara, CA, USA) with a 10 mm quartz cell; calculation of
the total phenolic index by multiplying the measured absorbance by the dilution factor. The analysis
was performed in duplicate.

2.2.3. Analysis of Low Molecular Weight Compounds

The phenolic and furanic compounds, namelyphenolic acids (gallic acid, vanillic acid, syringic acid
and ellagic acid), phenolic aldehydes (vanillin, syringaldehyde, coniferaldehyde, and sinapaldehyde),
coumarins (umbelliferone and scopoletin), and furanic aldehydes (5-hydroxymethylfurfural,
5-methylfurfural, and furfural) were quantified by liquid chromatography according to the method
proposed by Canas et al. [28]. All compounds were quantified to calculate the total content of
low molecular weight compounds. Only some compounds—furfural, ellagic acid, vanillin, and
coniferaldehyde—were considered in the discussion of individual compounds contents based on their
highly discriminant power.
The chromatographic analysis was carried out in a HPLC Lachrom Merck Hitachi system (Merck,
Darmstadt, Germany) composed of a quaternary pump L-7100, a column oven L-7350 equipped with
a 250 mm × 4 mm i.d. LiChrospher RP 18 (5 μm) column (Merck, Darmstadt, Germany), a UV–Vis
detector L-7400, a fluorescence detector L-7480 (connected to the UV–Vis detector), and an autosampler
L-7250. The HSM D-7000 software (Merck, Darmstadt, Germany) was used for the management,
acquisition, and treatment of the data. The following chromatographic conditions were used: (i) binary
gradient consisting of solvent A, water: formic acid (98:2 v/v) and solvent B, methanol:water:formic
acid (70:28:2 v/v/v) as follows: 0% isocratic B in 3 min, linear gradient from 0% to 40% B in 22 min,
from 40% to 60% B in 18 min, 60% isocratic B in 12 min, linear gradient from 60% to 80% B in 5 min,
80% isocratic B in 5 min; (ii) column temperature of 40 ◦ C; (iii) flow rate of 1 mL/min. Phenolic acids
and furanic aldehydes were detected at 280 nm, phenolic aldehydes were detected at 320 nm, and
coumarins were detected at 325 nm (excitation)/454 nm (emission).
Samples were added with an internal standard (20 mg/L of 4-hydroxybenzaldehyde), filtered
through 0.45 μm membrane (Titan, Scientific Resources Ltd., Gloucester, UK) and analysed by the direct
injection of 20 μL. The identification of chromatographic peaks was made through comparison
of their retention time and UV-Vis spectra with those commercial standards. The assessment of
chromatographic purity of the peaks and their UV–Vis spectra (200–400 nm) were performed in Waters
system composed of a photodiode-array detector (Waters 996), in the same chromatographic conditions
and managed by ‘Millennium 2010 software (Waters, Milford, NA, USA). The analysis was performed
in duplicate. The quantification of each compound was based on a calibration curve made with
the corresponding commercial standard.

16
Mathematics 2020, 8, 896

2.2.4. Spectroscopic Analyses

Spectra were acquired by the Fourier transform infrared spectroscopic method with platinum
Attenuated Total Reflectance (FTIR-ATR) with a Bruker spectrometer (Alpha, Bruker Optic GmbH,
Ettlingen, Germany) using a diamond crystal. Four spectra per sample were obtained with 128 scans
per spectrum at a spectral resolution of 8 cm−1 in the range of 4000 to 450 cm−1 .
The FTIR-ATR used was equipped with a flow-through cell with controlled temperature.
The cleaning of the cell was done by the injection of water in the flow-through cell, and the background
was also measured with distilled water.

2.3. Statistical Analysis

2.3.1. Statistical Treatment of Analytical Data

A two-way analysis of variance (ANOVA) was carried out to examine the influence of the two
fixed factors—ageing technology (two levels: alternative versus traditional) and kind of wood (two
levels: Limousin oak versus chestnut)—on the chromatic characteristics, total phenolic index, and low
molecular weight compounds’ contents of the wine spirits after 6, 12, and 18 months of ageing. For
each significant factor or interaction, the variance percentage was calculated. Fisher’s least significant
difference (LSD) test was used to compare the average values observed for each sample group. All
the calculations were carried out using Statistica vs. 5 (Statsoft Inc., Tulsa, OK, USA).

2.3.2. Functional Analysis

The similarity between the different samples of ageing technology and the different samples of
ageing time was contrasted. This analysis was carried out from a vectorial and functional approach.
Vectorially, the tests used were the classical ANOVA [25], comparing the mean levels of the groups, and
Kruskal’s non-parametric test [26], which studies whether the observations of each group come from
the same distribution. In addition, from the functional approach, the functional ANOVA (FANOVA)
was performed.

Functional Data Analysis (FDA)

The analyses from a functional approach study functions, based on sets of observations, were
defined in a determined interval I = [a,b]. One of its strengths is its structure of infinite dimensions
that allows to extend the possibilities of data analysis [29,30]. A random variable, measured at a set
 G
of discrete points t g ∈ [a, b], has to take values in metric or semi-metric spaces to be considered
g=1
functional. Thus, functional data take the form of a matrix with n rows, one for each individual studied,
and G columns representing the points of evaluation of the functions [31,32].
Smoothing is the most used process to convert discrete observations into continuous functions,
x(t), t ∈ X ⊂ F ; where F is the functional space. Specifically, assuming that the functions are observed
with error, a functional basis expansion can be adopted as follows:


W
x(t) = cw φw (t) (1)
w=1

 W
w=1 is the w-th basis coefficients, φw (t) w=1 is the w-th basis function, and W is the number
where {cw }W
of basis functions under consideration [29,32]. Thus, the basis functions used in this work were
splines [33] due to their specific properties such as the possibility of generating large basis sets easily or
their flexibility [34]. On the other hand, to select the number of bases of each sample, the determination
coefficient R2 was taken into account. The number of bases is the minimum number at which R2

17
Mathematics 2020, 8, 896

stops improving significantly or surpasses the value of 0.99 (see Martínez et al. [15]). Moreover,
the smoothing process involves solving the following problem:


G
minx∈F {z g − x(t g )}2 + λΓ(x) (2)
g=1

where z g = x(t g ) +  g is the value obtained by evaluating x at point t g with  g being a random noise
with zero mean, λ is a parameter controlling the intensity of regularisations, and Γ is a parameter that
makes it costly to reach complex solutions. Then, the basis coefficients can be expressed as the solution
of the smoothing process as follows [29,35]:

−1
c = (Φt Φ + λR) ΦT z (3)

 a GxW matrix formed by Φ gw = φw (t g ) and R being a WxW matrix of the elements

being Φ
Rwg = T D2 φw (t)D2 φ g (t)dt where Dn φw (t) is the nth-order differential operator of φw .

Functional Depths
The depth concept, in classical multivariate statistics, was used to measure the centrality of
a point x ∈ Rd within a data set. The points nearest to the centre obtain a higher depth value [36].
With a functional approach, depths measure the centrality of a curve in relation to the other curves
of the sample x1 , . . . , xn , coming from a stochastic process X(·) evaluated at a specific interval
[a, b] ∈ R [37,38].
Although there exist different functional types of depths (Fraiman-Muniz [37], h-modal [39] or
Random Projections [38]), the most widely used is the h-modal depth due to its better performance in
the correct detection of outliers [36]. Therefore, the functional mode of the sample will be the curve
most densely surrounded by other curves. The functional depth of a certain curve in a specific sample
is calculated as follows:

n  
||x − xw ||
MDn (xi , h) = K i (4)
h
w=1

where ||·|| is the norm in a functional space, K : R+ → R+ is a kernel function, and h represents
the bandwidth parameter [39].
Functional depths, which are considered as a measure of dispersion, are essential in the detection
of outliers. In any data analysis, the identification of these atypical data is crucial because they could
affect the subsequent estimations. In addition, examining them is important to discover the causes that
give these observations a different behaviour from the rest. Besides, in functional analyses, it is even
more important because it is possible that individually, the values that form the curve are not outliers
in a vectorial way but, from a functional point of view, the entire curve could be [36,40]. Martínez et
al. [29] explained in detail how to detect functional outliers within a functional sample.

Functional ANOVA (FANOVA)

Functional data Analysis of Variance, similar to the vector version, contrasts the distance between
the mean levels of the factor variables. The aim of this contrast is to find out if the set of functions studied
are statistically distinguishable [41]. There will also be Q independent samples X g j (t), j = 1, . . . , n g ;
t ∈ I = [a, b]. But these samples are extracted from L2 (l) processes X g (t), g = 1, . . . , Q and their mean
 n
function is E(X g (t)) = m g (t) [42,43]. If the functional sample is divided into groups like X j , A j ∈F
j=1
xA = {1, . . . , A}, being A the factor variable, the hypothesis contrast has the following form:

H0 : X1 (t) = X2 (t) = . . . = XA (t)
(5)
H1 : ∃ h, e s.t. Xh (t)  Xe (t)

18
Mathematics 2020, 8, 896

The model for the j-th observation belonging to the g-th group has the following form [41]:

X jg (t) = X(t) + α g (t) +  jg (t) (6)

where X jg (t) is the functional value ofgroup g, α g (t) is the effect of being part of a determined group
and  jg (t) represents the unexplained variability for the i-th observation of group g.Furthermore,
the model in Equation (6) can be represented in its matrix notation:

X(t) = Zγ(t) + (t) (7)

T
being X(t) a N-dimensional vector, γ(t) = (X(t), α1 (t), . . . , αQ (t)) a (Q + 1)-dimensional vector, (t)
a vector of N residual functions, and Z the design matrix with dimension (N, Q + 1).
Thus, to assure the indentification of the functional effects α g (t), the sum to zero constraint is
introduced [41,44]:

Q +1
γ(t) = 0, ∀t (8)
g=2

The parameter vector γ(t) in Equation (7) can be estimated minimising the standard least squares:

LMSSE(γ) = [X(t) − Zγ(t)]T [X(t) − Zγ(t)]dt

subject to the constraint (8) [41,44].

Regarding the contrast in Equation (5), most tests are based on F test statistic [32,45]:

SSRn (t)/(Q − 1)
Fn ( t ) =
SSEn (t)/(n − Q)

where

Q
2
SSRn (t) = n g (X g (t) − X(t))
g=1


Q 
ng
SSEn (t) = (X g j (t) − X g (t))2
g=1 j=1

represents the variations between groups and within groups, respectively. In addition, for these
Q ng
calculations, the sample mean function X(t) = (1/n) X gj (t) and the sample group mean function
g=1 j=1
ng
X g (t) = (1/n g ) X gj (t), g = 1, . . . , Q were taken into account.
j=1
In this work, two specific tests were used to contrast the similarity between samples. On the one
hand, the F test with the reduced bias estimation method (FB) [46]. This test uses both point variations
between groups and variations within groups. Specifically, the test statistic has the form:

Q
SSRn (t)dt/(Q − 1)
Fn =  (9)
Q
SSEn (t)/(n − Q)

The distribution of this statistic is approximated by F(Q−1)k,(n−Q)k , where k is estimated by

the bias-reduced method. The p-value taken into account comes from P(F(Q−1)k,(n−Q)k > Fn ) [45,46].
On other hand, a permutation test based on a representation of the base function (FP) was
used. This test is based on the basis representation procedure presented by Górecki and Smaga [47].

19
Mathematics 2020, 8, 896

The functional observations are be represented by a finite number of basis functions ϕm ∈ L2 (I ),

m = 1, . . . , K as follows:

K
X gj (t) ≈ c gjm ϕm (t), t ∈ I (10)
m=1

where c gjm are random variables with a significantly large K. Moreover, this test uses the following
approximation of Equation (9) [45]:
(a − b) / (Q − 1)
(11)
(c − a) / (n − Q)
where

Q
1 T T
a= 1 C J C g 1ng
ng ng g ϕ
g=1

1 T T
Q Q
b= 1ng C g Jϕ C j 1n j
n
g=1 j=1


Q
c= trace(CTg Jϕ C g )
g=1

being 1a a vector of ones with dimension ax1, C g = (c g jm ) j=1,...,n ;m=1,...,K , i = 1, . . . , Q, and

 g
Jϕ := I
ϕ(t)ϕT (t)dt is the matrix of cross-products with dimensions KxK based on ϕ(t) =
(ϕ1 (t), . . . , ϕK (t))T .

3. Results
The two-way ANOVA results (Table 1) show that the ageing technology and the kind of wood
had a highly significant effect on the colour and total phenolic content acquired by the wine spirits
during the ageing process (after 6, 12, and 18 months). Among these factors, greater influence was
exerted by the ageing technology (higher percentage of the variance explained) on the chromatic
characteristics, while a similar weight of both factors was observed in the phenolic content. Regardless
of the sampling time, the wine spirits aged with micro-oxygenation and chestnut wood staves (AC)
exhibited a significantly lower value of lightness (L *) and significantly higher values of saturation (C *)
and chromaticity coordinates (a * and b *) than the others. This set of chromatic characteristics reflects
a more evolved colour, since lower L * and higher C * correspond to a more intense/darker colour, and
the combination of higher a * and b * (yellow and red hues, respectively) is associated with a greater
intensity of amber or orange hue, which made these spirits look older than the others. The colour of
wine spirits from Limousin wooden barrels (TL) was on the opposite side, while the wine spirits aged
with micro-oxygenation and Limousin wood staves (AL) and those aged in chestnut wooden barrels
(TC) presented intermediate characteristics.

20
Mathematics 2020, 8, 896

Table 1. Effect of the ageing technology and kind of wood on the chromatic characteristics and total
phenolic index acquired by the wine spirits after 6, 12, and 18 months of ageing. AC: Alternative
Chestnut, AL: Alternative Oak, TC: Traditional Chestnut, TL: Traditional Oak.

Ageing
Code L *(%) A* B* C* TPI
Months
TC 85.41 ± 1.41 b 3.37 ± 1.08 b 50.96 ± 2.68 b 51.08 ± 2.74 b 24.94 ± 1.98 b
TL 93.73 ± 0.42 c −1.25 ± 0.17 a 26.76 ± 2.21 a 26.79 ± 2.20 a 11.99 ± 1.18 a
6
AC 77.14 ± 1.26 a 11.79 ± 1.22 c 70.00 ± 1.59 c 70.99 ± 1.77 c 47.79 ± 3.80 c
AL 87.55 ± 0.23 b 1.75 ± 0.14 b 46.24 ± 0.14 b 46.27 ± 0.15 b 26.88 ± 0.87 b
Technology (S) 61.9 *** 52.7 *** 60.2 *** 60.2 *** 42.1 ***
Variance Wood (W) 36.7 *** 31.9 *** 38.8 *** 38.8 *** 52.4 ***
origin SxW NS 14.0 ** NS NS 4.2 *
Residual 1.5 1.3 1.0 1.0 1.3
TC 79.11 ± 1.38 b 9.89 ± 1.31 b 69.41 ± 1.64 b 70.11 ± 1.80 b 37.90 ± 1.94 b
TL 90.62 ± 1.04 c −0.49 ± 0.65 a 40.66 ± 3.4 a 40.67 ± 3.40 a 17.55 ± 1.81 a
12
AC 65.58 ± 1.97 a 25.63 ± 1.78 c 87.25 ± 0.56 c 90.95 ± 1.04 c 65.72 ± 1.37 c
AL 79.17 ± 0.60 b 10.38 ± 0.53 b 70.87 ± 0.61 b 71.63 ± 0.68 b 37.86 ± 0.23 b
Technology (S) 49.2 *** 49.9 *** 49.4 *** 50.7 *** 48.6 ***
Variance Wood (W) 49.7 *** 46.3 *** 43.5 *** 44.9 *** 48.8 ***
origin SxW NS 3.1 * 6.2 *** 3.6 * 2.2 *
Residual 1.1 0.8 0.9 0.8 0.5
TC 77.33 ± 1.25 b 12.06 ± 1.24 b 73.97 ± 1.22 b 74.95 ± 1.40 b 40.98 ± 2.46 b
TL 89.61 ± 1.17 c −0.02 ± 0.84 a 44.35 ± 3.74 a 44.36 ± 3.74 a 18.15 ± 1.63 a
18
AC 62.29 ± 1.94 a 28.97 ± 1.62 c 89.27 ± 0.12 c 93.86 ± 0.61 c 71.60 ± 2.50 c
AL 76.59 ± 0.52 b 13.19 ± 0.45 b 75.87 ± 0.44 b 77.01 ± 0.51 b 40.55 ± 1.25 b
Technology (S) 52.2 *** 52.9*** 47.7 *** 50.0 *** 47.9 ***
Variance Wood (W) 46.9 *** 45.2*** 40.2 *** 42.3 *** 49.5 ***
origin SxW NS 1.4* 11.1 ** 6.8 *** 2.1 *
Residual 0.9 0.6 0.9 0.9 0.6
L *—lightness; a *, b *—chromaticity coordinates; C *—saturation; TPI—total phenolic index. For each sampling
time (6, 12 and 18 months), mean values with the same letter in a column are not statistically different. NS, p > 0.05;
* 0.01 < p < 0.05; ** 0.001 < p < 0.01; *** p < 0.001.

In a previous work, this research team have already identified furfural, ellagic acid, vanillin and
coniferaldehyde as markers of the ageing technology and the kind of wood used [8]. The results
obtained for these compounds are shown in Table 2. It should be stressed that most of the low molecular
weight compounds contents were mainly dependent on the ageing technology (higher percentage of
the variance explained), as aforementioned for the chromatic characteristics and TPI, but ellagic acid
content was closely related to the kind of wood. Indeed, significantly higher levels of furfural, vanillin,
and coniferaldehyde were found in the wine spirits aged by the alternative technology; the first two
were higher in the modality comprising chestnut wood staves (AC), whereas the latter was higher in
the modality comprising Limousin wood staves (AL). Slight differences were found between the wine
spirits from chestnut barrels (TC) and Limousin oak barrels (TL). Regarding ellagic acid, higher content
was promoted by the chestnut wood (AC and TC), especially in the alternative technology. The contact
with Limousin oak together with micro-oxygenation (AL) induced a slightly lower content of this
phenolic acid, and a weak performance was showed by the Limousin oak barrels (TC).
Besides the differences between the two ageing technologies and the two kinds of wood, the results
also reveal a remarkable role of the ageing time, as indicated by the percentage of variance observed
for each factor in ANOVA (Table 2). Regardless of the ageing modality, there was a gradual decrease of
lightness and a gradual increase of saturation and chromatic coordinates that correspond to the colour
development over the ageing process.

21
Mathematics 2020, 8, 896

Table 2. Effect of the ageing technology and kind of wood on the contents of low molecular weight
compounds (mg/L absolute ethanol) of the wine spirits after 6, 12, and 18 months of ageing.

Ageing
Code Furfural Ellagic Acid Vanillin Coniferaldehyde sumHPLC
Months
TC 38.31 ± 6.90 a
8.12 ± 1.41 b 2.03 ± 0.01 b 6.17 ± 0.63a
163.10 ± 24.08 b
TL 31.73 ± 6.38 a 3.43 ± 0.20 a 1.49 ± 0.08 a 5.21 ± 0.04 a 78.99 ± 9.71 a
6
AC 127.05 ± 5.07 c 15.35 ± 0.38 c 4.62 ± 0.20 d 10.60 ± 0.65 b 296.05 ± 15.91 c
AL 87.74 ± 4.11 b 6.28 ± 0.70 a,b 3.26 ± 0.16 c 12.20 ± 0.52 b 195.23 ± 3.04 b
Technology(S) 82.7 *** 30.4 *** 79.1 *** 98.5 *** 63.1 ***
Variance Wood(W) 8.1 ** 56.9 *** 15.1 *** NS 34.6 ***
origin SxW 8.0 *** 10.8 *** 5.3 ** NS NS
Residual 1.2 2.0 0.5 1.5 2.3
TC 35.85 ± 6.03 a 12.86 ± 1.16 b 3.61 ± 0.22 b 7.00 ± 0.57 a 231.68 ± 26.34 b
TL 31.36 ± 5.80 a 5.64 ± 0.34 a 2.66 ± 0.23 a 6.49 ± 0.58 a 95.09 ± 13.67 a
12
AC 131.17 ± 4.91 c 24.89 ± 1.26 c 8.68 ± 0.02 d 13.97 ± 0.17 b 369.24 ± 8.57 d
AL 96.08 ± 1.93 b 11.94 ± 0.88 b 6.77 ± 0.09 c 19.61 ± 0.41 c 275.32 ± 4.56 c
Technology(S) 87.9 *** 41.3 *** 89.2 *** 79.7 *** 64.5 ***
Variance Wood(W) 5.2 ** 50.0 *** 8.6 *** 5.1 *** 33.8 ***
origin SxW 6.1 *** 7.8 *** 1.8 *** 14.8 *** NS
Residual 0.8 0.9 0.3 0.4 1.7
TC 36.55 ± 7.28 a 15.48 ± 1.41 b 4.43 ± 0.34 b 6.85 ± 0.75 a 271.85 ± 38.84 b
TL 31.63 ± 6.26 a 6.81 ± 0.49 a 3.13 ± 0.23 a 6.41 ± 0.60 a 104.10 ± 16.10 a
18
AC 113.35 ± 4.27 c 28.17 ± 1.15 c 8.61 ± 0.07 d 11.16 ± 0.15 b 347.46 ± 4.33 c
AL 86.72 ± 0.21 b 13.81 ± 0.23 b 7.49 ± 0.24 c 17.93 ± 0.07 c 275.63 ± 4.79 b
Technology(S) 89.4 *** 39.1 *** 92.0 *** 63.3 *** 43.6 ***
Variance Wood(W) 4.8 *** 53.7 *** 7.2 *** 10.0 *** 40.9 ***
origin SxW 4.3 * 6.2 *** NS 26.0 *** 11.9 *
Residual 1.5 0.8 0.7 0.7 3.6
Furf—furfural; Ellag—ellagic acid; Vanil—vanillin; Cofde—coniferaldehyde; sumHPLC—total content of low
molecular weight compounds determined by HPLC. For each sampling time (6, 12, and 18 months), mean values
with the same letter in a column are not. NS, p >0.05; * 0.01 < p <0.05; ** 0.001 < p <0.01; *** p <0.001.

The levels of significance shown in Table 3 reveal that the variation of chromatic characteristics,
phenolic content (TPI and sumHPLC), and most of individual compounds contents were mostly
significant between 6 months and 12 months. Significant differences between the three sampling times
marked the wine spirits aged by the alternative technology with Limousin staves. In general, the L
* parameter decreased over the ageing time, while the other parameters (a *, b *, and C *) increased.
Furfural, ellagic acid, vanillin, coniferaldehyde and total content of low molecular weight compounds
increased. Similar results were observed in the first months of ageing time of this kind of wine spirit [8].

Table 3. Differences observed for each samples group (technology vs. wood) during the ageing time.

TC TL AC AL
6 12 18 6 12 18 6 12 18 6 12 18
L *(%) b a a b a a b a a c b a
a* a b b a a a a b b a b c
b* a b c a b b a b b a b c
C* a b c a b b a b b a b c
TPI a b b a b b a b b a b b
Furf a a a a a a a ab b a ab b
Ellag a b b a b c a b b a b c
Vanil a b c a b c a b b a b c
Cofde a a a a b b a b a a b c
sumHPLC a b b a a a a b b a b b

The spectra obtained for wine spirit are similar to those reported by others authors [28].

22
Mathematics 2020, 8, 896

The representative absorbance spectra of wine spirit samples studied is plotted in Figure 2; their
spectral information is in accordance with previous reports of other authors [48]. The IR region from
2990 to 3626 cm−1 has a very strong influence due to water present in the samples [49]. Nevertheless,
for these analyses, the background was measured with water.

Figure 2. Fourier transform infrared spectroscopy–Attenuated Total Reflection (FTIR-ATR) absorbance

spectra of wine spirit samples.

There is IR information in the regions from 3000 to 2900 cm−1 due to the O–H stretching of alcohols
and C–H stretching of CH3 and CH2 , and consequently related to the presence of ethanol and methanol
in the alcoholic beverages [49].
Regarding the region from 1500 to 860 cm−1 , it corresponds to C–C and C–O vibrations in volatile
compounds [12,14].
The small peak at 1450 cm−1 was assigned to C–OH bending deformation, and the peak at 1275
was assigned to C–O stretching in the acid molecules [11,50].
According to Stuart (2004), the region from 1300 to 840 cm−1 shows other absorption bands
assigned to the C=O and C=C groups present in furanic compounds. The highest peaks at 1086 and
1044 cm−1 are ascribed to the C–O stretch absorption bands, which are important regions for ethanol
and methanol identification and quantification respectively, and C–C absorption bands, which are
related to ethanol and some organic compounds such as sugars, phenols, alcohols, and esters [14,49,51].
The peak at 879 cm−1 could be ascribed to out-of-plane C–H bending of aromatic compounds [10],
and to CH–OH, C–C, C–O, and C–H bond stretching due to water, sugars, and phenolic compounds [51].
According to these regions of the FTIR-ATR absorption spectra (with baseline correction) of
the wine spirit, a mathematical analysis has been performed to determine the differences between
the groups studied. On the one hand, a functional ANOVA (FANOVA) model has been considered
using two different tests: FP, the permutation test based on a representation of the base function
(Equation (9)); and FB, the F test with the reduced bias estimation method (Equation (11)). On the other
hand, a vector analysis based on the classical Analysis of Variance and the non-parametric Kruskal test
have been carried out. Although all the areas of the curves have been analysed following the same
process, only two were plotted each time to show the differences between the samples (space problem).
First, the different groups based on the ageing technology and kind of wood (AC, AL, TC, and TL)
were tested. The contrast is different for each ageing time (18 months, 12 months, and 6 months). With
an 18-month period, the hypothesis of similarity between all groups is rejected in all areas of the whole
curve and from both points of view (Table 4). The samples of wine spirits aged by the alternative
technology, on average, have higher absorbance units than the aged by the traditional one. Within
the alternative technology, the wine spirit aged in oak wood always showed higher absorbance units

23
Mathematics 2020, 8, 896

than the one aged in chestnut wood (Figure 3). With an ageing period of 12 months, the similarity
hypothesis is also rejected in all areas by the two analyses (Table 4). Even so, in this case, the differences
between alternative and traditional ageing technologies are small. The spectrometric curves can hardly
be differentiated in the functional part of Figure 4, and the p-values obtained are higher (Table 4).
With an ageing time of 6 months, significant reasons were found to reject the similarity hypothesis
in all areas through functional and vector analysis (Table 4). With this sample, the difference is more
significant than with the 12-month sample but less than with the 18-month sample. Moreover, it can be
seen that the wine spirit that gets higher absorbance units in the alternative sample is the one aged
in oak, but with the traditional sample, it is the one aged in chestnut (Figure 5). These results are in
accordance with those previously observed for the colour and analytical parameters (Tables 1 and 2).

Table 4. Numerical results of the similarity contrast between the groups AC, AL, TC, and TL, depending
on the ageing time of the wine spirit samples. Functional results (FDA) are based on FANOVA using
two different tests (FP: permutation test based on a representation of the base function, FB: F test
with a reduced bias estimation method). In addition, the results of the ANOVA and Kruskal’s
test representing the vectorial results (VA) are shown. All the results are p-values based on a 5%
significance level.

3050–2750 cm−1 1525–120 cm−1 1150–960 cm−1 910–750 cm−1

TEST\SAMPLE
18 months
FP ≈0 ≈0 ≈0 ≈0
FDA FANOVA
FB ≈0 ≈0 ≈0 ≈0
ANOVA ≈0 ≈0 ≈0 ≈0
VA
Kruskal ≈0 ≈0 ≈0 ≈0
12 months
FP 0.001 ≈0 ≈0 ≈0
FDA FANOVA
FB 0.003 ≈0 1.31 × 10−5 ≈0
ANOVA 0.007 ≈0 0.003 ≈0
VA
Kruskal 0.008 2.23 × 10−6 0.003 ≈0
6 months
FP ≈0 ≈0 ≈0 ≈0
FDA FANOVA
FB ≈0 ≈0 ≈0 1 × 10−4
ANOVA ≈0 ≈0 ≈0 ≈0
VA
Kruskal ≈0 ≈0 ≈0 ≈0

Figure 3. Plots of two of the four significant areas of the curves with an ageing time of 18 months. In
the first row, vectorial analysis by means of boxplots. In the second row, functional data analysis (FDA)
through curves in the studied interval. The wine spirit sample is divided into four groups (AC, AL, TC,
and TL).

24
Mathematics 2020, 8, 896

Figure 4. Plots of two of the four significant areas of the curves with an ageing time of 12 months.
In the first row, vectorial analysis by means of boxplots. In the second row, FDA through curves in
the studied interval. The wine spirit sample is divided into four groups (AC, AL, TC, and TL).

Figure 5. Plots of two of the four significant areas of the curves with an ageing time of 6 months.
In the first row, vectorial analysis by means of boxplots. In the second row, FDA through curves in
the studied interval. The wine spirit sample is divided into four groups (AC, AL, TC, and TL).

Secondly, the similarity between the three different ageing times was contrasted. This contrast is
different depending on the ageing technology and kind of wood used (AC, AL, TC, and TL). Figure 6
shows the boxplots and curves of the sample resulting from alternative technology with chestnut wood.
It can be seen that there is little difference between the three ageing times. Especially at the spectral
region of 1150–960 cm−1 (second d column of Figure 6), in which the similarity hypothesis is rejected
in the vector analysis. Instead, FDA is able to detect these differences between the curve samples
(Table 5). This region is characteristic of the absorption bands assigned to C=O and C=C groups
existing in furanic compounds, C–O stretch absorption bands related to ethanol and methanol, and
C–C absorption bands also related to ethanol and some organic compounds such as sugars, phenols,
alcohols, and esters previously reported, and all of them are important to differentiate the samples in
this study. They are mainly identified at the peaks of the 1044 cm−1 and 1086 cm−1 , which are chiefly
related to the presence of ethanol and methanol but also related to some organic compounds such as
sugars, phenols, alcohols, and esters existing in minor concentration in the beverages. In addition,
similarity in the other areas of the entire curve within the sample resulting from alternative technology
with chestnut wood is rejected. In the case of the alternative technology, the wine spirits aged with

25
Mathematics 2020, 8, 896

Limousin oak wood are very similar to those aged with chestnut wood but with more distance between
the three ageing times (Figure 7), as observed in the chemical analysis. The similarity hypothesis is
rejected in all areas and from both points of view (Table 5). Figure 8 shows the differentiation within
the samples resulting from chestnut barrels. In this case, the two areas drawn from the whole curve
are closer, but the differences between the three ageing times are greater. The hypothesis of similarity
between the samples is rejected in all areas and from the vectorial and functional approach (Table 5).
Finally, regarding the wine spirits aged in Limousin oak barrels, the 18 and 12-month samples show
higher absorbance units than the 6-month sample (Figure 9). The spectrometric curves of the functional
graph can be easily distinguished. There are significant reasons to reject the similarity between the three
samples in all areas of the full curve from the two analyses (Table 5).

Figure 6. Plots of two of the four significant areas of the Alternative Chestnut (AC) curves. In the first
row, vectorial analysis by means of boxplots. In the second row, FDA through curves in the studied
interval. The wine spirit sample is divided into three groups depending on the ageing time (18, 12, and
6 months of ageing).

Figure 7. Plots of two of the four significant areas of the Alternative Oak (AL) curves. In the first row,
vectorial analysis by means of boxplots. In the second row, FDA through curves in the studied interval.
The wine spirit sample is divided into three groups depending on the ageing time (18, 12, and 6 months
of ageing).

26
Mathematics 2020, 8, 896

Table 5. Results of the similarity contrast between the three ageing times (18 months, 12 months, and
6 months), depending on the ageing technology. Functional results (FDA) are based on functional
ANOVA (FANOVA) using two different tests (FP: permutation test based on a representation of the base
function, FB: F test with a reduced bias estimation method). In addition, the results of the ANOVA and
Kruskal’s test representing the vectorial results (VA) are shown. All the results are p-values based on
a 5% significance level.

3050–2750 cm−1 1525–120 cm−1 1150–960 cm−1 910–750 cm−1

TEST\SAMPLE
Groups within AC
FP ≈0 ≈0 ≈0 ≈0
FDA FANOVA
FB ≈0 ≈0 ≈0 ≈0
ANOVA ≈0 ≈0 0.032 ≈0
VA
Kruskal ≈0 6.72 × 10−5 0.214 1 × 10−4
Groups within AL
FP ≈0 ≈0 ≈0 ≈0
FDA FANOVA
FB ≈0 ≈0 ≈0 ≈0
ANOVA ≈0 ≈0 ≈0 ≈0
VA
Kruskal 2.07e-06 4.14 × 10−6 3.34 × 10−5 3.71 × 10−6
Groups within TC
FP ≈0 ≈0 ≈0 ≈0
FDA FANOVA
FB ≈0 ≈0 ≈0 ≈0
ANOVA ≈0 ≈0 ≈0 ≈0
VA
Kruskal ≈0 7.32 × 10−6 ≈0 ≈0
Groups within TL
FP ≈0 ≈0 ≈0 ≈0
FDA FANOVA
FB ≈0 ≈0 ≈0 ≈0
ANOVA ≈0 ≈0 ≈0 ≈0
VA
Kruskal 1 × 10−4 1.95 × 10−6 ≈0 ≈0

Figure 8. Plots of two of the four significant areas of the Traditional Chestnut (TC) curves. In the first
row, vectorial analysis by means of boxplots. In the second row, FDA through curves in the studied
interval. The wine spirit sample is divided into three groups depending on the ageing time (18, 12, and
6 months of ageing).

27
Mathematics 2020, 8, 896

Figure 9. Plots of two of the four significant areas of the Traditional Oak (TL) curves. In the first row,
vectorial analysis by means of boxplots. In the second row, FDA through curves in the studied interval.
The wine spirit sample is divided into three groups depending on the ageing time (18, 12, and 6 months
of ageing).

4. Discussion
The colour and total phenolic content acquired by the wine spirits during the ageing process
(Table 1) are in line with the wine spirits’ features observed in the first six months of ageing [8], and
they are also in accordance with those of recent ageing studies of brandy [2,3]. The colour acquisition
has been assigned to extraction, oxidation, and other physicochemical phenomena occurring during
ageing [8,52]. Hence, more marked color development in the wine spirits aged by the alternative
technology, and with chestnut wood, may have resulted from the direct application of oxygen [8]
and/or from more oxygen transferred by this kind of wood due to its higher porosity [53], which may
have favored such phenomena. Regarding the total phenolic content (TPI), a consistent behavior with
the chromatic characteristics was found, reinforcing previous results [8]: there is a correlation between
the enrichment in phenolic compounds and the colour acquired, which are induced by the ageing
technology and the kind of wood.
Accordingly, the total content of low molecular weight compounds (phenolic acids, phenolic
aldehydes, coumarins, and furanic aldehydes) concentrations determined by HPLC (Table 2) acted
as distinctive feature of the aged wine spirits, which can be ordered as follows: AC > AL ~ TC >
TL. Therefore, the pattern reported for the first six months of ageing [8] prevailed until the end of
the ageing process (18 months).
Low molecular weight compounds, such as furfural, ellagic acid, vanillin, and coniferaldehyde,
as aforementioned, stood out as markers of the ageing technology (Table 2). Similar results for
these wood-derived phenolic compounds [1,54] were obtained in the beginning of ageing [8]. Their
higher accumulation in wine spirits resulting from the alternative technology has been ascribed to
a faster extraction from the wood, as described by other authors for the ageing of red wine with wood
pieces [55,56]; the behavior of coniferaldehyde is explained by its sensitivity towards oxidation.
The evolution observed for the colour parameters over the ageing process was supported
by the progressive increase of the total phenolic content (TPI) and particularly the low molecular
weight compounds quantified by HPLC (sumHPLC), which reflected the release of the wood-derived
compounds into the wine spirit together with a positive balance between their formation/degradation [4]
in the liquid medium. The kinetics of individual compounds shows a continuous increment of ellagic
acid content, regardless of the assay modality, which may have resulted from the release of this
phenolic acid existing in the wood [1] and from the oxidation and hydrolysis of ellagitannins in
the wine spirit during ageing [57,58]. The kinetics of vanillin presented a similar pattern, which was
likely due to its release from the wood [59], lignin’s hydroalcoholysis, and subsequent oxidation

28
Mathematics 2020, 8, 896

reactions that converted coniferaldehyde into vanillin in the wine spirit [4]. Interestingly, there was
an increase of coniferaldehyde content between 6 and 12 months followed by a decrease until 18
months. This suggests that after 12 months of ageing, oxidation reactions prevailed over extraction and
lignin’s hydroalcoholys is causing an unbalance between coniferaldehyde formation and degradation.
Similar results were previously noticed for wine spirits aged by the alternative technology but without
micro-oxygenation and also by the traditional ageing in wooden barrels [4]. The evolution of furfural
content was different from those of phenolic compounds. Indeed, its kinetics was closely related
to the ageing technology: there was an increase from 6 to 12 months and then a decrease in wine
spirits aged by the alternative technology; the opposite behavior occurred in wine spirits aged by
the traditional one (TC and TL).
The statistical analyses performed during the ageing time reveals significant differences between
6 months and 12 months in a higher number of parameters than the observed between 12 months and
18 months. However, for the wine spirits aged by the alternative technology with Limousin staves
significant differences between the three sampling times were observed. Regarding the alternative
technology, these outcomes suggest that one year is enough/adequate to obtain a high-quality wine
spirit aged with micro-oxygenation and chestnut staves as a consequence of higher pool of compounds
and specific anatomical properties of this kind of wood compared with the Limousin oak wood [2,60].
The mathematical analysis performed with functional and vectorial approach corroborates
the findings about colour parameters, total phenolic content, and low molecular weight compounds.
According to the mathematical analysis, two main conclusions have been reached. First, there are
significant reasons to reject the similarity between the wine spirits samples differentiated according
to the ageing technology used. It was rejected in all areas of the full spectrometric curve for each of
the three ageing times tested, both in vector and functional data analysis. Secondly, evidence was
found to state that the samples of wine spirits with different ageing time are not similar. The similarity
hypothesis between the three ageing times for each of the different technologies (traditional vs.
alternative) used was rejected in all the areas of the spectrometric curve analysed.
On the other hand, FDA shows higher consistency than the vector analysis as observed by Martínez
et al. [15,20]. In addition, to provide more information and accuracy, FDA can detect significant
differences between groups that vector analysis cannot. This can be seen in the sample obtained by
alternative technology with chestnut wood, specifically in the spectral region of 1150–960 cm−1 of
the entire curve. It is attributed to the functional groups present in furanic compounds and also related
to ethanol, methanol, and sugars, phenols, and esters existing in the wine spirit [1,4,5,8,14,49,51]. Vector
analysis found no evidence to reject the similarity between the samples based on ageing time, but FDA,
taking into account all correlated observations measured in the specific area, did.

5. Conclusions
The aged wine spirit composition and quality depend on the raw material, but mainly on the ageing
technology, the wood species used in the ageing process, and the ageing time. Monitoring this quality
over time with a fast methodology is very important for the industry.
This study showed a remarkable congruence between the analytical determinations (colour, total
phenolic index, and low molecular weight compounds) and the FTIR-ATR/mathematical approach in
the differentiation of wine spirits aged by the alternative (micro-oxygenation combined with wood
pieces in stainless steel tanks) and the traditional (wooden barrels) technologies, using two kinds of
wood (chestnut and oak), over the ageing time (6, 12, and 18 months). Additionally, the analytical
results suggest that one year is enough/adequate to achieve a high-quality wine spirit aged with
micro-oxygenation and chestnut staves.
FTIR-ATR with appropriate chemometric techniques, specifically functional data analysis, and
vector analysis proved to be a powerful tool for an easier and faster monitoring of the wine spirit’s
ageing process. FDA showed higher consistency than the vector analysis, providing more information
and accuracy; it detected significant differences between groups that the vector analysis did not detect.

29
Mathematics 2020, 8, 896

It was clearly demonstrated in the sample resulting from the alternative technology with chestnut
wood, specifically in the area 1150–960 cm−1 of the entire curve. Vector analysis found no evidence
to reject similarity between the time ageing samples, but FDA, taking into account all correlated
observation measured in the specific area, found it.
It was also possible to identify the more accurate spectral region to perform a calibration model to
be applied by the wine spirit industry.

