Journal of the Taiwan Institute of Chemical Engineers 156 (2024) 105292

Contents lists available at ScienceDirect

Journal of the Taiwan Institute of Chemical Engineers

journal homepage: www.journals.elsevier.com/journal-of-the-taiwan-institute-of-chemical-engineers

Enhanced dynamic latent variable analysis for dynamic process monitoring

Xinrui Wang, Hongbo Shi *, Bing Song, Yang Tao, Shuai Tan
The Key Laboratory of Smart Manufacturing in Energy Chemical Process, Ministry of Education, East China University of Science and Technology, Meilong Road 130,
Shanghai, 200237, China

A R T I C L E I N F O A B S T R A C T

Keywords: Background: Process dynamic, also known as temporal correlation, is widespread in industrial processes and can
Temporal correlation greatly affect process monitoring results. In dynamic process monitoring, dynamic latent variable (DLV) mainly
Process monitoring considers the autocorrelation and cross-correlation of variables, while slow feature analysis (SFA) only considers
Dynamic latent variable
the varying speed of variables. Complex dynamic information needs to be fully considered.
Varying speed
Methods: This paper proposes an enhanced dynamic latent variable (EDLV) analysis. First, EDLV focuses on both
the varying speed and correlation of variables when extracting dynamic latent variables. Therefore, The pro­
posed method achieves the distinction between the normal change of operating conditions and the occurrence of
faults. Second, the process data is broken into dynamic and static subspaces for monitoring respectively, which
benefits the accurate detection of different faults.
Significant Findings: Tennessee Eastman (TE) process and three-phase flow facility are used to verify the effec­
tiveness of the proposed method. It is proved that EDLV can divide dynamic process more reasonably and obtain
better detection results.

1. Introduction features by augmenting the data using time windows. Based on the same
idea, DPLS [11] and DICA [12] were also widely used. However, these
In modern industrial processes, the requirements for operational dynamic methods still have shortcomings. First, as the lag increases, the
safety and product quality are constantly increasing as the scale of matrix dimensionality and computational complexity also increase.
production grows. Therefore, process monitoring has received more Second, these methods mix static and dynamic information, which may
attention in recent years. Process monitoring techniques can generally not identify different process statuses usefully [13].
be classified into three categories: model-based methods, knowledge- In order to avoid the shortcomings, Li et al. [14] proposed dynamic
based methods, and data-driven methods [1]. The vast amount of data latent variable (DLV). By constructing a weighted sum of multiple lag­
stored in modern industrial production provides the basis for ged latent variables, DLV considers the dynamic autocorrelation and
data-driven modeling and monitoring techniques [2,3]. Therefore, cross-correlations within the data, and finally decomposes the process
multivariate statistical process monitoring (MSPM) has grown data into a dynamic subspace and a static subspace for process moni­
exponentially. toring. Based on this, Dong and Qin proposed dynamic-inner PCA
The most commonly multivariate statistical methods include prin­ (DiPCA) [15] and dynamic-inner CCA (DiCCA) [16]. One drawback of
cipal component analysis (PCA) [4], partial least squares (PLS) [5], DiPCA is that it doesn’t maximize correlation, but rather the covariance
canonical correlation analysis (CCA) [6], independent component between the latent variables and their predictions. These methods model
analysis (ICA) [7] et al. These methods assume that the process is static the dynamics in the latent domain through a lagged or time-shifting
where the data in the current sample time is independent from the data technique and have received a lot of attention in recent years [17].
in previous moment [8]. Since the measurement of modern industrial Chen et al. [18] proposed a hierarchical dynamic monitoring method
plants can be highly dependent on time, it is necessary to consider the combined with DiPCA for nonlinear dynamic processes. Dong et al. [19]
time-serial correlated characteristic in the industrial data [9]. To better proposed numerically efficient algorithms using autoregressive inte­
focus on the dynamic information, Ku et al. [10] first proposed dynamic grated moving average (ARIMA) models to improve the computational
principal component analysis (DPCA) which can extract dynamic efficiency of DiCCA models.

* Corresponding author.
E-mail address: [email protected] (H. Shi).

https://doi.org/10.1016/j.jtice.2023.105292
Received 20 July 2023; Received in revised form 12 November 2023; Accepted 8 December 2023
Available online 1 February 2024
1876-1070/© 2023 Taiwan Institute of Chemical Engineers. Published by Elsevier B.V. All rights reserved.
X. Wang et al. Journal of the Taiwan Institute of Chemical Engineers 156 (2024) 105292

Algorithm 1 occurrence of faults to improve the monitoring effect. To achieve the

EDLV Algorithm. above objectives, an enhanced dynamic latent variable (EDLV) analysis
1) Perform the normalization operation on the input matrix X, and setXi = X, i = 1, 2, is proposed. The main contributions are as follows.
…, a. a is the number of extracted dynamic latent variables.
2) Initialize wi as a unit vector. 1. An enhanced dynamic latent variable (EDLV) analysis for dynamic
3) Perform eigenvalue decomposition on Eq. (23) and set the eigenvector
process monitoring is proposed. By defining a new objective func­
corresponding to the largest eigenvalue to βi, and calculate
βi = βi /‖ βi ‖ tion, both temporal correlation and process variation speed are
4) Perform eigenvalue decomposition on Eq. (24) and set the eigenvector considered.
corresponding to the largest eigenvalue to wi, and calculate 2. By extracting dynamic latent variables more comprehensively, the
wi = wi /‖ wi ‖ monitoring strategy achieves a successful distinction between faults
Then verify whether wi converges. If it does not, return to step 3. The convergence
and normal changes in operating conditions.
condition is experimentally set as
iterative-error = |wi − wi − 1| < 10− 7 3. The proposed algorithm monitors dynamic and static subspaces
iterative-nums > 10000 separately, which contributes to accurate process monitoring.
5) Calculate the latent variable ti and the loading vector pi according to Eqs. (27) and
(28).
The rest of this work is organized as follows. In Section 2, slow
6) Set Xi+1 = Xi − ti pTi and return to step 2 until i > a.
feature analysis and dynamic latent variable algorithms are briefly
introduced. Then, in Section 3, the EDLV is described in detail, as well as
the parameter selection strategy. In Section 4 and Section 5, the pro­
posed algorithm is applied to TE process and three-phase flow facility,
and compared with other methods to verify the effectiveness of the
proposed algorithm. Finally, the conclusions of this work are given in
Section 6.

2. Preliminaries

2.1. Dynamic latent variable

In dynamic processes, current information can be predicted from

past information [27]. Therefore, the dynamic VAR model could be built
as follows
tk = β1 tk− 1 + ⋯ + βs tk− s + rk (1)

where β = [ β1 , β2 , ⋯, βs ]T denotes the weight coefficients of the

latent variable tk at time k, s is the dynamic order of the model, and rk is
the residual representing white noise.
In addition, the latent variable tk can be obtained as follows

tk = xTk w (2)
Fig. 1. The value of γ of different value of μ.
where w is the weight vector, xk is the original variables at time k, k = 1,
2, …, N + s. And the predicted values of the dynamic-inner model can be
In addition, these methods cannot differentiate normal shifts be­
obtained from Eqs.(1) and (2) as
tween different operating conditions from real faults [20]. In recent
years, slow feature analysis (SFA) [21–23] was proposed that it can ̂t k = xTk− 1 wβ1 + ⋯ + xTk− s wβs
measure both the change in latent variables and the varying speed, [ ] (3)
= xTk− 1 ⋯ xTk− s (β ⊗ w)
which is generally related to the occurrence of faults. Therefore, SFA can
determine whether a fault has occurred or a normal change in the where β⊗w is the Kronecker product. Based on the predicted values, thus
operating conditions by measuring the varying speed. On this basis, yielding the objective function of DLV algorithm as
Zhao et al. [24] proposed a full-condition monitoring strategy based on
CA and SFA to distinguish between changes in operating conditions and ̂T T
max T ̂ (4)
actual faults by examining deviations from the smooth relationship.
Wang et al. [25] proposed a robust SFA (RSFA) based on M-estimation, where T
̂ is the matrix form of ̂t k .
which can eliminate steady-state and dynamic anomalies caused by
outliers and improve the robustness of SFA. However, under closed-loop 2.2. Slow feature analysis
control, the process not only has the characteristics of slow varying, but
also has serial correlations [26]. The extracted features should contain SFA aims to find an input-output function g(x) = [g1(x),g2(x),⋅⋅⋅,
the most information in the data, so the autocorrelation and gm(x)] that, after inputting the original variables x(t) = [x1(t),x2(t),⋅⋅⋅,
cross-correlation of the residual parts will be as small as possible. The xm(t)], makes the output variables s(t) = [s1(t),s2(t),⋅⋅⋅, sm(t)] change as
SFA simply constrains the varying speed, that is, first-order autocorre­ slowly as possible, but contains the main information in the original
lation. Therefore, the residual features extracted by the SFA may be data. These extracted latent variables are called slow features. The
autocorrelated and may contain useful data information. objective function for SFA is set as follows
In conclusion, the realization of process accurate monitoring of dy­ 〈 〉
namic processes lies in two aspects. First, extract the dynamic infor­ min ṡ2i (t) (5)
gi (⋅)
mation and separate the dynamic subspace from the static subspace,
which helps to achieve accurate monitoring in both spaces. Second, to Three constraints are included as
distinguish the normal change of operating conditions and the 〈si (t)〉 = 0 (6)