Author Contributions: Conceptualisation, O.A. and S.C.; methodology, O.A., S.C., M.M.C., P.E.O.; formal analysis,
S.C., O.A., I.C., S.I.P.; Statistical analysis: M.M.C., P.E.O., O.A.; investigation, S.C., O.A., I.C., S.I.P.; resources, S.C.;
O.A.; writing, O.A., M.M.C., S.C.; supervision, O.A.; project administration, S.C.; funding acquisition, S.C. All
authors have read and agreed to the published version of the manuscript.
Funding: This research was funded by the Project CENTRO-04-3928-FEDER 0000281/Line “Validation of a new
ageing technology for wine spirit from Lourinhã”. This work is also funded by National Funds through
FCT—Foundation for Science and Technology under the Project UIDB/05183/2020 (MED—Mediterranean Institute
for Agriculture, Environment and Development) and the Project UIDB/00239/2020 (Centro de Estudos Florestais).
Acknowledgments: The authors thank João Pedro Catela, Nádia Santos, Manuela Gomes, Eugénia Gomes and
Inês Antunes from Adega Cooperativa da Lourinhã, and Tanoaria J.M. Gonçalves for the technical support.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Canas, S. Phenolic Composition and Related Properties of Aged Wine Spirits: Influence of Barrel
Characteristics. A Review. Beverages 2017, 3, 55. [CrossRef]
2. García-Moreno, M.V.; Sánchez-Guillén, M.M.; Mier, M.R.D.; Delgado-González, M.J.;
Rodríguez-Dodero, M.C.; García-Barroso, C.; Guillén-Sánchez, D.A. Use of Alternative Wood for
the Ageing of Brandy de Jerez. Foods 2020, 9, 250. [CrossRef] [PubMed]
3. Schwarz, M.; Rodríguez-Dodero, C.M.; Jurado, S.M.; Puertas, B.G.; Barroso, C.; Guillén, A.D. Analytical
Characterization and Sensory Analysis of Distillates of Different Varieties of Grapes Aged by an Accelerated
Method. Foods 2020, 9, 277. [CrossRef] [PubMed]
4. Canas, S.; Caldeira, I.; Belchior, A.P. Extraction/oxidation kinetics of low molecular weight compounds in
wine brandy resulting from different ageing technologies. Food Chem. 2013, 138, 2460–2467. [CrossRef]
5. Caldeira, I.; Santos, R.; Ricardo-da-Silva, J.M.; Anjos, O.; Mira, H.; Belchior, A.P.; Canas, S. Kinetics of odorant
compounds in wine brandies aged in different systems. Food Chem. 2016, 211, 937–946. [CrossRef]
6. Rodríguez-Solana, R.; Rodríguez-Freigedo, S.; Salgado, J.M.; Domínguez, J.M.; Cortés-Diéguez, S.
Optimisation of accelerated ageing of grape marc distillate on a micro-scale process using a Box–Benhken
design: Influence of oak origin, fragment size and toast level on the composition of the final product. Aust. J.
Grape Wine Res. 2017, 23, 5–14. [CrossRef]
7. Gómez-Plaza, E.; Bautista-Ortín, A.B. Chapter 10—Emerging Technologies for Aging Wines: Use of Chips
and Micro-Oxygenation. In Red Wine Technology; Morata, A., Ed.; Academic Press: Cambridge, MA, USA,
2019; pp. 149–162. [CrossRef]
8. Canas, S.; Caldeira, I.; Anjos, O.; Belchior, A. Phenolic profile and colour acquired by the wine spirit in
the beginning of ageing: Alternative technology using micro-oxygenation vs traditional technology. LWT
Food Sci. Technol. 2019, 111, 260–269. [CrossRef]
9. Llario, R.; Iñón, F.A.; Garrigues, S.; de la Guardia, M. Determination of quality parameters of beers by the use
of attenuated total reflectance-Fourier transform infrared spectroscopy. Talanta 2006, 69, 469–480. [CrossRef]
10. Yadav, P.K.; Sharma, R.M. Classification of illicit liquors based on their geographic origin using Attenuated
total reflectance (ATR)—Fourier transform infrared (FT-IR) spectroscopy and chemometrics. Forensic Sci. Int.
2019, 295, e1–e5. [CrossRef]
11. Nagarajan, R.; Mehrotra, R.; Bajaj, M.M. Quantitative analysis of methanol, an adulterant in alcoholic
beverages, using attenuated total reflectance spectroscopy. J. Sci. Ind. Res. 2006, 65, 416–419.
12. Teodora Emilia, C.; Carmen, S.; Florinela, F.; Floricuţa, R.; Raluca Maria, P.O.P.; Mira, F. Rapid Quantitative
Analysis of Ethanol and Prediction of Methanol Content in Traditional Fruit Brandies from Romania, using
FTIR Spectroscopy and Chemometrics. Not. Bot. Horti Agrobot. Cluj-Napoca 2013, 41. [CrossRef]

30
Mathematics 2020, 8, 896

13. Wiśniewska, P.; Boqué, R.; Borràs, E.; Busto, O.; Wardencki, W.; Namieśnik, J.; Dymerski, T. Authentication
of whisky due to its botanical origin and way of production by instrumental analysis and multivariate
classification methods. Spectrochim. Acta. Part A Mol. Biomol. Spectrosc. 2017, 173, 849–853. [CrossRef]
[PubMed]
14. Anjos, O.; Santos, A.J.A.; Estevinho, L.M.; Caldeira, I. FTIR–ATR spectroscopy applied to quality control of
grape-derived spirits. Food Chem. 2016, 205, 28–35. [CrossRef] [PubMed]
15. Martínez Comesaña, M.; Martínez Mariño, S.; Eguía Oller, P.; Granada Álvarez, E.; Erkoreka González, A.
A Functional Data Analysis for Assessing the Impact of a Retrofitting in the Energy Performance of a Building.
Mathematics 2020, 8, 547. [CrossRef]
16. Martínez, J.; Ordoñez, C.; Matìas, J.M.; Taboada, J. Determining noise in an aggregates plant using functional
statistics. Hum. Ecol. Risk Assess. 2011, 17, 521–533. [CrossRef]
17. Müller, H.-G.; Sen, R.; Stadtmüller, U. Functional data analysis for volatility. J. Econom. 2011, 165, 233–245.
[CrossRef]
18. López, M.; Martínez, J.; Matías, J.M.; Taboada, J.; Vilán, J.A. Shape functional optimization with restrictions
boosted with machine learning techniques. J. Comput. Appl. Math. 2010, 234, 2609–2615. [CrossRef]
19. Sancho, J.; Pastor, J.J.; Martínez, J.; García, M.A. Evaluation of Harmonic Variability in Electrical Power
Systems through Statistical Control of Quality and Functional Data Analysis. Procedia Eng. 2013, 63, 295–302.
[CrossRef]
20. Martínez, J.; Pastor, J.; Sancho, J.; McNabola, A.; Martínez, M.; Gallagher, J. A functional data analysis
approach for the detection of air pollution episodes and outliers: A case study in Dublin, Ireland. Mathematics
2020, 8, 225. [CrossRef]
21. Di Battista, T.; Fortuna, F. Functional confidence bands for lichen biodiversity profiles: A case study in
Tuscany region (central Italy). Stat. Anal. Data Min. ASA Data Sci. J. 2017, 10, 21–28. [CrossRef]
22. Ruiz-Bellido, M.A.; Romero-Gil, V.; García-García, P.; Rodríguez-Gómez, F.; Arroyo-López, F.N.;
Garrido-Fernández, A. Data on the application of Functional Data Analysis in food fermentations. Data Brief
2016, 9, 401–412. [CrossRef] [PubMed]
23. Ordoñez, C.; Martínez, J.; Matías, J.M.; Reyes, A.N.; Rodríguez-Pérez, J.R. Functional statistical techniques
applied to vine leaf water content determination. Math. Comput. Model. 2010, 52, 1116–1122. [CrossRef]
24. Crawley, M.J. The R Book; John Wiley & Sons: Hoboken, NJ, USA, 2013.
25. Montgomery, D.C. Design and Analysis of Experiments; John Wiley & Sons, Inc.: Hoboken, NJ, USA, 2013.
26. Ostertagová, E.; Ostertag, O.; Kovác, J. Methodology and application of the Kruskal-Wallis test. Appl. Mech.
Mater. 2014, 611, 115–120. [CrossRef]
27. Cetó, X.; Gutiérrez, J.M.; Gutiérrez, M.; Céspedes, F.; Capdevila, J.; Mínguez, S.; Jiménez-Jorquera, C.;
del Valle, M. Determination of total polyphenol index in wines employing a voltammetric electronic tongue.
Anal. Chim. Acta 2012, 732, 172–179. [CrossRef]
28. Canas, S.; Belchior, A.; Spranger, M.; Sousa, R. High-performance liquid chromatography method for analysis
of phenolic acids, phenolic aldehydes, and furanic derivatives in brandies. Development and validation. J.
Sep. Sci. 2003, 26, 496–502. [CrossRef]
29. Martínez, J.; Saavedra, Á.; García, P.J.; Piñeiro, J.I.; Iglesias, C.; Taboada, J.; Sancho, J.; Pastor, J. Air quality
parameters outliers detection using functional data analysis in the Langreo urban area (Northern Spain).
Appl. Math. Comput. 2014, 241, 1–10. [CrossRef]
30. Wang, J.L.; Chiou, J.M.; Müller, H.G. Functional Data Analysis. Annu. Rev. Stat. Appl. 2016, 3, 257–295.
[CrossRef]
31. Kramosil, I.; Michálek, J. Fuzzy metrics and statistical metric spaces. Kybernetika 1975, 11, 336–344.
32. Ramsay, J.O.; Silverman, B.W. Functional Data Analysis, 2nd ed.; Springer: New York, NY, USA, 2005.
33. Cormier, E.; Genest, C.; Nešlehová, J.G. Using B-splines for nonparametric inference on bivariate
extreme-value copulas. Extremes 2014, 17, 633–659. [CrossRef]
34. Kwok, W.Y.; Moser, R.D.; Jiménez, J. A Critical Evaluation of the Resolution Properties of B-Spline and
Compact Finite Difference Methods. J. Comput. Phys. 2001, 174, 510–551. [CrossRef]
35. Piñeiro, J.I.; Torres, J.M.; García, P.J.; Alonso, J.R.; Muñiz, C.D.; Taboada, J. Analysis and detection of functional
outliers in waterquality parameters from different automated monitoring stationsin the Nalón River Basin
(Northern Spain). Environ. Sci. Pollut. Res. Int. 2015, 22, 387–396. [CrossRef] [PubMed]

31
Mathematics 2020, 8, 896

36. Febrero, M.; Galeano, P.; Wenceslao, G.M. Outlier detection in functional data by depth measures, with
application to identify abnormal NOx levels. Environmetrics 2008, 19, 331–345. [CrossRef]
37. Fraiman, R.; Muniz, G. Trimmed means for functional data. TEST 2001, 10, 419–440. [CrossRef]
38. Cuevas, A.; Febrero, M.; Fraiman, R. Robust estimation and classification for functional data via
projection-based notions. Comput. Stat. 2007, 22, 481–496. [CrossRef]
39. Cuevas, A.; Febrero, M.; Fraiman, R. On the use of bootstrap for estimating functions with functional data.
Comput. Stat. Data Anal. 2006, 51, 1063–1074. [CrossRef]
40. Millán-Roures, L.; Epifanio, I.; Martínez, V. Detection of Anomalies in Water Networks by Functional Data
Analysis. Math. Probl. Eng. 2018, 2018, 13. [CrossRef]
41. Maturo, F.; Battista, T.; Fortuna, F. Parametric functional analysis of variance for fish biodiversity assessment.
J. Environ. Inform. 2016, 28, 101–109. [CrossRef]
42. Cuesta-Albertos, J.A.; Febrero, M. A simple multiway ANOVA for functional data. TEST 2010, 19, 537–557.
[CrossRef]
43. Zhang, J.-T. Analysis of Variance for Functional Data. In A Chapman & Hall Book; Press, C.R.C., Ed.; Taylor &
Francis Group: Abingdon, UK, 2013; p. 412.
44. Aguilera, A.; Fortuna, F.; Escabias, M.; Battista, T. Assessing Social Interest in Burnout Using Google Trends
Data. Soc. Indic. Res. 2019, 1–13. [CrossRef]
45. Górecki, T.; Smaga, Ł. fdANOVA: An R software package for analysis of variance for univariate and
multivariate functional data. Comput. Stat. 2019, 34, 571–597. [CrossRef]
46. Zhang, J.-T. Statistical inferences for linear models with functional responses. Stat. Sin. 2011, 21, 1431–1451.
[CrossRef]
47. Górecki, T.; Smaga, Ł. A comparison of tests for the one-way ANOVA problem for functional data. Comput.
Stat. 2015, 30, 987–1010. [CrossRef]
48. Moreira, J.; Santos, L. Spectroscopic interferences in Fourier transform infrared wine analysis. Anal. Chim.
Acta 2003, 513, 263–268. [CrossRef]
49. Shurvell, H.F. Spectra—Structure Correlations in the Mid- and Far-Infrared. In Handbook of Vibrational
Spectroscopy; John Wiley & Sons: Hoboken, NJ, USA, 2001.
50. Tarantilis, P.A.; Troianou, V.E.; Pappas, C.S.; Kotseridis, Y.S.; Polissiou, M.G. Differentiation of Greek red
wines on the basis of grape variety using attenuated total reflectance Fourier transform infrared spectroscopy.
Food Chem. 2008, 111, 192–196. [CrossRef]
51. Öztürk, B.; Yucesoy, D.; Ozen, B. Application of Mid-infrared Spectroscopy for the Measurement of Several
Quality Parameters of Alcoholic Beverages, Wineand Raki. Food Anal. Methods 2012, 5, 1435–1442. [CrossRef]
52. Fujieda, M.; Tanaka, T.; Suwa, Y.; Koshimizu, S.; Kouno, I. Isolation and Structure of Whiskey Polyphenols
Produced by Oxidation of Oak Wood Ellagitannins. J. Agric. Food. Chem. 2008, 56, 7305–7310. [CrossRef]
53. Carvalho, A. Identificação anatómica e caracterização física e mecânica das madeiras utilizadas no fabrico
de quartolas para produção de aguardentes velhas de qualidade—Denominação Lourinhã. Ciência Técnica
Vitivinícola 1998, 13, 71–105.
54. Martínez-Gil, A.; del Alamo-Sanza, M.; Sánchez-Gómez, R.; Nevares, I. Alternative Woods in Enology:
Characterization of Tannin and Low Molecular Weight Phenol Compounds with Respect to Traditional Oak
Woods. A Review. Molecules 2020, 25, 1474. [CrossRef]
55. Oberholster, A.; Elmendorf, B.L.; Lerno, L.A.; King, E.S.; Heymann, H.; Brenneman, C.E.; Boulton, R.B. Barrel
maturation, oak alternatives and micro-oxygenation: Influence on red wine aging and quality. Food Chem.
2015, 173, 1250–1258. [CrossRef]
56. Laqui-Estaña, J.; López-Solís, R.; Peña-Neira, Á.; Medel-Marabolí, M.; Obreque-Slier, E. Wines in contact
with oak wood: The impact of the variety (Carménère and Cabernet Sauvignon), format (barrels, chips and
staves), and aging time on the phenolic composition. J. Sci. Food Agric. 2019, 99, 436–448. [CrossRef]
57. Karvela, E.; Makris, D.P.; Kefalas, P.; Moutounet, M. Extraction of phenolics in liquid model matrices
containing oak chips: Kinetics, liquid chromatography-mass spectroscopy characterisation and association
with in vitro antiradical activity. Food Chem. 2008, 110, 263–272. [CrossRef] [PubMed]
58. García-Estévez, I.; Alcalde-Eon, C.; Martínez-Gil, A.M.; Rivas-Gonzalo, J.C.; Escribano-Bailón, M.T.;
Nevares, I.; del Alamo-Sanza, M. An Approach to the Study of the Interactions between Ellagitannins and
Oxygen during Oak Wood Aging. J. Agric. Food. Chem. 2017, 65, 6369–6378. [CrossRef] [PubMed]

32
Mathematics 2020, 8, 896

59. Martínez-Gil, A.; Cadahía, E.; Fernández de Simón, B.; Gutiérrez-Gamboa, G.; Nevares, I.; del Álamo-Sanza, M.
Phenolic and volatile compounds in Quercus humboldtii Bonpl. wood: Effect of toasting with respect to
oaks traditionally used in cooperage. J. Sci. Food Agric. 2019, 99, 315–324. [CrossRef] [PubMed]
60. Anjos, O.; Carmona, C.; Caldeira, I.; Canas, S. Variation of Extractable Compounds and Lignin Contents in
Wood Fragments Used in the Aging of Wine Brandies. BioResources 2013, 8, 4484–4496. [CrossRef]

© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).

33
 mathematics
Article
Outliers Detection Models in Shewhart Control
Charts; An Application in Photolithography:
A Semiconductor Manufacturing Industry
Ishaq Adeyanju Raji 1 , Muhammad Hisyam Lee 1, *, Muhammad Riaz 2 , Mu’azu Ramat Abujiya 3
and Nasir Abbas 2
1 Department of Mathematical Sciences, Universiti Teknologi Malaysia, Skudai 81310, Malaysia;
[email protected]
2 Department of Mathematics and Statistics, King Fahd University of Petroleum and Minerals,
Dhahran 31261, Saudi Arabia; [email protected] (M.R.); [email protected] (N.A.)
3 Preparatory Year Mathematics Program, King Fahd University of Petroleum and Minerals,
Dhahran 31261, Saudi Arabia; [email protected]
* Correspondence: [email protected]; Tel.: +60-7-553-0264

Received: 10 April 2020; Accepted: 21 May 2020; Published: 25 May 2020

Abstract: Shewhart control charts with estimated control limits are widely used in practice. However,
the estimated control limits are often affected by phase-I estimation errors. These estimation errors
arise due to variation in the practitioner’s choice of sample size as well as the presence of outlying
errors in phase-I. The unnecessary variation, due to outlying errors, disturbs the control limits
implying a less efficient control chart in phase-II. In this study, we propose models based on Tukey
and median absolute deviation outlier detectors for detecting the errors in phase-I. These two outlier
detection models are as efficient and robust as they are distribution free. Using the Monte-Carlo
simulation method, we study the estimation effect via the proposed outlier detection models on the
Shewhart chart in the normal as well as non-normal environments. The performance evaluation is
done through studying the run length properties namely average run length and standard deviation
run length. The findings of the study show that the proposed design structures are more stable in the
presence of outlier detectors and require less phase-I observation to stabilize the run-length properties.
Finally, we implement the findings of the current study in the semiconductor manufacturing industry,
where a real dataset is extracted from a photolithography process.

Keywords: average run length; control chart; median absolute deviation; outlier; photolithography;
Shewhart; Tukey

1. Introduction
The two salient tools of statistical process control (SPC) are memory and memory-less control
charts. The memory-less control charts are most suitable for large shift, while the memory-control
charts are used to monitor moderate and small shifts. The prominent form of memory-less control
chart for location monitoring is the Shewhart X control chart. In general, control charts-irrespective of
the magnitude they measure-operate in two phases: phase-I, the prospective stage from which the
control limits are obtained; phase-II, where we monitor the process and correct the unnatural causes of
variation whenever they occur (cf. [1]). In phase-I we estimate the control limits using the parameters of
the process under study which, in reality, are seldom known. The amount of data employed in phase-I
for estimating process parameters varies from one practitioner to the other. As a result, this variability
affects the chart performance in the monitoring stage i.e., phase-II. (see for example [2–6]).
Furthermore, the amount of data employed in estimating the process parameters does affect the
accuracy of the chart, as well as its limits. As we all know, the larger the sample size, the closer we are

Mathematics 2020, 8, 857; doi:10.3390/math8050857 35 www.mdpi.com/journal/mathematics

Mathematics 2020, 8, 857

to the parameter. Therefore, increasing the sample size used for estimating the parameters should be
the remedy to this shortcoming, but there is a limit to which we can increase sample sizes in real-life
situations. As a result, the Shewhart chart, like any other chart, loses its performance and credibility.
The depth of the loss depends on the efficacy of the parameter estimation and sample size employed
in phase-I.
The presence of outlying/extreme values in the phase-I dataset can affect the performance of the
control chart. The insufficiency of the phase-I estimates could be a result of extreme sample points in
the sample, and not necessarily the size of the sample (see [5,7]). The easiest remedy for the extreme
values is to drop such a sample and pick another one, but this is not appropriate for small sample
data. Therefore, there is a need to screen the extreme values to improve the overall performance of the
control chart.
Over the years, researchers have studied different types of robust outlier detection models in
a series of control charts to enhance their performance. Examples include [8–12]. These outlier
detectors require the data to be from normal distribution such as the Student-type and Grubbs-type
detectors. However, for a non-normal dataset, the Tukey’s and median absolute deviation (MAD)
outlier detection models are more accurate and robust since they are independent of mean and
standard deviation. (see [13–19]). SPC is widely applied and implemented in various sectors; health,
industrial, manufacturing and every service-rendering sector. Control charts, however, are most
applied in manufacturing industry, with semiconductors as a case study. Semiconductor manufacturing
processes are prone to high chances of assignable cause of variations, due to machine breakdown,
multiple products, re-entrant flows, batching processes etc. [20]. Researchers have employed SPC in
solving these recurring challenges in this industry (see [21–24]). The proposed charts in this study are
applied in photolithography, a semiconductor manufacturing process.
In this article, we study the effects of parameter estimation on the Shewhart X chart for normal and
non-normal environments. We also study the effect of outliers on the reliability of the control charts and
the process parameters are estimated. Furthermore, we propose non-parametric outlier detectors, namely:
the robust Tukey and MAD outlier detection models in designing the basic control chart structure. A fair
comparison between the two-outlier detection models is also made. We achieve all of these using average
run length (ARL) and standard deviation run length (SDRL) as the performance measures.
The remainder of this article is as follows: the next Section entails the methodologies employed
for the study; briefing the overview of the Shewhart X control chart when the parameters are known
and unknown, alongside the performance measure properties adopted in this study; the variability in
Shewhart chart performance due to phase-I estimation; a scenario for the presence of outliers in the
design structure of Shewhart chart, and its effect; incorporating the Turkey and MAD outlier detection
models in the design structure of the Shewhart chart as remedies for rectifying the presence of outliers;
Section 3 gives a concise and precise description of the simulation results. In Section 4, a detailed
comparison of the results is presented; while Section 5 provides an illustrative example with a real life
dataset; finally a concluding remark and future recommendations are given in Section 6.

2. Methodology
In this section, we give details of the Shewhart control chart for normal and non-normal
environments. The known and unknown parameter scenarios, the practitioner–practitioner variation in
the estimation stage, the presence of outliers/extreme values in the estimation sample, and incorporating
some outlier detection models in the Shewhart chart are all discussed in the following subsections.

2.1. Overview of the Shewhart Control Chart

Let Yij i = 1, 2, . . . , n and j = 1, 2, . . . represent a ith observation from jth sample of an
ongoing (continuous) process. Further Yij follows a normal distribution with mean μ0 + δσ0 and
variance σ0 2 i.e., Yij ∼ N μ0 + δσ0 , σ0 2 . The process is said to be in the in-control (IC) state if δ = 0,

36
Mathematics 2020, 8, 857

and out-of-control (OoC) otherwise. A default Shewhart set-up monitors a process by plotting the
sample mean (Yi = 1/n nj=1 Yij ) of Yij against the following control chart limits.

σ0 σ0
UCL = μ0 + L √ , LCL = μ0 − L √ (1)
n n

where UCL and LCL denote the upper and lower control limits, respectively. Limits in (1) are useful
when the parameters (μ0 and σ0 2 ) of the process are known. However, when they are unknown,
their respective unbiased estimators from the phase-I are used, and the resulting control chart structures
will be in estimated form.
For phase-I, let Yil represents ith observation from lth random sample ∀ i = 1, 2, 3, . . . , n and
l = 1, 2, 3, . . . , m, regarded to be under statistically IC state. It is good to mention here that the choice
of m and n varies from one practitioner-to another. Therefore, it affects the accuracy of the control
limits implying an influenced ARL in phase-II. The unbiased estimators for the parameters μ and σ of
an IC process are defined as:
μ̂0 = (1/m) m l=1 Yl (2)
σ̂0 = (1/mC4 ) m l = 1 Sl

2 √
i=1 (Yil −Yl )
n n
Y 2/(n−1)Γ(n/2)
where Yl = l=n1 il , Sl = n−1 and c 4 = Γ[(n−1)/2]
is the bias correction constant.
Subsequently, the resulting control limits in (1) are modified to the following:

m m m m
= l=1 Yl l = 1 Sl  = l=1 Yl l = 1 Sl
UCL + L̂ √ , LCL − L̂ √ (3)
m mC4 n m mC4 n

In phase-II, Yl s are plotted against the control limits in (3) and the chart is said to have given
an OoC signal if any value of Yl is plotted outside the limits. Here, the sample number at which the
statistic is plotted outside the limits is recorded as run length (RL). RL is an important variable in
measuring the performance of control charts in general, and the Shewhart is not an exception. The most
widely used property of RL is ARL, which is the average number of samples observed before the chart
sends an OoC signal. Mathematically, ARL = sk=1 RLk /s where s is the number of RLs recorded.
In addition to ARL, standard deviation of the RL (SDRL) gives more information about the behavior of
the RL variable in evaluating the performance of a control chart. Furthermore, the ARL is of two types
i.e., the IC ARL, denoted as ARL0 and the OoC ARL, referred to as ARL1 . ARL0 is expected to be
sufficiently large enough to avoid false alarms. On the other hand, ARL1 is anticipated to be sufficiently
small to enable the process to send a signal as soon as there is a shift in the process parameter(s).

2.2. Variability in the Shewhart Chart Performance

In this section, we explain the effect of the practitioner to practitioner variability on the Shewhart
chart, both in normal and non-normal distribution, by using the Monte Carlo simulation approach.
See ([25–29]) for more information about the effect of sample size and practitioners’ variability.
To achieve this aim, we develop an algorithm in R programing language to simulate the Shewhart
chart environment, using the standard Shewhart chart as our benchmark and reference point. The X
chart has a control limits width determinant L that influences RL properties. We use the standard
L = 3, that corresponds to the ARL0 = 370 (see [1] for more details). Without any loss of generality,
we generate random samples from a standard normal distribution N (μ = 0, σ = 1), each of sample
size n = 5, assuming the process parameters are known. While for the non-normal distribution,
we considered the t-distribution with degrees of freedom v = 5, 25, and 100. Since all the three
categories of v exhibit the same pattern, we report only the results for v = 100. In both environments,
normal and t-distributions, we set up the chart limits as given in Equation (1) and plot the sample

37
Mathematics 2020, 8, 857

means against the UCL and LCL. As soon as a value of Y j is plotted outside the limits, RL is recorded
and saved. The process is iterated 105 times to get ARL and SDRL.
For the unknown parameters, we estimate the parameter from phase-I. The number of samples
employed for the estimation differs from on practitioner to another and so does the accuracy of the
charts in phase-II. To depict that, we estimated both μ0 and σ0 from different number of in-control
phase-I samples i.e., m = 25, 50, 100, 250, 500 and 1000 each of sample size n = 5. The estimated
parameters μ̂0 and σ̂0 from the phase-I IC stage are, therefore, used in the same algorithm instead of μ0
and σ0 respectively. Subsequently the parameter L, changes as the amount of phase-I samples changes.
The corresponding L s for the different m s are L = 2.962, 2.983, 2.9925, 2.997, 2.999, and 3 respectively
for the normal distribution, and L = 2.974, 2.995, 3.005, 3.010, 3.012, and 3.012 respectively for
the t-distribution of v = 100. These L’s are determined through simulations to obtain ARL0 = 370.
We carry out the simulation with different level of shifts δ ranging from 0 to 5 i.e., δ ∈ (0, 0.5, 5),
as shown in Tables 1 and 2.

Table 1. Average run length (ARL) of the Shewhart chart with estimated parameters for standard
normal and t (v = 100) distributions.

ARL Standard Normal Distribution T-Distribution v = 100

δ/m 25 50 100 250 500 1000 25 50 100 250 500 1000
0 370.93 369.66 370.43 369.09 371.10 370.98 370.67 370.00 369.58 370.89 369.63 370.00
0.5 190.76 173.92 165.35 159.25 157.00 156.54 194.06 177.76 167.85 161.75 160.48 159.04
1 53.91 48.74 46.37 44.73 44.41 44.21 55.14 50.13 47.75 46.26 46.02 45.32
1.5 17.19 16.05 15.52 15.20 15.09 15.04 17.70 16.53 16.06 15.65 15.65 15.52
2 6.87 6.59 6.46 6.35 6.33 6.33 7.05 6.80 6.65 6.57 6.54 6.57
2.5 3.40 3.33 3.29 3.26 3.25 3.25 3.52 3.44 3.40 3.36 3.35 3.35
3 2.05 2.03 2.02 2.01 2.00 2.00 2.10 2.08 2.06 2.06 2.05 2.05
3.5 1.47 1.46 1.45 1.45 1.45 1.45 1.50 1.49 1.48 1.48 1.48 1.47
4 1.20 1.19 1.19 1.19 1.19 1.19 1.21 1.21 1.21 1.21 1.20 1.20
4.5 1.08 1.08 1.07 1.07 1.07 1.07 1.09 1.08 1.08 1.08 1.08 1.08
5 1.03 1.02 1.02 1.02 1.02 1.02 1.03 1.03 1.03 1.03 1.03 1.03
L 2.962 2.983 2.9925 2.997 2.999 3 2.974 2.995 3.005 3.010 3.012 3.012

Table 2. Standard deviation of the run length (SDRL) of the Shewhart chart with estimated parameters
for standard normal and t (v = 100) distributions.

SDRL Standard Normal Distribution T-Distribution v = 100

δ/m 25 50 100 250 500 1000 25 50 100 250 500 1000
0 601.54 473.20 420.75 387.76 379.99 375.44 601.69 468.55 416.69 389.69 376.91 373.30
0.5 333.88 232.92 193.19 169.15 161.43 158.10 340.19 241.98 197.21 172.49 164.11 160.46
1 88.35 61.71 52.00 46.39 45.00 44.24 91.10 64.04 53.93 47.96 46.61 45.29
1.5 24.05 18.52 16.40 15.20 14.85 14.67 25.00 19.21 16.92 15.71 15.33 15.20
2 8.24 6.88 6.32 5.97 5.90 5.85 8.73 7.10 6.48 6.22 6.11 6.11
2.5 3.40 3.04 2.86 2.76 2.72 2.71 3.60 3.16 2.99 2.86 2.84 2.83
3 1.66 1.54 1.48 1.44 1.43 1.42 1.70 1.59 1.54 1.49 1.48 1.47
3.5 0.90 0.85 0.83 0.81 0.81 0.81 0.94 0.89 0.86 0.84 0.84 0.84
4 0.51 0.49 0.48 0.48 0.48 0.48 0.54 0.51 0.51 0.50 0.50 0.50
4.5 0.30 0.29 0.28 0.28 0.28 0.28 0.32 0.31 0.30 0.29 0.29 0.29
5 0.17 0.16 0.16 0.16 0.16 0.16 0.18 0.17 0.17 0.17 0.17 0.17
L 2.962 2.983 2.9925 2.997 2.999 3 2.974 2.995 3.005 3.010 3.012 3.012

2.3. Presence of Outliers in the Shewhart Chart with Estimated Parameters

Although the estimation of the unknown parameters in phase-I samples plays its role on the
efficiency of the control chart in phase-II. The drop in the efficacy of the chart performance is not limited
to this fact alone, rather it extends to presence of outlying/extreme values in the phase-I samples.

38
Mathematics 2020, 8, 857

In this Section, we study the effect of outliers in the phase-I samples on the performance and
accuracy of the Shewhart chart. Here, through Monte Carlo simulation, we generate the m phase-I
samples from a mixture distribution i.e., (1 − α)100% from assumed (normal or t-distribution) and
the remaining α100% from a chi-square distribution with n degrees of freedom denoted by χ2(n) .
Subsequently, the estimated parameters emerging from the m samples have an extreme values effect
on the control chart in phase-II. That is, each observation of the phase-I sample is generated from the
following expression:  
(1 − α)N μ, σ2 + α N μ, σ2 + w χ2(n) or
  (4)
(1 − α)t(v) + α t(v) + w χ2(n)

where α > 0, is the probability of having a multiple of χ2(n) added to the assumed distribution,
serving as the outliers in the samples. In addition, w ≥ 1 is the magnitude of the outlier. We develop an
algorithm from the R language, similar to that in Section 2.2, but the samples are from the environment
described in (4). We set μ = 0, σ2 = 1, v = 100, w = 3, and α [0, 0.01]. We design the Shewhart chart
using the same parameters L and m as in Section 2.2.
In general, the pattern exhibited by the RL properties implies the following:
• Increasing the m phase-I samples in the presence of outliers, gets the ARL0 ’s closer to the
theoretical values.
• Reducing the value of α, the percentage of outliers present in the m samples also brings the ARL0 ’s
closer to the theoretical values.
Unfortunately, neither of the two suggested remedies is practicable in real life. Thus, we propose
outliers detecting structures through the robust Turkey and MAD detection models.

2.4. Shewhart Chart with Outlier Detection Models

In the section, we propose two outlier-detecting models as remedy to the issues raised in
Sections 2.2 and 2.3. The Tukey and the MAD model-based Shewhart charts. Their procedures applied
in parallel to the Shewhart chart are described in the sub sections below:

2.4.1. The Tukey Shewhart Control Chart


 For the phase-I samples, Y be the median of all m × n observations. For any observation yo
 > p × IQR, then yo is declared an outlier. Here IQR = Q3 − Q1 is the inter-quartile range of
if  yo − Y
the sample. Q3 and Q1 are the third and first quartiles, respectively, of all m × n phase-I observations.
The constant p on the other hand is the confidence factor of the Tukey’s detector, commonly chosen
between 1.5 and 3.0. The confidence factor should be carefully chosen, and not too small, to avoid
over detection. Also it should not be too large, to prevent under detection [18]. In this study,
we choose p = 2.2. Applying the same algorithm, parameters and limits employed in Section 2.2,
we incorporate the Tukey outlier-detector model on the phase-I samples to screen out the extreme
values present there in. Then we compute the IC ARL and SDRL values for the Shewhart chart based
on the Tukey model in phase-II, when the parameters are estimated.

2.4.2. The Median Absolute Deviation (MAD) Shewhart Control Chart

We define median
 absolute
 deviation (MAD) as the deviation of the dataset about the median
 /0.6574. Then it follows, that any observation yo from the sample that falls
as MAD = median Yil − Y
outside the expression 
Y ± b ∗ MAD , is declared an outlier. Here b is the outlier detecting constant
and chosen 3.642 so that the percentage of screening by MAD is the same as Tukey. This has been done
to keep the comparison between two outlier detectors valid [19].
Furthermore, it is worth distinguishing between outlying and OoC sample points. The former
emerges from mphase-I samples, which are used to construct the control limits for the monitoring

39
Mathematics 2020, 8, 857

stage; phase-II; while the latter are the sample points that fall beyond the control limits in phase-II.
Therefore, the presence of outlying sample points in phase-I leads to wider control limits, rendering
the control charts less effective. A flowchart summarizing the procedure is depicted in Figure 1.

Figure 1. Flowchart of the procedures of proposed Shewhart control chart.

3. Results
In this section, we provide the results of the methodologies discussed in Section 2. These results
are presented in three folds, so is the discussion in the next section.

3.1. Practitioners’ Estimation Variability

Here, through the simulation results of the algorithm explained in Section 2.2, we observe the
variability that appears in the Shewhart control chart due to different choices of sample size m,
amongst practitioners. Tables 1 and 2 depict the Shewhart chart whose parameters, both mean and
variance, are estimated from m phase-I samples for both normal and non-normal distributions. It is
evident from the result, the effect of parameter estimation on the performance of the chart. The ARL0 s
when δ = 0, are clustering around the target 370 with their respective L’s. However, when δ  0,
we observe that the smaller m becomes, the less effective the Shewhart chart performance. The ARL1 ’s
are expected to be sufficiently small in order to detect any drift in the ongoing process, but as m gets
smaller, ARL1 ’s get bigger. Which implies the chart is less sensitive in identifying the presence of shifts
in the ongoing process early enough. Another noticeable effect of the parameter estimation on the
Shewhart chart is the decrement in the limits L, as m reduces. This should be recorded as an edge if the
corresponding phase-II charts detects shift earlier than when the parameters are known.

3.2. Effect of Outliers on the Shewhart Control Charts

In Tables 3 and 4, we present the simulation results of environment (4) discussed in Section 2.3.
From these results, the gross impact of outliers in the phase-I samples on the performance of the
Shewhart chart cannot be over emphasized. Having seen the pattern of the IC and OoC RL properties
in Tables 1 and 2, in order to save space, we restrict the performance evaluation to the IC RL properties.
That is, considering the case when δ = 0 only. From Tables 3 and 4, when α = 0, in the absence
of outlier, the ARL0 ’s are clustering around its target 370, irrespective of the amount of phase-I

40
Mathematics 2020, 8, 857

sample m. However, when α > 0, the ARL0 ’s deviate from the target, vigorously. As the amount of
phase-I samples m reduces, and the percentage of outliers present in the samples α increases, the more
the ARL0 ’s deviate from the target. Similarly the pattern of the SDRL, even more.

Table 3. ARL of the Shewhart chart in the presence of outliers with estimated parameters for standard
normal and t (v = 100) distributions.

ARL Standard Normal Distribution T-Distribution v = 100

α/m 25 50 100 250 500 1000 25 50 100 250 500 1000
0 371.20 369.51 370.29 369.59 370.72 370.85 370.65 369.57 369.28 370.85 369.11 370.87
0.001 424.76 401.49 392.98 389.92 388.74 389.13 418.21 394.60 386.19 385.76 385.04 381.04
0.002 484.38 435.14 417.35 410.86 409.61 408.88 478.03 424.90 408.83 405.04 403.56 398.94
0.003 540.09 469.66 440.42 431.91 429.47 428.70 534.47 458.93 435.77 424.67 424.75 422.44
0.004 602.97 507.51 469.33 452.76 450.30 449.72 587.12 500.66 457.98 443.64 444.15 437.50
0.005 664.58 545.85 494.67 477.03 471.98 471.09 650.79 528.62 489.31 469.04 461.71 461.22
0.006 731.55 588.72 525.81 500.69 495.46 493.34 718.10 569.85 512.77 491.40 486.42 481.51
0.007 788.48 629.88 554.41 525.13 519.74 516.88 758.54 604.64 539.10 514.95 509.78 504.97
0.008 861.51 673.43 586.11 551.83 544.06 541.56 837.89 652.55 578.41 541.64 531.29 528.75
0.009 931.14 726.12 618.88 579.38 569.55 565.85 916.99 693.94 606.71 568.85 560.07 552.75
0.01 996.35 773.78 654.18 607.70 596.56 592.55 953.61 745.34 634.00 593.84 586.09 575.08
L 2.962 2.983 2.9925 2.997 2.999 3 2.974 2.995 3.005 3.010 3.012 3.012

Table 4. SDRL of the Shewhart chart in the presence of outliers with estimated parameters for standard
normal and t (v = 100) distributions.

SDRL Standard Normal Distribution T-Distribution v = 100

α/m 25 50 100 250 500 1000 25 50 100 250 500 1000
0 607.93 473.75 419.35 388.27 379.67 375.44 595.89 468.00 418.16 382.43 375.42 369.72
0.001 1238.32 664.83 467.76 414.51 400.57 394.71 1150.79 635.34 463.74 411.63 395.24 385.68
0.002 1704.83 838.46 515.41 441.77 423.19 415.18 1802.06 760.16 508.22 433.04 416.23 406.03
0.003 2041.64 991.95 568.33 468.17 446.51 436.26 2096.30 913.63 558.95 460.42 441.81 431.04
0.004 2410.92 1176.26 630.64 497.30 469.60 459.66 2354.27 1184.84 606.26 486.65 462.63 447.94
0.005 2689.45 1294.24 687.44 528.00 494.90 481.60 2626.16 1192.76 683.26 517.24 485.89 473.05
0.006 2994.88 1502.07 756.39 559.42 521.35 505.08 2871.17 1449.52 699.03 546.74 512.57 494.30
0.007 3193.80 1637.48 822.18 592.02 549.29 530.76 2956.37 1514.95 757.01 573.97 540.59 519.76
0.008 3480.50 1799.40 888.29 627.68 578.43 557.35 3370.66 1664.42 912.65 618.19 564.09 544.14
0.009 3772.38 1994.86 959.76 664.49 607.34 582.18 3703.48 1766.87 955.78 655.56 594.72 571.46
0.01 4012.72 2139.41 1028.74 702.52 638.62 612.00 3823.24 2009.14 980.24 685.94 628.22 591.55
L 2.962 2.983 2.9925 2.997 2.999 3 2.974 2.995 3.005 3.010 3.012 3.012

3.3. Improvement of Tukey and MAD Outlier Detection Models on Shewhart Chart Performance
While incorporating the procedures in Sections 2.4.1 and 2.4.2, the simulation results are presented
in Tables 5–8 respectively. Tables 5 and 7 represents the ARL result for Tukey and MAD outlier detection
models respectively, as Tables 6 and 8 are the corresponding SDRL results. The effect of these detection
models are noticed as ARLs and SDRLs are closer to when there is an absence of outliers or even better.

41
Mathematics 2020, 8, 857

Table 5. ARL of the Shewhart chart with Tukey outlier detection for standard normal and t
(v = 100) distributions.

ARL Standard Normal Distribution T-Distribution v = 100

α/m 25 50 100 250 500 1000 25 50 100 250 500 1000
0 363.24 363.44 364.17 364.36 365.34 366.80 358.67 358.71 357.97 358.10 359.64 357.52
0.001 365.71 366.33 366.48 366.83 367.56 368.22 361.59 360.26 357.98 362.04 359.85 360.39
0.002 367.81 367.30 368.51 368.71 369.89 366.67 364.59 363.25 359.73 359.63 363.42 361.61
0.003 370.30 370.04 370.91 371.14 370.90 371.03 365.85 365.75 364.12 361.29 362.85 362.73
0.004 372.35 371.48 372.44 372.76 372.72 374.13 368.84 365.44 365.73 365.64 365.87 366.25
0.005 376.13 374.05 374.49 373.95 375.73 377.23 370.27 368.13 366.95 366.59 366.89 367.05
0.006 377.97 376.06 376.96 376.61 376.88 379.44 372.27 368.44 368.28 371.20 371.40 370.77
0.007 380.13 378.38 379.03 378.98 379.15 379.44 374.78 372.44 371.54 370.82 370.45 371.76
0.008 383.48 380.17 380.51 380.63 380.56 381.29 378.04 372.69 372.06 373.31 376.35 374.45
0.009 386.35 383.02 383.50 382.73 383.29 382.84 377.58 374.74 371.66 375.88 375.09 372.29
0.01 388.82 384.62 385.57 383.80 384.95 386.39 381.84 377.80 378.17 377.29 378.48 374.09
L 2.962 2.983 2.9925 2.997 2.999 3 2.974 2.995 3.005 3.010 3.012 3.012

Table 6. SDRL of the Shewhart chart with Tukey outlier detection for standard normal and t
(v = 100) distributions.

SDRL Standard Normal Distribution T-Distribution v = 100

α/m 25 50 100 250 500 1000 25 50 100 250 500 1000
0 596.32 468.02 414.04 383.16 374.68 368.66 607.33 467.10 411.17 377.81 369.81 360.34
0.001 600.30 472.43 417.55 385.96 377.25 372.35 612.79 467.20 406.53 380.03 366.57 366.34
0.002 607.28 476.91 419.54 388.33 379.39 371.64 627.84 471.38 410.10 376.44 372.50 365.67
0.003 615.99 477.95 422.41 389.95 380.10 375.45 617.88 474.03 415.42 378.88 374.34 366.23
0.004 615.39 481.58 423.16 392.96 382.24 377.77 636.02 476.02 416.09 385.14 377.04 371.24
0.005 625.95 486.00 427.29 393.54 385.71 380.37 650.31 480.25 420.69 387.36 377.24 370.45
0.006 636.60 488.82 430.22 396.23 386.61 383.77 637.39 483.89 423.95 390.68 383.17 376.49
0.007 646.05 490.89 433.11 399.86 389.30 384.98 630.84 488.73 429.10 390.52 381.42 377.88
0.008 656.27 496.12 435.44 400.84 391.44 385.54 654.64 489.53 430.19 391.87 386.60 378.02
0.009 654.98 500.43 439.24 403.24 393.19 387.01 675.27 499.78 423.47 395.23 385.19 380.11
0.01 669.83 502.93 441.89 404.61 395.16 390.51 672.71 493.77 436.38 399.23 390.17 380.26
L 2.962 2.983 2.9925 2.997 2.999 3 2.974 2.995 3.005 3.010 3.012 3.012

Table 7. ARL of the Shewhart chart with median absolute deviation (MAD) outlier detection for
standard normal and t (v = 100) distributions.

ARL Standard Normal Distribution T-Distribution v = 100

α/m 25 50 100 250 500 1000 25 50 100 250 500 1000
0 363.24 363.44 364.17 364.36 365.34 366.80 358.67 358.71 357.97 358.10 359.64 357.52
0.001 365.71 366.33 366.48 366.83 367.56 368.22 361.59 360.26 357.98 362.04 359.85 360.39
0.002 367.81 367.30 368.51 368.71 369.89 366.67 364.59 363.25 359.73 359.63 363.42 361.61
0.003 370.30 370.04 370.91 371.14 370.90 371.03 365.85 365.75 364.12 361.29 362.85 362.73
0.004 372.35 371.48 372.44 372.76 372.72 374.13 368.84 365.44 365.73 365.64 365.87 366.25
0.005 376.13 374.05 374.49 373.95 375.73 377.23 370.27 368.13 366.95 366.59 366.89 367.05
0.006 377.97 376.06 376.96 376.61 376.88 379.44 372.27 368.44 368.28 371.20 371.40 370.77
0.007 380.13 378.38 379.03 378.98 379.15 379.44 374.78 372.44 371.54 370.82 370.45 371.76
0.008 383.48 380.17 380.51 380.63 380.56 381.29 378.04 372.69 372.06 373.31 376.35 374.45
0.009 386.35 383.02 383.50 382.73 383.29 382.84 377.58 374.74 371.66 375.88 375.09 372.29
0.01 388.82 384.62 385.57 383.80 384.95 386.39 381.84 377.80 378.17 377.29 378.48 374.09
L 2.962 2.983 2.9925 2.997 2.999 3 2.974 2.995 3.005 3.010 3.012 3.012

42
Mathematics 2020, 8, 857

Table 8. SDRL of the Shewhart chart with MAD outlier detection for standard normal and t
(v = 100) distributions.

SDRL Standard Normal Distribution T-Distribution v = 100

α/m 25 50 100 250 500 1000 25 50 100 250 500 1000
0 596.32 468.02 414.04 383.16 374.68 368.66 607.33 467.10 411.17 377.81 369.81 360.34
0.001 600.30 472.43 417.55 385.96 377.25 372.35 612.79 467.20 406.53 380.03 366.57 366.34
0.002 607.28 476.91 419.54 388.33 379.39 371.64 627.84 471.38 410.10 376.44 372.50 365.67
0.003 615.99 477.95 422.41 389.95 380.10 375.45 617.88 474.03 415.42 378.88 374.34 366.23
0.004 615.39 481.58 423.16 392.96 382.24 377.77 636.02 476.02 416.09 385.14 377.04 371.24
0.005 625.95 486.00 427.29 393.54 385.71 380.37 650.31 480.25 420.69 387.36 377.24 370.45
0.006 636.60 488.82 430.22 396.23 386.61 383.77 637.39 483.89 423.95 390.68 383.17 376.49
0.007 646.05 490.89 433.11 399.86 389.30 384.98 630.84 488.73 429.10 390.52 381.42 377.88
0.008 656.27 496.12 435.44 400.84 391.44 385.54 654.64 489.53 430.19 391.87 386.60 378.02
0.009 654.98 500.43 439.24 403.24 393.19 387.01 675.27 499.78 423.47 395.23 385.19 380.11
0.01 669.83 502.93 441.89 404.61 395.16 390.51 672.71 493.77 436.38 399.23 390.17 380.26
L 2.962 2.983 2.9925 2.997 2.999 3 2.974 2.995 3.005 3.010 3.012 3.012

For better visuals of the results, we depict the ARL results (Tables 3, 5 and 7) in Figures 2 and 3
and the SDRL results (Tables 4, 6 and 8) in Figures 4 and 5.

Figure 2. In-control ARL values for the Shewhart chart from standard normal distribution in the
presence of outliers with and without outlier screening.

43
Mathematics 2020, 8, 857

Figure 3. In-control ARL values for the Shewhart chart from t-(v = 100) distribution in the presence of
outliers with and without outlier screening.

Figure 4. In-control SDRL values for the Shewhart chart from standard normal distribution in the
presence of outliers with and without outlier screening.

44
Mathematics 2020, 8, 857

Figure 5. In-control SDRL values for the Shewhart chart from t-(v = 100) distribution in the presence of
outliers with and without outlier screening.

4. Discussion
We summarize the findings of the study under the following subsections: (a) parameter estimation
effect on the Shewhart control chart, (b) effect of outliers on Shewhart X chart performance, and (c)
improvement of outliers screening models on the Shewhart X chart performance. Through the
discussion, we use the run length properties as a yardstick for measuring the performance of the charts.

4.1. Parameter Estimation Effect on the Shewhart X Control Chart

Theoretically, when the Shewhart charts parameters are known, the limit L corresponding to the
IC ARL0 = 370 is L = 3. When the parameters are estimated from phase-I samples, the first effect
of the estimation is the change in L. The control limit L deviates from its theoretical value as much
as the sample size m reduces. That implies, the smaller the sample size m, the farther the control
limit from the theoretical value. This is noticeable in Tables 1 and 2, as L changes as the sample size
does. We compute Ls based on 100,000 iterations of simulation. Secondly, in the introduction of shifts,
which makes the process OC, the RL properties values of the estimated parameters are bigger than
the theoretical values. This indicates that the chart with estimated parameters are slower in detecting
shifts in the process as compared to the chart with known parameters. For instance, (cf. Tables 1 and 2),
with m = 1000, δ = 0.5 the resulting ARL1 and SDRL1 are 156.42 and 158.84 for normal distribution
and 150.92 and 160.46 for t-distribution respectively. However, with m = 25, δ = 0.5 ARL1 and SDRL1
are 190.12 and 333.88 for normal distribution and 194.06 and 340.19 for t-distribution respectively.

4.2. Effect of Outliers on Shewhart X Control Chart performance

Haven noticed the effect of parameter estimation on Shewhart chart performance; one major
cause could be the presence of outliers in the dataset. The results in Tables 3 and 4 prove that extreme
values in the sample causes great havoc to the performance of the process. As discussed earlier in
Section 4, α = 0 indicates absence of outliers, and the presence of outliers if otherwise. We observe

45
Mathematics 2020, 8, 857

jumps in the values of IC ARL and SDRL from Tables 3 and 4. With different combinations of α and m,
we say the bigger the value of α and the smaller the value of m, the gross the effect of the outliers on the
chart. Take for instance, in the normal environment, the ARL and SDRL values of just 1% of outliers
(α = 0.01) for when m = 1000 as against when m = 25. It shocks to see the ARL and SDRL jumped
from 592.55 and 612.00 to 996.3 and 4012.72 respectively. However, in the t-distribution, ARL and
SDRL values of 1% of outliers (α = 0.01) for when m = 1000 as against when m = 25, are 575.08 and
591.55 to 953.61 and 3823.24 respectively.

4.3. Improvement of Outliers Screening Models on Shewhart Chart Performance

The proposed remedy for the effect of outliers on the Shewhart chart works perfectly.
The incorporation of Tukey and MAD outlier-screening models in the Shewhart chart normalizes the
outlier effects and restores the performance even much better than it was. To access the effect
of these two screening methods, we present Figures 2–5, displaying the IC ARL values with
m = 25, 50, 100, 250, 500 and 1000, and the magnitude w = 3, without outliers screening, alongside the
IC ARL whose outliers are screened with the Tukey and MAD-based models. The IC ARL that are
supposed to be around the target 370 has jumped to more than 250% increment due to the effect of
outliers. However, with our proposed screening models; both Tukey and MAD-based models; the IC
ARL is returned back to its target with less than 5% increment and decrement. The IC SDRL also
exhibits the same pattern; in fact, its improvement is more appreciable as compared to the ARL’s.

5. Illustrative Example
In the manufacturing industry, semiconductor lithography (photolithography) refers to the
formation of three-dimensional images on the substrate for subsequent transfer of the pattern to the
substrate. A keynote aspect of this process is the bake process, both the pre (soft)-bake and post
(hard)-bake. In this section, we implement the Shewhart chart with the proposed outlier detection
models on the flow width measurement of a hard bake process. In the subsequent subsections, we give
a brief overview of the hard-bake process and then application of the Shewhart chart on the dataset
extracted from such a process (the Basics of Microlithography n.d.).

5.1. The Post (Hard) Bake Process

A typical photolithography process consist of the following sequence of operation:
substrate preparation, photoresist spin coat, pre-bake, exposure, post-exposure bake, development
and finally the post-bake. The hard-bake process, as the name implies, is used to harden the final
resist image so that it will withstand the harsh environments of etching. This post-bake ensures
complete removal of solvent, improving adhesion in wet etch processes and resistance to plasma
etches. Practitioners use different temperatures depending on the material under study. However,
the temperature should be carefully chosen and not more than 200 ◦ C. A major characteristic of
this process is the wafer. Recall that the word lithography is a combination of two Greek words:
lithos meaning stones and graphia, meaning to write. Our stones in this case are silicon wafers and the
patterns are written with photoresist, which are sensitive polymers. Figures 6 and 7 depict a typical
photolithography flowchart and the hard-bake process.

46
Mathematics 2020, 8, 857

Figure 6. A flowchart of a photolithography process of semiconductor manufacturing industry.

Figure 7. Illustration of hard-bake process.

5.2. Application of Shewhart Control Charts with Outlier

In this section, we implement the findings of this study on a set of data generated from a
semiconductor manufacturing of a hard-bake process, which monitors the flow width measurement
of wafers [1]. The variable of interest is the flow width measurement (in microns) for the hard-brake
process. The data consist of 25 IC phase-I samples and 10 phase-II samples each of sample size 5.
The process mean and standard deviation of the phase-I samples are 16.7163 and 3.5167, respectively.
Therefore, we use these estimates to setup Shewhart chart control limits for monitoring phase-II
samples. Figure 8 shows all phase-I sample points staying within the limits and 3 of the phase-II
sample points stretching beyond the LCL making them OoC due to some assignable cause of variation.
Prior to setting the limits, we test the data for possible autocorrelation. The data is autocorrelation-free
as the Durbin–Watson (DW) test result proves. The value of the DW test statistics is DW = 1.7564 and the
critical values at 1% level of significance are dL = 1.19, and dU = 1.31. By the interpretation explained in
Table 9, we fail to reject the null hypothesis and conclude that there is no evidence of autocorrelation in
the data.

47
Mathematics 2020, 8, 857

Figure 8. Scatter plot of phase-I sample and the Shewhart chart with estimated parameters.

Table 9. Interpretation of Durbin–Watson autocorrelation test.

Categories Decision Rules Decisions

0 ↔ dL = 1.19 Reject H0 : positive autocorrelation
dL = 1.19 ↔ dU = 1.31 Inconclusive
dU = 1.31 4 ↔ dL = 2.69 Do not reject H0 : no autocorrelation 1.31 < DW = 1.7564 < 2.69
4 − dL = 2.69 ↔ 4 − dU = 2.81 Inconclusive
4 − dU = 2.81 ↔ 4 Reject H0 : negative autocorrelation

Furthermore, we introduce a 5% of outliers to the phase-I samples, to illustrate the argument that
the presence of outliers affects the performance of control charts. This subsequently increased the
mean and standard deviation by 4% and 25% respectively resulting to an increased UCL and decreased
LCL. The changes in the control limits implies a wider range of the boundaries. Therefore the resulting
control charts is less efficient as compared to the previous one without outliers. Figure 9 depicts this.

Figure 9. Scatter plot of phase-I sample and the resulting Shewhart chart with estimated parameters
and 5% of outliers with magnitude 3.

48
Mathematics 2020, 8, 857

5.3. Application of Shewhart Outlier Detection Model

Having established the deficiency of the Shewhart chart with outliers on the dataset; we employ
our proposed outlier detection model with the Shewhart chart explained in Section 2.4 to rectify this
shortcoming. Figure 10 shows the application of the Shewhart Tukey-based model. It is evident
there in that the chart was not only able to restore the efficiency of the chart as there were no outliers,
detecting 3 OoC sample points, but also to identify the outliers in the phase-I sample points. Similarly,
Figure 11 portrays the scenario when the Shewhart MAD-based model is applied on the monitoring
stage. Despite the presence of outlier in the dataset, the chart is able to detect the OC sample points as
much as it does when there were no outliers.

Figure 10. Scatter plot of phase-I sample and the resulting Shewhart chart with Tukey outlier
detection screening.

Figure 11. Scatter plot of phase-I sample and the resulting Shewhart chart with MAD-outlier
detection screening.

6. Conclusions
In this article, we evaluate the performance of the Shewhart control chart for location monitoring
with estimated parameters. The study substantiates the effect of estimation error and the variability
in the practitioners’ choice of phase-I samples on the chart, especially when the samples are prone
to outliers. Increasing the phase-I sample size (although not practicably) will to some extent reduce
the gross impact on the Shewhart chart. The results of this study further prove that incorporation of

49
Mathematics 2020, 8, 857

the non-parametric outlier screening models, Tukey and MAD, in the design of the Shewhart chart is
more practicable as it requires less phase-I samples and yields better results. Another advantage of this
study lies in the simplicity of its design and ease of usage. The study rounds up with an illustrative
example with a photolithography real data. A comparison of the two detection models, Tukey and
MAD, reveals that duo relatively efficient. The study is limited to operate within the univariate setup,
while focusing on multivariate setup will be a great advantage and we plan a future study for that. Also,
proposed charts are memory-less, which implies they are suitable for monitoring large shift. However,
the idea of the study is not only applicable in Shewhart multivariate setup, but also extendable to
other control charts, like exponentially weighted moving average and cumulative sum charts both
univariate and multivariate setups.

Author Contributions: Conceptualization, I.A.R. and N.A.; Data curation, I.A.R. and M.R.A.; Investigation,
M.H.L.; Methodology, N.A.; Resources, M.H.L.; Supervision, M.R. and M.R.A.; Writing—original draft, I.A.R.;
Writing—review & editing, M.H.L., M.R. and N.A. All authors have read and agreed to the published version of
the manuscript.
Funding: This study was partially funded by the Ministry of Higher Education, Malaysia through the
Industry-International Incentive Grant for Universiti Teknologi Malaysia (grant number 01M30).
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Montgomery, D.C. Introduction to Statistical Quality Control, 6th ed.; John Wiley & Sons Inc.: Hoboken, NJ,
USA, 2009.
2. Chen, G. The mean and standard deviation of the run length distribution of X charts when control limits are
estimated. Stat. Sin. 1997, 7, 789–798.
3. Jensen, W.A.; Jones-Farmer, L.A.; Champ, C.W.; Woodall, W.H. Effects of parameter estimation on control
chart properties: A literature review. J. Qual. Technol. 2006, 38, 349–364. [CrossRef]
4. Psarakis, S.; Vyniou, A.K.; Castagliola, P. Some Recent Developments on the Effects of Parameter Estimation
on Control Charts. Qual. Reliab. Eng. Int. 2014, 30, 1113–1129. [CrossRef]
5. Saleh, N.A.; Mahmoud, M.A.; Keefe, M.J.; Woodall, W.H. The difficulty in designing shewhart X and X
control charts with estimated parameters. J. Qual. Technol. 2015, 47, 127–138. [CrossRef]
6. Goedhart, R.; Da Silva, M.M.; Schoonhoven, M.; Epprecht, E.K.; Chakraborti, S.; DoesA, R.J.M.M.
Veiga Shewhart control charts for dispersion adjusted for parameter estimation. IISE Trans. 2017, 49,
838–848. [CrossRef]
7. Dasdemir, E.; Weiß, C.; Testik, M.C.; Knoth, S. Evaluation of Phase I analysis scenarios on Phase II performance
of control charts for autocorrelated observations. Qual. Eng. 2016, 28, 293–304. [CrossRef]
8. Anscombe, F.J. Rejection of Outliers. Technometrics 1960, 2, 123–146. [CrossRef]
9. Grubbs, F.E.; Beck, G. Extension of Sample Sizes and Percentage Points for Significance Tests of Outlying
Observations. Technometrics 1972, 14, 847–854. [CrossRef]
10. Hawkins, D.M. Identification of Outliers; Chapman and Hall: London, UK, 1980.
11. Beckman, R.J.; Cook, R.D. Outlier . . . . . . . . . . s. Technometrics 1983, 25, 119–149.
12. Barnett, V.; Lewis, T. Outliers in Statistical Data. 3rd edition. J. Wiley & Sons 1994, XVII. 582 pp., £49.95.
Biom. J. 1995, 37, 256.
13. Torng, C.-C.; Lee, P.-H. ARL Performance of the Tukey’s Control Chart. Commun. Stat. Simul. Comput. 2008,
37, 1904–1913. [CrossRef]
14. Tercero-Gomez, V.; Ramirez-Galindo, J.; Cordero-Franco, A.; Smith, M.; Beruvides, M. Modified Tukey’s
Control Chart. Commun. Stat. Simul. Comput. 2012, 41, 1566–1579. [CrossRef]
15. Fan, S.-K.S.; Huang, H.-K.; Chang, Y.-J. Robust Multivariate Control Chart for Outlier Detection Using
Hierarchical Cluster Tree in SW2. Qual. Reliab. Eng. Int. 2013, 29, 971–985. [CrossRef]
16. Khaliq, Q.-U.-A.; Riaz, M. Robust Tukey-CUSUM Control Chart for Process Monitoring. Qual. Reliab. Eng. Int.
2016, 32, 933–948. [CrossRef]

50
Mathematics 2020, 8, 857

17. Fitriyah, H.; Budi, A.S. Outlier Detection in Object Counting based on Hue and Distance Transform using
Median Absolute Deviation (MAD). In Proceedings of the 2019 4th International Conference on Sustainable
Information Engineering and Technology, Lombok, West Nusa Tenggara, Indonesia, 28–30 September 2019;
pp. 217–222.
18. Abbas, N. A robust S2 control chart with Tukey’s and MAD outlier detectors. Qual. Reliab. Eng. Int. 2020, 36,
403–413. [CrossRef]
19. Abbas, N.; Abujiya, M.R.; Riaz, M.; Mahmood, T. Cumulative Sum Chart Modeled under the Presence of
Outliers. Mathematics 2020, 8, 269. [CrossRef]
20. Dequeant, K.; Vialletelle, P.; Lemaire, P.; Espinouse, M.L. A literature review on variability in semiconductor
manufacturing: The next forward leap to Industry 4.0. In Proceedings of the Winter Simulation, Arlington,
VA, USA, 11–14 December 2016; pp. 2598–2609.
21. Spanos, C.J. Statistical Process Control in Semiconductor Manufacturing. Proc. IEEE 1992, 80, 819–830.
[CrossRef]
22. Huh, I. Multivariate EWMA Control Chart and Application to a Semiconductor Manufacturing Process.
Ph.D. Thesis, Department of Statistics, McMaster University, Hamilton, ON, Canada, 2010.
23. Baud-Lavigne, B.; Bassetto, S.; Penz, B. A broader view of the economic design of the X-bar chart in the
semiconductor industry. Int. J. Prod. Res. 2010, 48, 5843–5857. [CrossRef]
24. Higashide, M.; Nishina, K.; Kawamura, H.; Ishii, N. Statistical process control for semiconductor
manufacturing processes. Front. Stat. Qual. Control 2010, 9, 71–84.
25. Quesenberry, C.P. The Effect of Sample Size on Estimated Limits for X and X Control Charts. J. Qual. Technol.
1993, 25, 237–247. [CrossRef]
26. Albers, W.; Kallenberg, W.C.M. Estimation in Shewhart control charts: Effects and corrections. Metrika 2004,
59, 207–234. [CrossRef]
27. Dovoedo, Y.H.; Chakraborti, S. Effects of Parameter Estimation on the Multivariate Distribution-free Phase II
Sign EWMA Chart. Qual. Reliab. Eng. Int. 2017, 33, 431–449. [CrossRef]
28. Maleki, M.R.; Castagliola, P.; Amiri, A.; Khoo, M.B.C. The effect of parameter estimation on phase II
monitoring of poisson regression profiles. Commun. Stat. Simul. Comput. 2019, 48, 1964–1978. [CrossRef]
29. Jardim, F.S.; Chakraborti, S.; Epprecht, E.K. Two perspectives for designing a phase II control chart with
estimated parameters: The case of the Shewhart X Chart. J. Qual. Technol. 2020, 52, 198–217. [CrossRef]

© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).

51
 mathematics
Article
A Functional Data Analysis for Assessing the Impact
of a Retrofitting in the Energy Performance of
a Building
Miguel Martínez Comesaña 1, *, Sandra Martínez Mariño 1 , Pablo Eguía Oller 1 ,
Enrique Granada Álvarez 1 and Aitor Erkoreka González 2
1 Department of Mechanical Engineering, Heat Engines and Fluid Mechanics, Industrial Engineering School,
University of Vigo, Maxwell s/n, 36310 Vigo, Spain; [email protected] (S.M.M.);
[email protected] (P.E.O.); [email protected] (E.G.Á.)
2 ENEDI Research Group, Department of Thermal Engineering, University of the Basque Country,
48013 Bilbao, Spain; [email protected]
* Correspondence: [email protected]

Received: 13 February 2020; Accepted: 19 March 2020; Published: 7 April 2020

Abstract: There is an increasing interest in reducing the energy consumption in buildings and in
improving their energy efficiency. Building retrofitting is the employed solution for enhancing
the energy efficiency in existing buildings. However, the actual performance after retrofitting
should be analysed to check the effectiveness of the energy conservation measures. The aim
of this work was to detect and to quantify the impact that a retrofitting had in the electrical
consumption, heating demands, lighting and temperatures of a building located in the north of Spain.
The methodology employed is the application of Functional Data Analyses (FDA) in comparison
with classic mathematical techniques such as the Analysis of Variance (ANOVA). The methods
that are commonly used for assessing building refurbishment are based on vectorial approaches.
The novelty of this work is the application of FDA for assessing the energy performance of renovated
buildings. The study proves that more accurate and realistic results are obtained working with
correlated datasets than with independently distributed observations of classical methods. Moreover,
the electrical savings reached values of more than 70% and the heating demands were reduced more
than 15% for all floors in the building.

Keywords: retrofitting; refurbishment; functional data analysis; vectorial analysis; energy efficiency

1. Introduction
The building sector is considered the largest energy consumer in the European Union,
representing 40% of the final energy consumption [1]. Globally, the energy consumption of this
sector accounts for 20% of the total delivered energy [2]. Thus, there is increasing interest in improving
the energy efficiency of buildings [3–5]. Furthermore, the potential of saving energy by renovation
in Europe is considerable as two-thirds of European buildings were constructed before 1980 [6].
Building retrofitting can contribute to reduce the energy consumption of existing buildings with lower
energy efficiencies. In this context, it is important to develop methodologies that can evaluate the actual
impact of refurbishment on renovated buildings in terms of energy consumption, thermal comfort
and lighting.
Therefore, building retrofitting is essential to prove the effectiveness of the applied energy
conservation measures to check if the energy efficiency of the building has certainly improved.
However, there are very few studies that actually evaluate the retrofit of buildings [7]. Most of
the studies evaluate measures for building refurbishment based on energy simulation [8–10],

Mathematics 2020, 8, 547; doi:10.3390/math8040547 53 www.mdpi.com/journal/mathematics

Mathematics 2020, 8, 547

mathematical models [11], artificial neural networks (ANN) [12] and building information modelling
(BIM) [13]. Thus, most studies analyse the energy conservation measures based on model outputs and
not proving the real effect on the building with monitored data. There is a performance gap between
simulated and measured energy consumption. Some studies have found that the calculated heating
energy consumption levels in the design phase were much lower than the measured values [14].
Some authors have evaluated the effect of refurbishment and renovations made in buildings with
real data. Ardente et al. [15] presented the results of an energy and environmental assessment of retrofit
actions implemented in six public buildings by using life cycle analysis (LCA). It is a common approach
to evaluate the decrease of energy consumption, operational cost and environmental impact in building
retrofitting [16–18]. Another approach for energy diagnosing of existing buildings is U-value in-situ
measurement [19] that characterises the heat losses through the building envelope and that can be
used to evaluate retrofitting actions [20]. Zavadskas et al. [21] proposed an approach to assess indoor
environmental conditions before and after retrofitting of dwellings with multiplicative optimality
criteria and experimental data. Hamburg et al. [14] analysed how well the energy performance targets
of building refurbishment are reached by collecting energy consumption and indoor measurements
after renovation and constructing simulations.
All of the previous presented methodologies to evaluate the retrofitting made in buildings use
vector-based data approaches. The presented methodology in this work for evaluating the impact of
a building retrofitting is based on functional analysis of monitored data before and after retrofitting.
By comparison of vector and functional analysis, we demonstrate that functional analysis provides
more realistic and accurate evaluations of the studied variables.
The methods for assessing building renovation have been applying vectorial analysis to the data.
These methods do not take into account the observations within a day as a set when evaluating the
daily behaviour of the data. As a consequence, the correlation between observations is missed.
In this context, Functional Data Analysis (FDA) can be useful because it is able to detect days
that do not have individual outliers, but may be far from the mean behaviour [22–26]. A proof
of its application is that FDA has expanded to a great number of scientific fields related with
continuous-time monitoring processes such as the environment [24,26–29], health and medical
research [30,31], industrial processes [32,33], sensor technology [34,35] or even econometrics [36].
Moreover, it has also been applied with machine learning techniques in optimisation and classification
problems [37,38]. Nowadays, FDA continues to expand its applications in more fields such as quality
control or sports [39,40]
In this work, we propose the use of FDA for assessing the impact that building retrofitting has in
the energy performance, indoor temperatures and lighting conditions of a building. The methodology
was applied to a case study of the renovated building of the Rectorate of the University of the Basque
Country (Spain). The novelty of this work is the application of FDA to statistically contrast the
differences in the energy performance of a building before and after a retrofitting. The literature
review shows that just few studies actually evaluated the retrofit of buildings with monitored data
and functional analysis was not used for this application.
The samples are composed of daily curves of variables such as electricity, heating demands and
temperatures. FDA allows making the contrast between samples taking into account the average
behaviour of the group throughout all day [24,37,41], which would not be possible with a vectorial
approach. In vectorial analysis, the data of a whole day have to be summarised in a single value to
work with daily observations. To classify a day as outlier, it has to move away from the sample mean,
in this case calculated with simplified daily observations [42,43].
On the one hand, to carry out the functional analysis, a functional ANOVA (FANOVA) was used
to evaluate whether there are differences between monitored data in the building before and after
retrofitting. On the other hand, a classical analysis of variance (ANOVA) was also used to study the
differences between the samples before and after the retrofitting [43–46]. To complement the vectorial
analysis, Kruskal’s non-parametric test was applied to contrast if the two samples come from the

54
Mathematics 2020, 8, 547

same initial distribution [47–49]. In addition, some variables such as the rate of change in sample
variance or the functional L2 distance between curves are also presented to measure the impact of
the refurbishment. The vectorial method is based on the differences between the medians, and the
functional method consists of measuring the distance between the curves that represent the mean
functions [50–52].
The results show that FDA efficiently demonstrates that the heating demands in the building were
reduced thanks to the envelope insulation, although ventilation was increased, indoor temperatures
were increased and internal lighting loads were reduced. The results also show that a significant
reduction in lighting consumption was achieved with the installation of LED lighting. Moreover, it is
demonstrated that taking into account the correlation of the data from a functional approach is a more
realistic and informative way to study how different two or more samples are.

2. Materials and Methods

2.1. Functional Data Analysis (FDA)

Functional Data Analysis (FDA) studies observations which form functions defined over
a determined set T. The infinite-dimensional structure of the data enlarges the possibilities of
research [23,27,53]. A random variable X is defined such a functional variable if it takes values
np
in a complete metric or semi-metric space and is observed in a discrete set of points {t j } j=1 ∈ [ a, b]
(not necessarily equispaced) for each of the n individuals studied [54,55]. Thus, the data consist of a X
matrix with n rows representing the different individuals and n p columns representing the different
discrete points where the functions are evaluated [55].
The functional model, through a process known as smoothing, converts the initial discrete values
into a set of continuous functions over time x (t) ∈ X ⊂ F , being F a functional space. To estimate
these functions, F is F = span{φ1 , ..., φnb }, where φk is a base function and nb the number of basis
functions necessary to build a functional sample. Although there are other types, the basis functions
used commonly are spline or Fourier functions [56], and the expansion considered is [24,25,27,57]:
nb
x (t) = ∑ ck φk (t) (1)
k =1
n
where {ck }k=
b
1 represent the coefficients that shape the function x ( t ) with respect to the chosen set of
basis functions. In this way, the smoothing process consist on solving the following regularisation
problem [24,25,27,57]:
np
min ∑ {z j − x (t j )}2 + λΓ( x ) (2)
x∈ F
j =1

where z j = x (t j ) + j (being j a value of the zero-mean random noise) is the result of observing x at
the point t j , λ a regularisation parameter that controls the intensity of the regularisation, and Γ an
operator that penalises the complexity of the solution. Taking into account the expansion, Equations (1)
and (2) can be expressed as [24,25,27,57]:

min{(z − Φc) T (z − Φc) + λc T Rc} (3)

c

where z = (z1 , ..., zn p ) T is the observation vector, c = (c1 , ..., cnb ) T the vector coefficients of the
functional expansion, Φ the n p × nb matrix of Φ jk = φk (t j ) elements, and R the matrix formed by nb
× nb elements [24,25,57,58]:

Rkl = D2 φk , D2 φl L2 ( l ) = D2 φk (t) D2 φl (t)dt (4)

T

55
Mathematics 2020, 8, 547

where D n φk (t) represents the nth-order differential operator of the function φk . After this, it is easy to
know that the solution can be calculated as follows:

c = (Φt Φ + λR)−1 Φ T z (5)

2.1.1. Functional Depths

Initially, the depth concept appeared in the multivariate statistics for measuring the centrality
of a point x ∈ Rd within a specific dataset: giving greater value to the points near the center [22].
Then, some authors extended this measured to FDA [59,60]. The functional depth give us a centrality
measure of a specific curve xi with respect a set of curves x1 , ..., xn that comes from a stochastic process
X (·) in a defined interval [ a, b] ∈ R.
There are several functional depths in the statistical literature, but the three main are:
Fraiman–Muniz [59], h-modal [61] and Random Projections [60]. The most used is the h-modal depth
because it has a better frequency of correct detection than the others [22]. The h-modal depth defines the
functional mode as the curve most densely surrounded by other curves of the dataset. In this manner,
the functional depth of a curve xi with respect the other curves in the sample is given by:
 
n
|| xi − xk ||
MDn ( xi , h) = ∑ K
h
(6)
k =1

with || · || being a norm in the functional space, K : R+ → R+ a kernel function, and h a bandwidth
parameter [61]. Thus, the curve that gets the maximum value in Equation (6) is considered the
functional mode. Moreover, some authors [60,61] recommended the use of L2 (l ) norms and a truncated
Gaussian kernel:
 b
1/2  2
2 t
|| xi − xk ||2 = ( xi (t) − xk (t))2 dt K (t) = √ exp − ,t > 0 (7)
a 2π 2

The principal aim of functional depths, viewed as functional dispersion measure, is the detection
of outliers. As in the classical analysis, detecting and examining these curves is important because
they may bias our functional estimations and because it allows us to discover the reasons that make
these curves deviate from the mean. Furthermore, from a functional approach, it is essential because it
may occur that the individual values of a curve are not outliers vectorially, but, instead, the complete
curve is a functional outlier [22,58]. If we assume that every curve in the data come from the same
stochastic process, a curve would be considered such an outlier for two reasons: it is at a significant
distance from the expected function of the stochastic process or its shape represents a very different
behaviour from the other curves. Therefore, the curves with functional depth below a specific C value
would be considered atypical and would be removed from the sample (see [23–26]). On the other
hand, it would be convenient to choose a C that provides a controlled type I error level. It should be
a value that, in absence of outliers, the probability of mislabelling a correct data as outlier would be
approximately a 1% [23–26]:

P( Dn ( xi ) ≤ C ) = 0.01, i = 1, ..., n (8)

In this way, the chosen C will be first percentile of the depths distribution chosen. Since this
distribution is unknown this, percentile must be estimated using the sample data. For this purpose,
there are two different bootstrap techniques: trimming bootstrap [62] and weighting bootstrap [63].
Some studies demonstrate that, despite having a larger incorrect outlier detection, the trimming
bootstrap has a better performance detecting the curves that are actually outliers [22,64].

56
Mathematics 2020, 8, 547

2.1.2. Functional Test ANOVA (FANOVA)

Any test or contrast that can be made in a vectorial analysis can have a functional version
that usually provides more relevant information. An example of this is the classical ANOVA.
Its functional version, although also contrasting the mean levels of a variable, is based on k
independent samples Xij (t), j = 1, ..., ni t ∈ [ a, b] drawn from L2 (l ) processes Xi , i = 1, ..., k such
that E( Xi (t)) = mi (t) [65–68]. If we have a functional sample classified in several groups such as
{Xi , Gi }in=1 ∈ F × G = {1, ..., G }, where G is a discrete variable that tell us the membership group,
the contrast will be:

H0 : X 1 = X 2 = ... = X G
(9)
H1 : ∃k, j s.t. X k = X j

After a few operations, as shown in [50,51], it is possible to go from classic (Fn ) to functional
statistic (Vn ).
ni
∑iG ni. (Y i. − Y )2 /( G − 1)
Fn = n =⇒ Vn = ∑ ||Yi. − Y j. ||2 (10)
∑iG ∑ j i (Yij − Y i. )2 /(n − G ) i< j

In addition, according to Cuevas et al. [50] and Tarrío-Saavedra et al. [51], the asymptotic
distribution of Vn under H0 is the same as the following statistic:

G
V := ∑ ||Zi (t) − Cij Zj (t)||2 (11)
i< j

where Cij = ( pi /p j )1/2 and Z1 (t), ..., ZG (t) are independent Gaussian processes with mean 0 and
covariance functions Ki (s, t).
Finally, H0 will be rejected, at a level α, whenever Vn > V where PH0 {V > Vα } = α [52].
Because in practise it is not easy to estimate the distribution of V, usually, it is necessary to implement
a Monte Carlo procedure through which we get for each i = 1, ..., G, N iid observations

Zil∗ = ( Zil∗ (t1 ), ..., Zil∗ (tm )), l = 1, ..., N (12)

from a m-dimensional Gaussian random variable with mean 0 and covariance matrix (K̂i (t p , tq ))1≤ p, q≤m .
The functional L2 (l )-distances || Zi (t) − Cij Zj (t)||2 are approximated by the Rm -Euclidean distances
|| Zil∗ − Cij Zjl∗ (t)||2 . Ultimately, the replications Ṽl of V are

G
Ṽl = ∑ ||Zil∗ − Cij Zjl∗ || (13)
i< j

and the distribution of V is approximated from the empirical distribution to a sample Ṽ1 , ..., ṼN [50,51].

2.1.3. Functional Strengths

FDA has numerous advantages but the following are the ones that make FDA suitable for studying
daily behaviours in energy variables [24,28,58]:

• It is not mandatory to have prior information on data distribution. The study does not depend on
or is not limited to certain distributions.
• The analysis takes into account time intervals as a unit. The sample analysed focusses on complete
time units such as days, months or years.
• Analysis of homogeneity. The definition of outliers is different; it is based on the idea that,
even though data do not surpass the cut-off, if they show constant deviations, they will be
identified as outlier.

57
Mathematics 2020, 8, 547

• Possibility of study trends. Besides calculating mean functions or detect outliers, it is also possible
to study slight variations from the normal data behaviour of the data without outliers.
• Complete analysis of the time spectrum. Before this approach, most analyses were based on the
values obtained in a given grid of discrete points. On the contrary, with FDA, it is possible to
work with the entire time set in a continuous way.

2.2. Building Description

The case study of this work is the Rectorate building of the University of the Basque Country
(Spain) constructed in the 1970s. It is a large building divided in three blocks (west, central and east),
as shown in Figure 1. For the purpose of this work, only the west block was considered, which is
an office building that includes a nursery. Four storeys form the block: floor 0 (Ground Floor) and
Floor 2 (2F) consist of rooms and offices while Floor 1 (1F) and Floor 3 (3F) are mainly open spaces
(see Figure 2). As the building use is office work, the occupation takes place only during weekdays,
being reduced during summer. There is lighting consumption from 7 a.m. to 8 p.m., except on the
second floor, where lighting consumption ends at 6 p.m. The building has a centralised heating system
with hot water radiators powered by a campus district heating. The heating works the whole year
following a control program according to a schedule from 6 a.m. to 7 p.m. except in July and August,
when the heating remains off. A refurbishment of the building was made in the summer of 2016 on the
ground floor and in the summer of 2017 in the rest of the building to reduce its energy consumption.

(a))

(b) (c)
Figure 1. (a) View of the whole rectorate building that is formed by three blocks (west, central and east).
This work focussed in the study of the retrofitting measures carried out in the west block.
(b) West block before retrofitting. (c) West block after retrofitting.

2.2.1. Building Description before Retrofitting

Before the retrofitting, the building did not have any insulation. Most of the façade was built with
precast reinforced concrete panels with non-ventilated air gap. Regarding the windows, some were
single-glazed with wooden frame, and others were double-glazed but with aluminium frame without
thermal break. Before retrofitting, there was no air conditioning or mechanical ventilation system.

2.2.2. Building Description after Retrofitting

The refurbishment consisted in several measures to improve the building’s envelope and energy
system. The façade was insulated with vacuum insulated panels (VIPs) to reduce heat losses through
the envelope. Some windows were replaced by high-performance windows. To improve the lighting

58
Mathematics 2020, 8, 547

consumption of the building, LED lighting was installed. A ventilation system with heat recovery was
added on each floor. Furthermore, thermostatic valves were installed to improve the control of the
hot water radiators. A detailed description of the refurbishment and an assessment of the heat loss
coefficient of the building is provided in [20].

Figure 2. Plans of the four floors in the west block of the building with their space distribution
and dimensions. The use of each space is represented with colours: offices (brown), nursery (green),
storage rooms (grey), corridors (pink) and server room (blue). In addition, the distribution of the sensors
is also displayed: circles, illuminance, temperature and relative humidity; triangles, calorimeters;
and square, lighting consumption.

2.2.3. Monitoring Description of the Building

The monitoring started before the refurbishment to investigate the necessary energy conservation
measures. Thus, monitored data before and after retrofitting from sensors located all around the
building are available. Monitored data before retrofitting correspond to 2016 and 2017, and data
after retrofitting correspond to s 2018 and 2019; all data are minute-by-minute values. In particular,
on the ground floor, the renovation started a year earlier with the insulation of a false ceiling and
the installation of LED lighting. In this case, the years considered before retrofitting are 2015 and
2016. The monitored variables include outdoors conditions, indoors conditions and building heating
and lighting consumption. Table 1 presents all the monitored variables in the building used to assess
the impact that the retrofitting had on the building’s energy performance, comfort and lighting.
Indoor condition sensors are located in several points at each floor, as shown in Figure 2. The electrical
consumption and the heating demand is provided per floor. Further information about the monitoring
of the building can be found in the work carried out by Erkoreka et al. [69].

2.3. Pre-Processing Data

Before presenting the results, it is necessary to explain the specific smoothing process performed
in this study. On the one hand, because there are parts of the day in which the variables are constant,
the basis functions chosen are splines. On the other hand, to select the optimal number of basis,
the determination coefficient R2 was taken into account to measure the smoothing adjustment in
relation to the raw data. As shown in Figure 3, the criterion was to select the minimum number of
basis (in a given grid) where the R2 surpasses the value of 99%.

59
Mathematics 2020, 8, 547

Table 1. Monitored variables in the building to assess the impact of the refurbishment in the energy
performance, illuminance and comfort of the building.

Type of Measurement Monitored Variable Units Sensor

Illuminance [LUX] Siemens 5WG1 255-4AB12
Indoor conditions
Temperature [o C] ARCUS SK04-S8-CO2-TF
Before Lighting W
Power meters ABB
retrofitting + elec. equipment
EM/S and ABB a41/43 per floor
Electrical Consumption
W
Lighting
After
retrofitting
Ventilation + elec. equipment W Power meters ABB EM/S and
ABB a41/43 per floor
W Calorimeter: Kamstrup Multical 602
Heating consumption Thermal energy of the heating water
and ZENNER Zelsius (DN20)

Figure 3. Example of the process of selecting the optimum number of basis for smoothing the original
sample. The red line represents the minimum number of basis where the R2 is higher than 0.99.

In addition, it is important highlight the data cleaning that was carried out using a functional
approach. It is known that throughout all the analysed years there were many days when the building
was closed and the information that these days provide is not only not useful, but can also distort the
final results. To solve this problem, an algorithm that searches for these days and deletes them from
the sample was developed (see Algorithm 1).
After an exploratory analysis of the data, the values of the chosen parameters for Algorithm 1 are:
β 1 = 500, β 2 = 250, α = 0.25, and θ = 0.5. Moreover, because there are parts of the day in which the
variables are constant, the chosen basis functions are spline. With the application of the algorithm to
the data, the sample becomes smaller but only with relevant days that take into account the normal
behaviour of the building. Figure 4 illustrates the performance of the algorithm. It can be seen that
it is capable of deleting the days with an abnormal behaviour without affecting the bulk of the data.
The days when the building was unoccupied and closed, and therefore with very small or no electrical
consumption, are detected and eliminated in the picture on the right of Figure 4.
Figure 4 also presents the mean functions (in form of curves) and the change that they suffer
after deleting non-representative days. Working with non-representative days will produce erroneous
results in any study; for example, if an ANOVA test is performed, the test may not reject the equality
of means even if the groups are different.

60
Mathematics 2020, 8, 547

Figure 4. Performance of the Functional cleaning algorithm with an example variable: Electrical demand
on the second floor of the building in the representative months (in dark gray the data before the
refurbishment, in light gray after). The number of days taken into account in each sample is also shown:
(left) the raw data and the mean functions separated between before and after the refurbishment;
and (right) the data after applying the algorithm and the mean functions separated between before
and after the refushbishment.

Algorithm 1: Functional cleaning.

Input: Data divided in groups and the parameters: β 1 , β 2 , α, θ.
Output: Data without inappropriate days.
1 Transform the data to funcional format: 1440 minute data each day.
2 Searching for missing values (NAs). The daily limits are:
β 1 NAs per day β 2 consecutive NAs per day
3 Delete the days that exceeded the daily limits.
4 Approximation, with an interpolation technique, of the remaining missing values.
5 Calculate the variability of every daily curve in the sample.
6 Delete the curves that:
• Have a variability less than or equal to a percentage α ∈ R of the average sample variability.
• Be below the sample mean function for at least a percentage θ ∈ R of the day.

3. Results and Discussion

The effects of the refurbishment carried out in 2017 in the Rectorate building of the University
of the Basque Country were analysed. In this analysis, lighting consumption, illuminance,
indoor temperatures and heating demand were studied. These variables were measured every minute
between 2016 and 2019 (changing 2016 for 2015 in the case of the ground floor) and only taking into
account those months in which the heating systems operate significantly (October to March). As the
retrofitting started in the summer of 2017, the months of this year after the summer are not suitable for
analysis. In this way, data were divided into nine months before retrofitting (six months in 2016 and
three months in 2017) and nine months after retrofitting (six months in 2018 and three months in 2019).
Section 3.1 presents the lighting analysis of the study and Section 3.2 the same analysis for heating
demands. The numerical results were based on the p-values of the ANOVA and Kruskal tests in the
vectorial analysis, and on the p-values of the FANOVA in the functional analysis. Different measures
are also shown to illustrate the differences between the sample groups: Dvec , the difference between
the medians from the vectorial approach; Dfunc , the average minute difference between the mean

61
Mathematics 2020, 8, 547

functions; and Ddist , the L2 (l ) distance between the mean functions. Additionally, to measure the
functional smoothing adjustment to the raw data, the coefficient of determination R2 is shown in
the tables. The change in the variance of monitored data (Var) and the savings obtained with the
retrofitting, calculated in relation to the initial energy demands, are also shown from both approaches.
Lastly, all figures presented here were made with the R-programming software.