2
X. Wang et al. Journal of the Taiwan Institute of Chemical Engineers 156 (2024) 105292

Fig. 2. Autocorrelation and cross-correlation of tk of TEP data.

〈 〉
s2i (t) = 1 (7) Zs = [X1 , X2 , ⋯, Xs+1 ] (13)
〈 〉 Substitute Eq.(13) into Eqs.(3) and (4), the first objective function we
= j, si (t)sj (t) = 0
∀i ∕ (8)
need can be obtained as follows
where ṡi denotes the first-order derivative of slow features si and 〈 ⋅ 〉t maxJ1 = (β ⊗ w)T ZsT Zs (β ⊗ w)
denotes the time average.
w,β (14)
The objective function in Eq.(5) requires that slow features change at s.t. ‖ w ‖= 1, ‖ β ‖= 1
the minimum rate. Eqs.(6) and (7) requires that each slow feature has
zero mean and unit variance. Eq.(8) ensures that different slow features where T̂ = Zs (β ⊗ w) denotes the predicted dynamic latent variable.
are uncorrelated with each other and can retain different information. For a more comprehensive description of dynamic features, in
Slow features are combined by the original variables as follows addition to considering the autocorrelation, it is also necessary to
consider their rate of change, thus leading to the second objective
S = WX (9) function as

where W = [w1,w2,⋅⋅⋅, wm]T denotes the weight matrix.

T
minJ2 = ̂t˙ .i ̂t˙ i (15)

3. The proposed method Since industrial processes are sampled with mostly discrete data, the
derivative of the latent variable ̂t i with respect to time is given as
3.1. Enhanced dynamic latent variable
̂t˙ i(k+1) = ̂t i(k+1) − ̂t ik = Zi(k+1) (βi ⊗ wi ) − Zik (βi ⊗ wi )
Given the input matrix X ∈ Rn × m, where n is the number of samples = (Zi(k+1) − Zik )(βi ⊗ wi ) (16)
and m is the number of variables. Define it as follows
where ̂t ik denotes the value of the i-th predicted dynamic latent variable
X = [x1 , x2 , ⋯, xN+s ]T (10) at the time of k. Thus, Eq.(15) can be rewritten as follows

where s is the dynamic order of the model and maxw,β J2

T
= (β ⊗ w)T Ż s Ż s (β ⊗ w)
(17)
n=N+s (11) s.t. ‖ w ‖ = 1, ‖ β ‖= 1

The data matrix is then constructed as follows In extracting dynamic latent variables, autocorrelation and varying
speed should be considered to enable a comprehensive description of the
Xi = [xi , xi+1 , ⋯, xN+i− 1 ]T , i = 1, 2, ⋯, s + 1 (12) dynamic process. The first objective of EDLV is the same as that of DLV,

3
X. Wang et al. Journal of the Taiwan Institute of Chemical Engineers 156 (2024) 105292

Fig. 3. Autocorrelation and cross-correlation of Xe of TEP data.