3.1. Lighting Analysis

The results of the electrical lighting consumption are shown in Figure 5. On the one hand,
the vectorial analysis by means of box plots is shown in the first row. It can be seen that, after the
refurbishment, the lighting demands of the building decreased and became more homogeneous on all
floors. This is demonstrated in Table 2, where the change is quantified, and the statistical tests used
(ANOVA and Kruskal) corroborate the change. In this case, the floor with the highest reduction was
the third floor (2004 W less) and the floor with the lowest reduction was the second floor (489 W less).
On the other hand, in the second row of Figure 5, the functional analysis, through the daily curve
graph, is presented. In this case, thanks to this approach, it can be seen that, despite the decrease in
the consumption, the retrofitting hardly affects the daily behaviour of the lighting consumption. As it
is observable in Figure 5 that the curves have very similar shapes, with the exception on the third
floor where a change in the lighting schedule was implemented. Table 2 presents the FANOVA results.
The similarity of the samples is rejected on every floor. However, the impact varies among floors,
as shown in Table 2. The first floor was the most benefited (2032 W less per minute on average) and
the third floor the least benefited (400 W less per minute on average). The differences between the two
analyses, in absolute terms, have been noticed: with the vectorial analysis the third floor obtained the
highest reduction, while with functional analysis it obtained the lowest reduction. As shown in the
second row of Figure 5, on the third floor, the reduction was concentrated on the last hours of the day;
however, on average during the day, this reduction was lower. The vectorial approach does not take
this fact into account because it distorts the sample with the calculation of daily averages.

Figure 5. Analysis of the electrical lighting consumption on each floor measured in W. In the first row,
the vectorial results (in form of box plots) are presented. In the second row, instead, the functional data
are represented with the respective mean functions. The data are divided into winter days before and
after the refurbishment.

62
Mathematics 2020, 8, 547

Table 2. Numerical results on each floor for lighting consumption. The vectorial results are presented
with Dv measuring the difference between medians. The functional results, accompanied with the
average minute difference (Dfunc ), the L2 (l ) distance between the mean functions (Ddist ) and the
smoothing adjustment (R2 ), are also presented. For both analyses, the change in the variability of the
data (Var) and the electrical savings are displayed. Lastly, the p-values for the tests, both vectorial
and functional, are displayed in this table.
Electrical Consumption
Vectorial Analysis Functional Analysis
panova pkruskal Dvec (W) Var Savings pfanova Dfunc (W) Ddist (L2 (l )) Var Savings R2
Floor 0 ≈0 ≈0 −587.48 −72.63% 36.96% ≈0 −609.69 34536.02 −71.80 % 38.11% 0.9909
Floor 1 ≈0 ≈0 −1987.68 −94.47% 72.68% ≈0 −2032.31 116400.04 −89.54% 73.15% 0.9903
Floor 2 ≈0 ≈0 −489.05 −43.49% 24.83% ≈0 −434.68 29537.58 −55.54% 22.50% 0.9929
Floor 3 ≈0 ≈0 −2004.80 −95.32% 53.82% ≈0 −400.82 35603.07 +17.29% 18.18% 0.9915

On the other hand, the effects of the refurbishment on the illuminance conditions were studied.
Figure 6 shows the analysis for the illuminance levels. Through vectorial analysis, an impact is also
detected, but is different on each floor. In general, the illuminance level was improved. Table 3
indicates that the floors with the biggest increase in illuminance were the first and third floors
(356 lx and 414 lx more, respectively). Ground floor was the only one with an illuminance reduction
from this point of view (270 lx less). In this case, this reduction is related to the shading that was
installed to have a better protection against natural light (see Table 3). Furthermore, observing the
functional illuminance curves floor by floor shown in the second row of Figure 6, the conclusion
is the same: there was also an improvement in the illuminance levels, and the biggest increase of
illuminance took place on the first and third floors. From this point of view, Table 3 shows that the
increments were about 174 lx on the first floor and 204 lx on the third floor, on average, every minute.
Once again, FDA makes it possible to see that the behaviour of the illuminance was maintained on all
floors. As shown in Table 3, FDA detects that the illuminance changes on the second floor, while the
vectorial approach fails in this detection (the p-values obtained are bigger than 0.05).

Figure 6. Analysis of the illuminance on each floor measured in lx. In the first row, the vectorial results
(in form of box plots) are presented. In the second row, instead, the functional data represented with the
respective mean functions. The data are divided into winter days before and after the refurbishment.

63
Mathematics 2020, 8, 547

Table 3. Numerical results on each floor for illuminance. The vectorial results are presented with Dvec
measuring the difference between medians. The functional results, accompanied with the average
minute difference (Dfunc ), the L2 (l ) distance between the mean functions (Ddist ) and the smoothing
adjustment (R2 ), are also presented. For both analyses, the change in the variability of the data is
displayed (Var). Lastly, the p-values for the tests, both vectorial and functional, are displayed in
this table.
Illuminance
Vectorial Analysis Functional Analysis
panova pkruskal Dvec (lx) Var pfanova Dfunc (lx) Ddist (L2 (l )) Var R2
Floor 0 ≈0 ≈0 −266.22 −42.27% ≈0 −286.81 15274.23 −12.51% 0.9909
Floor 1 ≈0 ≈0 +356.12 +62.88% ≈0 +174.78 9379.46 +59.44% 0.9913
Floor 2 0.217 0.278 +9.46 −27.78% ≈0 +23.33 1608.27 −9.29% 0.9904
Floor 3 ≈0 ≈0 +414.86 +81.24% ≈0 +204.32 11140.84 +66.30% 0.9914

Analysing the graphs shown in Figures 5 and 6, it can be seen that the retrofitting reduced
the electrical consumption of lighting while the illuminance levels were improved or maintained in
proper levels. The level of illuminance on the first and third floors was improved mainly due to the
replacement by LED technology, as shown in Figure 6 and Table 3. In the case of the second floor,
although it obtained an average reduction of 435 W, the illuminance level was maintained according to
European illuminance regulations. This regulation states that the optimal lighting level in offices is
500 lx, a level already reached before the renovation. In the case of the ground floor, as mentioned
above, it had to be analysed individually. The mean function of illuminance before retrofitting on this
floor had peaks above 1500 lx (see Figure 6), which indicates a high influence of natural light. After the
retrofitting, the shading that was installed to protect from sunlight achieved a reduction, on average,
of 286 lx each minute, as it can be seen in Table 3 (266 lx, observing the vectorial results). Furthermore,
the electrical consumption associated to lighting on the ground floor is also reduced. Table 2 shows
that the vectorial reduction was 587 W, and the functional reduction 609 W, on average, every minute.
Generally, monitored lighting consumption data after retrofitting have less dispersion
(see Var in Table 2). This means that the monitored lighting data are more homogeneous. In this
case, the reduction of the data variability reaches values higher than 30% from both methods. However,
the functional analysis detects an increase of 17% in the data dispersion of the third floor, as shown
in Table 2. The functional graph of this floor in Figure 5 supports this result. The vectorial approach,
which summarises the days with the mean, distorts the results by considering false conclusions
such as, in this case, that the data variance decreased for all floors (see Table 2). On the other hand,
the homogeneity of the monitored illuminance data after the renovation is different depending on the
floor. Table 3 shows that the floors where the data dispersion was reduced are the ground and the
second floors.
From an energetic point of view, the relative electrical consumption savings after retrofitting
are also calculated. Table 2 shows that savings of more than 18% were obtained on every floor with
the functional analysis and more than 24% with the vectorial analysis. The first floor is the most
benefited floor with savings around 73% from the two approaches applied. However, there are
differences between the results of the vectorial analysis and the functional analysis. This is shown in
Figures 5 and 6 and, specifically, in Table 2 where the savings are presented. The savings calculation of
functional analysis are more accurate because it is based on the areas under the curves representing
the mean functions, taking into account the entire daily behaviour. The vectorial results, instead,
only quantify the differences between the medians of the samples.

3.2. Heating Analysis

The results show that the refurbishment had a significant impact in the heating demands of the
building. First, observing the box plots from the vectorial analysis, as shown in the first row of Figure 7,
it is clear that the heating demand was reduced on all floors. This fact is supported by vectorial tests

64
Mathematics 2020, 8, 547

that reject similarity between samples on all floors (see Table 4). After that, with the functional graphs,
as shown in the second row of Figure 7, it is also appreciated that the heating demand curves after
refurbishment are below the initial curves. This is also proved by the functional tests that reject all
the sample similarity hypothesis (see Table 4). The results shown in Table 4 change from vectorial
to functional analysis. Both analyses come to the same conclusions, but the magnitude of change is
different for each one. With vectorial approach, the reduction per floor ranges from 3667 W as the
highest reduction on the third floor to 1057 W as the lowest on the first floor. In contrast, with the
functional analysis, although the highest and lowest reduction were on the same floors, the values
of the reduction are not the same. The heating demand, each minute, was reduced on average 3918
W on third floor and 1456 W on first floor. Therefore, the relative heating savings were significant
(see Table 4). The floor most benefited was the third floor; both analysis obtained savings higher than
a 30% on this floor. Again, the ground floor had to be analysed individually because, in addition to the
installation of shading in 2017 that prevents solar gains, a false ceiling was already insulated on this
floor in 2015. Thus, it was possible to obtain a saving of about 12% of the initial heating demands from
vectorial results and about 17% from functional results (see Table 4).

Figure 7. Analysis of the heating demands of each floor measured in W. In the first row, the vectorial
results (in form of box plots) are presented. In the second row, instead, the functional data are
represented with the respective mean functions. The data are divided into winter days before and after
the refurbishment.

Table 4. Numerical results on each floor results for heating demands. The vectorial results are presented
with Dvec measuring the difference between medians. The functional results, accompanied with the
average minute difference (Dfunc ), the L2 (l ) distance between the mean functions (Ddist ) and the
smoothing adjustment (R2 ), are also presented. For both analyses, the change in the variability of the
data (Var) and the heating savings are displayed. Lastly, the p-values for the tests, both vectorial and
functional, are calculated.
Heating Demand
Vectorial Analysis Functional Analysis
panova pkruskal Dvec (W) Var Savings pfanova Dfunc (W) Ddist (L2 (l )) Var Savings R2
Floor 0 1.761 × 10−6 7.047 × 10−6 −1838.12 −35.68% 11.86% ≈0 −1455.83 158722.42 −30.66% 17.36% 0.9901
Floor 1 0.018 0.02 −1057.67 +24.71% 7.46% ≈0 −1975.15 95981.90 −23.89% 16.97% 0.9908
Floor 2 ≈0 ≈0 −2457.87 −43.99% 22.73% ≈0 −2158 96002.01 −43.48% 23.60% 0.9914
Floor 3 ≈0 ≈0 −3667.06 −60.22% 31.49% ≈0 −3917.87 185968.48 −51.68% 35.51% 0.9911

65
Mathematics 2020, 8, 547

In the case of the vectorial analysis, Table 4 shows that, in general, the measurements are less
variable in general (between 35% and 60% less). On the first floor, instead, this method detects an
increase in data dispersion after the retrofitting. However, this fact is not supported by the functional
approach (see Table 4). Figure 7 shows in its second row that the heating curves on this floor are
similar or even less variable. In the case of functional analysis results, as shown in Table 4, the variation
of the measurements is lower on all floors (between 23% and 51% less). The functional method is
demonstrated to be more accurate and provides more information. For instance, functional analysis
detects the demand peaks on ground floor in the morning when the heating starts, as it can be seen
in second row of Figure 7. With the vector analysis, this information is lost as shown in first row of
Figure 7.
The possible consequences of the retrofitting on the building temperatures are also studied.
Both the vectorial approach and the functional approach conclude that the temperatures increased on
every floor except on the third floor (see Figure 8 and Table 5). The tests do not detect any change in
the average indoor temperature on this floor. This is probably because this floor is a large space with
low occupancy where the temperatures were stable before and after refurbishment. This is supported,
on the one hand, by the functional analysis in Figure 8 where the curves are almost overlapping.
On the other hand, Table 5 shows that the temperature on the third floor increased very slightly and
the FANOVA test does not detect a significant change (p-value = 0.17). On the contrary, the ground,
first and second floors had higher temperatures after retrofitting, as shown in Figure 8 and Table 5.
The increase, depending on the floor and method, was between 0.5 and 2 ◦ C. The reason is that after
the refurbishment the building is more insulated, heat losses are reduced and it is easier to keep it
warmer. Moreover, the temperature set point have been increased in the common zones. Only with
FDA it is detected that the retrofitting succeeds to reduce the influence of natural light on the ground
floor temperatures. In the second row of Figure 8, it is observed that temperatures on the ground
floor after refurbishment do not have peaks at the end of the day due to solar radiation. Finally,
both analyses show that the homogeneity of monitored data is significantly improved. Table 5 shows
that, on each floor and from both approaches, there is more homogeneity in the measurements related
to the building’s indoor temperatures.

Figure 8. Analysis of the temperatures on each floor measured in ◦ C. In the first row, the vectorial
results (in form of box plots) are presented. In the second row, instead, the functional data are
represented with the respective mean functions. The data are divided into winter days before and after
the refurbishment.

66
Mathematics 2020, 8, 547

Table 5. Numerical results on each floor indoor temperatures. The vectorial results are presented with
Dvec measuring the difference between medians. The functional results, accompanied with the average
minute difference (Dfunc ), the L2 (l ) distance between the functional mean (Ddist ) and the smoothing
adjustment (R2 ), are also presented. For both analyses, the change in the variability of the data is
displayed (Var). Lastly, the p-values for the tests, both vectorial and functional, are calculated.
Indoor Temperatures
Vectorial Analysis Functional Analysis
panova pkruskal Dvec (W) Var pfanova Dfunc (W) Ddist (L2 (l )) Var R2
Floor 0 ≈0 ≈0 +1.92 −62.75% ≈0 +1.95 75.09 −40.10% 0.9918
Floor 1 0.001 0.004 +0.55 −45.33% ≈0 +0.35 13.43 −37.77% 0.9970
Floor 2 ≈0 ≈0 +0.76 −53.31% ≈0 +0.92 35.29 −60.68% 0.9976
Floor 3 0.39 0.274 −0.072 −50.65% 0.17 +0.002 7.97 −63.15% 0.9959

As in the electrical analysis, the relative savings obtained in the heating analysis were calculated,
as shown in Table 4. These savings reached values of more than 30% (in particular on third floor), and,
in this case, with the functional approach are higher than with the vectorial approach. The floor with
the smallest relative saving, observing the results of both methods, was the first floor, but with the
vectorial method the reduction was almost 8% and with the functional method almost 17%. In this
case, the results are more homogeneous among floors; there is not much difference from floor to
floor. As expected, the form of the heating demands curves, before and after retrofitting, is the same
although the values after retrofitting decreased (see Figure 7). Similar behaviour is appreciated in the
temperature curves (see Figure 8).
After retrofitting the indoors temperature increased on all floors, a ventilation system with
exterior air was installed and the internal gains were reduced with LED lighting. These changes
should contribute to an increase of the heating demand. However, Figure 7 and Table 4 show that the
heating demand on each floor has been reduced, demonstrating the effectiveness of façade insulation.
Besides the decrease on the heating demands, indoor temperatures in the building are maintained or
even increased, as demonstrated in Figure 8 and Table 5.

4. Conclusions
A new application of FDA and a new methodology to assess the impact of retrofitting in buildings
are presented in this paper. The study was conducted by analysing monitored data of lighting
consumption, heating demands, illuminance levels and indoor temperatures of the Rectorate building
of the University of the Basque Country (Spain). These analyses aimed to detect an impact on
the measured variables and to quantify the changes achieved with retrofitting. The methodology
used in this study is based, on the one hand, in the functional analysis contrasting the distance
between mean functions of monitored samples before and after retrofitting. On the other hand, as a
comparison, the classical or vectorial approach was carried out measuring the dissimilarities between
sample medians.
The presented method contributes with advantages over the already existing research in the
topic of building retrofitting evaluation. Some research evaluated the effect of the refurbishment
based on environmental indicators. Other studies give a measure of the heat losses before and after
retrofitting, and other studies focus on the heating demands. One of the advantages of the proposal
of using functional analysis is that it can be applied to evaluate different building variables such as
temperatures, lighting levels, electrical consumption or heating demands. Consequently, it can be
used to search for relationship or effects between variables. For instance, if the heating demand is not
reduced as expected, the evolution of other variables can be observed to look for the cause. Moreover,
a daily based analysis can be done, evaluating the peculiarities of some days in the performance of the
studied variables.

67
Mathematics 2020, 8, 547

The research contribution of this paper is the application of a mathematical method such as
functional analysis to evaluate the impact that a building retrofitting had in its energy performance.
There are few studies that present evaluations of retrofitting actions in buildings with monitored
data and the employed methods are based in vector-based data approaches. The benefits of applying
FDA to contrast and measure the similarity between samples of monitored data before and after a
retrofitting were demonstrated. An advantage of FDA is that it is not restricted to certain characteristics
of the data distribution, thus it is not necessary to test its normality. It considers complete time units,
working with the time set in a continuous way, without having to summarise them, which is beneficial
to evaluate monitored building variables. In addition, the outlier detection takes into account constant
deviations as a reason to identify an outlier, even though it does not surpass the cut-off criterion.
Furthermore, the variables used in this study are commonly analysed in different types of
buildings. The methodology presented here can be applied to assess the energy and thermal
performance of different buildings, such as industrial, residential, educational or office buildings.
Functional analysis can be applied, with variables or tools different from those used here, to evaluate
different aspects of the studied building. Thus, this method could also be used to evaluate monitored
variables in thermal facilities within an energy efficiency framework. The effectiveness and usefulness
of the functional approach to evaluate variables that affect the energy efficiency was demonstrated in
this study.
The results illustrate the greater accuracy of FDA in detecting if there was a significant change
in the studied variables in comparison with vectorial analyses results. In the illuminance analysis of
the second floor, the FANOVA detected a change (p-value ≈ 0) while the vectorial ANOVA did not
(p-value = 0.22). Additionally, FDA could detect that most of the lighting consumption reduction on
the third floor was concentrated in the last hours of the day. In addition, it has been demonstrated that
FDA can provide trustable information about the dispersion of the data. In the lighting consumption
analysis of the third floor, the greater dispersion of data after retrofitting was only identified with
FDA. Moreover, the representation of the monitored variables in a continuous way throughout the
day shows that the FDA allows a greater accuracy and a better adjustment to reality than the vectorial
methods. It allows identifying average patterns of the studied variables and it has the potential for
detecting anomalous behaviours of monitored variables. In the analysis of the heating demand of the
ground floor, the FDA reported a peak in the early hours of the day (around 7 p.m.), both before and
after the retrofitting, which the vectorial analysis did not notice.
From an energetic point of view, the conclusion is that the refurbishment carried out in the
building under study had a significant impact on its energy performance. The main aim of the
retrofitting was to reduce the heat losses. This goal was achieved, as demonstrated by the decrease in
the heating demands of the entire building even though the temperatures were increased, a ventilation
system being installed and the LED lighting reducing the internal heat gains. The analysis also shows
that the illuminance levels were improved on all floors, only decreasing on the ground floor as a
shading was installed to prevent from direct sunlight. According to FDA results, the heating demands
were reduced 17% on the ground and first floor, 24% on the second floor and 36% on the third floor.
Moreover, the reduction in lighting consumptions were 38% for the ground floor, 73% for the first floor
and around 20% for the second and third floors.

Author Contributions: Mathematical methodology, M.M.C.; engineering knowledge, S.M.M. and E.G.Á.; and data
collection, P.E.O. and A.E.G. All authors have read and agreed to the published version of the manuscript.
Funding: This paper was funded by the Spanish Government (Science, Innovation and Universities Ministry)
under the project RTI2018-096296-B-C21.
Acknowledgments: This paper was supported by the Spanish Government (Science, Innovation and Universities
Ministry) under the project RTI2018-096296-B-C21.
Conflicts of Interest: The authors declare no conflict of interest.

68
Mathematics 2020, 8, 547

References
1. Foucquier, A.; Robert, S.; Suard, F.; Stéphan, L.; Jay, A. State of the art in building modelling and energy
performances prediction: A review. Renew. Sustain. Energy Rev. 2013, 23, 272–288. [CrossRef]
2. Energy Information Administration. International Energy Outlook 2019. Available online: https://www.eia.gov/ieo
(accessed on 21 March 2020).
3. RCP policy: public health and health inequality. Every breath we take: the lifelong impact of air pollution.
Royal College of Physicians (RCPCH). 2016. Available online: https://www.rcplondon.ac.uk/projects/
outputs/every-breath-we-take-lifelong-impact-air-pollution (accessed on 21 March 2020).
4. Wang, L.; Pereira, N.; Hung, Y. Air Pollution Control Engineering; Humana Press: New York, NY, USA, 2004.
5. EEA. Air Quality in Europe; European Environmental Agency: Copenhagen, Denmark, 2019.
6. Frances Bean, J.; Dorizas, V.; Bourdakis, E.; Staniaszek, D.; Pagliano, L.; Roscetti, A. Future-Proof Buildings
for all Europeans—A Guide to Implement the Energy Performance of Buildings Directive; Buildings Performance
Institute Europe (BPIE): Brussels, Belgium, 2019.
7. Cabeza, L.F.; Rincón, L.; Vilariño, V.; Castell, A. Life cycle assessment (LCA) and life cycle energy analysis
(LCEA) of buildings and the building sector: A review. Renew. Sustain. Energy Rev. 2014, 29, 394–416.
[CrossRef]
8. Santamouris, M.; Pavlou, C.; Doukas, P.; Mihalakakou, G.; Synnefa, A.; Hatzibiros, A.; Patargias, P.
Investigating and analysing the energy and environmental performance of an experimental green roof
system installed in a nursery school building in Athens, Greece. Energy 2007, 32, 1781–1788. [CrossRef]
9. Chidiac, S.E.; Catania, E.J.; Morofky, E.; Foo, S. Effectiveness of single and multiple energy retrofit measures
on the energy consumption of office buildings. Energy 2011, 36, 5037–5052. [CrossRef]
10. Zmeureanu, R. Assessment of the energy savings due to the building retrofit. Build. Environ. 1990, 25,
95–103. [CrossRef]
11. Asadi, E.; Da Silva, M.G.; Antunes, C.H.; Dias, L. Multi-objective optimization for building retrofit strategies:
A model and an application. Energy Build. 2012, 44, 81–87. [CrossRef]
12. Yalcintas, M. Energy-savings predictions for building-equipment retrofits. Energy Build. 2008, 40, 2111–2120.
[CrossRef]
13. Tobias, L.; Vavaroutsos, G. Retrofitting Buildings to be Green and Energy-Efficient: Optimizing Building
Performance, Tenant Satisfaction, and financial Return; Urban Land Institute: Washington, DC, USA, 2012.
14. Hamburg, A.; Kalamees, T. How well are energy performance objectives being achieved in renovated
apartment buildings in Estonia? Energy Build. 2019, 199, 332–341. [CrossRef]
15. Ardente, F.; Beccali, M.; Cellura, M.; Mistretta, M. Energy and environmental benefits in public buildings as
a result of retrofit actions. Renew. Sustain. Energy Rev. 2011, 15, 460–470. [CrossRef]
16. Mohammadpourkarbasi, H.; Sharples, S. Eco-retrofitting very old dwellings: current and future energy
and carbon performance for two UK cities, Plea 2013. In Proceedings of the 29th Conference, Munich,
Germany, 10–12 September 2013.
17. Beccali, M.; Cellura, M.; Fontana, M.; Longo, S.; Mistretta, M. Energy retrofit of a single-family house:
Life cycle net energy saving and environmental benefits. Renew. Sustain. Energy Rev. 2013, 27, 283–293.
[CrossRef]
18. Famuyibo, A.A.; Duffy, A.; Strachan, P. Achieving a holistic view of the life cycle performance of existing
dwellings. Build. Environ. 2013, 70, 90–101. [CrossRef]
19. Ficco, G.; Iannetta, F.; Ianniello, E.; dAmbrosio Alfano, F.R.; DellIsola, M. U-value in situ measurement for
energy diagnosis of existing buildings. Energy Build. 2015, 104, 108–121. [CrossRef]
20. Uriarte, I.; Erkoreka, A.; Giraldo-Soto, C.; Martin, K.; Uriarte, A.; Eguia, P. Mathematical development of
an average method for estimating the reduction of the Heat Loss Coefficient of an energetically retrofitted
occupied office building. Energy Build. 2019, 192, 101–122. [CrossRef]
21. Zavadskas, E.; Kaklaukas, A.; Turskis, Z.; Kalibatas, D. An approach to multi-attribute assessment of indoor
environment before and after refurbishment of dwellings. J. Environ. Eng. Landsc. Manag. 2009, 17, 5–11.
[CrossRef]
22. Febrero, M.; Galeano, P.; Wenceslao, G. Outlier detection in functional data by depth measures,
with application to identify abnormal NOx levels. Environmetrics 2008, 19, 331–345. [CrossRef]

69
Mathematics 2020, 8, 547

23. Horváth, L.; Kokoszka, P. Inference for Functional Data with Applications; Springer: Berlin/Heidelberg,
Germany, 2012. [CrossRef]
24. Martínez, J.; Saavedra, A.; García, P.; Piñeiro, J.; Iglesias, C.; Taboada, J.; Sancho, J.; Pastor, J. Air quality
parameters outliers detection using functional data analysis in the Langreo urban area (Northern Spain).
Appl. Math. Comput. 2014, 241, 1–10. [CrossRef]
25. Pi neiro, J.; Torres, J.; García, P.; Alonso, J.; Mu niz, C.; Taboada, J. Analysis and detection of functional
outliers in waterquality parameters from different automated monitoring stationsin the Nalón River Basin
(Northern Spain). Environ. Sci. Pollut. Res. 2015, 22, 387–396. [CrossRef]
26. Martínez, J.; Pastor, J.; Sancho, J.; McNabola, A.; Martínez, M.; Gallagher, J. A functional data analysis
approach for the detection of air pollution episodes and outliers: A case study in Dublin, Ireland. Mathematics
2020, 8, 225. [CrossRef]
27. Sancho, J.; Martínez, J.; Pastor, J.; Taboada, J.; Piñeiro, J.; García Nieto, P. New methodology to determine air
quality in urban areas based on runs rules for functional data. Atmos. Environ. 2014, 83, 185–192. [CrossRef]
28. Sancho, J.; Iglesias, C.; Piñeiro, J.; Martínez, J.; Pastor, J.; Araújo, M.; Taboada, J. Study of water quality in
a spanish river based on statistical process control and functional data analysis. Math. Geosci. 2016, 48,
163–186. [CrossRef]
29. Iglesias, C.; Sancho, J.; Piñeiro, J.I.; Martínez, J.; Pastor, J.J.; Taboada, J. Shewhart-type control charts and
functional data analysis for water quality analysis based on a global indicator. Desalin. Water Treat. 2016, 57,
2669–2684. [CrossRef]
30. Dombeck, D.; Graziano, M.; Tank, D. Functional clustering of neurons in motor cortex determined by cellular
resolution imaging in awake behaving mice. J. Neurosis. 2009, 29, 13751–13760. [CrossRef]
31. Martínez, J.; Ordoñez, C.; Matìas, J.M.; Taboada, J. Determining noise in an aggregates plant using functional
statistics. Hum. Ecol. Risk Assess. 2011, 17, 521–533. [CrossRef] [PubMed]
32. Ordóñez, C.; Martínez, J.; Saavedra, A.; Mourelle, A. Intercomparison Exercise for Gases Emitted by a
Cement Industry in Spain: A Functional Data Approach. J. Air Waste Manag. Assoc. 1995 2011, 61, 135–141.
[CrossRef]
33. Sancho, J.; Pastor, J.; Martínez, J.; García, M. Evaluation of Harmonic Variability in Electrical Power
Systems through Statistical Control of Quality and Functional Data Analysis. In Procedia Engineering;
The Manufacturing Engineering Society: Southfield, MI, USA, 2013; Volume 63, pp. 295–302. [CrossRef]
[PubMed]
34. Wu, D.; Huang, S.; Xin, J. Dynamic compensation for an infrared thermometer sensor using
least-squares support vector regression (LSSVR) based functional link artificial neural networks (FLANN).
Meas. Sci. Technol. 2008, 19, 105202.1–105202.6. [CrossRef]
35. Ordoñez, C.; Martínez, J.; Cos Juez, J.; Sánchez Lasheras, F. Comparison of GPS observations made in a
forestry setting using functional data analysis. Int. J. Comput. Math. 2012, 89, 402–408. [CrossRef]
36. Müller, H.G.; Sen, R.; Stadtmüller, U. Functional data analysis for volatility. J. Econometr. 2011, 165, 233–245.
[CrossRef]
37. López, M.; Martínez, J.; Matías, J.M.; Taboada, J.; Vilán, J.A. Functional classification of ornamental stone
using machine learning techniques. J. Comput. Appl. Math. 2010, 234, 1338–1345. [CrossRef]
38. López, M.; Martínez, J.; Matías, J.M.; Taboada, J.; Vilán, J.A. Shape functional optimization with restrictions
boosted with machine learning techniques. J. Comput. Appl. Math. 2010, 234, 2609–2615. [CrossRef]
39. Flores, M.; Naya, S.; Fernández-Casal, R.; Zaragoza, S.; Raña, P.; Tarrío-Saavedra, J. Constructing a Control
Chart Using Functional Data. Mathematics 2020, 8, 58. [CrossRef]
40. Warmenhoven, J.; Harrison, A.; Robinson, M.A.; Vanrenterghem, J.; Bargary, N.; Smith, R.; Cobley, S.;
Draper, C.; Donnelly, C.; Pataky, T. A force profile analysis comparison between functional data analysis,
statistical parametric mapping and statistical non-parametric mapping in on-water single sculling. J. Sci.
Med. Sport 2018, 21, 1100–1105. [CrossRef]
41. Ordoñez, C.; Martínez, J.; Matías, J.M.; Reyes, A.N.; Rodríguez-Pérez, J.R. Functional statistical techniques
applied to vine leaf water content determination. Math. Comput. Model. 2010, 52, 1116–1122. [CrossRef]
[PubMed]

70
Mathematics 2020, 8, 547

42. Eisenhart, C. The Assumptions Underlying the Analysis of Variance. Int. Biometr. Soc. 1947, 3, 1–21.
[CrossRef]
43. Montgomery, D. Design and Analysis of Experiments, 8th ed.; John Wiley & Sons, Inc.: Hoboken, NJ, USA, 2013.
[CrossRef]
44. Kotz, S.; Johnson, N.L. Breakthroughs in Statistics. Volume II. Methodology and Distribution; Springer:
Berlin/Heidelberg, Germany, 1993; Volume 2.
45. Vikneswaran. An R companion to “Experimental Design”; Vikneswaran: Duxbury, MA, USA, 2005.
46. Crawley, M. The R Book, 2nd ed.; John Wiley & Sons: Hoboken, NJ, USA, 2013.
47. Theodorsson-Norheim, E. Kruskal-Wallis test: BASIC computer program to perform nonparametric one-way
analysis of variance and multiple comparisons on ranks of several independent samples. Comput. Meth.
Prog. Biomed. 1986, 23, 57–62.
48. Ostertagová, E.; Ostertag, O.; Kovác, J. Methodology and application of the Kruskal-Wallis test. Mech. Mater.
2014, 611, 115–120. [CrossRef]
49. Van Hecke, T. Power study of anova versus Kruskal Wallis test. Stat. Manag. Syst. 2012, 15, 241–247.
[CrossRef]
50. Cuevas, A.; Febrero, M.; Fraiman, R. An anova test for functional data. Comput. Stat. Data Anal. 2004, 47,
111–122.
51. Tarrío-Saavedra, J.; Naya, S.; Francisco-Fernández, M.; Artiaga, R.; Lopez-Beceiro, J. Application of functional
ANOVA to the study of thermal stability of micro–nano silica epoxy composites. Chemometr. Intell. Lab. Syst.
2011, 105, 114–124. [CrossRef]
52. Kaufman, C.G.; Sain, S.R. Bayesian functional (ANOVA) modeling using Gaussian process prior distributions.
Bayes. Anal. 2010, 5, 123–149. [CrossRef]
53. Wang, J.L.; Chiou, J.M.; Müller, H.G. Functional Data Analysis. Ann. Rev. Stat. Appl. 2016, 3, 257–295.
[CrossRef]
54. Kramosil, I.; Michálek, J. Fuzzy metrics and statistical metric spaces. Kybernetika 1975, 11, 336–344. [CrossRef]
55. Ramsay, J.; Silverman, B. Applied Functional Data Analysis: Methods and Cae Studies; Springer:
Berlin/Heidelberg, Germany, 2002.
56. Walz, M.; Zebrowski, T.; Küchenmeister, J.; Busch, K. B-spline modal method: A polynomial approach
compared to the Fourier modal method. Opt. Express 2013, 21, 14683–14697. [CrossRef]
57. Muñiz, C.; García, P.; Alonso, J.; Torres, J.; Taboada, J. Detection of outliers in water quality monitoring
samples using functional data analysis in San Esteban estuary (Northern Spain). Sci. Total Environ. 2012, 439,
54–61. [CrossRef] [PubMed]
58. Martínez, J.; García, P.; Alejano, L.; Reyes, A. Detection of outliers in gas emissions from urban areas using
functional data analysis. J. Hazard. Mater. 2011, 186, 144–149. [CrossRef] [PubMed]
59. Fraiman, R.; Muniz, G. Trimmed means for functional data. TEST Off. J. Spanish Soc. Stat. Operat. Res. 2001,
10, 419–440. [CrossRef]
60. Cuevas, A.; Febrero, M.; Fraiman, R. Robust estimationand classification for functional data via
projection-based notions. Comput. Stat. 2007, 22, 481–496. [CrossRef]
61. Cuevas, A.; Febrero, M.; Fraiman, R. On the use of bootstrap for estimating functions with functional data.
Comput. Stat. Data Anal. 2006, 51, 1063–1074. [CrossRef]
62. Léger, C.; Romano, J. Boostrap adaptive estimation. The trimmed-mean example. Can. J. Stat. 1990, 18,
297–314. [CrossRef]
63. Hall, P.; Maesono, Y. A Weighted Bootstrap Approach to Bootstrap Iteration. J. R. Stat. Soc. Ser. B Stat. Methodol.
2000, 62, 137–144. [CrossRef]
64. Millán-Roures, L.; Epifanio, I.; Martínez, V. Detection of Anomalies in Water Networks by Functional Data
Analysis. Math. Problems Eng. 2018, 2018, 13. [CrossRef]
65. Dette, H.; Derbort, S. Analysis of Variance un Non Parametric regression Models. J. Multivar. Anal. 2001, 76,
110–137. [CrossRef]
66. Maldonado, Y.; Staniswalis, J.; Irwin, L.; Byers, D. A similarity analysis of curves. Can. J. Stat. 2002, 30,
373–381. [CrossRef]

71
Mathematics 2020, 8, 547

67. Cuesta-Albertos, J.A.; Febrero, M. A simple multiway ANOVA for functional data. TEST 2010, 19, 537–557.
[CrossRef]
68. Zhang, J.T. Analysis of Variance for Functional Data. In A Chapman & Hall Book; Press, C., Ed.;
Taylor & Francis Group: Abingdon, UK, 2013; Chapter 5, p. 412. [CrossRef]
69. Erkoreka, A.; García, K.; Teres-Zubiaga, J.; Del Portillo, L. In-use office building energy characterization
through basic monitoring and modelling. Energy Build. 2016, 119, 256–266. [CrossRef]

c 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).

72
 mathematics
Article
Classical Lagrange Interpolation Based on General
Nodal Systems at Perturbed Roots of Unity
Elías Berriochoa 1, *,† , Alicia Cachafeiro 1, *,† , Alberto Castejón 1,† and José Manuel García-Amor 2,†
1 Departamento de Matemática Aplicada I, Universidad de Vigo, 36201 Vigo, Pontevedra, Spain;
[email protected]
2 Departamento de Matemáticas, Instituto E. S. Valle Inclán, 36001 Pontevedra, Spain;
[email protected]
* Correspondence: [email protected] (E.B.); [email protected] (A.C.);
Tel.: +34-988-387216 (E.B.); +34-986-812138 (A.C.)
† These authors contributed equally to this work.

Received: 28 February 2020; Accepted: 29 March 2020; Published: 2 April 2020

Abstract: The aim of this paper is to study the Lagrange interpolation on the unit circle taking only
into account the separation properties of the nodal points. The novelty of this paper is that we do
not consider nodal systems connected with orthogonal or paraorthogonal polynomials, which is
an interesting approach because in practical applications this connection may not exist. A detailed
study of the properties satisfied by the nodal system and the corresponding nodal polynomial is
presented. We obtain the relevant results of the convergence related to the process for continuous
smooth functions as well as the rate of convergence. Analogous results for interpolation on the
bounded interval are deduced and finally some numerical examples are presented.

Keywords: lagrange interpolation; unit circle; nodal systems; separation properties; perturbed roots
of the unity; convergence

1. Introduction
The polynomial interpolation is a classical subject that has been widely studied under different
approaches like Lagrange, Hermite, Birkhoff, Pál-type interpolation and some others. Although it is
obvious that the subject is important by itself, its numerous numerical applications like numerical
integration or numerical derivation are not less important and indeed the polynomial interpolation
continues being a subject of current research.
Lagrange interpolation is a very good tool although it is known that for this interpolation scheme
and for good nodal systems such as the classical Chebyshev nodes there exists a continuous function
on [−1, 1] for which the Lagrange interpolation polynomial diverges (see [1]). A similar problem has
been posed for arbitrary arrays and it was proved in [2] that for each nodal array in [−1, 1], there exists
a continuous function such that the Lagrange polynomial interpolation diverges almost everywhere.
In any case, recalling the words written by Trefethen in his paper [3] we can say that there is nothing
wrong with Lagrange polynomial interpolation. "Yet the truth is, polynomial interpolants in Chebyshev
points always converge if f is a little bit smooth". As a consequence, to obtain better results one needs to
assume better properties on the function to be interpolated like bounded variation or a condition on its
modulus of continuity. Thus one of the most important questions in relation with the interpolation of
functions is the choice of the interpolation arrays or nodal systems for which one can expect to obtain
pointwise or uniform convergence to the function to be interpolated and another important issue is to
determine the class of functions for which some type of convergence can be guaranteed. The nodal
systems strongly normal or normal, introduced by Fejér, play an important role in the interpolation
theory, although from a practical point of view, the difficulty of testing the definition makes the

Mathematics 2020, 8, 498; doi:10.3390/math8040498 73 www.mdpi.com/journal/mathematics

Mathematics 2020, 8, 498

applications of these systems quite limited. For these systems Grünwald studies in [4] the convergence
of the polynomials of Lagrange interpolation for functions satisfying a Lipschitz condition.
Most of the research obtain results on the convergence from the distribution properties of the
nodal points. Indeed it was Fejér, who was the first to invert the problem, trying to deduce separation
properties of the nodal systems from the interpolation results. The importance of this idea is avaled by
the fact that the required interpolation properties are easily verified.
In [5] it is proved that strongly normal distributions on [−1, 1] give quasi uniformly nodal systems
on the unit circle, that is the length of the arcs between two consecutive nodes has the order of n1 .
Although the situation more widely studied corresponds to the bounded interval, there are important
results in some other situations in the complex plane like the unit circumference, (see [6]). By taking
into account that continuous functions on the unit circle can be approximated by Laurent polynomials,
the interpolation polynomials on the unit circumference are constructed in this Laurent space. In this
field of research, [7] deserves to be highlighted. There, the nodal points are the roots of complex
numbers with modulus 1 and in this case it is obtained a result about convergence of the interpolants
for continuous functions with a suitable modulus of continuity. Moreover, in the same paper the
problem of extending the results to general nodal systems. Indeed, since the roots of complex numbers
with modulus 1 can be interpreted like the zeros of the para-orthogonal polynomials with respect
to the Lebesgue measure, now the natural extension is to consider the zeros of the para-orthogonal
polynomials with respect to other measures.
In [8] we have generalized the results given in [7] for these new nodal systems. First we work
with nodal systems characterized by fulfilling some properties of boundedness, which are suggested
by those fulfilled by the roots of unimodular complex numbers, obtaining a result of convergence
for continuous functions with a suitable modulus of continuity. Next, by taking into account that the
zeros of the para-orthogonal polynomials with respect to measures in the Szegő class (see [9]) with
analytic extension up to |z| > 1 satisfy the properties that we need, we obtain a similar result about
convergence for these type of nodal systems.
In [10] we have studied the Lagrange interpolation process for piecewise continuous functions
with suitable properties and by using as nodal points the zeros of the para-orthogonal polynomials with
respect to analytic weights, which constitutes a novel approach to the Lagrange interpolation theory.
Another extension to more general nodal systems is given in [11] where it has opened a new
trend to interpolation at perturbed roots of unity and the functions to be interpolated belong to the
disc algebra.
Now, in the present paper we assume a distribution for the nodes that can be obtained through
a perturbation of the uniform distribution and, in particular of the roots of the unity, and which is
more general than that given in [11]. Thus in the present paper we start from a different point of
view because we base it on properties satisfied by the nodal systems and we do not need to consider
orthogonality nor para-orthogonality with respect to any measure on the unit circle. The interpolation
arrays are described by a separation property and the main goal is to obtain the properties that play a
role in the Lagrange process, as well as to present some relevant examples.
The organization of the paper is the following. In Section 2 we introduce the nodal systems
that we use throughout all the paper and we prove the main properties that they satisfy in several
propositions. Section 3 is devoted to the Lagrange interpolation problem. First we obtain the Lebesgue
constant of the process and then we study the convergence of the Lagrange interpolation polynomials
related to continuous functions with appropriate modulus of continuity. Secondly we analyze the rate
of convergence when we deal with smooth functions, (see [12]) and we also deduce analogous behavior
for interpolation on the bounded interval. The last section is devoted to give some numerical examples.

2. Some General Nodal Systems on the Unit Circle

The aim of this paper is to study interpolation problems on the unit circle T = {z ∈ C : |z| = 1}
by using nodal systems satisfying some suitable properties.

74
Mathematics 2020, 8, 498

We denote the nodal polynomials by Wn (z) and their zeros by {α j,n }nj=1 , that is, we assume that
n
Wn (z) = ∏ (z − α j,n ), where |α j,n | = 1 for j = 1, · · · , n, and α j,n = αk,n for j = k. For simplicity, in the
j =1
sequel we omit the subscript n and we write α j instead of α j,n for j = 1, · · · , n. We denote the length
of the shortest arc between any two points of the unit circle, z1 and z2 , by z1 − z2 , and we use the
an
Landau’s notation for complex sequences, denoting by an = O(bn ) if | | is bounded.
bn
Throughout all the paper we assume that the zeros {α j }nj=1 of the nodal polynomials Wn (z) satisfy
A
the following separation property: there exists a positive constant A such that for n > the length of
π
the shortest arc between two consecutive nodes α j and α j+1 , satisfies:

2π A( j)

αj − α j +1 = + 2 with | A( j)| ≤ A ∀ j = 1, · · · , n, (1)
n n
2π 1

where αn+1 = α1 , that is, α j − α j +1 =
+ O( 2 ).
n n
We use the same O to denote different sequences. Unless we mention otherwise, the bounds we
obtain from (1) will be uniform.
We also consider other nodal polynomials, W  n (z), well connected with Wn (z). If we denote
n
 n (z) = zn − λ, with λ = αn , then W
W  n (z) = ∏ (z − β j ), where
1
j =1

√
n
βj = λ, j = 1, · · · , n, and α1 = β 1 .

 n (z) is
Hence it is clear that the separation property satisfied by the zeros { β j } of W

 2π
βj − β j +1 = ∀ j = 1, · · · , n. (2)
n
In this section we obtain in several propositions the main properties of the nodal polynomials
Wn (z). First we recall the following well known relations between arcs and strings that we are going
to use throughout the whole paper and which is based on the convex character of the arcsin function:
If z1 and z2 belong to T then
2
(z 
− z2 ) ≤ |z1 − z2 | ≤ (z1
− z2 ). (3)
π 1

Proposition 1. If {α j }nj=1 and { β j }nj=1 , with α1 = β 1 , are the nodal points satisfying the separation
properties (1) and (2) and we assume they are numbered in the clockwise sense, then

(i)
2π A 2π A
( j − 1)( − 2 ) ≤ α1
− α j ≤ ( j − 1)( + 2 ), for j ≥ 1.
n n n n
(ii)
2π A 2π A
( j + 1)( − 2 ) ≤ αn−
j − α1 ≤ ( j + 1)( + 2 ), for j ≥ 0.
n n n n
(iii)
A
α j
− β j ≤ ( j − 1) , for j ≥ 1.
n2
(iv)
A
αn− j
− β n− j ≤ ( j + 1) 2 , for j ≥ 0.
n

75
Mathematics 2020, 8, 498

2π A 2π A
Proof. (i) By applying (1) we have − 2 ≤ α1 − α2 ≤ + 2 and for j ≥ 3 it holds
n n n n
2π A 2π A
− 2 ≤ α j− 1 − αj ≤ + 2 . Then if we sum, it is straightforward (i).
n n n n
2π A 2π A
(ii) Proceeding in the same way we get − 2 ≤ αn − α1 ≤ + 2 and for j ≥ 0 it holds
n n n n
2π A 2π A
− 2 ≤ αn− j− 1 − αn− j ≤ + 2 . Hence, by computing the sums we have (ii).
n n n n

(iii) We know that α1 = β 1 and we distinguish two possibilities depending on the position of β j
related to α j . If
α1
− β j + β j
− α j = α1
− αj,
that is,
β 1
− β j + β j
− α j = α1
− αj
2π
and we use that β 1
− β j = ( j − 1) , which is a consequence of (2), and we also take into account
n
(i) we get
A
α j
− β j ≤ ( j − 1) .
n2
− α j + α j
The second case corresponds to α1 − β j = β 1
− β j and it can be deduced in the same way.

(iv) We proceed like in (iii) distinguishing the following cases αn−  

j − α1 = α n − j − β n − j + β n − j − α1
or β n− j
− αn− j + αn−j − α 1 = β 
n− j − β 1 .

1
Notice that we can write (iii) and (iv) as follows α j
− β j = ( j − 1)O( ) and αn− j
− β n− j =
n2
1
( j + 1)O( ).
n2

Proposition 2. Let us assume that the zeros of the nodal polynomials Wn (z) satisfy the separation property (1).
Then it holds
|Wn (z)| < 2e A , ∀z ∈ T. (4)

Moreover, it also holds

|Wn (z)| |Wn (z)|
< 2e A and < 2e A , ∀z ∈ T.
n n2

Proof. Since Wn (α j ) = 0 for every j, let us take z ∈ T, such that z is not a nodal point. In order to
obtain the result we renumber the nodes in the clockwise sense in such a way that α1 is the nodal point
nearest to z. We distinguish two cases depending on whether the node closest to z is turning in the
clockwise sense or in the counter clockwise sense from z. If we assume that the situation is given in
Figure 1, that is, α1 is turning in the counter clockwise sense from z, then we have

α1
− α2 π A
z
− α1 < ≤ + 2.
2 n 2n
n
 n (z) = ∏ (z − β j ), with β 1 = α1 and satisfying (2), introduced
Now we consider the polynomial W
j =1
π A 2π
at the beginning of the section. Using property (1) we have + 2 < = α1
− β 2 and then
n 2n n
z − β 2 = 0.

76
Mathematics 2020, 8, 498

z − αj β j − αj
Now, for every j it holds = 1+ and therefore, by using Proposition 1,
z − βj z − βj

  A
z−α  π β j
− αj π ( j − 1) n2
 j
  ≤ 1+ ≤ 1+ .
 z − βj  2 z − βj 2 z − βj

π π ( j − 2)2π ( j − 2)2π
Besides, from property (1) we have z  − β2 > and for j ≥ 3, z 
− βj > + > .
  2n 2n n n
z−α 
 j A
Hence   ≤ 1+ .
 z − βj  n
Proceeding in the same way and taking into account that for j ≥ 0 it holds that β n− j − αn− j ≤
A 2π
( j + 1) 2 and z − β n − j > ( j + 1) we obtain
n n
  A
z−α  π β n− j
− αn− j π ( j + 1) n2
 n− j  A
  ≤ 1+ ≤ 1+ = 1+ .
 z − β n− j  2 z − β n− j 2 2π n
( j + 1)
n
Therefore we have that  
n z−α 
|Wn (z)|  j
= ∏  < e A ∀z,
 n (z)|
|W j =2
 z − βj 

 n (z)| ≤ 2, then we get |Wn (z)| < 2e A .

and since |W
Notice that if the node closest to z, α1 , is in the clockwise sense from z, we can proceed in a similar
α − αn π A π A 2π
way. Indeed z − α1 < 1 ≤ + 2 and since + 2 < β n − α1 = then z − β n = 0.
2 n 2n n 2n n
The second statement, related to the first and second derivatives of the nodal polynomial, is a
consequence of Bernstein’s theorem, (see [13]).







 





 






Figure 1. An arbitrary point z and the nodal system.

Proposition 3. Let us assume that the zeros of the nodal polynomials Wn (z) satisfy the separation property (1).
Then there exists a positive constant C > 0 such that for n large enough and for every j = 1, · · · , n, it holds that

|Wn (α j )|
> C. (5)
n

77
Mathematics 2020, 8, 498

|Wn (α1 )|
Proof. For simplicity we take j = 1 and we try to bound from below .
n
n
 n (z) satisfying (2), that is, W
Thus we consider the polynomial W  n (z) = zn − αn = ∏ (z − β j )
1
j =1
with α1 = β 1 .

 n (α1 )| = n and so our aim is to prove that |Wn (α1 )| > C.
 n (z) = nzn−1 then |W
Since W

|Wn (α1 )|
Now, by taking into account that
 
n α −α 
|Wn (α1 )|  1 j
= ∏ ,
 n (α1 )|
|W j =2
 α1 − β j 
   
α −α  α −α 
 1 j  1 n− j 
we study the quotients   and  .
 α1 − β j   α1 − β n − j 
On the one hand,      
α −α   β j − α j  β −α 
 1 j   j j
  = 1 +  ≥ 1− ,
 α1 − β j   α1 − β j   α1 − β j 

and by applying (3) and Proposition 1, we have for j ≥ 2,

  π A
β −α  π β j
− αj ( j − 1) 2
 j j 2 n A
 ≤ ≤ = ,
 α1 − β j  2 α1− βj 2π 4n
( j − 1)
n
 
α −α 
 1 j A
and therefore   ≥ 1− .
 α1 − β j  4n
On the other hand,
     
α −α   β n− j − αn− j  β 
 1 n− j    n− j − αn− j 
 =
  1 +  ≥ 1 −  
 α1 − β n − j   α1 − β n − j   α1 − β n − j 

and since for j ≥ 0 we have

  π A
β  π β n− j ( j + 1) 2
 n− j − αn− j  − αn− j 2 n A
 ≤ ≤ = ,
 α1 − β n − j  2 α1 − β n− j 2π 4n
( j + 1)
n
 
α −α 
 1 n− j  A
then   ≥ 1− . Hence
 α1 − β n − j  4n
 
n α −α 
|Wn (α1 )|  1 j A n −1
= ∏  ≥ (1 − ) ,
 n (α1 )|
|W j =2
 α 1 − β j  4n

A
A n −1 −
that is, |Wn (α1 )| ≥ (1 − ) n. Thus, given ε > 0 if C = e 4 − ε > 0, then for n large enough it
4n
holds that |Wn (α1 )| > Cn.
Notice that for another value of j there is no any significant difference. Indeed to obtain that
n
 n (z) = zn − αn = ∏ (z − β j ) with αi = β i , we
|Wn (αi )| > Cn, we take the auxiliary polynomial W i
j =1
renumber the nodes as in the previous proof and we proceed in a similar way.

78
Mathematics 2020, 8, 498

Proposition 4. Let us assume that the zeros of the nodal polynomials Wn (z) satisfy the separation property (1).
Then there exists a positive constant D > 0 such that

|Wn (z)|2 n
1
n2 ∑ | z − α j |2 < D, ∀z ∈ T. (6)
j =1

Proof. Following the same steps of the proof of Lemma 2 in [14] we have
n
1
|Wn (z)|2 ∑ | z − α j |2 = |zWn (z)Wn (z) + z2 (Wn (z)Wn (z) − (Wn (z))2 )|
j =1

and therefore by using (4) and its consequences in Proposition 2 we get

|Wn (z)|2 n
1 |Wn (z)| |Wn (z)| |Wn (z)| |W  (z)|2
n2 ∑ | z − α j |2 ≤
n n
+
n2
|Wn (z)| + n 2
n
<
j =1
B
B + B2 + B2 , where B = 2e A .
n

Remark 1. The nodal systems considered in [15] satisfy condition (1). Indeed they are the para-orthogonal
polynomials related to measures in the Baxter class, (see [16]). In that work it is also assumed the additional
condition that the sequence {(φn∗ ) } is uniformly bounded on T, where {φn } is the sequence of monic orthogonal
polynomials related to the measure and {φn∗ } is the sequence of the reciprocal polynomials, (see [9]). In that
situation studied in [15], properties (4)–6) also hold. Now, in the present paper we start from a different point of
view because we base it on properties satisfied by the nodal systems and we do not need to consider orthogonality
nor para-orthogonality with respect to any measure.

3. Lagrange Interpolation. Convergence in Case of Smooth Continuous Functions

To compute the interpolation polynomials, first we recall some well known definitions related
to interpolation problems on the unit circle. We work in the space of Laurent polynomials and, in
particular, in the subspaces Λ p,q [z] = span{zk : p ≤ k ≤ q}, with p and q integers p ≤ q.
Let us continue denoting by {α j }nj=1 the zeros of the the nodal polynomial Wn (z). If {u j }nj=1
are arbitrary complex numbers, the Laurent polynomial of Lagrange interpolation L− E[ n ],E[ n−1 ] (z) ∈
2 2
Λ− E[ n ],E[ n−1 ] [z] characterized by satisfying
2 2

L−E[ n ],E[ n−1 ] (α j ) = u j , for j = 1, · · · , n,

2 2

has the following expression

E[ n2 ]
Wn (z) n αj
L−E[ n ],E[ n−1 ] (z) =
2 2 z E[ 2 ]
n ∑ W  (α j )(z − α j ) u j .
j =1 n

If F is a function and u j = F (α j ) we denote the corresponding Laurent polynomial

L−E[ n ],E[ n−1 ] ( F, z). If n is odd, since E[ n−
2 ] = E [ 2 ], then the interpolation polynomial L− E[ 2 ],E[ 2 ] ( z ) ∈
1 n n n
2 2
Λ− E[ n ],E[ n ] [z] and when n is even, taking into account that E[ n− 2 ] = E [ 2 ] − 1, then the Laurent
1 n
2 2
polynomial of Lagrange interpolation L− E[ n ],E[ n ]−1 (z) ∈ Λ− E[ n ],E[ n ]−1 [z].
2 2 2 2
Without loss of generality, to fix ideas and to simplify the notation we assume that the number
of nodes is even, 2n, in which case the interpolation polynomial Ln,n−1 belongs to the space Λ−n,n−1
and it can be written in terms of the fundamental polynomials as follows:

79
Mathematics 2020, 8, 498

W2n (z) 2n αnj

L−n,n−1 (z) = ∑ W uj. (7)
zn j =1 2n ( α j )( z − α j )

In order to compute the interpolation polynomials for applications and examples it is more convenient
to use the barycentric expression, which is given by

2n wj
∑ u
j =1 z − αj j
L−n,n−1 (z) = , (8)
2n wj
∑
j =1 z − αj

αnj
with w j =  ( α ) , (see [17]).
W2n j
This last expression has some advantages due to its numerical stability in the sense established
in [18]. In this article author claims literally:
The Lagrange representation of the interpolating polynomial can be rewritten in two more computationally
attractive forms: a modified Lagrange form and a barycentric form. We give an error analysis of the evaluation of
the interpolating polynomial using these two forms. The modified Lagrange formula is shown to be backward
stable. The barycentric formula has a less favourable error analysis, but is forward stable for any set of
interpolating points with a small Lebesgue constant. Therefore the barycentric formula can be significantly less
accurate than the modified Lagrange formula only for a poor choice of interpolating points.
So with a good Lebesgue constant (see next Theorem 1) we have good accuracy, at least as good
as the intensively used Lagrange interpolation on the Chebyshev nodal systems.
Following [10] we can obtain the Lebesgue constant, (see [19]), and the convergence of this
interpolatory process. Notice that this is a novelty result for our general nodal systems satisfying
property (1), although the techniques that we use are the same as in [10].

Theorem 1. There exists a positive constant L > 0 such that for every function F bounded on T it holds that

|L−n,n−1 ( F, z)| ≤ L  F ∞ log n,

for every z ∈ T, where  ∞ denotes the supremum norm on T.

Proof. Let z be an arbitrary point of T and assume that z is not a nodal point. Then, if we continue
assuming the even case, from (7) we get

2n  W (z) F (α ) 
 2n j 
|L−n,n−1 (z)| ≤ ∑  W ,
j =1 2n ( α j )( z − α j )

and by our hypothesis about F and by Proposition 3 we have

 F ∞ 2n  W (z) 
 2n 
|L−n,n−1 (z)| ≤
2nC ∑  z − αj .
j =1

If we assume that the nodal points closest to z are α1 and α2n then by applying (1) we obtain that for
j > 1 it holds
2π 1
z− α j > ( j − 1)( + O( 2 )).
2n 4n
Thus, by using (3) we obtain
1 π 2n 1 nE
< = ,
|z − α j | 2 ( j − 1) 1 j−1
(2π + O( ))
2n

80
Mathematics 2020, 8, 498

for some positive constant E.

2π 1 1 nE
Proceeding in the same way we get z −
α2n− j > j( + O( 2 )) and therefore < .
2n 4n |z − α2n− j | j

We also obtain  W (z)   W (z) 

 2n   2n 
 ,   < 2nK,
z − α1 z − α2n
for some positive constant K.
 ( z )|| θ − θ | ≤
Indeed |W2n (z)| = |W2n (z) − W2n (α1 )| = |W2n (eiθ ) − W2n (eiθ1 )| ≤ 2 max |W2n 1
z∈T
 ( z )| π | z − α | ≤ 2nK | z − α |.
max |W2n 1 1
z∈T
Hence

 F ∞  W (z)  n   n −1  W ( z )   W ( z ) 
 2n   W2n (z)   2n   2n 
|L−n,n−1 (z)| ≤  + ∑ + ∑  +  ≤
2nC z − α1 j =2
z − α j j =1
z − α 2n − j z − α 2n
 
 F ∞ n
2e A nE 2  F ∞ n
eA E
4nK + 2 ∑ = K+ ∑ ≤ 2  F ∞ P(1 + Hn−1,1 ),
2nC j =2
j−1 C j =2
j−1

n −1 1
with Hn−1,1 the harmonic number equal to ∑ and P a positive constant.
j =1 j

Remark 2. When the values of F (αi ) are affected by any type of error, which we can suppose is bounded by
some > 0, then the previous result ensures us, taking into account the linearity of the interpolation process,
that the final result is affected by an error bounded by L log(n) , that is, it is at least so good as the intensively
used Lagrange interpolation on the Chebyshev nodal systems.

For applying the interpolation it could be very useful the following results concerning the
convergence and the rate of convergence for smooth continuous functions (see [10,12]).

Theorem 2. (i) Let F (z) be a function defined on T. If F is continuous with modulus of continuity ω ( F, δ) =
o (| log δ|−1 ), then L−n,n−1 ( F, z) converges uniformly to F on T.
∞ 1
(ii) Let F (z) be a function defined on T. If F (z) = ∑ Ak zk with | Ak | ≤ K c for k = 0, with c > 1
−∞ |k| 
log n
then L−n,n−1 ( F, z) uniformly converges to F on T and the rate of convergence is O .
n c −1
(iii) If F (z) is an analytic function in an open annulus containing T, then L−n,n−1 ( F, z) uniformly
converges to F on T. Besides, the rate of convergence is geometric.

Proof. The results are consequence of the preceding Theorem 1 and they are also based on the
properties satisfied by our nodal systems. Thus one can obtain these results following the same steps
as in the proof of Theorems 3 and 4 in [10], where one can see the details.

The Case of the Bounded Interval

We recall that the Lagrange interpolation polynomial n−1 ( x ) related to a nodal system { x j }nj=1 in
[−1, 1] and satisfying the conditions {v j }nj=1 is given by

n
wn ( x )
 n −1 ( x ) = ∑ w (x j )(x − x j ) v j ,
j =1 n

n
where wn ( x ) = ∏ ( x − x j ). When v j = f ( x j ) for a function f , we write n−1 ( f , x ).
j =1

81
Mathematics 2020, 8, 498

n
In this subsection we consider the nodal polynomial wn ( x ) = ∏ ( x − x j ) with { x j }nj=1 ⊂ [−1, 1]
j =1
and numbered in the following way: 1 ≥ x1 > x2 > · · · > xn−1 > xn ≥ −1.
We also assume that the nodes satisfy the following separation property:
A
There exists a positive constant A such that for n >
π
π a( j)
(i) If x1 = 1 and xn = −1 then arccos x j+1 − arccos x j = + 2 , with | a( j)| ≤ A, ∀ j = 1, · · · , n − 1.
n n

π a( j)
(ii) If x1 < 1 and xn = −1 then arccos x j+1 − arccos x j = + 2 , with | a( j)| ≤ A, ∀ j = 1, · · · , n − 1,
n n
π a (0)
and 2 arccos x1 = + 2 , with | a(0)| ≤ A.
n n
π a( j)
(iii) If x1 = 1 and xn > −1 then arccos x j+1 − arccos x j = + 2 , with | a( j)| ≤ A, ∀ j = 1, · · · , n − 1,
n n
π a(n)
and 2(π − arccos xn ) = + 2 , with | a(n)| ≤ A.
n n
π a( j)
(iv) If x1 < 1 and xn > −1 then arccos x j+1 − arccos x j =
+ 2 , with | a( j)| ≤ A, ∀ j = 1, · · · , n − 1,
n n
π a (0) π a(n)
and 2 arccos x1 = + 2 , with | a(0)| ≤ A, and 2(π − arccos xn ) = + 2 , with | a(n)| ≤ A.
n n n n
Under the above assumptions we obtain the following results about the convergence and the rate of
convergence for the interpolation polynomials with these nodal systems.

Theorem 3. If f is a continuous function on [−1, 1] and ω ( f , δ) = o (| log δ|−1 ), then the interpolation
polynomial n−1 ( f , x ) fulfilling the interpolation conditions

n−1 ( f , x j ) = f ( x j ) for j = 1, · · · , n (9)

converges uniformly to f on [−1, 1].

1
z+
Proof. Let us define a continuous function F on T by F (z) = F (z) = f ( x ) with x = z . Then it is
2
clear that its modulus of continuity satisfies

ω ( F, δ) = sup | F (z1 ) − F (z2 )| ≤ sup | f ( x1 ) − f ( x2 )| = ω ( f , δ).

z1 ,z2 ∈T,|z1 −z2 |<δ x1 ,x2 ∈[−1,1],| x1 − x2 |<δ

To fix ideas we assume that x1 = 1 and xn = −1, that is, case (iv). By applying Szegő
1
z+ n n αj + αj
z) = 1
transformation wn (
2 n n ∏
2 z j =1
(z − α j ) ∏ (z − α j ), where
2
= x j , that is, α j = eiθ j
j =1
1
z+
with θ j = arccos x j . Hence we consider the nodal polynomial W2n (z) = 2n zn wn ( z) =
2
n n
∏(z − α j ) ∏(z − α j ) .
j =1 j =1
Now our nodal system is constituted by {α j }nj=1 ∪ {α j }nj=1 and the arguments are {θ j }nj=1 ∪
{−θ j }nj=1 . If we renumber the arguments in such a way that −θn = θn+1 , · · · , −θ1 = θ2n , then it
holds that
2π A( j)
θ j +1 − θ j = + 2 ,
2n n
with | A( j)| ≤ A for j = 1, · · · , 2n and θ2n+1 = θ1 .

82
Mathematics 2020, 8, 498

Let L−n,n−1 ( F, z) be the Lagrange interpolation polynomial satisfying the conditions

L−n,n−1 ( F, α j ) = L−n,n−1 ( F, α j ) = f ( x j ), for j = 1, · · · , n.

Since F satisfies the hypothesis of Theorem 2 (i) then L−n,n−1 ( F, z) converges uniformly to F on T.
Analizying the expression of L−n,n−1 ( F, z) and by taking into account that W2n (z) as well
 ( z ) have real coefficients we get that L
as W2n −n,n−1 ( F, z ) has real coefficients. Since it is clear
1
that L−n,n−1 ( F, ) ∈ Λ−(n−1),n satisfies the same interpolation conditions, now we consider the
z  
1 1
algebraic polynomial L−n,n−1 ( F, z) + L−n,n−1 ( F, ) , which has real coefficients and satisfies
2 z
the interpolation conditions (9).  Since the polynomial satisfying (9) is uniquely determined, then
1 1
L−n,n−1 ( F, z) + L−n,n−1 ( F, ) = n−1 ( f , x ) and it converges uniformly to f on [−1, 1].
2 z
When x1 = 1 (case (iii)), or xn = −1 (case (ii)), or x1 = 1 and xn = −1 (case (i)), one proceeds
1
2n z n z+
in a similar way and the auxiliary nodal polynomials are given by W2n−1 (z) = wn ( z) =
z−1 2
1
n n
2n z n z+ n −1 n −1
(z − 1) ∏ (z − α j ) ∏(z − α j ) or W2n−1 (z) = z ) = ( z + 1)
z+1
wn (
2 ∏ (z − α j ) ∏ (z − α j ) or
j =2 j =2 j =1 j =1
1
2n z n z+ n −1 n −1
z ) = (z − 1)(z + 1)
W2n−2 (z) =
(z − 1)(z + 1)
wn (
2 ∏ (z − α j ) ∏ (z − α j ), respectively. Notice
j =2 j =2
that in the four cases the nodal polynomials have real coefficients.

Notice that some well known results related to Lagrange interpolation with Chebyshev and
Chebyshev extended nodes are particular cases of the above theorem (see [9,20]).
Notice that from the proof of the above theorem and by applying Theorem 1 we can obtain an
analogous bound, the Lebesgue constant, for this interpolatory process, that is, there exists a positive
constant M such that
|n−1 ( f , x )| ≤ M  f ∞ log n. (10)

In order to obtain information concerning the rate of convergence, first we recall the following
result about the expansion of an analytic function in a Jacobi series (see [9,21]). For a more actual
version see [12].

Theorem 4. Let f ( x ) be analytic on the closed segment [−1, 1]. The expansion of f in a Jacobi series, f ( x ) ∼
∞ (α,β)
∑ an Pn ( x ), is convergent in the interior of the greatest ellipse with foci at ±1, in which f is regular.
n =0
The expansion is divergent in the exterior of this ellipse and the sum R of the semi-axes of the ellipse of
1
convergence is R = lim inf  .
n
| an |

Thus, in our conditions we have the following results, which are in concordance with Theorem 2.

Theorem 5. (i) If f is a function defined on [−1, 1] by f ( x ) = ∑∞

k =0 ak Tk ( x ), where Tk ( x ) is the Chebyshev
K
polynomial of degree k and with | ak | ≤ s , with k = 0, K > 0 and s > 1, then the Lagrange interpolation
k  
log n
polynomial n−1 ( f , x ) converges to f with rate of convergence O .
n s −1
(ii) If f is analytic on the closed segment [−1, 1], then the Lagrange interpolation polynomial n−1 ( f , x )
converges to f with rate of convergence geometric.

83
Mathematics 2020, 8, 498

n −1 ∞
Proof. (i) If we decompose f like f ( x ) = f 1,n−1 ( x ) + f 2,n−1 ( x ) = ∑ ak Tk ( x ) + ∑ ak Tk ( x ), we have
k =0 k=n
∞ 1 ∞ K
that n−1 ( f 1,n−1 , x ) = f 1,n−1 ( x ) and | f 2,n−1 ( x )| ≤ ∑ | ak | ≤ K ∑
< , where the
k=n k=n k s ( s − 1 )( n − 1) s −1
last inequality follows from the integral criterion. Thus, by applying the Lebesgue constant obtained
K
before in (10), we get |n−1 ( f 2,n−1 , x )| ≤ M log n. Hence
(s − 1)(n − 1)s−1

| f ( x ) − n−1 ( f , x )| = | f 1,n−1 ( x ) − n−1 ( f 1,n−1 , x ) + f 2,n−1 ( x ) − n−1 ( f 2,n−1 , x )| =

K log n
| f 2,n−1 ( x ) − n−1 ( f 2,n−1 , x )| ≤ (1 + M log n) ≤ T s−1 ,
(s − 1)(n − 1)s−1 n

for some T > 0.

(ii) Since f is analytic on [−1, 1], then it can be analytically continued to a neighborhood of [−1, 1]
∞
in the complex plane. Hence the expansion in Chebyshev series ∑ an Tn ( x ) converges to f in the
n =0
interior of the greatest ellipse with foci at ±1, known as Bernstein ellipse ER and the sum R of the
1
semi-axes of the ellipse of convergence is R = lim inf  . Then it holds that | an | ≤ Pr n , for some
n
| an |
0 < r < 1 and P > 0.
n −1
Proceeding in the same way as before we have f ( x ) = f 1,n−1 ( x ) + f 2,n−1 ( x ) = ∑ ak Tk ( x ) +
k =0
∞ ∞ rn
∑ ak Tk ( x ), n−1 ( f 1,n−1 , x ) = f 1,n−1 ( x ) and | f 2,n−1 ( x )| ≤ ∑ | ak | ≤ P . Thus, by applying the
k=n k=n 1−r
rn
Lebesgue constant we get |n−1 ( f 2,n−1 , x )| ≤ MP log n and therefore
1−r

| f ( x ) − n−1 ( f , x )| = | f 2,n−1 ( x ) − n−1 ( f 2,n−1 , x )| ≤ Qr n log n ≤ Qr1n ,

for some Q > 0 and 0 < r < r1 < 1.

4. Numerical Examples
We have carried out different numerical experiments to visualize the main contributions of this
article. The first examples correspond to the three cases of Theorem 2 and in all of them we work in
the following way:

1. We construct the nodal systems in a quite random way. We consider four arcs or sections in
n
the unit circumference T. The first one begins in α1 = 1 and its nodes are constructed in
4
2π
counter clockwise sense separated by an angular length + , where the are random errors
n
A 2A
determined by using the uniform distribution in [ 2 2π, 2 2π ]. The fourth section begins in
n n
n 2π
α1 = 1 and its nodes are constructed in clockwise sense with arcs of angular length + ,
4 n
A 2A
where the are random errors determined by using the uniform distribution in [ 2 2π, 2 2π ].
n n
n
The second section begins after the first one and its nodes are constructed in counter clockwise
4
2π
sense with arcs of angular length + , where the are random errors determined by using the
n
2A A
uniform distribution in [− 2 2π, − 2 2π ]. Finally, in the third section the arcs between the nodes
n n
are all equal.
Obviously we use different values of n and we must remark that we obtain always the same
results, really we must say similar results because due to our random choice we never have the
same nodal system.

84
Mathematics 2020, 8, 498

2. We consider a test function F (z), that we detail in each example, and we always plot F (z) in black.
3. We consider the Lagrange interpolation polynomial L−n,n−1 ( F, ), which interpolates the test
function at the nodal system. We always plot (L−n,n−1 ( F, )) in red.

These examples are devoted to visualize the items (i), (ii) and (iii) respectively of Theorem 2.
   
z + z −1 2
Example 1. In this example we work with F (z) = 1 + 20 sin −
for z ∈ T, which
2 z+z 1
satisfies the hypotheses of Theorem 2 (i). We take n = 1000, A = 2 and we use (8) to obtain L−n,n−1 ( F, ).
We represent the function F (eiθ ) which takes real values and, as we have said, the real part of the
interpolation polynomial. Notice that due to its variability, F is a quite difficult function to interpolate.
π
Indeed, it is easy to check that F (ei θ ) is not differentiable at .
2
We present in Figure 2 two graphics. On the left we have a general panoramic of the interpolation along
T and we have added the interpolation points in green. We must point up that the interpolatory process is
successful where the function has no variability. However, we have an unsuccessful situation where the function
has great variability.
π
In the graphic on the right we have a detailed situation between 1.2 and 2, that is near , which can help
2
us to understand the problem. According to the theory presented, we must increase the number of nodes to obtain
better results in this region.









 



    

z + z −1 2
Figure 2. F (z) and (L−n,n−1 ( F, z)) with F(z)= 1 + 20( ) sin( ), z = eiθ , θ ∈ [0, 2π ],
2 z + z −1
θ ∈ [1.2, 2] and n = 1000.
∞
1
Example 2. Now we consider the function defined on T by F (z) = ∑ (zk + z−k ), which satisifies
k =1
k6
the hypotheses of Theorem 2. In the next Figure 3 we plot on the left F (e )
iθ and (L−n,n−1 ( F, eiθ )) for
θ ∈ [0, 2π ] and n = 60. Notice that they are indistinguishable. On the right we plot the errors given by
(L−n,n−1 ( F, eiθ )) − F (eiθ ) with θ ∈ [0, 2π ]. We point out that the errors are less or equal than 2 × 10−8 .

In the next example we also construct an alternative interpolation polynomial based on the
equispaced nodal system on T, but using the values of the function on our nodal system. We do this
because a natural criticism to our method could be that with errors as O(1/n2 ) we can be so close to
the equispaced nodal system to accept this approximation. We denote by A−n,n−1 ( F, ) this alternative
interpolation polynomial.

85
Mathematics 2020, 8, 498

 


    

    

 



 


∞ 1 k
Figure 3. F (z) and (L−n,n−1 ( F, z)) and (L−n,n−1 ( F, z)) − F (z) with F (z) = ∑ ( z + z − k ),
k =1 k6
z = eiθ , θ ∈ [0, 2π ] and n = 60.

Example 3. In this example we take F (z) = ez , n = 24, A = 2 and we use (8) to obtain the interpolation
polynomials L−n,n−1 ( F, ) and A−n,n−1 ( F, ). Taking into account that F is analytic we know that F and
L−n,n−1 ( F, ) must be close. In Figure 4 we plot ( F ) in black, (L−n,n−1 ( F, )) in red and (A−n,n−1 ( F, ))
in green for z = eiθ with θ ∈ [0, 2π ]. On the left hand side we have a global vision with θ ∈ [0, 2π ] and we can
observe that ( F ) and (L−n,n−1 ( F, )) are indistinguishable; in fact for this example the maximum error was
3 × 10−9 .
Although (A−n,n−1 ( F, )) has a similar shape, see that it can drive us to catastrophic errors. On the right
hand side we present a detail of the previous one, which give us an idea of the error. Notice that in general we
cannot know the details of the nodal distribution.
We have done the same with the imaginary part and we obtain the same results.


















        

Figure 4. ( F (z)), (L−n,n−1 ( F, z)) and (A−n,n−1 ( F, z)) with F (z) = ez , z = eiθ , θ ∈ [0, 2π ], θ ∈
[4.5, 4.9] and n = 24.

Example 4. Finally we choose F (z) = χS (z) defined on T as the characteristic function of the superior arc
S of T, we take n = 2000, A = 2 and we use expression (8) to obtain L−n,n−1 ( F, ). We know the behavior
when the nodal system is related to para-orthogonal polynomials with respect to an analytic positive measure
(see [10]), but we do not have a theory about the behavior of L−n,n−1 ( F, ) in our situation. We plot the results
in Figure 5. Notice that the basic ideas of the Gibbs–Wilbraham phenomenon are present in this graphic, that is,
the convergence of the interpolator to the function in regions which are far enough from the discontinuities and a
heavy oscillation near the discontinuities. A representation of the oscillation and its amplitude, maybe, deserves
a detailed study.

86
Mathematics 2020, 8, 498







 

 

 

 

      
 
   




Figure 5. F (eiθ ) and (L−n,n−1 ( F, eiθ )) with F (z) = χS (z), z = eiθ , θ ∈ [0, 2π ], θ ∈ [3, 3.3] and
n = 2000.

5. Discussion
Usually, the nodal systems used for interpolation problems are strongly connected with measures
on the bounded interval and on the unit circle and their corresponding orthogonal or paraorthogonal
polynomials. We must point out that these choices are very suitable to construct the whole theory but
in some numerical applicatons it is possible that the nodal systems do not satisfy this requisite. So, the
starting point of the paper is a distribution for the nodes that can be obtained through a perturbation
of the uniform distribution and, in particular of the roots of the unity, and which is more general than
that related to measures and orthogonality.
The results of this article contribute to elaborate a theory over these type of nodal systems, as well
as to the Lagrange interpolation theory based on these interpolatory arrays. Moreover, a theory about
the rate of convergence for some types of smooth functions is given. Finally, we translate the results to
perturbed Chevyshev nodal systems and to Lagrange interpolation on the bounded interval.
We think that this research could be of interest for some mechanical models that generate these
types of nodal systems. As an example we consider the next problem.
Let us suppose that we are studying a equatorial characteristic F (eiθ ) of a planet which depends
on the angle θ and we have a theory which establishes that F (eiθ ) is an analytic function. We observe
the phenomenon using an observatory in the boundary of a spatial station in an elliptic orbit of
period T which rotates over itself with period T1 (with T = n T1 and n large enough). Moreover, we
take our observations when the center of the planet, the observatory and the center of the station
are aligned. We can translate the problem thinking that the planet is our Sun, the spatial station is
our Earth and the observatory is our city. So the time between observations is the equivalent of a
solar day. It is well known that the duration of a solar day is not constant (see https://en.wikipedia.
org/wiki/Equation_of_time for a brief introduction about the so called Equation of time), in our
case have a little oscillation on T1 and our observations are taken on a nodal system which satisfies
(1). Notice that in this case we do not have a equispaced distribution nor the support of the theory
of Orthogonal Polynomials. Therefore, before this paper we did not know how to use our data to
reconstruct F (eiθ )and after this paper we can be confident about the use of Lagrange interpolation.
Some future research directions could be the study of other types of interpolation on the unit
circle and on the bounded interval by using these general interpolatory arrays; as well as to study the
correspondig Gibbs–Wilbraham phenomena.

6. Materials and Methods

The experiments given in the section Numerical examples were obtained by using personal codes
elaborated with Mathematica R
12 (Wolfram Research Europe Ltd, Long Hanborough Oxfordshire,
United Kingdom). These programs to obtain the nodal points and to compute the interpolation
polynomials as well as the plots of the test functions and their interpolators are available at the public

87
Mathematics 2020, 8, 498

repository https://www.dropbox.com/sh/0cx9chq3jfzov2w/AAA_SvL2i7HlC7ChMGpuG-Ata?dl=0
There one can find the program related to Example 2. To obtain the other examples some minor
changes must be done.

Author Contributions: Conceptualization, E.B., A.C. (Alicia Cachafeiro), A.C. (Alberto Castejón) and J.M.G.-A.;
Investigation, E.B., A.C. (Alicia Cachafeiro), A.C. (Alberto Castejón) and J.M.G.-A.; Software, E.B., A.C. (Alicia
Cachafeiro), A.C. (Alberto Castejón) and J.M.G.-A.; Writing—original draft, E.B., A.C. (Alicia Cachafeiro), A.C.
(Alberto Castejón) and J.M.G.-A. All authors have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Szili, L.; Vértesi, P. On the theorem of Géza Grünwald and Józef Marcinkiewicz. Banach Cent. Publ. 2011,
95, 251–259. [CrossRef]
2. Erdös, P.; Vértesi, P. On the almost everywhere divergence of Lagrange interpolation of polynomials for
arbitrary systems of nodes. Acta Math. Acad. Sci. Hung. 1980, 36, 71–89. [CrossRef]
3. Trefethen, L.N. Six Myths of Polynomial Interpolation and Quadrature. Math. Today (Southend-on-Sea) 2011,
47, 184–188.
4. Grünwald, G. On the theory of interpolation. Acta Math. 1942, 75, 219–245. [CrossRef]
5. Erdös, P.; Turán, P. On Interpolation II: On the Distribution of the Fundamental Points of Lagrange and
Hermite Interpolation. Ann. Math. 1938, 9, 703–724. [CrossRef]
6. Walsh, J.L. Interpolation and Approximation by Rational Functions in the Complex Plane, 5th ed.; American
Mathematical Society Colloquium Publications: Providence, RI, USA, 1969; Volume 20.
7. Darius, L.; González-Vera, P. Some results about interpolation with nodes on the unit circle. Indian J. Pure
Appl. Math. 2000, 31, 1273–1296.
8. Berriochoa, E.; Cachafeiro, A.; García-Amor, J.M. About nodal systems for Lagrange interpolation on the
circle. J. Appl. Math. 2012, 2012. [CrossRef]
9. Szegő, G. Orthogonal Polynomials, 4th ed.; American Mathematical Society Colloquium Publications:
Providence, RI, USA, 1975; Volume 23.
10. Berriochoa, E.; Cachafeiro, A.; García-Amor, J.M. Gibbs–Wilbraham phenomenon on Lagrange interpolation
based on analytic weights on the unit circle. J. Comput. Appl. Math. 2020, 365, 112376. [CrossRef]
11. Chui, C.K.; Shen, X.-C.; Zhong, L. On Lagrange interpolation at disturbed roots of unity. Trans. Am. Math. Soc.
1993, 336, 817–830. [CrossRef]
12. Trefethen, L.N. Approximation Theory and Approximation Practice; Society for Industrial and Applied
Mathematics (SIAM): Philadelphia, PA, USA, 2013.
13. Milovanovic, G.V.; Mitrinovic, D.S.; Rassias, T.M. Topics in Polynomials: Extremal Problems, Inequalities, Zeros;
World Scientific Publishing Co.: Singapore, 1994.
14. Berriochoa, E.; Cachafeiro, A.; Martínez, E. About measures and nodal systems for which the Hermite
interpolants uniformly converge to continuous functions on the circle and interval. Appl. Math. Comput.
2012, 218, 4813–4824. [CrossRef]
15. Berriochoa, E.; Cachafeiro, A.; García-Amor, J.M. Gibbs–Wilbraham oscillation related to an Hermite
interpolation problem on the unit circle. J. Comput. Appl. Math. 2018, 344, 657–675. [CrossRef]
16. Simon, B. Fine structure of the zeros of orthogonal polynomials, I. A tale of two pictures. Electron. Trans.
Numer. Anal. 2006, 25, 328–368.
17. Corless, R.M.; Fillion, N. A Graduate Introduction to Numerical Methods; Springer: New York, NY, USA, 2013.
18. Higham, N.J. The numerical stability of barycentric Lagrange interpolation. IMA J. Numer. Anal. 2004,
24, 547–556. [CrossRef]
19. Berrut, J.-P.; Mittelmann, H.D. Lebesgue constant minimizing linear rational interpolation of continuous
functions over the interval. Comput. Math. Appl. 1997, 33, 77–86. [CrossRef]

88
Mathematics 2020, 8, 498

20. Szabados, J.; Vértesi, P. Interpolation of Functions; World Scientific: Singapore, 1990.
21. Rivlin, T.H. The Chebyshev Polynomials; John Wiley & Sons: New York, NY, USA, 1974.

c 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).

89
 mathematics
Article
A Functional Data Analysis Approach for the
Detection of Air Pollution Episodes and Outliers:
A Case Study in Dublin, Ireland
Javier Martínez Torres 1, *, Jorge Pastor Pérez 2 , Joaquín Sancho Val 3 , Aonghus McNabola 4 ,
Miguel Martínez Comesaña 5 and John Gallagher 4
1 Department of Applied Mathematics I. Telecommunications Engineering School, University of Vigo,
36310 Vigo (Pontevedra), Spain
2 Centro de Evaluación, Formación y Calidad de Aragón, 50018 Zaragoza, Spain; [email protected]
3 Centro Universitario de la Defensa. Academia General Militar, 50090 Zaragoza, Spain; [email protected]
4 Department of Civil, Structural and Environmental Engineering, Trinity College Dublin, University of
Dublin, Dublin D02 PN40, Ireland; [email protected] (A.M.); [email protected] (J.G.)
5 Escuela de Ingeniería Industrial, University of Vigo, 36310 Vigo (Pontevedra), Spain; [email protected]
* Correspondence: [email protected]

Received: 14 January 2020; Accepted: 4 February 2020; Published: 10 February 2020

Abstract: Ground level concentrations of nitrogen oxide (NOx) can act as an indicator of air quality in
the urban environment. In cities with relatively good air quality, and where NOx concentrations rarely
exceed legal limits, adverse health effects on the population may still occur. Therefore, detecting
small deviations in air quality and deriving methods of controlling air pollution are challenging.
This study presents different data analytical methods which can be used to monitor and effectively
evaluate policies or measures to reduce nitrogen oxide (NOx) emissions through the detection of
pollution episodes and the removal of outliers. This method helps to identify the sources of pollution
more effectively, and enhances the value of monitoring data and exceedances of limit values. It will
detect outliers, changes and trend deviations in NO2 concentrations at ground level, and consists
of four main steps: classical statistical description techniques, statistical process control techniques,
functional analysis and a functional control process. To demonstrate the effectiveness of the outlier
detection methodology proposed, it was applied to a complete one-year NO2 dataset for a sub-urban
site in Dublin, Ireland in 2013. The findings demonstrate how the functional data approach improves
the classical techniques for detecting outliers, and in addition, how this new methodology can
facilitate a more thorough approach to defining effect air pollution control measures.