which is to maximize the correlation of latent variables, such that latent

∂L [ ]
variables with high temporal correlation can be viewed as representing (21)
T
= 2(Is ⊗ w)T μZsT Zs − (1 − μ)Ż s Ż s (Is ⊗ w)β − 2λβ β = 0
∂β
dynamic relationships. The second objective is the varying speed of the
process data, that is, the first-order autocorrelation of the variables. [ ]
∂L
(22)
T
Variables that change more slowly and are more correlated can repre­ = 2(β ⊗ Im )T μZsT Zs − (1 − μ)Ż s Ż s (β ⊗ Im )w − 2λw w = 0
∂w
sent internal changes in the process [28]. The goal of limiting varying
speed is to isolate latent variables from rapidly changing noise to avoid The above equation can be organized to
noise overwriting abnormal dynamic behavior [13]. Therefore, [ ]
(23)
T
(Is ⊗ w)T μZsT Zs − (1 − μ)Ż s Ż s (Is ⊗ w)β = λβ β
combining the two objectives of J1 and J2, the objective function of
EDLV is given as [ ]
(24)
T
(β ⊗ Im )T μZsT Zs − (1 − μ)Ż s Ż s (β ⊗ Im )w = λw w
[ T ]
max(β ⊗ w)T μZsT Zs − (1 − μ)Ż s Ż s (β ⊗ w)
w,β (18) From the constraint in Eq.(18), the lengths of β and w are both unit.
s.t. ‖ w ‖= 1, ‖ β ‖= 1 Therefore, multiplying both sides of Eqs.(23) and (24) by βT and wT
respectively, we can get
where μ ∈ [0, 1] is the weighting factor to measure the relative impor­ [ ]
(25)
T
tance of autocorrelation and varying speed. These constraints are (Is ⊗ w)T μZsT Zs − (1 − μ)Ż s Ż s (Is ⊗ w) = λβ = J
designed to constrain the range of latent variables, so as not to affect the
[ ]
result too much. The above problem can be solved by applying the (26)
T
(β ⊗ Im )T μZsT Zs − (1 − μ)Ż s Ż s (β ⊗ Im ) = λw = J
Lagrange multipliers. Define Lagrange functions as follows
The above results show that J is equal to the values of λβ and λw.
[ ]
However, from Eqs.(23) and (24), it can be seen that β and w are coupled
T
L = (β ⊗ w)T μZsT Zs − (1 − μ)Ż s Ż s (β ⊗ w)
(19)
( ) (
+λw 1 − wT w + λβ 1 − βT β
) with each other, so they cannot be solved separately and need to be
solved in an iterative manner. After solving w, the dynamic latent var­
In addition, there are properties of the Kronecker product as follows iables and loading vectors can be obtained as follows
( )
β ⊗ w = Iq ⊗ w β = (β ⊗ Im )w (20) t = Xw (27)
/
where Is and Im are the unit vectors of dimensions s and m. According to p = X T t tT t (28)
Eq.(19) and Eq.(20), the partial derivatives of β and w, can be obtained
as follows In summary, the EDLV algorithm is calculated as follows
Thus, we can get W = [w1,w2,⋅⋅⋅, wa], Td = [t1,t2,⋅⋅⋅, ta], Pd = [p1,p2,⋅⋅⋅,

4
X. Wang et al. Journal of the Taiwan Institute of Chemical Engineers 156 (2024) 105292

Fig. 4. Fault detection results of DPCA(a)-(b), SFA (c)-(f), EDLV(g)-(j) and DiCCA (k)-(m) for fault 5 of TEP.

pa].
X = Td PTd + Xe
(30)
= Td PTd + Ts PTs + Es
3.2. EDLV model
where Ts and Ps are the score matrix and the loading matrix in static
subspace, and Es represents the residual part with small variance.
After calculating the dynamic latent variables, EDLV can divide the
original variables into dynamic subspace and static subspace, yielding Then, the monitoring statistics S2d and Td2 are designed for the dy­
Algorithm 1. namic subspace. S2d is used to monitor the change rate of dynamic latent
variables, and Td2 is used to monitor the change in magnitude of dynamic
Xe = X − Td PTd (29)
latent variables as
{ 2
where Xe denotes the remaining static subspace. In order to improve the Td = Td Ω−d 1 TdT
detection rate of the static part, PCA is used for its further decomposi­ T (31)
Sd2 = Ṫ d Ω−dd1 Ṫ d
tion. The final decomposition of the original variable X can be expressed
as

5
X. Wang et al. Journal of the Taiwan Institute of Chemical Engineers 156 (2024) 105292

Fig. 5. Fault detection results of DPCA(a)-(b), SFA (c)-(f), EDLV(g)-(j) and DiCCA (k)-(m) for fault 10 of TEP.