Keywords: air pollution; functional data analysis; non-normal data; statistical process control; outlier

1. Introduction
Nowadays, most cities have an increasing environmental problem related to air pollution [1–4].
This specific pollution is a continuing threat to human health and welfare, with a range of different
sources generating different pollutants which have distinct health effects on urban populations [5–7].
Detailed air quality monitoring data for pollutants, such as carbon monoxide (CO), nitrogen oxides
(NO and NO2 ), sulphur dioxide (SO2 ), ozone (O3 ) and particulate matter (PM10 and PM2.5 ), are
becoming more important because of the health problems said pollutants can cause in living beings [6].
The measurements of pollutants provide real-time data to inform the public and provide a mechanism
of alerting local residents of a possible hazard. In particular, pollutant sources from traffic emissions,
such as NOx, which represents a combination of nitrogen oxide (NO) and nitrogen dioxide (NO2 ), are
typically emitted at ground level from vehicles and are associated with health-related problems [8].

Mathematics 2020, 8, 225; doi:10.3390/math8020225 91 www.mdpi.com/journal/mathematics

Mathematics 2020, 8, 225

Despite a reduction in emissions from the transport sector, an increasing trend in NO2 concentrations
has been observed in a number of different European countries; for example, the United Kingdom and
Ireland [9,10]. Therefore, meeting the standards and air quality guidelines by European and national
environmental agencies for pollutants such as NO2 is becoming more challenging [9], as exceedances
of pollutant concentrations can lead to short-term, chronic human health problems [11].
On the other hand, it is understandable that occasional values in polluted air samples behave as
outliers in an urban environmental database. They can be classified as local outliers [12,13] or global
outliers. Unlike global outliers, local outliers can be detected by comparison with near neighbours.
For the purpose of air pollution studies in urban areas, global outliers that deviate from the guide
values indicate that there may be a significant source of pollution. Observations which are not
excessively high but are different from neighbouring values may also contain information on unusual
processes such as pollution. Outliers may merely be noisy observations, or alternatively, they may
indicate atypical behaviour in the system. These abnormal values are very important and may lead to
useful information or significant discoveries, but also contribute to the selection of the most suitable
mitigation techniques or measures [14].
Different techniques of functional data analysis (FDA) have been used in vectorial problems. This
new methodology appeared due to the inefficiency of the classical data mining techniques treating
vector data [15]. FDA is applicable in a multitude of fields, such as environmental [16–19] and medical
research [20], and is applicable for sensors [21,22] and industrial methods [23,24]. The functional model
is based on two ideas that make it unique: it takes into account the time correlation structure of the
data and leads to a global view of the problem through curves analysis instead individual observations.
This analysis is focused on the comparison of the curves using the functional depths, a variable that
measures the centrality of a given curve within a set of curves [25]. Functional depth has already been
used in several environmental problems [26,27].
The aspiration of this research is to create a model to detect air pollution episodes and identify
outliers in gaseous emissions, and to validate this method using real world data from a suburban air
quality monitoring site in Dublin, Ireland. Although many methods are known to identify outliers
(from the classical Grubbs test [28] to a test proposed in 2019 by [29]), they are all based on the vector
approach. This study was carried out, on one hand, with conventional methods, and on the other hand,
with a functional approach; a comparative study between the two methodologies is presented. Each
method will be presented and the findings will outline the most effective method for detecting outliers
in air pollution monitoring data to enhance its capacity for informing new measures to improve local
air quality.

2. Methods

2.1. Case Study—A Sub-Urban Air Quality Monitoring Station in Dublin, Ireland
Ireland has a range of air quality monitoring stations across the country, which are part of
the national ambient air quality monitoring programme (AAMP). The data collated from these
monitoring sites are used to inform on air quality at the local and national levels, and are being used
for forecast modelling. The Blanchardstown sub-urban site is one of the 17 national sites, managed
by the Environmental Protection Agency (EPA), which monitors NO2 and is classified as a suburban
monitoring site. It is located to the west of Dublin city centre in Ireland [30]. The Blanchardstown air
quality monitoring station was selected, as it provided continuous, high-resolution NO2 data emissions
over a 1-year period. Its location is adjacent to the major arterial carriageway around Dublin city
centre, and as a monitoring location, is therefore affected by traffic emissions. In this manner, NO2
hourly data was collected throughout 2013, with 96% of data capture and availability from [31], and the
information needed about weather conditions in Dublin in 2013 was obtained on [32].
EU legislation for NO2 limit values (2008 CAFE Directive and S.I. number 180 of 2011) align with
the World Health Organisation (WHO) guidelines, with 1-hour and 1-year limit values of 200 μg/m3

92
Mathematics 2020, 8, 225

and 40 μg/m3 respectively. In circumstances in which the hourly value is exceeded on three consecutive
hours, short-term action plans must be implemented by local authorities to mitigate against continued
pollution events (limiting traffic flows, restricting construction work, industrial processes, etc.). Despite
there being no daily NO2 limit value in the EU and WHO, some countries have set an average daily
limit of 100 μg/m3 (range from 80–150 μg/m3 ) [31].
The sources and trends of NO2 emissions over the last 20 years has seen a recent increasing trend
due to growth in the transport sector and a recovery since the 2008 economic downturn. As such,
Ireland’s air quality in relation to NO2 is considered to be deteriorating, as measurement data suggests
it may reach limit values and the national emissions ceiling in the coming years.

2.2. Analysis Methods

A range of systems are available to analyse environmental data, such as air quality measurements.
These systems can be used to detect uncharacteristic data points by taking into account trends,
variations between neighbouring network stations and expected values with respect to the sampling
location. An example of these expert systems of data and environmental parameter validation would
be the trends analysis throughout R-programming (openair) [33].
With classical analysis, the data are only analysed statically. The proposed methodology includes
using a large amount of existing data to extract conclusions. Today, the amount of data that has been
stored in environmental databases requires automated analysis techniques. Actually, the analysis
methodology presented here is oriented to knowledge discovery in databses (KDD) [34], which
provides a complete process for extracting information and also provides a clear methodology for the
preparation of data and interpreting the results obtained. KDD involves an iterative and interactive
process of searching for models, patterns and parameters that are useful for detection, classification
and/or prediction in order to generate knowledge and help in decision making.

2.2.1. Classical Analysis

The classical monitoring strategy for air quality uses individual time series, descriptive statistics,
box plots, autocorrelation analysis, etc., to determine if any of the values fall outside of the limits and
to analyse trends [35,36]. In general, classical statistical analysis seeks to describe the distribution of a
measurable variable (descriptive statistics) and to determine the consistency of a sample drawn from
an initial population (inferential statistics). In addition, classical analysis is based on repetition; one
must measure properties of objects and try to predict the frequency of occurrence of results when the
measuring operation is repeated at random or stochastically.
This type of analysis determines the empirical frequency distribution that yields the absolute or
relative frequency of the occurrence of the different possible results of the repeated measurement of a
property of an object (discrete case). Instead, if the case is an infinitely repeated and arbitrarily precise
measurement and every outcome is diffferent, the relative frequency of a single outcome would not be
very instructive; the distribution function is used, which, for every numerical value x of the measured
variable, yields the absolute (or relative) frequency of the occurrence of all values smaller than x [37].

2.2.2. Statistical Process Control

By applying statistical process control (SPC) methods to the monitoring of a system, it is possible
to detect outliers. This study is concentrated on significantly high and low measurements, even
in situations where the values do not exceed the established limit. These methodologies can be used to
study individual observations, using individual or average charts.
The dataset should be partitioned into rational subgroups, minimising the probability of large
differences between subgroups [38]. The formation of rational subgroups is important, because
variation within subgroups can be clustered and the presence of special causes of variability can
be easily detected. However, sometimes it is not practical to use rational subgroups; for example,

93
Mathematics 2020, 8, 225

when repeated measurements in the process differ only by laboratory or analytical error. Even when
automated inspection is used, because every unit manufactured is analysed.
The rational subgroups represent the way of collecting the data. Usually, they should be gathered
so that each of them shows only the inherent variation that is natural for the process (common cause
variation). Because they contribute to identifying any other source of variation (special cause variation)
that may badly affect the process, the subgroups should avoid special-cause variation where possible.
Moreover, the limits on a control chart, which mark the boundary to identify if a process is too volatile,
are calculated on the basis of the variability within each subgroup. For this, only subgroups that
reproduce the common cause variation in a process should be selected.
Once the data are correctly structured, a normality test has to be done. If the hypothesis of
normality is rejected, there are two possibilities: use modified classical techniques to non-normal
distributions [39], or transform the data to normalise the dataset [40]. The second option applies a
Box-Cox transformation [41], which smoothes the data structure. The most widely used and known
transformation is the Box-Cox transformation, defined as follows:
⎧ λ
⎨ X j −1
(λ) λ , if λ = 0
Xj =
⎩log( X ), if λ =0
j

where λ maximises the profile likelihood function of the data X j .

A classical analysis process can be divided in two main stages: the learning stage, when a test of
normality is performed and atypical measurements are deleted from the data; and the control stage,
when the trends are analysed to encounter out-of-control situations. At the first one, it is when the
centreline (CL) and control limits are defined. Specifically, the CL is defined with the control sample
and represents the objective value. In addition, the warning limits are set at a distance of ±2σ from it,
and control limits at ±3σ, with σ being the standard deviation of the process [42].
Shewhart control charts have been the most widely used due to their good performance in
detecting large changes in a process. However, because these charts use the most recent samples, they
do not efficiently detect small or progressive changes in a process. In this regard, complementary
rules are needed; multiple authors have defined different rules to detect specific deviations [43,44]
and to complement the initial control rules [45]. The use of these supplementary rules makes
Shewhart’s control charts more sensitive and leads to an improvement in one’s capacity to detect
non-random patterns.
A widely used way to quantify the potential of a control chart with supplementary rules is
through the average run length (ARL). The ARL, in control charts, is the average number of points
that should be analysed before showing an alert warning that the process is not under control. When
this occurs, the efficient thing to do would be to detect it as soon as possible. On the contrary, when
the process is statistically stable, it would be appropriate to have few false alarms. This term is directly
related to a Type I error (also known as α) and a Type II error (also known as β), which also describe
the sensitivity of the method, and it is highly related to the number of false alarms. For that reason, it
must be contemplated that if the capacity of this methodology to detect out-of-control situations is
high, there will be a lot of false alarms [43].

2.2.3. Functional Data Analysis

The functional data analysis works with the observations that come from a continuous random
process that is evaluated at discrete points [46]. Starting from vector samples, the dataset will be
transformed into a functional sample. The first step is to construct the most appropriate curves from
the initial points that come from the discrete values measured in the study. This process, known as
smoothing, converts the vector values into continuous functions over time. This structure of data is
essential in the air pollution context because it is taking into account all the values in the day as a
set. In this way, a day in which NO2 values are obtained with a lot of variability but which has an

94
Mathematics 2020, 8, 225

average similar to the other days, is not detected from any vectorial approach. Functional analysis
would identify these types of days as candidates for outliers. Furthermore, in several similar studies in
which they also tried to detect outliers with data from certain gases (see [23,47]), the superiority of
functional approaches was demonstrated.
In a situation where the initial observations x (t j ) are contained in a set of n p points, t j ∈ R
represents the time steps and n p is the number of observations ( j = 1, 2, . . . , n p ). They can be watched
as the individual values of the function x (t) ∈ X ⊂ F, F being a functional space. The estimation of
x (t) takes into account a functional space F = span{φ1 , . . . , φnb }, where φk is the set of basis functions
(k = 1, 2, . . . , nb ) and nb is the number of basis functions required to build a functional sample. There
are several types of basis in statistics, but the most used one is the Fourier basis [26,48]. Moreover,
with periodic data such as we have in this study, Fourier bases are the most appropriate [49]. Smoothing
consists of finding a solution to the regularisation problem [46],
np
min ∑ {z j − x (t j )}2 + λΓ( x ) (1)
X∈ F j=1

with z j = x (t j ) + j being the result of observing x at the point t j ; j the random noise with zero mean,
Γ being a penalising operator focused on obtaining the simplest possible solutions; and λ being a
parameter that defines the level of the regularisation. The initial step is the following expansion
nb
x (t) = ∑ ck φk (t) (2)
k =1

n
where {ck }k=
b
1 are coefficients that multiply the selected basis functions. The smoothing problem can
be written as follows:
min{(z − Φc) T (z − Φc) + λc T Rc} (3)
c
with a vector of observations z = (z1 , . . . , zn p ) T ; a vector of coefficients of the expansion c =
(c1 , . . . , cnb ) T ; a (n p , nb )-matrix Φ whose elements are Φ jk = φk (t j ); and a (nb , nb )-matrix R whose
elements are:
Rkl = D2 φk , D2 φl L2 (T ) = D2 φk (t) D2 φl (t)dt (4)
T
Finally, the problem is solved [46] according to:

c = (Φ Φ + λR)−1 Φ z (5)

As soon as the data are in functional form, they can be analysed to identify pollution episodes
and detect outliers. The functional data allow us to identify whether different periods of time such as
days, weeks or months are above the mean function and how much they are deviating. Moreover, it
permits the elimination of outliers which are not real; they are due to system fails. The depth concept
provides a way for ordering a set of data, contained in a euclidian space, according their proximity to
the centre of the sample.
The concept of depth appeared in multivariate analyses, and was created to measure the centrality
of a point among a cloud of them [50,51]. Over the years, this concept began to be introduced into
functional data analysis [52]. In this field, depth represents the centrality of a certain curve xi and
the mean curve is the centre of the sample. Two depth measurements very common in the context of
functional data are: Fraiman-Muniz depth (FMD) [25] and H-modal depth (HMD) [52].
Therefore, the identification of outliers with a functional approach is possible with the calculation
of depths. In this case, elements that have different behavioural patterns than the rest will be taken
into account. The concept of depth makes it possible to work with observations, defined in a given
time interval, in the form of curves, instead of having to summarise the observations of the curve into
a single value, such as the mean. This method of outlier detection is based on depth measures and

95
Mathematics 2020, 8, 225

centrality: an element that is far from the centre of the sample will have a low depth. Thus, the curves
with the least depth are the functional outliers.
Firstly, the cumulative empirical distribution function Fn,t ( xi (t)) of the values of the curves
{ xi (t)}in=1 in a certain time t ∈ [ a, b] it is contemplated. It can be defined as
n
1
Fn,t ( xi (t)) =
n ∑ I (xk (t) ≤ xi (t)) (6)
k =1

with I (·) being a indicator function. Subsequently, the FMD for any curve xi within a set of curves
x1 , . . . , xn is calculated as
b
FMDn ( xi (t)) = Dn ( xi (t))dt (7)
a

where Dn ( xi (t)) is the depth of the curve xi (t), ∀t ∈ [ a, b], expressed as

 
1 
Dn ( xi (t)) = 1 −  − Fn,t ( xi (t)) (8)
2

On the other hand, for HMD the functional mode is the element or curve most densely surrounded
by the other curves in the dataset. HMD is defined as
 
n
 Xi − X k 
HMDn ( xi , h) = ∑K h
(9)
k =1

with a kernel function K : R+ → R+ , a bandwidth parameter h and  ·  as the norm in a functional

space. Among all the norms, in the most cases, it is used the norm L2 , expressed as
 b
1/2
 xi (t) − x j (t)2 = ( xi (t) − x j (t))2 dt (10)
a

In addition, also exist several options for the kernel functions K (·). A widely used one is the
truncated Gaussian kernel, expressed as
 2
2 t
K (t) = √ exp − , t>0 (11)
2π 2

In this paper, the depth chosen to identified outliers is the HMD. The bandwidth h is the value that
leaves, below it, 15% of the data coming from the distribution of { xi (t) − xt (t)2 , i, j = 1, . . . , n} [15],
and the cut-off C is selected, specifically, to have a 1% Type I error [50], according to

Pr ( HMDn ( xi (t)) < C ) = 0.01, i = 1, . . . , n (12)

The cut-off C has to be estimated because the distribution of the functional depth is not known.
There are several ways to carry out this estimation; however, the bootstrapping method is the most
appropriate for the purpose of this research [52,53]. The steps to follow are:

1. Extract, with replacement, a new sample of the original.

2. Estimate the study parameter through the statistic of this new sample.
3. Repeat the above steps a large number of times. This repetition is also known as Monte Carlo
simulation; it relies on repetition to extract information from the data (see examples in [54,55]).
4. Obtain the empirical distribution of the statistic.

96
Mathematics 2020, 8, 225

2.2.4. Functional Strengths

The functional data analysis (FDA) has many strong points, but the ones that allow the FDA to
have a better performance in contamination analysis are the following: [18,19,24]:

• It is not necessary to know anything in advance about the distribution of the data.
• The analysis of the time sets as a unit. The sample analysed is structure in complete time units
like days or years. Individually distributed values are not taken into account.
• Analysis of homogeneity. Outliers are defined differently; data which do not exceed the limit but
which constantly have small deviations should be classified as outliers.
• Trend analysis. With these techniques, besides detect outliers, also it is possible to analyse
situations where there are no outliers but small deviations from the normal data behaviour
are observed.
• Complete analysis of the time spectrum. In classic studies, generally, the analyses are based on
specific values measured in a determined set of points. The FDA, on the other hand, made it
possible to work with the entire time spectrum of a continuous mode.

3. Results and Discussion

The results of the three methodologies presented in this document are presented below. All
figures were obtained with the R-programming [56] and Python [57] software.

3.1. Classical Analysis

To perform the classical monitoring strategy on air quality the individual time series, descriptive
statistics, box plots and autocorrelation analysis were calculated to determine if any of the values fell
outside of the limits, and to analyse trends. The descriptive statistical parameters of the dataset are
shown in Table 1:

Table 1. Summary descriptive statistics of hourly NO2 concentrations from Blanchardstown air quality
monitoring station in Dublin, Ireland. The statistical quartiles (Q1, Q2, Q3) and the interquartile range
(IQR) are also displayed. Take into account that 0 μg/m3 represents a missing or wrong value.

max 153.48 μg/m3 Q1 9.56 μg/m3

min 0 μg/m3 Q2 20.61 μg/m3
mean 28.47 μg/m3 Q3 42.33 μg/m3
mode 42.33 μg/m3 IQR 32.77 μg/m3
std 23.91 μg/m3 var 571.79μg/m3
n 8760

The descriptive statistical parameters in Table 1 show that the limit values are not exceeded.
The next step of classical data analysis is present a time series of the hourly data in 2013 (Figure 1),
ranging from the maximum value 153.48 μg/m3 to the minimum value 0 μg/m3 . From here it is
possible to say that in any moment, the hourly upper limit (200 μg/m3 ) is not exceed and that the data
have a high variability.

97
Mathematics 2020, 8, 225

Figure 1. Individual time series of hourly NO2 concentrations from Blanchardstown air quality
monitoring station in Dublin, Ireland. Software: Python [57].

Figure 2 presents a boxplot which graphically characterises the data groups of the NO2
concentration by quartiles. The diagram graphically displays the values of the first quartile
(9.56 μg/m3 ); third quartile (42.33 μg/m3 ); the interquartile range (32.77 μg/m3 ); and some, in red,
that are considered atypical.

Figure 2. Box-plot of hourly NO2 concentrations from Blanchardstown air quality monitoring station.
The central and blue line represents the median, and the end of the whiskers are the quartiles (25% for
the lower part and 75% for the upper part). The red dots represent the outliers. Software: Python [57].

Figure 3 presents the frequency of hourly concentrations of NO2 , which, as can be seen, are biased
by 0 values. Another weakness of this analysis is that, when data are poorly collected or no data are
available, only two options remain: either delete these observations (data are lost) or replace them
with 0 values.
Figure 4 shows the normal probability plot of the data, again affected by 0 values.
A Kolmogorov-Smirnov test and Anderson Darling test were applied to compare NO2 concentrations
to a standard normal distribution [54]. The null hypothesis is that the values have a standard normal

98
Mathematics 2020, 8, 225

distribution. The alternative hypothesis is that the values do not have that distribution. The results
obtained for both tests were p-values very close to 0, so, with a 5% significance level, statistical
evidence of the non-normality of the data has been found. The test statistic is: max( F ( x ) − G ( x )),
where F(x) is the empirical cumulative distribution function and G(x) is the standard normal cumulative
distribution function.

Figure 3. Frequency of hourly NO2 concentrations from Blanchardstown air quality monitoring station.
Comparison of data distribution with normal. Software: Python [57].

Figure 4. Normal probability plot of hourly NO2 (QQ-plot) concentrations from Blanchardstown air
quality monitoring station. Software: Python [57].

Other tests have been performed to check whether the data approaches any type of distribution:
normal, generalised extreme value or Weibull and Rayleigh, but none have been acceptable with a null
hypothesis at 5% significance. From the classic analysis of the data it must be concluded that there
are no data that are outside the limit values. This classical method is limited to a time series analysis
with regard to the assessment of trends (Figure 1), and although it allows for the identification of the
main parameters within the data and how the data are distributed, is an incomplete method because it
provides us with information that is too simple and does not take into account the correlation between
hourly observations.

99
Mathematics 2020, 8, 225

3.2. Statistical Process Control

3.2.1. Control I-MR Charts with Individual Mean

To analyse the data using the SPC method, an individual-moving range chart (IMR chart) of
hourly NO2 concentrations was made. With the examination of the results shown in Figure 5, it can be
observed that the number of false alarms, i.e., outliers, is significant. This problem is attributable to:

• The non-normality of the data, which is clear from the analysis shown in Figure 4.
• The effect of autocorrelation in time series data (Figure 6).
• The existence of greater variability with data of different rational subgroups than within the data
inside each analised subgroup.

Figure 5. Individual-moving range chart (IMR chart) X/R with mobile range of hourly NO2
concentrations from Blanchardstown air quality monitoring station. Software: R-programming [56].

By performing an autocorrelation analysis, it can be observed from Figure 6 that the data are very
autocorrelated. This is very common in environmental data and shows that the autocorrelation has
24-hour cycles and decreases with time.

Figure 6. Sample autocorrelation function of hourly values for NO2 concentrations from
Blanchardstown air quality monitoring station. Software: Python [57].

100
Mathematics 2020, 8, 225

In Figure 6 the correlation of all data for the year is shown, while Figure 7 only shows the data of
the first 86 h in order to see in more detail, the 24 h cycles. Due to the non-normality of the data and
the data’s autocorrelation, the control chart has a large number of false alarms. Therefore, SPC is not a
very suitable method with which to detect outliers for NO2 concentration.

Figure 7. Sample autocorrelation function of hourly values for NO2 concentrations from
Blanchardstown air quality monitoring station over the initial 86 h period. Software:
R-programming [56].

3.2.2. Control Charts with Daily Rational Subgroups

The study of datasets choosing days as the rational subgroup of the X/s chart (every day is
summed up by one point), is not under control due to the non-normality and the autocorrelation (see
Figures 8 and 9). Although the chart is not under control, and there is much variability, none of the 365
days exceed the limit value (100 μg/m3 ).

Figure 8. Xbar-chart of hourly NO2 concentrations with the daily rational subgroup of data. Software:
R-programming [56].

101
Mathematics 2020, 8, 225

Figure 9. S-chart of hourly NO2 concentrations with the daily rational subgroup of data. Software:
R-programming [56].

3.2.3. Trend Analysis

Using the hourly NO2 concentration data, a trend analysis was undertaken to examine the diurnal
patterns and identify outliers. Figure 10 shows the box plots of hourly emissions over 365 days and
represents the mean, the confidence interval of this mean, quartiles and abnormal values in red that.
This way it is possible to studying individually the distribution of the NO2 emissions in each hour of
the day.

Figure 10. Hourly box plot of NO2 concentrations from Blanchardstown air quality monitoring station.
Software: Python [57].

102
Mathematics 2020, 8, 225

In Figure 11, where only the mean values have been represented, two maximum values were
analysed, at 9 a.m. and 7 p.m., which correspond to heavy traffic hours in this area of Dublin. These
two hours, as can be seen in Figure 10, also have high variability (wide boxes), and around them there
are numerous outliers.

Figure 11. Mean hourly NO2 concentrations from Blanchardstown air quality monitoring station.
Software: Python [57].

It can also be seen in these figures that the absolute minimum is 4 a.m., which also corresponds to
the hour with the least variability in the NO2 concentration. If a daily analysis approach is considered
throughout the year (Figure 12), it is clear that Wednesday, Thursday and Friday are the days with the
highest NO2 concentrations.

Figure 12. Daily analysis of NO2 concentrations at Blanchardstown air quality monitoring station.
Software: R-programming [56].

103
Mathematics 2020, 8, 225

Blanchardstown contains the largest shopping centre in Ireland and the most major motorway
nearby; because of this, there is a lot of traffic when rush hour and evening shopping are combined
in this area. On the other hand, there are lower concentrations during the weekend, although not as
low as expected. This is because the largest shopping centre in the country is very busy at weekends,
especially during Christmas period.
In this case, a SPC gives us more information than the previous analysis, such as the differences
in the NO2 levels between days or hours. However, it still does not take into account the complete
daily behaviour of NO2 emissions from correlated hourly measurements.

3.3. Functional Analysis of NO2 in Dublin

The subsequent step in the functional methodology is to compare the results between the classical
analysis and the SPC. In the functional methodology, the first thing to do is to build a sample of curves
based on the discrete values measured every hour. Figure 13 shows the 365 functions generated with
24 hourly data. Once the data are transformed into functional data, i.e., daily curves of 24 values,
each of them takes into account the correlation between the hourly NO2 values and can be analysed
for outliers.

Figure 13. Data represented in functional form (functions): 365 daily curves of NO2 emissions.
Software: R-programming [56].

The results obtained with the functional analysis, taking into account the depths, allow us to
identify days with abnormal functional values, even if, discreetly, they are not outliers. Despite not
exceeding the daily limit values, the concentration of NO2 over a whole day may have an abnormal
behaviour. For this reason the vectorial analysis, like SPC, does not get to detect these days. In a
different way, the functional approach detects any deviation from normal daily behaviour in the
emissions of NO2 , without relying on any distribution restrictions. This is shown in Figure 14 where
the functional outliers found in this case study are presented.

104
Mathematics 2020, 8, 225

Figure 14. NO2 functional data; and in the dark grey and dotted line, the functional outliers detected.
Highlighted is the 11th day which is also an outlier. Software: R-programming [56].

The data analysed were not discretely outside the limits values; however, functionally excessive
variations were observed on specific days. It can be deduced that there were no NO2 pollution
problems in 2013 because hourly, daily and annual NO2 concentration limits were not exceeded.
But it is also important to analyse whether there are hours or days with anomalous trends of NO2
concentrations, although they remain within legal limits. The NOx (NO, NO2 ) is an pollutant not
coming from a natural source ([58]), and for this reason also, its involvement in the mechanisms of
depleting the ozone layer are very unfortunate. The FDA methodology has proven to be very effective
in detecting days with trends that are not the same as the rest of the data. It is important not only to
analyse whether the contamination is within the allowed limits, but also to find days that are different
than expected.
For example, on the 11th day, detected by the functional approach and highlighted in the Figure 14,
with the results obtained through the SPC, as can be seen in Figure 8, this day has a higher mean,
but it is within the limit values. Neither using the classical analysis (Figure 1) with individual time
series nor the one with the corresponding statistical parameters, it could be considered an outlier.
Figure 14 shows a strength of the functional approach by detecting this curve as an atypical day; it will
be possible to study the reasons that lead the NO2 to behave this way on this particular day. In fact,
there are several studies that demonstrate the greater power of the functional approach for detecting
outliers than other methodologies (see [17,59]). There are also studies that, specifically, show that the
depth measure used here (h-modal) is the one that achieves the lowest error rates [50].
To find a reason that explains the anomalous behaviour on those days, it would be necessary
to have greater traceability of the most important sources of NO2 emissions. It would be necessary
to have data relating to the weather conditions, traffic movements, industry sources affecting the
study area, etc. For example, incorporating weather conditions, i.e., temperatures, sunshine hours and
precipitation, could improve the assessment of outliers. The detection of outliers and air pollution
episodes can help to separate the causes of normal and specific variability, and is a first step towards
the effective design and implementation of mitigation measures. Although the only reason for these
outliers is possibly not the weather conditions, as can be seen in Table 2, those days were colder than

105
Mathematics 2020, 8, 225

usual; had very little precipitation; and in general, had fewer sunshine hours than the average for
the month.

Table 2. Sample of other environmental characteristics from Dublin (temperature, sunshine and
precipitation) which may impact hourly NO2 concentrations measured at Blanchardstown air quality
monitoring station. The averages shown represent the monthly average of each variable.
No Date Day Lowest Ta Average Lowest Ta Sunshine Hours Average Sunshine Hours Precipitations
10 10/01 Thursday −2.5 ◦ C 2.6 ◦ C 0h 1.2 h 2 mm
11 11/01 Friday −0.5 ◦ C 2.6 ◦ C 1h 1.2 h 0.5 mm
59 28/02 Thursday 1 ◦C 2 ◦C 0h 2.6 h 0 mm
60 01/03 Friday −3.1 ◦ C 1.2 ◦ C 3h 2.1 h 0 mm
64 5/03 Tuesday −3,8 ◦ C 1.2 ◦ C 5h 2.1 h 0 mm
78 19/03 Tuesday 0 ◦C 1.2 ◦ C 0h 2.1 h 5.6 mm
122 02/05 Thursday −0.4 ◦ C 6.3 ◦ C 5h 6.3 h 0 mm
193 12/07 Friday 8.7 ◦ C 11.9 ◦ C 14 h 7.7 h 0 mm
305 01/11 Friday 2 ◦C 4.2 ◦ C 4.20 h 2.7 h 0 mm

4. Conclusions
In this paper, three different analytical methods were adopted and compared to determine their
effectiveness to identify pollution episodes and outliers. The data used come from a sub-urban air
quality monitoring site in Dublin, Ireland, and cover the whole year 2013 with hourly measurements.
Firstly, a classical vectorial approach was applied by analysing the data through time series, boxplots
and frequency plots. Secondly, a statistical process control was adopted to study the data grouped by
days or hours, and with different control charts (Xbar-chart, S-chart). Finally, to identify air pollution
episodes and outliers, a functional data analysis approach was adopted to analyse the daily patterns of
NO2 concentrations.
To effectively support the mitigation of air pollution and provide true air quality conditions,
a new approach and set of tools are required to effectively assess local air pollution. With this in
mind, the classical or vectorial approach is too simplistic, despite providing significant information for
decision making. It has several weaknesses related to the time correlation structure of the data, and the
inability to detect real outliers, days with behaviours far from the standard, just because they do not
exceed limit values. As such, more advanced and modern techniques can provide deeper insights to
support controlling air pollution episodes. Statistical process control presents similar problems. Even
though it manages to take into account the correlation between data, the lack of normality causes a
significant increase in false alarms for days that are within the legal limits; this method labels them as
outliers. This is because it works with discrete observations and it is not able to extract continuous
information from the data.
Alternately, the functional data analysis method, adapted to study pollution episodes, as shown
in this paper, presents important improvements as a method that can be extrapolated to any city in the
world. In short, it is not restricted to certain characteristics of the data distribution; takes into account
complete time units; works with the entire time spectrum of a continuous mode; and the detection
of outliers is more effective, which is very important in pollution issues. On the contrary, working
from a functional point of view can also have its drawbacks, such as the need for large amounts of
data (it is not always possible to get them), or other data constraints, such as continuity, positivity and
monotonicity. But as explained, the fact that all the hourly values of a day are not outliers does not
mean that that the day is not an outlier. The outlier search with the FDA allows one to detect days
that, for some reason, have different sets of measurements of NO2 than the rest (usually higher or with
strange peaks). Being able to study the reasons that make these daily functions behave differently from
others makes it possible to prevent them; to try to reduce the variability in emissions; and in short,
to improve air pollution control. With a classical method there would be no possibility of relating
certain events to lost hours that are labelled as outliers.

106
Mathematics 2020, 8, 225

The FDA’s approach is presented in this document as a methodology for more effective assessment
of air pollution, which is hazardous to the health of living beings, and to inform effective mitigation
measures in the future.

Author Contributions: The tasks of this paper were divided as follows: on the one hand, J.P.P., J.S.V. and M.M.C.
focused on the development of the mathematical methodology; on the other hand, J.M.T., A.M. and J.G. focused on
the environmental part of the work. All authors have read and agreed to the published version of the manuscript.
Funding: This paper was funded by the Spanish Government (Industry and Competitiveness Ministry) under the
project RTI2018-096296-B-C21.
Acknowledgments: This paper was supported by the Spanish Government (Industry and Competitiveness
Ministry) under the project RTI2018-096296-B-C21.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. García Nieto, P. Parametric study of selective removal of atmospheric aerosol by coagulation, condensation
and gravitational settling. Int. J. Environ. Health Res. 2001, 11, 151–162. [CrossRef] [PubMed]
2. Akkoyunku, A.; Ertürk, F. Evaluation of air pollution trends in Istanbul. Int. J. Environ. Health Res. 2003,
18, 388–398.
3. Karaca, F.; Alagha, O.; Ertürk, F. Statistical characterization of atmospheric PM10 and PM2.5 concentrations
at a non-impacted suburban site of Istanbul, Turkey. Chemosphere 2005, 59, 183–190. [CrossRef] [PubMed]
4. García Nieto, P. Study of the evolution of aerosol emissions from coal-fired power plants due to coagulation,
condensation, and gravitational settling and health impact. J. Environ. Manag. 2006, 79, 372–382. [CrossRef]
[PubMed]
5. RCPCH. Every Breath We Take: The Lifelong Impact of Air Pollution; Royal College of Physicians: London, UK,
2016. Available online: https://www.rcplondon.ac.uk/projects/outputs/every-breath-we-take-lifelong-
impact-air-pollution (accessed on 7 February 2020).
6. WHO. Review of Evidence on Health Aspects of Air Pollution—REVIHAAP Project; World Health Organization:
Geneva, Switzerland, 2013. Available online: http://www.euro.who.int/__data/assets/pdf_file/0004/
193108/REVIHAAP-Final-technical-report.pdf (accessed on 7 February 2020).
7. Kumar, P.; Druckman, A.; Gallagher, J.; Gatersleben, B.; Allison, S.; Eisenman, T.S.; Hoang, U.; Hama, S.;
Tiwari, A.; Sharma, A.; et al. The nexus between air pollution, green infrastructure and human health.
Environ. Int. 2019, 133, 105181. [CrossRef]
8. EPA. United States Environmental Protection Agency. 2019. Available online: https://www.epa.gov/
(accessed on 19 December 2019).
9. AQEG. Trends in Primary Nitrogen Dioxide in the UK. Air Quality Expert Group. 2007. Available
online: https://uk-air.defra.gov.uk/assets/documents/reports/aqeg/primary-no-trends.pdf (accessed
on 7 February 2020).
10. EPA. Ireland’s Transboundary Gas Emissions; Environmental Protection Agency: Washington, DC, USA,
2018. Available online: http://www.epa.ie/pubs/reports/air/airemissions/Irelands%20Air%20Pollutant%
20Emissions%202016.pdf (accessed on 7 February 2020).
11. Costa, S.; Ferreira, J.; Silveira, C.; Costa, C.; Lopes, D.; Relvas, H.; Borrego, C.; Roebeling, P.; Miranda, A.I.;
Teixeira, J.P. Integrating Health on Air Quality Assessment—Review Report on Health Risks of Two Major
European Outdoor Air Pollutants: PM and NO2 . J. Toxicol. Environ. Heal. Part B 2014, 17, 307–340. [CrossRef]
12. Cooper, C.; Alley, F. Air Pollut. Control; Waveland Press: New York, NY, USA, 2002.
13. Lutgens, F.; Tarbuck, E. The Atmosphere: An Introduction to Meteorology; Prentice Hall: New York, NY,
USA, 2001.
14. Jeanjean, A.; Gallagher, J.; Monks, P.; Leigh, R. Ranking current and prospective NO2 pollution mitigation
strategies: An environmental and economic modelling investigation in Oxford Street, London. Environ.
Pollut. 2017, 225, 587–597. [CrossRef]
15. Cuevas, A.; Fraiman, R. A plug-in approach to support estimation. Ann. Stat. 1997, 25, 2300–2312. [CrossRef]
16. Matías, J.; Ordóñez, C.; Taboada, J.; Rivas, T. Functional support vector machines and generalized linear
models for glacier geomorphology analysis. Int. J. Comput. Math. 2009, 86, 275–285. [CrossRef]

107
Mathematics 2020, 8, 225

17. Martínez, J.; Garcia Nieto, P.; Alejano, L.; Reyes, A. Detection of outliers in gas emissions from urban areas
using functional data analysis. J. Hazard. Mater. 2011, 186, 144–149.
18. Martínez, J.; Saavedra, A.; García Nieto, P.; Piñeiro, J.; Iglesias, C.; Taboada, J.; Sancho, J.; Pastor, J. Air quality
parameters outliers detection using functional data analysis in the Langreo urban area (Northern Spain).
Appl. Math. Comput. 2014, 241, 1–10. [CrossRef]
19. Sancho, J.; Iglesias, C.; Piñeiro, J.; Martínez, J.; Pastor, J.; Araújo, M.; Taboada, J. Study of water quality
in a spanish river based on statistical process control and functional data analysis. Math. Geosci. 2016,
48, 163–186. [CrossRef]
20. Dombeck, D.; Graziano, M.; Tank, D. Functional clustering of neurons in motor cortex determined by cellular
resolution imaging in awake behaving mice. J. Neurosci. 2009, 29, 13751–13760. [CrossRef] [PubMed]
21. Wu, D.; Huang, S.; Xin, J. Dynamic compensation for an infrared thermometer sensor using least-squares
support vector regression (LSSVR) based functional link artificial neural networks (FLANN). Meas. Sci.
Technol. 2008, 19, 105202.1–105202.6. [CrossRef]
22. Ordoñez, C.; Martínez, J.; Cos Juez, J.; Sánchez Lasheras, F. Comparison of GPS observations made in a
forestry setting using functional data analysis. Int. J. Comput. Math. 2011, 89, 402–408. [CrossRef]
23. Ordóñez, C.; Martínez, J.; Saavedra, A.; Mourelle, A. Intercomparison Exercise for Gases Emitted by a
Cement Industry in Spain: A Functional Data Approach. J. Air Waste Manag. Assoc. (1995) 2011, 61, 135–141.
[CrossRef] [PubMed]
24. Sancho, J.; Pastor, J.; Martínez, J.; García, M. Evaluation of Harmonic Variability in Electrical Power Systems
through Statistical Control of Quality and Functional Data Analysis. Procedia Eng. 2013, 63, 295–302.
[CrossRef]
25. Fraiman, R.; Muniz, R. Trimmed means for functional data. Test 2001, 10, 419–440. [CrossRef]
26. Piñeiro, J.; Martínez, J.; García Nieto, P.; Alonso, J.; Díaz, C.; Taboada, J. Analysis and detection of outliers in
water quality parameters from different automated monitoring stations in the Miño river basin (NW Spain).
Ecol. Eng. 2013, 60, 60–66.
27. Sancho, J.; Martínez, J.; Pastor, J.; Taboada, J.; Piñeiro, J.; García Nieto, P. New methodology to determine air
quality in urban areas based on runs rules for functional data. Atmos. Environ. 2014, 83, 185–192. [CrossRef]
28. Grubbs, F.E. Procedures for Detecting Outlying Observations in Samples. Technometrics 1969, 11, 1–21.
[CrossRef]
29. Jäntschi, L. A test detecting the outliers for continuous distributions based on the cumulative distribution
function of the data being tested. Symmetry 2019, 11, 835. [CrossRef]
30. EPA. Air Quality in Ireland 2018; Environmental Protection Agency: Washington, DC, USA, 2019.
Available online: http://www.epa.ie/pubs/reports/air/quality/Air%20Quality%20In%20Ireland%202018.
pdf (accessed on 7 February 2020).
31. EPA. Air Quality in Ireland 2013: Key Indicators of Ambient Air Quality; Environmental Protection Agency:
Washington, DC, USA, 2014. Available online: https://www.epa.ie/pubs/reports/air/quality/Air%
20Quality%20Report%202013.pdf (accessed on 7 February 2020).
32. Romer, U. Weather Online(Ireland). 2013. Available online: https://www.weatheronline.co.uk/weather/
maps/current?TYP=tmin&KEY=IE&LANG=en&ART=tabelle&JJ=xxxx&SORT=2&INT=24 (accessed on 19
December 2019).
33. Carslaw, D.; Ropkins, K. Openair—An R package for air quality data analysis. Environ. Model. Softw. 2012,
27–28, 52–61. [CrossRef]
34. Piatesky-Shapiro, G.; Frawley, W. Knowledge Discovery in Databases; MIT Press: Cambridge, MA, USA, 1991.
35. Takeuchi, J.; Yamanishi, K. A unifying framework for detecting outliers and change points from time series.
IEEE Trans. Knowl. Data Eng. 2006, 18, 482–492. [CrossRef]
36. Sim, C.; Gan, F.; Chang, T. Outlier Labeling With Boxplot Procedures. J. Am. Stat. Assoc. 2005, 100, 642–652.
[CrossRef]
37. Montgomery, D. Design and Analysis of Experiments; John Wiley & Sons, Inc.: Hoboken, NJ, USA, 2013;
Chapter 1–2, pp. 1–65.
38. Shewhart, W. Economic Control of Quality of Manufactured Product; Van Nostrand Company: New York, NY,
USA, 1931.
39. Chen, Y.K. An evolutionary economic-statistical design for VSIXcontrol charts under non-normality. J. Adv.
Manuf. Technol. 2003, 22, 602–610. [CrossRef]

108
Mathematics 2020, 8, 225

40. Freeman, J.; Modarres, R. Inverse Box-Cox: the power-normal distribution. Stat. Probab. Lett. 2006,
76, 764–772. [CrossRef]
41. Box, G.; Cox, D. An analysis of transformations. J. R. Stat. Soc. Ser. B (Stat. Methodol.) 1964, 26, 211–252.
[CrossRef]
42. Grant, E.; Leavenworth, R. Statistical Quality Control; McGraw-Hill: New York, NY, USA, 1998.
43. Champ, C.; Woodall, W. Exact results for Shewhart control charts with supplementary runs rules.
Technometrics 1987, 29, 393–399. [CrossRef]
44. Zhang, M.; Lin, W.; Klein, S.; Bacmeister, J.; Bony, S.; Cederwall, R.; Del Genio, A.; Hack, J.; Loeb, N.;
Lohmann, U.; et al. Comparing clouds and their seasonal variations in 10 atmospheric general circulation
models with satellite measurements. J. Geophys. Res. 2005, 110. [CrossRef]
45. Electric, W. Statistical Quality Control Handbook; AT&T Technologics: Indianapolis, Indiana, 1956.
46. Ramsay, J.; Silverman, B. Functional Data Analysis; Springer: New York, NY, USA, 2005.
47. Sánchez-Lasheras, F.; Ordóñez, C.; Garcia Nieto, P.J.; García-Gonzalo, E. Detection of outliers in pollutant
emissions from the Soto de Ribera coal-fired plant using Functional Data Analysis: A case study in northern
Spain. Proceedings 2018, 2, 1473. [CrossRef]
48. Muñiz, C.D.; García Nieto, P.J.; Alonso Fernández, J.R.; Torres, J.V.; Taboada, J. Detection of outliers in water
quality monitoring samples using functional data analysis in San Esteban estuary (Northern Spain). Sci.
Total. Environ. 2012, 439, 54–61. [CrossRef] [PubMed]
49. Kamada, M.; Toraich, K.; Mori, R. Periodic spline orthonormal bases. J. Approx. Theory 1988, 55, 27–34.
[CrossRef]
50. Febrero, M.; Galeano, P.; González-Manteiga, W. Outlier detection in functional data by depth measures,
with application to identify abnormal NOx levels. Environmetrics 2008, 19, 331–345. [CrossRef]
51. Zuo, Y.; Serfling, R. General notions of statistical depth function. Ann. Stat. 2000, 28, 461–482. [CrossRef]
52. Cuevas, A.; Febrero, M.; Fraiman, R. On the use of the bootstrap for estimating functions with functional
data. Comput. Stat. Data Anal. 2006, 51, 1063–1074. [CrossRef]
53. Cuevas, A. A partial overview of the theory of statistics with functional data. J. Stat. Plan. Inference 2014,
147, 1–23. [CrossRef]
54. Jäntschi, L.; Bolboacă, S.D. Computation of Probability Associated with Anderson-Darling Statistic.
Mathematics 2018, 6, 88. [CrossRef]
55. Jäntschi, L.; Bolboacă, S.D. Rarefaction on natural compound extracts diversity among genus. J. Comput. Sci.
2014, 5, 363–367. [CrossRef]
56. R Core Team. R: A Language and Environment for Statistical Computing; R Foundation for Statistical Computing:
Vienna, Austria, 2014.
57. Van Rossum, G.; Drake, F.L., Jr. Python Reference Manual; Centrum voor Wiskunde en Informatica:
Amsterdam, The Netherlands, 1995.
58. Cosma, C.; Suciu, I.; Jäntschi, L.; Bolboaca, S. Ion-Molecule Reactions and Chemical Composition of
Emanated from Herculane Spa Geothermal Sources. Int. J. Mol. Sci. 2008, 9, 1024–1033. [CrossRef]
59. Yu, G.; Zou, C.; Wang, Z. Outlier Detection in Functional Observations With Applications to Profile
Monitoring. Technometrics 2012, 54, 308–318. [CrossRef]

c 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).

109
 mathematics
Article
Constructing a Control Chart Using Functional Data
Miguel Flores 1 , Salvador Naya 2 , Rubén Fernández-Casal 3 , Sonia Zaragoza 4 , Paula Raña 3 and
Javier Tarrío-Saavedra 5, *
1 MODES Group, Department of Mathematics, Escuela Politécnica Nacional, 170517 Quito, Ecuador;
miguel.fl[email protected]
2 MODES Group, CITIC, ITMATI, Department of Mathematics, Escola Politécnica Superior, Campus
Industrial, Universidade da Coruña, Mendizábal s/n, 15403 Ferrol, Spain; [email protected]
3 MODES Group, CITIC, Department of Mathematics, Faculty of Computer Science, Campus de Elviña,
Universidade da Coruña, 15008 A Coruña, Spain; [email protected] (R.F.-C.); [email protected] (P.R.)
4 PROTERM Group, Department of Naval and Industrial Engineering, Campus Industrial, Universidade da
Coruña, Mendizábal s/n, 15403 Ferrol, Spain; [email protected]
5 MODES Group, CITIC, Department of Mathematics, Escola Politécnica Superior, Campus Industrial,
Universidade da Coruña, Mendizábal s/n, 15403 Ferrol, Spain
* Correspondence: [email protected]

Received: 12 October 2019; Accepted: 23 December 2019; Published: 2 January 2020

Abstract: This study proposes a control chart based on functional data to detect anomalies and
estimate the normal output of industrial processes and services such as those related to the energy
efficiency domain. Companies providing statistical consultancy services in the fields of energy
efficiency; heating, ventilation and air conditioning (HVAC); installation and control; and big data
for buildings, have been striving to solve the problem of automatic anomaly detection in buildings
controlled by sensors. Given the functional nature of the critical to quality (CTQ) variables, this study
proposed a new functional data analysis (FDA) control chart method based on the concept of data
depth. Specifically, it developed a control methodology, including the Phase I and II control charts.
It is based on the calculation of the depth of functional data, the identification of outliers by smooth
bootstrap resampling and the customization of nonparametric rank control charts. A comprehensive
simulation study, comprising scenarios defined with different degrees of dependence between curves,
was conducted to evaluate the control procedure. The proposed statistical process control procedure
was also applied to detect energy efficiency anomalies in the stores of a textile company in the Panama
City. In this case, energy consumption has been defined as the CTQ variable of the HVAC system.
Briefly, the proposed methodology, which combines FDA and multivariate techniques, adapts the
concept of the control chart based on a specific case of functional data and thereby presents a novel
alternative for controlling facilities in which the data are obtained by continuous monitoring, as is the
case with a great deal of process in the framework of Industry 4.0.

Keywords: functional data analysis; statistical process control; control chart; data depth;
nonparametric control chart; energy efficiency

1. Introduction
Generally, univariate and multivariate control charts are applied to identify anomalies in the
industry and control the quality of products and services. However, the specific characteristics of
data obtained by continuous monitoring (which has become the main trend due to advancements
in sensoring and communications in the framework of Industry 4.0), require the use of increasingly
sophisticated tools that take into account the presence of autocorrelation in critical to quality (CTQ)
variables for the process under study.

Mathematics 2020, 8, 58; doi:10.3390/math8010058 111 www.mdpi.com/journal/mathematics

Mathematics 2020, 8, 58

Recently, diverse solutions have been provided that propose the modification and application of
exponentially weighted moving average chart (EWMA) control charts [? ]; control charts based on
the fitting of autoregressive integrated moving average (ARIMA) models [? ]; and the use of control
charts for profiles, which is understood as the control of the parameters that define the relationship
between two different CTQ variables [? ? ? ? ]. Additionally, there are solutions that propose the
use of techniques based on the control chart concept for anomaly detection such as machine learning
techniques (neural networks and support vector machines, among others) and time series [? ? ? ? ? ? ?
? ]. The increasingly common use of these tools is attributed to the fact that they adapt very well to the
new paradigm of data (in the framework of Industry 4.0) defined by the continuous monitoring of
numerous variables. Traditional control charts cannot be applied to many of the new cases due to their
non-compliance with their starting hypotheses. In the domain of statistical quality control, the use of
the concept of latent variables along with multivariate control charts is one of the most popular and
useful alternatives to solve the problem of process control in the Industry 4.0. domain [? ? ? ].
Many of these new data, usually curves, can be studied as functional data. This is the case of data
in the fields of energy consumption, indoor and machine temperatures, relative humidity, amount of
CO2 , among other variables, measured in all types of buildings. These new data, which currently
characterize industrial processes, require innovative solutions, developed by researchers in the field
of statistical quality control, based on the application of functional data analysis (FDA) techniques.
Until now, few works have been published (in relation to the importance of the topic and the frequent
monitoring of this type of data) on controlling the quality of processes when the data to be monitored
are functional CTQ variables. Among the most outstanding research in this context, Reference [? ]
compared the performances of different control charts used for monitoring functional data; these charts
are often identified as profiles in the statistical quality control literature. According to References [? ? ],
in the case of the monitoring of profiles, a set of statistical techniques (usually multivariate) is applied
for controlling processes when they are defined by the functional relationship between two variables.
However, a control chart for monitoring functional data, which is based on bootstrap resampling, is
also proposed in Reference [? ], whereas, in References [? ? ], two complete monographs are proposed
in which the concepts of control charts for profiles are adapted to the context of the functional data.
Moreover, recently, new alternatives for outlier detection based on interlaboratory studies and those
with application in industrial anomaly detection have been performed in the FDA framework [? ? ? ].
This study intends to provide an alternative solution to the detection of out-of-control anomalies
or states in the area of energy efficiency in buildings, specifically in commercial areas such as different
stores of textile companies. Therefore, an important aim of this work is to solve the problem of
such companies, specifically the ones providing lighting facilities; office automation; and heating,
ventilation and air conditioning (HVAC). One such company is Fridama SL. The group comprises
Fridama’s facilities, Σqus (web platforms for big data management) and Nerxus (statistical consultancy
for data analysis in the energy sector); the solution to the problem raised by this company is defined
by continuously monitored data over time that can be treated as profiles or functional data. Based on
this need, the present study proposes methodologies to build control charts that allow us to control the
aforementioned processes. in order to test the applied statistical methodologies for anomaly detection
properly, these companies have provided a database on a real HVAC installation, whose anomalies
and assignable causes are identified by its maintenance personnel.
These energy efficiency data are non-Gaussian and autocorrelated, as the main component of
continuously monitored data in the Industry 4.0 framework. Thus, the present proposal provides
alternative control charts for Phases I and II of the statistical process control that can deal with
non-Gaussian and autocorrelated data. As is well-known, using control charts, the process can be
controlled in the following two phases: Phase I involves the stabilization or the calibration of the
process and Phase II focuses on process monitoring. For Phase I, a control chart is proposed based
on the depth measurement of a functional datum and the idea of atypical detection; in Phase II,
a nonparametric range control chart is proposed, based on the calculation of the functional data depth,

112
Mathematics 2020, 8, 58

to monitor the process of interest. The procedure of building control charts corresponding to Phases
I and II are explained in the next section; it also shows the results of their performance through a
simulation study including diverse scenarios. The proposed methodology has been developed and
programmed in R through different functions integrated in the “Quality Control Review” package,
qcr, which can be freely and easily accessed by the practitioners.
Additionally, it provides a way to visualize the control charts for functional data, including the
original data and the curves corresponding to the estimated control limits. This visual tool allows
users and maintenance managers to relate each anomaly to an intuitively assignable cause. The results
of the simulation study and its application to the real data show the usefulness of this control chart
methodology in detecting anomalies when the process is defined by functional data, specifically the
daily curves of energy consumption in commercial areas.

1.1. Alternatives of the Statistical Quality Control When the Basic Assumptions of the Control Charts Are
Not Met
This control chart approximation for functional data is based on nonparametric multivariate
control charts, which are useful when the assumptions of Gaussianity are not met. Therefore, a brief
introduction of nonparametric control charts and statistical tools for dealing with autocorrelated data
are presented in the following section.
Recently, analyses of the robustness of different control charts against the non-compliance of
Gaussianity hypothesis have been developed. Particularly, several nonparametric control charts have
been proposed. In this domain, we should highlight the works of Regina Liu [? ? ? ], who has
developed the r, Q and S control charts based on the data depth and rank concepts and the rank or
ranges. In this regard, it is important to highlight the strong influence of Liu’s works; it must be noted
that several nonparametric alternatives for control charts are based on the concepts of data depth
and rank. Moreover, it should be emphasized that one of the most important lines of research of the
SQC, the profiles’ control charts, are based, in many cases, on the application of nonparametric or
semiparametric regression models [? ? ]. Additionally, it is also important to highlight the use of
resampling techniques for calculating natural control limits for different types of control charts [? ? ].
The work in Reference [? ] constitutes a complete monograph on the current trends for the development
of control charts.
Despite their important advantages (no assumptions on the probability distribution of CTQ
variables), nonparametric control charts are not predominantly used in the industrial and business
framework. As noted in Reference [? ], this can be attributed to several factors such as the lack of
specific software, both commercial and free; the lack of general training in nonparametric statistics that
generates insecurity and distrust in users and the lack of contrasted reference texts for the application
of nonparametric methods in SQC. However, the research activity in this field is growing, as illustrated
by the work in Reference [? ].
Another starting hypothesis of control charts is the independence of the observations. The data
continuously monitored over time by different sensors usually show a variable level of autocorrelation
(the greater the correlation the closer the observations are in time). The application of standard
techniques in the case of the violation of the independence hypothesis often results in the detection of
an unacceptable number of false alarms [? ]. Therefore, the development and analysis of techniques
that remove the sample autocorrelation is fully justified. Within these techniques, the most widespread
is the application of time series models (e.g., autoregressive moving average (ARMA) and ARIMA)
to remove the correlation between observations and the subsequent monitoring of the error variable
(difference between actual values and those estimated by the model) using control charts [? ? ? ].
Moreover, References [? ? ] propose the combination of control charts with adjustment algorithms.
Finally, the monographs of [? ? ] describe the most relevant lines of research for controlling and
monitoring autocorrelated data.

113
Mathematics 2020, 8, 58

It is important to note that this type of data is related to the control charts for functional data.
The use of FDA techniques allows us to consider the autocorrelation of the data as well as, by means
of resampling techniques, to circumvent parametric assumptions about the trend and dependence.

1.2. Control Charts for Phases I and II

The studied process is stabilized in the framework of the Phase I control charts, that is, the process
is left under control [? ]. This implies that there are no other assignable causes in the process, except for
those present due to randomness itself (non assignable causes). This is equivalent to stating that the
process remains stable—the parameters of the probability distribution of the CTQ characteristic remain
unchanged. This allows us to estimate the natural control limits for the variable that describes the
quality of the process. The natural control limits are estimated using a preliminary or calibration
sample. Thus, in this stage, the main assignable causes of variation are identified and rectified through
corrective actions. After the variations are removed, we estimate the natural control limits of the
CTQ variable corresponding to the process under control. While most of the literature in the field
has referred to Phase II control charts, within the new data paradigm for Industry 4.0, it would be
essential to develop methodologies for Phase I [? ]. In this regard, Reference [? ] presents an overview
of the recent contributions to the development of Phase I control charts. Among the latest proposals
for control charts, the work of Grasso et al. [? ] is especially interesting taking into account that is
a Phase I control chart proposal for profiles based on the use of functional data depth concept (as
well as the case of the present proposal). In fact, they proposed a Phase I control chart methodology
for profiles belonging to the “multi-modelling framework”, that includes the following stages: (1) a
classification stage of new profiles into different operating modes or profile patterns using functional
classification techniques from functional data depth measures (maximum depth approach based on
Mode depth); (2) even the identification of a novel operation mode is included; (3) Once the operation
mode corresponding to a new profile is identified, this is assigned to the corresponding control chart
and consequently a suitable control charting method is applied to determine if the process remained in
control over the period of time where those data were collected.
In this study, the atypical detection procedure presented in Reference [? ] is adapted to develop a
control chart for Phase I; this approach allows us to obtain a calibration sample from an under-control
process that can be monitored through new measurements using a Phase II rank control chart. In fact,
in Phase II, it is assumed that the process is under control [? ]; additionally, in each new sample
(monitored sample), the statistical rank is obtained and it is represented in the control chart with the
estimated lower control limit.

2. Methodology
In this section, methodologies of control charts for functional random variables, X, which take
values in a functional space E = L2 ( T ), with T ⊂ R, are developed.
Based on observations of the functional variable X, we obtain a sample each of calibration and
monitoring; they are functional datasets of sizes n and m, respectively, which allow us to build control
charts for Phase I (in the case of the calibration sample) and Phase II (from the monitoring sample).
In the case of designing the control charts, any unstable or out-of-control process is refereed to
the assignable causes of variation emerging from unusual and avoidable events that interrupt the
process, that is, when they cause a change in the parameters of the underlying model of the profile or
functional data [? ]; these variations can be eliminated from the data by identifying and acting on the
cause; this approach will avoid such variations in the future [? ].
Concerning a method for building quality control charts, the probability of process instability
(level of significance) presents a measure of its performance. This probability, provided that it is within
the H0 , allows the derivation of at least one measure (observed value of the statistic) outside the control
limits [? ]. Phase I involves the development of a method and the estimation of the level of significance;

114
Mathematics 2020, 8, 58

however, in the case of Phase II, is the process is assumed to be under control, that is, the level of
significance is fixed.

2.1. Procedure for Building a Control Chart for Phase I (Stabilization)

As mentioned above, in Phase I, the retrospective data corresponding to a calibration sample of
size n is analyzed for evaluating the stability of the process whose quality is ascertained over time and
for estimating the parameters of the control chart [? ].
In Phase I, a control chart is used to test the hypothesis that there is no change in the distribution
of observations of the variable ordered with respect to time {X1 (t), X2 (t), . . . , Xn (t)}.
These changes can be punctual (freaks or bunches) or they can be related to a change in process
that is evaluated (observable through patterns of sudden or gradual change in the mean of the process).
Concerning isolated changes, it refers to the occurrence of at least one observation of the observed
variable that deviates from the distribution of the other observations [? ]. The hypothesis tested in
Phase I is:
d
H0 : Xi (t) = X j (t), ∀i, j ∈ {1, . . . , n}
d
(1)
Ha : Xi (t) = X j (t), for some i, j ∈ {1, . . . , n}.

The stabilization phase of a process consists of applying an iterative method that allows the
detection and elimination of those observations (in this context curves) that have a deviation with
respect to the shape or magnitude of most of the observed curves. In other words, a curve is an
atypical value if it has been generated by a different stochastic process or there is a change in the trend
or variability of the stochastic process with respect to that corresponding to the remaining data [? ].
The quantity of outliers is assumed to be unknown, although small.
In this study, the proposed control charts for Phase I are based on the adaptation of outlier
detection methodologies from functional data, based on the data depth calculations [? ].
The method of the detection of outliers for functional data [? ] considers an atypical curve if its
depth is less than a specific quantile of the distribution of depths estimated by bootstrap. In other
words, an atypical curve will have a significantly low depth.
This procedure can be used with different types of functional depths. In the library fda.usc [? ],
the following alternatives are offered: Fraiman and Muniz (FM) [? ? ], mode depth [? ] and random
projections depth (RP) [? ? ]. In this work, the outlier detection procedure proposed by Febrero et al.
fda.usc has been adapted to estimate a specific quantile of the depth distribution that plays the role
of the lower control limit (LCL) for a Phase I control chart.
The control chart proposed for Phase I is estimated and plotted from a depth measurement (FM, RP
or mode) and only the lower control limit (LCL) is considered to detect if the process is out-of-control
(the depth of a curve is less than the LCL). In addition to this representation, an additional chart
showing the original curves is proposed to provide an intuitive idea about the cause behind the
identified anomaly (by checking its shape or magnitude) and thus to identify assignable causes.
For instance, in the case of the HVAC energy consumption in buildings, anomalies may include
stopping air handler, failure of the counter or sensor, a change in the regulation of the machines
and adverse weather conditions.
In Phase I, we consider the functional random variable X , from which a random sample is
drawn—{X1 (t), X2 (t), . . . , Xn (t)}. These data are used to formulate the following steps to build the
control chart for Phase I:

1. Calculate the depth corresponding to each observation of the dataset, D (Xi )in=1 and make a
control chart based on the depth of each datum with respect to time.
2. Choose the parameter LCL according to the significance level of the control chart, that
is, the percentage of false alarms (observations under control but erroneously detected as

115
Mathematics 2020, 8, 58

out-of-control) is small (for example, α = 1%). The following procedures are used to estimate
the LCL.

• Bootstrap procedure based on Trimmed:

– Reorder the curves according to their depths in a decreasing way. X(1) , . . . , X(n) .
– It is assumed that, at most, α% of the sample can be considered outliers.
– B samples are obtained by a smoothed bootstrap procedure from the dataset resulting
from discarding the α% of the less depth curves. Let Xi∗b , i = 1, . . . , n, b = 1, . . . , B these
bootstrap samples. To obtain each bootstrap resample:

∗ A uniform sampling is done, i∗ of 1, . . . , [n(1 − α)].

∗ Zi∗ is generated as a Gaussian process with zero mean and variance and
covariance matrix.
γΣX with γ ∈ [0, 1], where ΣX is the variance and covariance matrix of the
observations X(1) , . . . , X([n(1−α)]) .
∗ Finally, you get Xi∗b = X(i∗) + Zi∗ .
– For each b = 1, . . . , B, we obtain C b , which is the α% empirical quantile corresponding
to the distribution of the depths, D (Xi∗b ). The output value, C = LCL, is the median of
the values C b , b = 1, . . . , B.
• Bootstrap procedure based on weighing:

– The depths of the X1 , . . . , Xn curves are obtained.

– B samples are obtained through a smoothed bootstrap from the original dataset
weighted by their depths. Let Xi∗b , i = 1, . . . , n, b = 1, . . . , B these bootstrap samples.
These replicas would be obtained:

∗ Weighted sampling is performed, with i∗ of 1, . . . , n and with probability

proportional to D (X1 ), . . . , D (Xn ).
∗ Zi∗ is generated as a Gaussian process with zero mean and variance-covariance
matrix γΣX , with γ [0, 1], where ΣX is the variance and covariance matrix of the
observations X1 , . . . , Xn .
∗ Finally, we get Xi∗b = Xi∗ + Zi∗

– For each b = 1, . . . , B, we get C b , which is the empirical quantile corresponding to the

α% of the distribution of the depths, D (Xi∗b ). The final value C = LCL is the quantile β
of the values C b , with b = 1, . . . , B.

3. If there is any curve such that D (Xi ) ≤ LCL for a given LCL, then it would be considered atypical
and the process would be out-of-control.
4. Additionally, a control chart including the original curves and the functional envelope obtained
from 99% of the deeper bootstrap replicas is also developed.

Moreover, once the atypical curves are detected, they are removed and the procedure is repeated
until the process becomes stable (under control), namely, defined by a total absence of atypical data.

2.2. Procedure for Building a Control Chart for Process Monitoring (Phase II)
Phase II deals with process monitoring; it involves quick detection of changes from the calibrated
sample stabilized in Phase I [? ]. For the scalar and multivariate cases, the process is monitored by
taking the estimated control limits in Phase I [? ] as a reference. In this phase, the average run length
(ARL) is used to evaluate the performance of the control charts [? ].

116
Mathematics 2020, 8, 58

In this context, we test if there are deviations between the data obtained in Phase II, also called
monitoring sample, {Xn+1 (t), Xn+2 (t), . . . , Xm (t)} and the reference data {X1 (t), X2 (t), . . . , Xn (t)} or
calibration sample, taking into account its distribution.
In Phase II, in the univariate case, an F distribution for the under-control process is estimated from
a calibration sample or reference data. It is assumed that F is the distribution of the CTQ variable of an
under-control process (Phase I). This distribution is used to establish control limits that will be used to
monitor the process in Phase II. The limits comprise an interval that will cover new observations of the
process with a high probability, assuming that the process is under control. In Phase II, a sample of the
G distribution is monitored. Therefore, in this stage, the methods for constructing control charts are
based on contrasting the hypothesis:
H0 : F = G
(2)
H1 : F = G.
In the FDA context, we do not have a density function for a functional random variable X that
allows us to perform different tests corresponding to Phases I and II. Alternatively, we can estimate the
distribution of the depth corresponding to the curves that belong to a sample of functional data. Thus,
for Phase II, the use of the rank control charts [? ] is proposed in an FDA context. The calculation of
depths for functional data is proposed, which facilitate the calculation of ranks. These form the basis
for developing r control charts, called rank charts.
The adaptation of r control charts involves the calculation of the rank statistic from functional
depth measurements. The r chart plots the rank statistic as a function of time. The central control
line CL = 0.5 serves as a reference point for observing possible patterns or trends. The lower limit is
LCL = α, where α is the false alarm rate.
For the practical application of this proposal, the qcr [? ] and fda.usc [? ] R packages are used.
The fda.usc package provides tools for the calculation of functional data depth, while the rank control
chart, among other nonparametric charts proposed in Reference [? ], is applied using the qcr package,
which was developed by the authors.
As mentioned above, in Phase II, the curves corresponding to the calibration sample of
Phase I, {X1 (t), X2 (t), . . . , Xn (t)}, are used for detecting changes or deviations with respect
to the behavior of the process described in Phase I. The curves of the monitoring sample,
{Xn+1 (t), Xn+2 (t), . . . , Xn+m (t)}, are collected; additionally, we test the hypothesis of each new curve
belonging to the same distribution that corresponds to the calibration sample.
The procedure for estimating control charts for Phase II follows the same scheme presented in
Reference [? ]. Particularly, we assume that the rank statistic follows a uniform asymptotic distribution.
This result is applicable to the functional case because of the way the rank corresponding to each
observation is calculated (percentage of less deep curves than the observed ones). This fact provides a
computational advantage in the monitoring of continuous processes since it eliminates the need to
estimate the LCL. However, it is set as the quantile of a uniform distribution at a significance level α.
The procedure to develop the rank control chart for the functional univariate case (process
defined by one functional variable) is detailed below, which can be easily generalized to the functional
multivariate case (process defined by more than one functional variable).

1. From the reference sample {X1 (t), X2 (t), . . . , Xn (t)}, get the depths of the dataset, D (Xi )in=1
and the depths of the curves that make up the monitoring sample, D (X j )nj=+nm+1 . The depth of each
curve corresponding to the monitoring sample is calculated from the n curves of the calibration
sample, that is, with respect to n + 1 cases.
2. The rank statistic is calculated for each curve of the monitoring sample, rG (Xn+1 ), . . . , rG (Xm+n ),
considering the calibration sample {X1 (t), X2 (t), . . . , Xn (t)} as sample of reference.

#{Xi | D (Xi ) ≤ D (X ), i = 1, . . . , n}
rG (X ) = .
n

117
Mathematics 2020, 8, 58

3. The values of the rank statistic, the lower control limit LCL = α and central line CL = 0.5 (the
expected value of the rank statistic) are plotted, thereby generating the control chart.
4. Proceed to monitor the process. If at least the rank of a curve, X , is such that rG (X ) ≤ LCL,
then the process is considered out-of-control.
5. A functional control chart is developed. This is a graphical tool that allows us to identify
the possible assignable causes of the out-of-control states. The original curves are included,
those correspond to the reference and monitoring samples, in addition to the functional envelope
obtained from (1 − α)% of the deepest curves of the calibration sample.

3. Data Collection: Case Study of HVAC Installations in Commercial Areas

Here, the case study of HVAC installations’ control are considered for a clothing store of
a commercial area in the Panama City [? ]. The data stream has been obtained by using the
Σqus energy web platform. Sixteen CTQ variables are measured taking into account their ability
to provide information about the energy efficiency, air quality and the thermal comfort of the
store environment—indoor temperatures, overall energy consumption, HVAC energy consumption,
CO2 content in the air (ppm), relative humidity (%), temperatures of impulsion and return
temperatures of the chillers in different areas of the store (see Figure ??).

Figure 1. Plan of the case study store located in the Panama City.

Hourly measurements are obtained from 1 August 2017 to 31 October 2018. The operations of
the HVAC facilities of the store start at 9:00 a.m. or 10:00 a.m. At start-up, the energy consumption
peaks due to the characteristics of the HVAC installation. From 12:00 p.m., the consumption remains
relatively constant the store closure at 20:00 p.m., 21:00 p.m. or 22:00 p.m. The shutdown takes about 1
or 2 h, with consumption falling at a constant rate of change. The resulting data can be considered
functional data and thus FDA techniques can be applied. It is also important to note that this case
study is a controlled study in which the anomalies and their assignable causes have been previously
detected for the maintenance staff.
The data were obtained in the framework of a controlled environment where anomalies were
identified by the maintenance staff. They are briefly described as follows:

• On 11 September, there was a decline in the air conditioning consumption at about midday.
• On 21, 22 and 30 September, the shopping center was closed and thus there was no energy
consumption and the temperatures remained high.
• On 27 September, several maintenance tests were applied to the store’s HVAC installations.
• On 29 September, the store’s HVAC installations were stopped one hour earlier.

118
Mathematics 2020, 8, 58

• Additionally, on 19 September, the air conditioning was stopped half an hour before the usual
time. Particularly, there was a regulation change in the HVAC system.
• In mid-October, there was a leak in the air conditioning circuit. From that moment,
energy consumption began to rise.
• On 1 November, repairing activities were performed. Consequently, the consumption decreased
and the start-up consumption peak was removed.
• Between 17 and 20 November, the energy consumption in HVAC was again increased.

It is important to note that only working days have been studied in this work to evaluate the
performance of the proposed control charts.

4. Application to Real Data

This section shows the usefulness and performance of the new graphical methodology for quality
control using functional data, which is evaluated in the case study on the detection of energy efficiency
anomalies of an HVAC installation. Specifically, the case study considers a commercial area of
a well-known Galician apparel brand located in the Panama City. In this controlled case study,
the anomalies and their assignable causes were detected by the maintenance personnel.
The following section shows the need to develop and apply FDA methodologies for control charts,
considering the observations of the data of the present case study, particularly those corresponding to
August. As mentioned above, in August, no event destabilized the process due to assignable causes.
However, by using a methodology for scalar data (ignoring the autocorrelation between the variables),
an unacceptable number of false alarms could be detected.
In the scalar case, boxplot [? ] is commonly used to detect anomalous or atypical data. Figure ??
shows a traditional scalar approach for detecting outliers using boxplot. The left panel shows the
boxplot for each variable of energy consumption in HVAC systems per hour, while the right panel
shows the curves of daily energy consumption in HVAC systems, highlighting curves detected as
outliers by the descriptive procedure based on the application of boxplots to each hourly consumption.
In the usual procedure, atypical curves are those in which at least one point has been detected as an
outlier in some boxplot; however, the drawback of this approach is that it increases the probability of
type I error. It detects 12 daily energy consumption curves as outliers.

+RXUO\ER[SORWV $QRPDORXVFXUYHVGHWHFWHGE\DSSO\LQJKRXUO\ER[SORWV




íí
+9$&HQHUJ\FRQVXPSWLRQ N:
+9$&HQHUJ\FRQVXPSWLRQ N:

íí
íí
íí
íí




íí
íí
íí
íí
íí
íí




íí



 

 

                  

7LPH K 7LPH K

Figure 2. Detecting outliers in the heating, ventilation and air conditioning (HVAC) energy
consumption by developing a boxplot for energy consumption per h.

Based on the information described in the previous section, first, we apply the data depth control
chart for Phase I and, subsequently, the rank control chart to monitor the process during Phase II.
The application of these two statistical techniques, together with the contribution of an intuitive
graphic tool (to facilitate the detection of assignable causes for the anomalies), constitutes the new

119
Mathematics 2020, 8, 58

proposed procedure of control charts for functional data. Generally, the procedure can be summarized
as follows:

• The energy consumption curves in HVAC during August and September accounts for the
calibration sample. A broad temporal range has been taken to estimate the natural control
limits of energy consumption in HVAC accurately.
• For Phase I, the reference or calibration sample is stabilized by using the control chart developed
from the depth measures of the curves. Particularly, curves detected as atypical by FDA outlier
data detection methods are removed from the calculations to estimate the natural control limits.
• Data corresponding to October and November are monitored to confirm that there is no deviation
in the HVAC processes. If there is a change, that is, there are days with out-of-control HVAC
energy consumption, then it was recommended that the possible assignable causes should
be sought, which, when detected, should be removed in order to remove the corresponding
process variations.

In Figure ??, the black curves correspond to August (23 curves), whereas the gray curves account
for the HVAC energy consumption in September (21 curves). The days from Monday to Friday
are used in this study, taking into account that the work schedule is different on Saturday and
Sunday. In September, the actual anomalies were detected (see Figure ??, curves in red) using the
Phase I control chart based on the functional data depth outlier detection methods. Particularly,
the anomalies corresponding to 11, 21, 22, 27 and 29 September were identified (refer to Section ?? for
more information on the assignable causes of these anomalies).

&XUYHVEHORQJLQJWRWKHFDOLEUDWLRQVDPSOH


$XJXVW
+9$&HQHUJ\FRQVXPSWLRQ N:

íí
6HSWHPEHU
íí
íí


íí
íí


 

           

7LPH K

Figure 3. Daily curves of energy consumption in HVAC facilities of the Panama City’s store. The natural
control limits are estimated from the curves belonging to the calibration sample, which have been
shown which (Phase I).

From this reference sample and based on the simulation study corresponding to the control
chart based on the functional depths (see Section ??), the mode functional depth and the weighed
method for outlier detection are used. Additionally, a significance level of α = 0.025 is used for the
estimation of the LCL from B = 500 bootstrap resamples, a smoothing coefficient γ = 0.8, and a
percentage of Trimming trim = 0.05, which allows obtaining an envelope of 95% of the deepest curves.
The advantage of this procedure of detecting atypical curves in the Phase I control chart based on
functional data is its flexibility to adapt to a wide variety of real problems, by regulating its parameters
(γ and trim).
The left panel of the Figure ?? shows the original gray curves, the estimate of the median (blue
curve), the functional trimmed mean (red curve) and the envelope corresponding to the 95% of the

120
Mathematics 2020, 8, 58

deepest curves, which is plotted in red. The curves detected as anomalous are shown in gray and
dotted in black. After the mentioned atypical energy consumption curves are identified, they are
removed from any calculation related to the estimation of the natural control limits, given that their
assignable causes have been identified (causes apart from the randomness of the data). Subsequently,
the process is repeated according to an iterative scheme until atypical curves associated with assignable
causes are not identified.
Figure ?? shows the functional approximation of the rank control chart, where each datum or point
accounts for a daily curve of energy consumption in HVAC. The first 23 points represent the HVAC
energy consumption curves for August, whereas the next 21 points represent the energy consumption
curves of September.

&DOLEUDWLRQVDPSOH 3KDVH,GDWDGHSWKFRQWUROFKDUW
2ULJLQDOFXUYHV


'DWDGHSWKRIHQHUJ\FRQVXPSWLRQFXUYHV
0HGLDQ 'HHSHVW
(QYHORSHWKDWDFFRXQWVIRUWKHGHHSHUFXUYHV


2XWOLHUV
+9$&HQHUJ\FRQVXPSWLRQ N:























   













7LPH K
7LPH GDWH

Figure 4. Left panel: Curves of daily energy consumption in HVAC facilities of the Panama City’s
store. The figure identifies curves detected as anomalies as well as functional position measurements.
Right panel: Control chart corresponds to Phase I; they facilitate the detection of anomalous curves,
thus eliminating the assignable causes of variation. The depths of each of the daily consumption curves
and the natural control limit are shown.

In this first iteration, the previously identified curves are detected as anomalous. Their structure
corresponds to an assignable cause of variation (see Section ??). They correspond to 11, 21, 22 and 29
September. However, in this iteration, the anomalous curve corresponding to 27 September has been
not detected. Additionally, the curves corresponding to 23 August, 20 and 26 September are detected.
Moreover, the data depth corresponding to the curves from 19 September are very small—they are
outside the lower control limit and thus they are identified as out-of-control states. The assignable
cause corresponding to this behavior is that the air conditioning was being shut off half an hour earlier
than usual.
In the next iteration, the following daily consumption curves are detected as out-of-control or
anomalous: 4 August, 15, 19, 25 and 27 September. In this iteration, the HVAC energy consumption
curve of 27 September and those corresponding to the last days of September are identified; the timing
of HVAC shutdown was changed for the latter part of September. Finally, energy consumption curves
are not detected as anomalous after the second iteration (see Figure ??).

121
Mathematics 2020, 8, 58

&DOLEUDWLRQVDPSOH 3KDVH,GDWDGHSWKFRQWUROFKDUW

(QHUJ\FRQVXPSWLRQFXUYHVGDWDGHSWK


       

RULJLQDOFXUYHV
+9$&HQHUJ\FRQVXPSWLRQ N:

0HGLDQ 'HHSHVW
/LPLWWULP


7ULP0HDQ
RXWOLHUV






                  

7LPH K 7LPH GD\V

Figure 5. Under-control process (HVAC energy consumption is stabilized). Left panel: daily HVAC
energy consumption curves in the Panama City store, showing the envelope corresponding to 95% of
deeper curves. Right panel: control chart for Phase I based on functional data depth.

The process began with 44 curves, of which 9 (in the first iteration) and 5 (in the second iteration)
curves were identified as anomalous. Of these 14 energy consumption curves, 2 and 12 curves
correspond to August and September, respectively. Hence, the calibration sample comprises days in
August and September, until 18 September. The current performance of energy efficiency of HVAC
installation has been characterized through this reference or calibration sample. The next step is to
detect changes (from the reference HVAC performance) in the HVAC system. This task belongs to
Phase II control charts.
After Phase I, the Phase II control charts based on functional data, also called monitoring phase,
is performed. Therefore, the sample to be monitored comprises days corresponding to October in
which the HVAC facilities of the apparel store were operated at regular times. This is in consideration
of the fact that a leak in the air conditioning circuit was recorded in mid-October, especially after the
HVAC consumption began to rise. This behavior can be observed in Figure ??; it plots the monitored
sample as well as the curves of the calibration sample and their 95% envelope, which is estimated in
Phase I.
Figure ?? shows that the HVAC energy consumption, the CTQ variable for the energy efficiency
of HVAC installations, is out of control. The assignable cause is the leak in the air conditioning system.
In order to rectify the anomalous operation of the HVAC installations, on 1 November, a provisional
repair was performed. These actions produced a decrease in the HVAC energy consumption and
attenuated peak energy consumption corresponding to the start-up (see the energy consumption
curves before 11:00). Consequently, the consumption of the HVAC facilities in November is slightly
different from that of August and September, which account for the calibration sample. However,
this energy consumption resumes its rise between 17 and 20 October.
These changes can be observed in Figure ?? wherein it is observed that the HVAC energy
consumption corresponding to the monitored sample is characterized by a greater variability than
that corresponding to the calibration or reference sample. The results of the application of the rank
control chart show that the process is out-of-control both for the monitoring samples of October and
November. In October, the assignable cause is the fault in HVAC facilities, whereas the assignable
cause in November is the change in the process due to repair activities that deals on a decreasing in
HVAC energy consumption.

122
Mathematics 2020, 8, 58

 &DOLEUDWLRQVDPSOHDQGPRQLWRULQJVDPSOH

/LPLWWULPREWDLQHGIURPFDOLEUDWLRQ
+9$&HQHUJ\FRQVXPSWLRQ N:

VDPSOH0RQLWRULQJVDPSOHRI2FWRUEHU
2XWOLHUV






           

7LPH K

Figure 6. Reference of calibration and monitored samples corresponding to October.

&DOLEUDWLRQVDPSOHDQGPRQLWRULQJVDPSOH

/LPLWWULPREWDLQHGIURPFDOLEUDWLRQVDPSOH
       

0RQLWRULQJVDPSOHFRUUHVSRQGLQJWR1RYHPEHU
+9$&HQHUJ\VRQVXPSWLRQ N:

2XWOLHUV


           

7LPH K

Figure 7. Reference (August and September) and monitoring samples corresponding to November.

Given that the process has changed, a new reference sample should be obtained and studied.
The stabilization and monitoring process should begin for the next year, starting December. Figure ??
shows the rank control chart for the calibration (August and September) and monitoring samples of
November. In practice, monitoring can be efficiently performed using the proposed control charts
because you do not need to estimate the LCL through resampling procedures. Additionally, the rank
control chart can be adapted for monitoring multivariate functional data, that is, when there are
different types of curves that define the quality of a specific system or process simultaneously.
For example, a direct case of application considers their first derivative along with the original curves.
The application of rank control charts also allows us to monitor the process by studying more
than one functional variable. Particularly, it not only allows us to study the energy consumption
curves but also the curves of daily temperature, daily relative humidity, daily CO2 concentration,
among others; they completely characterize the energy efficiency of the facilities and their thermal
comfort and ambient air quality. Precisely, this procedure would allow us to perform a functional
multivariate monitoring.

123
 Mathematics 2020, 8, 58

3KDVH,,UDQNFRQWUROFKDUW
5DQNRIHDFKGD\O\HQHUJ\FRQVXPSWLRQFXUYH

     

     

7LPH GD\V

Figure 8. Rank control chart for Phase II of quality control. Both reference (corresponding to August
and September) and monitoring (corresponding to November) samples are shown.

5. Simulation Study
The control chart performance is many times measured by observing its power (1 − β). It is
defined as the probability of identifying an out-of-control state when the process state is actually
out-of-control [? ]. Moreover, assuming that the process is actually under control, the type I error (α) or
false alarm rate will be defined as the probability of detecting an out-of-control signal [? ].
In the case of the process is under control, the probability of identifying an out-of-control
observation should be small enough to prevent an unacceptable number of false alarms. Otherwise,
if the process is effectively out of control, the power should be high enough to detect the process
change as quickly as possible [? ].
Another common index for measuring the performance of a control chart is Average Run Length
(ARL), which is defined as the average number of observations plotted before a signal is out-of-control.
1
The ARL is equal to (if we can assume that the signals are independent, that the run length
p
distribution is geometric), where p is the probability of having an out-of-control signal [? ].
To evaluate the performance of a control chart in Phase II, the ARL0 and ARL1 are often used,
which are the average number of observations until the first out of control is detected, in cases where
the process is really under control (ARL0 = α1 ) or actually is not (ARL1 = 1−1 β ) [? ]. The ARL1 should
be at its lowest to increase the probability of quickly identifying events (1 − β), power of a test that
lead to the process being out-of-control [? ].
Given that the F distribution is not known, a Monte Carlo simulation is designed to calculate
the control charts power. The simulated scenarios allow us to estimate and compare the power of the
control charts for different data depth measurements and for the case of independent and dependent
functional data.
In this section, the performance of the control graphics proposed for Phases I and II will be
evaluated. First, the simulation scheme designed in Febrero et al. [? ] will be used to evaluate the
performance of the control chart proposal for Phase I. Realizations of a Gaussian stochastic process
have been proposed following the expression below [? ? ]:

X ( t ) = μ ( t ) + σ ( t ) · ( t ), (3)

124
Mathematics 2020, 8, 58

whereby σ2 (t) = 0.5 and

μ(t) = E (X (t)) = 30t(1 − t)3/2 , (4)

whereas (t) is a Gaussian process (t) ∼ GP(0, Σ) with 0 mean and variance-covariance matrix
equal to

  | ti − t j |
E (ti ) × (t j ) = e− 0.3 .
Additionally, Reference [? ] developed an alternative model to generate atypical curves, μ(t) =
30t3/2 (1 − t). In Figure ??a, the two functional means are presented. The black curve accounts for the
process mean without atypical curves, while the red curve is the mean of the process that generates
the atypical curves.
The control charts proposed in this work have been designed to monitor the functional mean to
detect two events—change in the mean of the process in terms of the magnitude and shape—which
reveal that the process is not under control. For designing control charts for Phase II, it is assumed that
the process is under control, that is, outliers are not detected. To generate simulation scenarios for each
of these events, the following functional means have been considered:

• Mean of the model with a change in the magnitude:

μ(t) = 30t(1 − t)3/2 + δ, (5)

by which δ denotes the change that goes from 0.4 to 2 in steps of 0.4.
• Mean of the model with a change in form:

μ(t) = (1 − η ) · 30t(1 − t)3/2 + η · 30t3/2 (1 − t), (6)

where η is the change from 0.2 to 1 in steps of 0.2.