TdT Td
T
Ṫ d Ṫd
0.99 is chosen. For the statistics Td2 , S2d , Ts2 and SPEs, the corresponding
where Ωd = n− 1 , Ωdd = n− 2s− 1. control limits are limd, limdd, lims and limQ.
The monitoring statistics Ts2
and SPEs are designed separately for the The final monitoring strategy is summarized as follows
static part. Ts2 is used to monitor the part with larger variance in the
static subspace and SPEs is used to monitor the part with smaller vari­ (1) If S2d > limdd and Td2 > limd , it means that a fault has occurred and
ance in the static subspace. the fault has affected both the dynamic relationship and the speed
{ between variables.
Ts2 = Ts Ω−s 1 TsT
(32) (2) If S2d < limdd and Td2 > limd , it means a slow-varying fault
SPEs = Es EsT
occurred and gradually evolve into serious faults over time. The
process may be affected subsequently.
TT T
where Ωs = n−s 1s . (3) If S2d first alarms and then returns to normal while Td2 > limd , it
For each of the above four monitoring statistics, the kernel density means that the control system compensates for the disturbance
estimation (KDE) method is used to calculate the control limits and α = that occurred and brings the system to a new equilibrium.

6
X. Wang et al. Journal of the Taiwan Institute of Chemical Engineers 156 (2024) 105292

Table 1
FDR(%) of DPCA, SFA, DiCCA and EDLV of TEP data.
DPCA SFA DiCCA EDLV
2 2
Fault T SPE T SPE S2d S2e T2v T2s SPEs S2d Td2 Ts2 SPEs

1 100 100 89.38 99.75 32.67 1.88 36.59 99.37 100 24.47 100 100 100
2 98.62 98.49 96.88 98.13 7.26 1.38 86.59 98.87 99.25 3.14 98.49 98.37 99.12
4 12.3 100 1 100 1.75 1.25 4.01 68.92 100 1.51 100 100 95.61
5 24.47 89.08 100 100 24.03 1.38 99.87 100 100 17.69 100 100 100
6 99.37 100 99.63 100 92.62 22.53 99.5 100 100 92.72 100 100 100
7 100 99.5 44.88 100 33.9 2.88 25.44 100 100 26.98 100 100 56.46
8 97.87 96.61 97.38 97.38 92.49 3.63 77.57 97.12 95.61 80.8 98.24 95.86 86.32
10 25.47 61.61 65.13 85.38 31.17 11.76 29.2 41.48 72.93 25.47 89.84 88.08 87.95
11 45.55 97.11 9 80.5 6.76 6.63 4.89 59.15 67.54 11.67 81.18 88.03 59.47
12 99.37 97.49 99.38 99.55 97.12 75.59 95.61 99.25 98.75 97.11 99.87 99.62 98.49
13 95.11 95.86 94.88 95.25 90.11 10.76 79.95 95.86 94.49 81.68 95.23 95.23 96.73
14 100 100 0.38 100 22.52 100 99.87 100 98.5 99.87 100 99.62 98.75
16 10.54 63.74 50.5 90.25 24.28 18.4 27.19 30.7 73.06 27.48 92.60 88.71 83.56
17 84.67 98.24 34.63 97.5 14.77 29.54 42.11 80.45 97.49 35.13 97.24 97.49 98.12
18 89.71 90.96 89.88 90.38 85.48 33.54 89.72 90.23 91.60 85.7 90.46 91.47 91.34
19 29.36 81.93 2.25 94.63 10.14 72.72 14.41 19.92 40.85 73.65 94.23 76.16 90.46
20 46.8 66.5 75.25 90.38 24.16 14.27 22.68 48.75 67.79 23.71 91.09 71.02 81.56
21 49.18 17.57 68.5 41 0.25 0.13 4.39 44.49 19.42 0.38 56.96 44.29 39.52

Fig. 6. Sketch of the three-phase flow facility.