In Figure ??b, the green curve accounts for the functional mean of a process when there is a change
in the magnitude (δ = 0.7), while the curve of blue denotes the mean of a process when there is a
change in the shape (η = 0.3).
In Febrero et al. [? ], the functional data X1 , . . . , Xn denote realizations of a stochastic process
X (·), assuming continuous trajectories in the [ a, b] = [0, 1] period and independence between the
curves. However, simulation scenarios in which the simulated curves are defined by a variable degree
of dependence have also been considered. This is because several practical applications of this type
of chart are related to continuously monitored data with respect to time, forming functional time
series, such as the curves of daily energy consumption in commercial areas. In this way, dependent
curves are generated from the model Ỹi (t) = μ(t) + σ(t) · ˜ (t), with ˜ (t) = ρ · ˜ i−1 (t) + (1 − ρ) · i (t),
where ρ is the correlation measure between curves and σ (t) = 0.5 and both (t) and ˜ (t) are Gaussian
processes [? ].
In order to compare the results of the simulations in the scenarios defined by independence and
dependence between the curves, the variance of is rescaled (we define the variance of the error ˜ to
(1− ρ2 ) (1+ ρ )
be one). Specifically, considering σ2 = (1−ρ)2 = (1−ρ) , you have σ2˜ = 1.
In Figure ??, different scenarios are presented considering the changes in the functional mean
of the process, in the shape and magnitude, in the cases of independence and dependence between
curves. The gray curves show the realizations of the process when it is under control (whose mean
is the Equation (??)). However, the red curve in each graph accounts for the scenarios in which the
presence of events that destabilize the process is considered, that is, the process is not under control.
In Figure ??a,b, the cases of independence between curves and the presence of events defined by
changes in the functional mean in terms of the magnitude and in shape, respectively, are shown.

125
Mathematics 2020, 8, 58

However, Figure ??c,d show two cases defined by the presence of dependence between curves,
including changes in the magnitude of the mean (panel c), with respect to its shape (panel d).

D E









MW

MW











     

W W
Figure 9. (a) Functional means and (b) changes in the shape and the magnitude in the mean of
the process.

In the building energy efficiency domain, the out of control signals of energy consumption,
temperature, CO2 proportion and humidity, among others, can be defined by a change of shape
and/or magnitude with respect to the under control signals analogous to those analyzed in the
simulation study. The change in magnitude is related with a change in the scale of the studied process,
for example, the increasing of energy consumption, temperature, humidity or concentration of CO2
in all the hours corresponding of a specific day with respect to the otherwise normal pattern. This
type of change is accounted by the addition of the δ term to the Equation (??) in order to obtain the
Equation (??). Thus, the amount of that change is controlled by δ parameter. The green curve of
Figure ?? accounts for an example of magnitude change. If we compare with the black one, we could
realize that the two curves have the same shape and the only difference is the scale. On the other hand,
changes in the shape of the curves are introduced by modifying the Equation (??) by the η parameter
resulting in the Equation (??). In the energy efficiency domain, this type of change can be related to a
change in the HVAC facilities programing (changes in temperature regulation, changes in the time
schedule), in a failure of the HVAC in just one interval of the day, an extremely high or low level of
occupation in a building and extreme changes in the weather, among other causes. The proposed
simulation study has performed taking into account the specific domain where the FDA control chart
approach is applied, namely energy efficiency in buildings. Thus, the shape of these types of profiles
is similar to that corresponding to CO2 , temperature, energy consumption and humidity curves in
buildings. In order to measure the performance of the proposed control charts for very different types
of profiles, new studies may be necessary.

126
Mathematics 2020, 8, 58

In the following section, a simulation study is performed to determine conditions under which
it can be verified that the smooth bootstrap procedure works when there is independence and
dependence between curves.

D E
í  

í  
;W

;W
           
W W

F G
í  

í  
;W

;W

           
W W
Figure 10. Scenarios in which independence between curves is studied. Changes in the functional
mean with respect to its magnitude (a) and shape (b) are shown. In the case of dependence, panels (c)
and (d) show the simulation scenarios in which changes in the magnitude and shape, respectively,
are observed in the functional mean.

5.1. Measurement and Comparison of the Performance of the Control Chart Proposed for Phase I
The performance of the control chart is estimated and compared from the generation of calibrated
samples of size 50 and 100 (curves). For each sample, different functional depth measurements
described in Section ?? are calculated and the outlier detection robust procedures (weighted and
trimmed) are applied for the estimation of type I error when the process is under control and the
power of the test when the process is out-of-control. For the estimation of type I error, each scenario
is replicated 1000 times (n = 50, 100, assuming independence and dependence between curves).
When the power of the test is estimated, in each scenario (assuming independence and dependence),
a curve within the alternative hypothesis is generated; this procedure is also repeated 1000 times.
Following the scheme described in Reference [? ], curves observed at equidistant points are
considered; the number of points that define each curve is 51 in the interval [0, 1]. From 1000 resamples
(B = 1000) and with a 2.5% trimming procedure (removing less deep curves), a smoothing bootstrap
procedure defined by a smoothing factor γ = 0.05 is applied to estimate the C = 0.01 quantile
representing the LCL.
First, a simulation study is performed to estimate and compare the type I error (α = 0.01 is fixed)
of the proposed control chart, assuming scenarios with independence and dependence between curves.
Subsequently, a similar study is carried out to estimate and evaluate the power of the control chart to
detect out-of-control signals in different situations (independence, dependence, different sample size
and a change in the shape or magnitude).

127
Mathematics 2020, 8, 58

In Table ??, the results of the estimation of the false alarm rate (type I error) in the independence
scenario are shown. The average of the percentage of false out-of-control signals (type I error) detected
by the procedure shown above are very close to the nominal 1% for the two considered sample sizes.
Furthermore, it can be observed that, when n increases, the type I error percentages are closer to the
nominal level. In general, for a sample of size n = 100, the results of applying the weighted method
are closer to α, especially when using the mode depth measurement. The results obtained in the
simulations are similar to those presented in Reference [? ].

Table 1. Estimation of the false alarm rate (%) for the case of independence between curves using a
nominal type I error of 1%.

n Method FM RP Mode
Weighted 1.94 1.89 1.49
50
Trimmed 1.34 1.95 1.36
Weighted 1.55 1.75 1.25
100
Trimmed 1.67 2.33 1.76

In any process, type I error increases the production cost. Hence, it is essential not to overestimate
this error rate when managing the quality.
Table ?? shows the results of the simulation to evaluate the ability of the control chart to detect a
change in the shape or magnitude of the functional mean of the process through the estimation of its
power (1 − β). The percentage of out-of-control signals (outliers) correctly detect when the population
defined by the Equation (??) is contaminated with curves belonging to the M1 model (Equation (??))
and M2 model (Equation (??)); it is denoted by pc ; however, the percentage of false alarms (false
states out of control) is p f . These parameters have been estimated, in all the scenarios assuming the
independence of curves, using the average of the corresponding empirical values, p̂c and p̂ f .
Table ?? shows that a better performance is achieved when the curves of model M1 (where changes
in the magnitude of the proposed control chart are simulated) are studied. Precisely, p̂ f and p̂c are
closer to the nominal α and (1 − β). When identifying changes in the shape of the process average, M2 ,
the mode depth provides the highest percentages of correctly detected out-of-control signals. However,
in the case of the M1 model, the use of RP depth provides percentages of the detection of the true
out-of-control states lower than those corresponding to the use of FM and mode depths. With respect
to a robust method for outlier detection, the performance is similar in all the scenarios. However,
an exception is the case wherein the RP depth is used; it reveals the low performance of the control
chart in detecting observations corresponding to actual out-of-control states.
Briefly, the detection rate of false out-of-control signals for the independence scenario is close to
1%. However, when using the trimmed method, the detection rate of false out-of-control signals is
overestimated but this percentage decreases when the sample size increases.
The results of the false alarm rate (type I error) for scenarios defined by dependence between
curves are shown in Table ??. It is important to note that, for different values of ρ, very similar
results with respect to the independence scenarios have been obtained. Precisely, the average of the
percentages of false out-of-control signals are close to the nominal 1% in the two studied sample sizes.
Additionally, when n increases, the type I error percentages are closer to the nominal level. However,
some differences are observed when the RP data depth measure is used to develop the control chart.
In this case, there is an overestimation of the percentage of false out-of-control signals.

128
Mathematics 2020, 8, 58

Table 2. Percentages of p̂c and p̂ f for the cases of curves simulated with M1 (Equation (??)) and M2
(Equation (??)) models, assuming independence between curves.

δ 0.4 0.8 1.2 1.6 2

Stage Displays Depth Method p̂ f p̂c p̂ f p̂c p̂ f p̂c p̂ f p̂c p̂ f p̂c
Weighted 1.80 8.30 1.63 27.80 1.47 56.50 1.48 85.30 1.60 95.70
FM
Trimmed 1.25 7.70 1.19 26.80 1.19 56.00 1.22 83.00 1.20 95.50
Weighted 1.88 6.80 1.68 20.10 1.49 45.10 1.42 71.90 1.42 87.70
50 RP
Trimmed 1.88 7.00 1.89 23.20 2.08 49.30 2.14 75.70 2.20 91.90
Weighted 1.41 5.70 1.27 19.90 1.09 45.10 1.08 76.70 1.17 94.30
Mode
Trimmed 1.33 6.80 1.32 22.20 1.33 48.60 1.38 77.50 1.39 95.20
M1
Weighted 1.51 5.70 1.40 23.30 1.34 53.50 1.34 82.00 1.38 96.20
FM
Trimmed 1.63 7.40 1.51 25.90 1.32 56.10 1.13 83.60 1.00 96.70
Weighted 1.68 5.00 1.57 18.60 1.50 45.40 1.45 71.40 1.52 90.10
100 RP
Trimmed 2.28 7.40 2.22 23.70 2.23 54.20 2.13 78.70 2.08 93.50
Weighted 1.22 4.20 1.15 17.90 1.08 46.80 1.06 78.90 1.12 94.70
Mode
Trimmed 1.70 5.80 1.62 22.90 1.46 54.20 1.28 83.50 1.13 95.90
η 0.2 0.4 0.6 0.8 1
Weighed 1.91 1.50 1.85 3.40 1.70 11.00 1.46 26.30 1.25 55.00
FM
Trimmed 1.28 1.90 1.23 4.10 1.19 11.60 1.13 26.70 1.18 50.60
Weighted 1.94 2.40 1.90 3.30 1.75 8.00 1.64 16.90 1.42 32.20
50 RP
Trimmed 1.89 2.30 1.83 4.60 1.73 10.10 1.84 22.50 1.92 41.40
Weighted 1.43 2.80 1.35 10.70 1.19 31.00 1.06 65.30 1.12 91.70
Mode
Trimmed 1.33 4.00 1.31 12.30 1.31 33.70 1.42 68.00 1.48 92.10
M2
Weighted 1.53 2.00 1.51 2.70 1.45 12.20 1.35 27.90 1.25 53.00
FM
Trimmed 1.66 2.50 1.63 4.30 1.57 14.40 1.51 33.40 1.43 58.00
Weighted 1.69 2.70 1.66 4.10 1.63 8.20 1.55 17.30 1.48 34.70
100 RP
Trimmed 2.25 2.90 2.19 5.60 2.15 13.60 2.15 25.40 2.21 45.50
Weighted 1.22 3.30 1.18 11.90 1.12 34.10 1.07 66.80 1.09 89.60
Mode
Trimmed 1.72 4.60 1.65 15.70 1.57 41.70 1.41 72.70 1.22 92.90

Table 3. Results for the scenarios with dependence between curves. The estimates of false alarm rate
(type I error) for values of ρ between 0.3 and 0.7 are shown.

ρ n Method FM RP Mode
Weighted 1.85 1.88 1.41
50
Trimmed 1.31 1.90 1.33
0.3
Weighted 1.54 1.71 1.20
100
Trimmed 1.65 2.23 1.71
Weighted 1.80 1.81 1.28
50
Trimmed 1.26 1.82 1.29
0.5
Weighted 1.46 1.68 1.15
100
Trimmed 1.61 2.21 1.65
Weighted 1.66 1.73 0.93
50
Trimmed 1.25 1.74 1.06
0.7
Weighted 1.42 1.61 0.99
100
Trimmed 1.56 2.15 1.57

Tables ??–?? show the results of the empirical estimation of pc and p f , assuming different values
of ρ (from 0.3 to 0.7). The power (estimated by p̂c ) of the control chart proposed for the model M1
(Equation (??)) performs better when the weighted method is applied and if the sample size is increased.
It is also observed that the performance of the control chart tends to be the same, independent of the
type of data depth measurement used. Certainly, the performance of control charts in detecting real
changes in the process, related to differences in the shape and mean, is better when the mode depth
is used.

129
Mathematics 2020, 8, 58

Table 4. Empirical values of p̂ f and p̂c , with ρ = 0.3 (assuming dependence between curves).

δ 0.4 0.8 1.2 1.6 2

Model Sample Size Depth Method p̂ f p̂c p̂ f p̂c p̂ f p̂c p̂ f p̂c p̂ f p̂c
Weighted 1.58 8.90 1.22 33.60 0.98 69.80 1.07 91.70 1.19 99.30
FM
Trimming 1.12 9.20 0.75 26.60 0.37 47.40 0.09 54.30 0.02 49.20
Weighted 1.75 7.10 1.38 26.30 0.97 53.80 0.74 77.60 0.67 90.70
50 RP
Trimmed 1.72 7.60 1.42 25.90 0.89 45.90 0.49 60.90 0.20 68.70
Weighted 1.21 5.80 0.90 24.90 0.67 59.80 0.70 88.30 0.86 98.50
Mode
Trimmed 1.21 6.40 0.89 24.00 0.50 49.40 0.19 73.50 0.04 89.80
M1
Weighted 1.47 7.30 1.23 29.60 1.07 67.80 1.13 92.30 1.24 99.20
FM
Trimmed 1.53 7.60 1.29 31.20 1.04 68.70 0.96 91.30 0.92 99.10
Weighted 1.69 5.50 14.67 24.40 1.21 58.00 1.06 83.90 1.02 95.90
100 RP
Trimmed 2.19 8.80 1.99 27.40 1.88 64.80 1.92 89.30 1.85 97.30
Weighted 1.17 4.60 0.98 23.00 0.86 62.40 0.89 91.00 1.01 98.50
Mode
Trimming 1.62 6.90 1.38 27.90 1.15 67.10 1.07 91.70 1.06 98.90
η 0.2 0.4 0.6 0.8 1
Weighted 1.75 1.30 1.61 1.95 1.31 7.70 0.94 16.85 0.64 26.05
FM
Trimmed 1.25 1.40 1.15 2.35 0.97 7.70 0.73 15.95 0.51 20.50
Weighted 1.86 1.20 1.75 3.00 1.43 5.50 1.24 10.85 0.96 16.55
50 RP
Trimmed 1.79 1.65 1.68 3.40 1.58 7.15 1.46 13.10 1.16 19.40
Weighted 1.31 3.70 1.08 16.00 0.78 44.90 0.62 79.30 0.81 96.90
Mode
Trimmed 1.22 4.70 1.11 17.50 0.74 42.20 0.29 65.00 0.07 82.50
M2
Weighted 1.53 0.90 1.46 2.05 1.30 7.45 1.07 19.50 0.92 32.35
FM
Trimmed 1.65 1.20 1.57 2.70 1.44 9.40 1.22 22.75 1.12 35.35
Weighted 1.76 0.60 1.73 2.20 1.57 5.10 1.41 12.10 1.27 21.80
100 RP
Trimmed 2.23 1.10 2.18 3.30 2.09 8.25 1.91 17.45 1.97 29.35
Weighted 1.17 3.55 1.07 14.70 0.91 47.20 0.86 83.10 0.98 97.50
Mode
Trimmed 1.64 4.70 1.48 20.50 1.25 52.60 1.13 85.10 1.12 98.20

Table 5. Empirical values of p̂ f and p̂c , with ρ = 0.5 (assuming dependence between curves).

δ 0.4 0.8 1.2 1.6 2

Model Sample Size Depth Method p̂ f p̂c p̂ f p̂c p̂ f p̂c p̂ f p̂c p̂ f p̂c
Weighted 1.46 11.50 1.04 43.90 0.84 80.80 1.06 97.40 1.14 99.90
FM
Trimmed 1.02 10.80 0.63 33.40 0.23 50.70 0.28 53.80 0.00 45.60
Weighted 1.68 9.10 1.21 31.10 0.80 63.00 0.65 85.60 0.60 94.60
50 RP
Trimmed 1.73 9.60 1.26 30.70 0.66 52.60 0.28 66.40 0.11 68.80
Weighted 1.10 7.50 0.76 34.40 0.55 73.00 0.69 95.20 0.82 99.90
Mode
Trimmed 1.05 8.10 0.72 31.40 0.32 62.60 0.06 86.50 0.01 96.40
M1
Weighted 1.20 10.00 1.08 39.60 1.01 79.40 1.10 97.30 1.18 99.80
FM
Trimmed 1.30 10.00 1.18 41.60 0.96 80.20 0.89 97.00 0.84 99.80
Weighted 1.40 10.00 1.39 32.30 1.09 69.20 0.97 90.60 0.97 97.90
100 RP
Trimmed 1.80 10.00 1.93 38.20 1.84 77.10 1.80 94.70 1.73 98.80
Weighted 0.90 10.00 0.86 32.10 0.79 75.80 0.91 96.00 0.98 99.90
Mode
Trimmed 1.30 10.00 1.30 38.30 1.10 79.40 1.06 97.20 1.02 99.90
η 0.2 0.4 0.6 0.8 1
Weighted 1.63 1.40 1.43 3.55 1.02 10.80 0.68 21.60 0.50 33.10
FM
Trimmed 1.16 1.50 1.02 3.60 0.80 11.20 0.54 18.00 0.32 21.70
Weighted 1.89 1.75 1.65 3.30 1.32 7.30 0.98 14.30 0.75 22.55
50 RP
Trimmed 1.79 2.30 1.62 3.80 1.54 9.05 1.26 16.40 0.95 22.90
Weighted 1.13 5.40 0.86 22.60 0.56 60.50 0.59 91.50 0.85 99.80
Mode
Trimmed 1.10 6.30 0.83 23.40 0.46 54.00 0.11 77.20 0.02 93.80
M2
Weighted 0.01 0.01 0.01 0.03 1.13 11.90 0.90 28.25 0.77 38.90
FM
Trimmed 0.02 0.01 0.01 0.04 1.28 15.25 1.11 30.90 1.04 39.90
Weighted 0.02 0.01 0.02 0.03 1.42 7.65 1.26 16.65 1.08 27.55
100 RP
Trimmed 0.03 0.02 0.02 0.05 2.02 12.60 1.91 23.25 1.96 35.00
Weighted 0.01 0.05 0.01 0.23 0.82 63.00 0.87 92.50 0.99 99.80
Mode
Trimmed 0.02 0.07 0.01 0.28 1.21 68.40 1.12 93.40 1.12 99.60

130
Mathematics 2020, 8, 58

Table 6. Empirical values of p̂ f and p̂c , with ρ = 0.7 (assuming dependence between curves).

δ 0.4 0.8 1.2 1.6 2

Model Sample Size Depth Method p̂ f p̂c p̂ f p̂c p̂ f p̂c p̂ f p̂c p̂ f p̂c
Weighted 1.32 16.00 0.87 57.70 0.85 91.40 1.01 99.30 1.01 100.00
FM
Trimmed 0.87 13.40 0.42 40.30 0.07 53.60 0.00 47.50 0.00 35.80
Weighted 1.41 13.00 0.97 43.40 0.61 76.70 0.56 92.00 0.53 98.00
50 RP
Trimmed 1.51 14.00 0.97 40.70 0.39 61.40 0.14 68.20 0.04 71.30
Weighted 0.71 14.00 0.44 52.20 0.41 88.90 0.51 99.30 0.53 100.00
Mode
Trimmed 0.82 14.30 0.43 47.40 0.09 80.20 0.01 96.00 0.00 99.90
M1
Weighted 1.29 12.00 0.99 54.90 1.05 90.40 1.14 99.50 1.13 100.00
FM
Trimmed 1.36 13.80 1.08 56.00 0.89 90.20 0.84 99.20 0.76 100.00
Weighted 1.50 9.90 1.20 48.30 0.97 80.80 0.93 96.40 0.96 99.50
100 RP
Trimmed 2.00 13.90 1.80 53.10 1.80 87.30 1.72 98.30 1.60 99.90
Weighted 0.90 9.80 0.68 51.00 0.69 90.00 0.81 99.40 0.83 99.90
Mode
Trimmed 1.43 13.40 1.14 58.10 1.01 92.40 0.97 99.60 0.92 99.90
η 0.2 0.4 0.6 0.8 1
Weighted 1.52 1.80 1.21 6.05 0.76 17.90 0.49 30.70 0.41 39.95
FM
Trimming 1.07 2.10 0.85 6.15 0.56 15.30 0.31 22.30 0.18 24.05
Weighted 1.62 2.10 1.36 5.25 0.98 11.40 0.68 19.95 0.56 28.30
50 RP
Trimmed 1.59 2.40 1.41 5.80 1.19 13.85 0.92 21.05 0.70 28.15
Weighted 0.75 8.60 0.47 39.90 0.37 82.80 0.50 99.10 0.61 100.00
Mode
Trimmed 0.87 10.30 0.51 38.30 0.14 73.10 0.01 93.80 0.00 99.30
M1
Weighted 2.02 0.00 1.24 6.35 0.99 21.20 0.81 36.45 0.75 45.10
FM
Trimmed 3.03 0.00 1.36 7.80 1.16 24.30 1.05 37.95 1.03 44.55
Weighted 3.69 0.00 1.50 5.25 1.25 13.00 1.08 24.50 0.99 34.85
100 RP
Trimmed 6.28 0.00 1.98 7.20 1.84 18.50 1.89 31.80 1.89 41.60
Weighted 1.63 0.00 0.72 37.90 0.69 82.70 0.80 98.60 0.90 100.00
Mode
Trimmed 3.48 0.00 1.24 46.70 1.10 87.10 1.05 99.00 1.04 100.00

With respect to the false out-of-control rate p̂ f , in the scenarios corresponding to the use of M1
model, when the trimmed method is also used, a lower rate is obtained. In the case of the M2 model,
there are similar results on the scenarios defined by independence between curves, that is, the p̂ f is
lower when the trimmed method for outlier detection is used.
In Reference [? ], new methods for the detection of outliers were proposed for the case in which
there is dependence between curves. From the simulation studies carried out in this study, at different
degrees of dependence, we can say that the outlier detection method proposed in Reference [? ] was
relatively robust against the presence of dependence between curves. The simulation study performed
in this section supports the results obtained in the work in Reference [? ] and, in conclusion, justifies
the use of this method within the new control charts proposed for Phase I, even in scenarios with
dependence between curves.
Although the application of the weighed outlier detection method to Mode data depth has
generally provided Phase I control charts with best performance, if the false alarm rate of Tables ??
and ?? are observed, the use of trimmed outlier detection method tends to provide values of p̂ f slightly
closer to α = 1% (with respect to the weighted method) when the process is actually under control,
the curves are independent and the sample is relatively small (n = 50). In addition, if the process is
out of control, the curves are independent and the outliers are generated by the Equation (??) (changes
in magnitude), the trimmed method applied to FM data depth provide a p̂ f close to α = 1% and the
highest p̂c , as shown in Table ??. In all the remaining scenarios, the use of weighed method applied to
Mode data depth tends to provide the closest to α% p̂ f and the highest p̂c (see Tables ??–??).

5.2. Measurement and Comparison of the Performance of the Control Chart Proposed for Phase II
For Phase II, the monitoring stage, the use of the rank control chart has been proposed.
The application of the rank control chart allows simultaneous monitoring of changes in the mean and
variability of a process. In the functional case, in order to calculate the rank statistic, the functional FM,
RP and mode depths are used.

131
Mathematics 2020, 8, 58

An ARL0 = α=0.025 1
(the monitoring sample is assumed under control) is assumed to evaluate
the performance of the control chart. The power of the control chart is estimated and compared for
an under-control process, based on the generation of a calibration sample of size n = 50 by a Monte
Carlo procedure.
Following the simulation scheme of Phase I, curves observed at equidistant points are assumed;
they are composed of 51 points at [0, 1] interval. A smoothed bootstrap with a smoothing factor
γ = 0.05, 1000 resamples (B = 1000), using a 2.5% trimmed procedure (removing the shallowest
curves), is applied to estimate and compare the power of the control chart to detect out-of-control
signals when a significance level of α = 0.025 is assumed. Additionally, in the same way as in Phase I,
the simulation of scenarios with independence and dependence between curves are assumed.
Table ?? shows the estimates of the power (%) of the control charts for the scenario of independence
between curves, whereas Table ?? shows power of the control chart for the scenario with dependence.
In both cases, the ability of the control chart to detect a change in the magnitude of the functional mean
of the process is evaluated by the estimation of its power (1 − β).

Table 7. Power of the control chart, 1 − β, for the case M1 (Equation (??)) model in the scenario of
independence between curves.

δ FM RP Mode
0.5 14.5 14.8 14.8
1 47.6 47.2 48.0
1.5 83.3 80.8 83.0
2 97.8 97.5 97.7

From the results of the Table ??, any depth measure can be used to detect a shift in the process
mean, since the same performance, in terms of power, is obtained.

Table 8. Power of the control chart, 1 − β, for the scenarios defined by the M1 (??) model, assuming
dependence between curves.

ρ δ FM RP Mode
0.5 20.28 23.73 21.84
1 62.95 64.75 64.06
0.3
1.5 92.98 92.33 93.64
2 99.56 99.16 99.75
0.5 25.59 29.41 28.41
1 73.84 75.12 76.24
0.5
1.5 97.15 96.56 97.94
2 99.91 99.77 99.98
0.5 35.47 40.23 43.55
1 86.14 86.87 90.38
0.7
1.5 99.48 99.11 99.78
2 100.00 99.99 100.00

The results of the detection of a shift in the process mean are shown in Table ??. A similar
performance is observed when using any depth measure for different values of ρ. Apparently,
the control chart for Phase II is robust against the existence of dependence between curves.
As observed in the simulation study and in the analysis of the case study with real data, the present
proposal of control charts for functional data, including Phase I and II control charts, can be useful to
detect anomalies in diverse scenarios. In the case of its application to real data, the set of proposed
techniques is being examined for implementation in the web platform Σqus and for its use by the
company Nerxus for detecting false alarms in facilities in commercial areas. The present control
chart methodology can be used for control tasks, monitoring, anomaly detection and continuous

132
Mathematics 2020, 8, 58

improvement in diverse industrial processes, monitoring of environmental variables, chemical industry

and, in general, any process involving continuous monitoring of functional data over time.
Regarding the use of our methodology in more complex case studies defined by different operation
modes, the application of the multi-modelling framework methodology in combination with our
proposal could be useful. Indeed, in the building energy efficiency domain, there are many different
operation modes of installations, each one defined by a specific operation pattern. Namely, the HVAC
installations can be operated in heating or ventilation modes (there are even different modes within
ventilation or heating). The automatic classification of each profile in the corresponding profile pattern
could be very useful in the building energy efficiency field and previously to the application of our
control chart proposal for Phase I and Phase II. With respect to the work of Grasso et al. [? ], it is also
interesting to mention that the proposed profile monitoring control charting scheme is that based on
functional PCA and described in Colosimo and Pacella [? ].

6. Conclusions
A new alternative of control charts has been proposed when CTQ variables of the process are
functional. The proposal includes alternatives to develop the the Phase I and II control charts for
stabilizing and monitoring the processes, respectively. In order to develop Phase I control charts
based on functional data, outlier detection methods are used based on a method of smooth bootstrap
resampling and the depth calculation of functional data is proposed. However, in order to implement
Phase II, the use of rank-type nonparametric control charts based on the concept of functional data
depth is proposed. This Phase II control chart is directly estimated assuming that the asymptotic
distribution of the rank statistic is a uniform distribution. The application of the control charts to the
two-process control phases and the development of a new graphic tool for visualizing functional data
(including an envelope with 95% of the deepest curves that facilitate the identification of the assignable
cause of each anomaly) give rise to the proposed methodology. It has been successfully applied
in real case studies belonging to the framework of anomaly detection in building energy efficiency.
Additionally, a simulation study is conducted to measure the performance (as the percentage of
rejection when the null hypothesis is not met) of the control charts, depending on the functional data
depth used, the sample size, the presence of dependence between curves and the use of different FDA
procedures for outlier detection.
In the simulation study, the use of different types of functional depths has been compared to
develop Phase II of the proposed control chart. In case of the univariate functional data (single type of
curves), for the three scenarios, a better performance is obtained with the mode depth measurement
combined with the weighted outlier detection method and moderately large samples. Additionally,
one of the final observations of the simulation study is that the control chart methodology is robust
against the presence of dependence between curves. Thus, this alternative tool can be applied to the
framework of continuously monitored data streams.
Generally, the authors recommend using the weighed method and the Mode functional data
depth for the case of Phase I taking into account the values of p̂ f and p̂c . Thus, when the Phase I
control chart is evaluated, both weighted method and Mode data depth are generally the best options
in those scenarios defined by under control assumption and even in those where out of control curves
are simulated. In the latter, when the change in magnitude or shape is very small, the corresponding
p̂c tends to be not higher to those obtained by the use of other combinations of data depth measure
and outlier detection method. Nevertheless, when the change in magnitude (δ) or shape (η) increases,
the power of the combination of weighted method and Mode depth tends to be higher than those
corresponding to the other combinations. Moreover, regarding the Phase II control chart and taking
into account the higher values of power estimates included in Tables ?? and ??, we also recommend
the use of Mode data depth for Phase II control charts.
The present proposal has been verified by its application in a real case study dealing with the
detection of energy efficiency anomalies in buildings. Specifically, all the previously identified real

133
Mathematics 2020, 8, 58

anomalies (by the maintenance personnel) have also been successfully identified by the application
of this functional approximation of control charts for Phases I and II of process control. Additionally,
the proposed graphical tool helps to intuitively identify the assignable causes corresponding to
each anomaly.
This procedure can be used in different industrial and scientific domains in which the control
procedures are defined by functional CTQ variables.

Author Contributions: Conceptualization, S.N., M.F. and J.T.-S.; methodology, M.F., R.F.-C. and S.N.; software,
M.F. and R.F.-C.; validation, M.F., R.F.-C. and J.T.-S.; formal analysis, M.F., R.F.-C., S.N. and J.T.-S.; investigation,
M.F., R.F.-C., S.N., J.T.-S., S.Z. and P.R.; resources, S.N., S.Z. and P.R.; data curation, S.Z. and P.R.; writing—original
draft preparation, M.F., J.T.-S., S.N. and R.F.-C.; writing—review and editing, J.T.-S., R.F.-C., M.F., S.N. and S.Z.;
visualization, Miguel Flores, R.F.-C. and S.N.; supervision, S.N., J.T.-S. and R.F.-C.; project administration, S.N.;
funding acquisition, S.Z. and S.N. All authors have read and agreed to the published version of the manuscript.
Funding: This study has been funded by the eCOAR project (PC18/03) of CITIC. The work of Salvador Naya,
Javier Tarrío-Saavedra, Miguel Flores and Rubén Fernández-Casal has been supported by MINECO grants
MTM2014-52876-R, MTM2017-82724-R, the Xunta de Galicia (Grupos de Referencia Competitiva ED431C-2016-015,
and Centro Singular de Investigación de Galicia ED431G/01 2016-19), through the ERDF. The research of Miguel
Flores has been partially supported by Grant PII-DM-002-2016 of Escuela Politécnica Nacional of Ecuador.
Acknowledgments: The authors strongly thank, on the one hand, Fridama, Σqus, and Nerxus companies, and on
the other hand CITIC, Campus Industrial and MODES group their valuable help and support.
Conflicts of Interest: The authors declare no conflict of interest. The founding sponsors had no role in the design
of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the
decision to publish the results.

References
. Lu, C.W.; Reynolds, M.R, Jr. EWMA control charts for monitoring the mean of autocorrelated processes.
J. Qual. Technol. 1999, 31, 166. [CrossRef]
. Alwan, L.C.; Roberts, H.V. Time-series modeling for statistical process control. J. Bus. Econ. Stat. 1988,
6, 87–95.
. Qiu, P.; Zou, C.; Wang, Z. Nonparametric profile monitoring by mixed effects modeling. Technometrics 2010,
52, 265–277. [CrossRef]
. Shiau, J.J.; Huang, H.L.; Lin, S.H.; Tsai, M.Y. Monitoring nonlinear profiles with random effects by
nonparametric regression. Commun. Stat. Methods 2009, 38, 1664–1679. [CrossRef]
. Noorossana, R.; Saghaei, A.; Amiri, A. Statistical Analysis of Profile Monitoring; John Wiley & Sons: Hoboken,
NJ, USA, 2011; Volume 865.
. Chipman, H.; MacKay, R.; Steiner, S. Comment on Nonparametric profile monitoring by mixed effects
modeling. Technometrics 2010, 52, 280–283. [CrossRef]
. Lavin, A.; Ahmad, S. Evaluating Real-Time Anomaly Detection Algorithms–The Numenta Anomaly
Benchmark. In Proceedings of the 2015 IEEE 14th International Conference on Machine Learning and
Applications (ICMLA), Miami, FL, USA, 9–11 December 2015; pp. 38–44.
. Kroll, B.; Schaffranek, D.; Schriegel, S.; Niggemann, O. System modeling based on machine learning
for anomaly detection and predictive maintenance in industrial plants. In Proceedings of the 2014 IEEE
Emerging Technology and Factory Automation (ETFA), Barcelona, Spain, 16–19 September 2014; pp. 1–7.
. Ahmad, S.; Lavin, A.; Purdy, S.; Agha, Z. Unsupervised real-time anomaly detection for streaming data.
Neurocomputing 2017, 262, 134–147. [CrossRef]
. Meshram, A.; Haas, C. Anomaly detection in industrial networks using machine learning: A roadmap.
In Machine Learning for Cyber Physical Systems; Springer: Berlin/Heidelberg, Germany, 2017; pp. 65–72.
. Basu, S.; Meckesheimer, M. Automatic outlier detection for time series: An application to sensor data.
Knowl. Inf. Syst. 2007, 11, 137–154. [CrossRef]
. Mosallam, A.; Medjaher, K.; Zerhouni, N. Nonparametric time series modelling for industrial prognostics
and health management. Int. J. Adv. Manuf. Technol. 2013, 69, 1685–1699. [CrossRef]
. Talagala, P.D.; Hyndman, R.J.; Smith-Miles, K.; Kandanaarachchi, S.; Muñoz, M.A. Anomaly detection in
streaming nonstationary temporal data. J. Comput. Graph. Stat. 2019, pp. 1–28. [CrossRef]

134
Mathematics 2020, 8, 58

. Hyndman, R.J.; Liu, X.A.; Pinson, P. Visualizing big energy data: Solutions for this crucial component of
data analysis. IEEE Power Energy Mag. 2018, 16, 18–25. [CrossRef]
. Golshan, M.; MacGregor, J.F.; Bruwer, M.J.; Mhaskar, P. Latent Variable Model Predictive Control (LV-MPC)
for trajectory tracking in batch processes. J. Process Control 2010, 20, 538–550. [CrossRef]
. Kourti, T.; MacGregor, J.F. Multivariate SPC methods for process and product monitoring. J. Qual. Technol.
1996, 28, 409–428. [CrossRef]
. Ferrer-Riquelme, A. 1.04-Statistical Control of Measures and Processes. In Comprehensive Chemometrics;
Elsevier: Amsterdam, The Netherlands, 2009; pp. 97–126.
. Colosimo, B.M.; Pacella, M. A comparison study of control charts for statistical monitoring of functional
data. Int. J. Prod. Res. 2010, 48, 1575–1601. [CrossRef]
. Megahed, F.; Jones-Farmer, L. Statistical Perspectives on “Big Data”. In Frontiers in Statistical Quality Control
11; Springer: Berlin, Germany, 2015; pp. 29–47.
. Woodall, W.; Montgomery, D. Some current directions in the theory and application of statistical process
monitoring. J. Qual. Technol. 2014, 46, 78. [CrossRef]
. Sheu, S.H.; Ouyoung, C.W.; Hsu, T.S. Phase II statistical process control for functional data. J. Stat.
Comput. Simul. 2013, 83, 2144–2159. [CrossRef]
. García, D.R. Cartas de Control Para Datos Funcionales. Master’s Thesis, Centro de Investigación en
Matemáticas (CIMAT), Guanajuato, Mexico, 2011.
. Rodrigo, O.P. Monitoreo de Datos Funcionales. Master’s Thesis, Centro de Investigación en Matemáticas
(CIMAT), Guanajuato, Mexico, 2013.
. Flores, M.; Tarrio-Saavedra, J.; Fernandez-Casal, R.; Naya, S. Functional extensions of Mandel’s h and
k statistics for outlier detection in interlaboratory studies. Chemom. Intell. Lab. Syst. 2018, 176, 134–148.
[CrossRef]
. Flores, M.; Fernández-Casal, R.; Naya, S.; Tarrío-Saavedra, J.; Bossano, R. ILS: An R package for statistical
analysis in Interlaboratory Studies. Chemom. Intell. Lab. Syst. 2018, 181, 11–20. [CrossRef]
. Flores, M.; Naya, S.; Tarrío-Saavedra, J.; Fernández-Casal, R. Functional data analysis approach of
Mandel’s h and k statistics in Interlaboratory Studies. In Functional Statistics and Related Fields; Springer:
Berlin/Heidelberg, Germany, 2017; pp. 123–130.
. Liu, R.Y. Control Charts for Multivariate Processes. J. Am. Stat. Assoc. 1995, 90, 1380–1387. [CrossRef]
. Liu, R.Y.; Tang, J. Control charts for dependent and independent measurements based on bootstrap methods.
J. Am. Stat. Assoc. 1996, 91, 1694–1700. [CrossRef]
. Liu, R. On a notion of data depth based on random simplices. Ann. Stat. 1990, 18, 405–414. [CrossRef]
. Chatterjee, S.; Qiu, P. Distribution-free cumulative sum control charts using bootstrap-based control limits.
Ann. Appl. Stat. 2009, 3, 349–369. [CrossRef]
. Jones, L.; Woodall, W. The performance of bootstrap control charts. J. Qual. Technol. 1998, 30, 362. [CrossRef]
. Chakraborti, S. Nonparametric (Distribution-Free) Quality Control Charts. In Encyclopedia of Statistical
Sciences; John Wiley & Sons, Inc.: Hoboken, NJ, USA, 2011.
. Qiu, P.; Li, Z. On nonparametric statistical process control of univariate processes. Technometrics 2011,
53, 390–405. [CrossRef]
. Barbeito, I.H.; Zaragoza, S.; Tarrío-Saavedra, J.; Naya, S. Assessing thermal comfort and energy efficiency in
buildings by statistical quality control for autocorrelated data. Appl. Energy 2017, 190, 1–17. [CrossRef]
. Montgomery, D. Introduction to Statistical Quality Control.; John Wiley & Sons: New York, NY, USA, 2009.
. Ledolter, J.; Bisgaard, S. Challenges in constructing time series models from process data. Qual. Reliab. Eng.
Int. 2011, 27, 165–178. [CrossRef]
. Box, G.E.; Luceño, A.; Paniagua-Quinones, M.D.C. Statistical Control by Monitoring and Adjustment;
John Wiley & Sons: Hoboken, NJ, USA, 2011; Volume 700.
. Del Castillo, E. Statistical process adjustment: A brief retrospective, current status, and some opportunities
for further work. Stat. Neerl. 2006, 60, 309–326. [CrossRef]
. Knoth, S.; Schmid, W. Control charts for time series: A review. Front. Stat. Qual. Control 2004, 7, 210–236.
. Prajapati, D.; Singh, S. Control charts for monitoring the autocorrelated process parameters: a literature
review. Int. J. Prod. Qual. Manag. 2012, 10, 207–249. [CrossRef]
. Tsai, T.R.; Hsieh, Y.W. Simulated Shewhart control chart for monitoring variance components. Int. J. Reliab.
Qual. Saf. Eng. 2009, 16, 1–22. [CrossRef]

135
Mathematics 2020, 8, 58

. Jones-Farmer, L.; Woodall, W.; Steiner, S.; Champ, C. An overview of phase I analysis for process
improvement and monitoring. J. Qual. Technol. 2014, 46, 265. [CrossRef]
. Grasso, M.; Colosimo, B.M.; Tsung, F. A phase I multi-modelling approach for profile monitoring of signal
data. Int. J. Prod. Res. 2017, 55, 4354–4377. [CrossRef]
. Febrero-Bande, M.; Galeano, P.; González-Manteiga, W. Outlier detection in functional data by depth
measures, with application to identify abnormal NOx levels. Environmetrics 2007, 19, 331–345. [CrossRef]
. Kazemzadeh, R.B.; Noorossana, R.; Ayoubi, M. Change point estimation of multivariate linear profiles under
linear drift. Commun. Stat.-Simul. Comput. 2015, 44, 1570–1599. [CrossRef]
. Woodall, W.H.; Spitzner, D.J.; Montgomery, D.C.; Gupta, S. Using control charts to monitor process and
product quality profiles. J. Qual. Technol. 2004, 36, 309–320. [CrossRef]
. Lanhede, D. Statistical Process Control: Evaluation and Implementation of Methods for Statistical Process
Control at GE Healthcare. Master’s Thesis, Umeå University, Umeå, Sweden, 2015.
. Flores Sánchez, M. Nuevas Aportaciones del Análisis de Datos Funcionales en el Control Estadístico de
Procesos. Ph.D. Thesis, Universidade da Coruña, A Coruña, Spain, 2019.
. Febrero-Bande, M.; Oviedo de la Fuente, M. Statistical computing in functional data analysis: The R package
fda.usc. J. Stat. Softw. 2012, 51, 1–28. [CrossRef]
. Fraiman, R.; Muniz, G. Trimmed means for functional data. Test 2001, 10, 419–440. [CrossRef]
. López-Pintado, S.; Romo, J. On the concept of depth for functional data. J. Am. Stat. Assoc. 2009, 104, 718–734.
[CrossRef]
. Cuevas, A.; Febrero, M.; Fraiman, R. Robust estimation and classification for functional data via
projection-based depth notions. Comput. Stat. 2007, 22, 481–496. [CrossRef]
. Cuesta-Albertos, J.A.; Nieto-Reyes, A. The random Tukey depth. Comput. Stat. Data Anal. 2008, 52, 4979–4988.
[CrossRef]
. Flores, M.; Fernández-Casal, R.; Naya, S.; Tarrío-Saavedra, J. qcr: Quality Control Review. R Package
Version 1.0. 2016. Available online: https://cran.r-project.org/web/packages/qcr/index.html (accessed on
26 December 2019).
. Hyndman, R.J.; Shang, H.L. Rainbow plots, bagplots, and boxplots for functional data. J. Comput. Graph.
Stat. 2010, 19, 29–45. [CrossRef]
. Raña, P. Pointwise Forecast, Confidence and Prediction Intervals in Electricity Demand and Price. Ph.D.
Thesis, Universidade da Coruña, A Coruña, Spain, 2016.
. Colosimo, B.M.; Pacella, M. On the use of principal component analysis to identify systematic patterns in
roundness profiles. Qual. Reliab. Eng. Int. 2007, 23, 707–725. [CrossRef]

c 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).

136
 MDPI
St. Alban-Anlage 66
4052 Basel
Switzerland
Tel. +41 61 683 77 34
Fax +41 61 302 89 18
www.mdpi.com

Mathematics Editorial Office

E-mail: [email protected]
www.mdpi.com/journal/mathematics
MDPI
St. Alban-Anlage 66
4052 Basel
Switzerland
Tel: +41 61 683 77 34
Fax: +41 61 302 89 18
www.mdpi.com ISBN 978-3-03943-964-5