(4) If Ts2 > lims or SPEs > limQ, it indicates a fault in the corre­ function only considers the varying speed of variables. Therefore, the
sponding static subspace of the static part. size of μ decides the effect of extracting dynamic latent variables, and
this paper defines a new index γ to help select the appropriate weighting
parameter as follows
3.3. Determination of parameters
∑a T
(βi ⊗ wi )T Ż s Ż s (βi ⊗ wi )
γ = ∑i=1 (33)
In this section, how to determine the lagged number s, the weighting a T T
i=1 (βi ⊗ wi ) Zs Zs (βi ⊗ wi )
parameter μ, and the number a of latent variables of the three param­
eters of EDLV will be discussed. where
∑a T
⊗ wi )T Żs Żs (βi ⊗ wi ) is the sum of the varying speed of all
i=1 (βi
First, for the lagged number s, the method used here was proposed by ∑
dynamic latent variables and ai=1 (βi ⊗ wi )T ZTs Zs (βi ⊗ wi ) measures the
Ku et al. [10]. Assuming that there are r1 static relations in m original
autocorrelation of all dynamic latent variables. The size of γ changes
variables X1, so there will be 2r1 static relations and r2 dynamic relations
when μ ∈ [0, 1] takes different values. And μ will be selected when γ
in [X1,X2]. When no new relations appear in (s + 1)m augmented matrix,
takes the minimum value.
it is considered that the lagged number s has covered all dynamic re­
The last parameter to be determined is the number a of dynamic
lations currently. Second, in Eq.(18), the weighting parameter μ is
latent variables. After considering autocorrelation and varying speed,
mainly used to measure the importance of autocorrelation and varying
the value of γ is calculated for each dynamic latent variable and ranked
speed, so the value of μ ranges from 0 to 1. When μ is set to 1, the
in descending order based on the size of γ− 1 as follows:
objective function only considers the autocorrelation of variables, which
is equivalent to the traditional DLV. And when μ is set to 0, the objective

7
X. Wang et al. Journal of the Taiwan Institute of Chemical Engineers 156 (2024) 105292

Fig. 7. Autocorrelation and cross-correlation of tk of three-phase flow facility.

(βi ⊗ wi )T ZsT Zs (βi ⊗ wi ) cross-correlation of the extracted dynamic latent variables tk. It contains
γ − 1 (i) = T T , i = 1, 2, ⋯A (34) a lot of dynamic information. Fig. 3 shows the autocorrelation and cross-
(βi ⊗ wi ) Ż s Ż s (βi ⊗ wi )
correlation of the remaining static part Xe. In static subspace, when the
where A is the maximum value of the number of dynamic latent vari­ lagged number is not zero, correlations should be close to zero. We can
ables. Finally, the dynamic latent variables satisfying
∑a − 1 see that a successful separation of the dynamic subspace and the static
i=1 γ (i)
∑A − 1 subspace from Figs. 2 and 3.
< 0.95 i=1 γ (i), (a < A) are retained.
All methods are tested for 18 faults in the TEP, and the results for
fault 5 and fault 10, are shown in Figs. 4 and 5.
4. TEP data
Fault 5 is a step fault in the inlet temperature of the condenser
cooling water, which in turn causes a step change in the condenser
The Tennessee Eastman Process is a simulation of an actual chemical
cooling water flow. This will lead to an increase in the flow rate and
process that has been widely used in the study of process monitoring
temperature from the condenser outlet to the vapor/liquid separator,
technology [29,30]. The control scheme of TE process mainly consists of
which increases the temperature of the separator cooling water outlet.
five parts: reactor, product condenser, gas-liquid separator, circulating
However, due to the system control, the final temperature of the sepa­
compressor, and stripper tower. In the TE process, there are 12 manip­
rator will return to the setting point.
ulated variables, 22 process variables and 19 component variables with
From Fig. 4(a)-(b), we can see that DPCA is not effective in detecting
21 types of faults. Since the component variables in the actual process
fault 5, with only 24.47 % and 89.08 % of the T2 and SPE statistics. In
are difficult to collect, 33 variables are selected as process variables for
addition, when the fault still exists, the T2 statistic drops below the
monitoring. Since faults 3, 9 and 15 are difficult to detect, the remaining
control limit, so the fault cannot be detected continuously. In Fig. 4(c)-
18 faults are selected for process monitoring in this paper. Firstly, 960
(F), both T2 and SPE of SFA are 100 %. In Fig. 4(k)-(m), although the
samples from normal operation are taken as the training set. The test set
FDR of three statistics of DiCCA are all 99.75 %, the distinction between
consists of 960 fault samples, and all faults occur at sample point 161.
the change of normal operating conditions and faults is not realized.
In order to show the advantages of EDLV, it is compared with three
Therefore, DiCCA cannot determine the control effect of the system. The
dynamic process monitoring algorithms, DPCA, SFA, and DiCCA. The
fault detection results of EDLV are shown in Fig. 4(g)-(j). The FDR of
lagged numbers are all set to s = 2 using the method mentioned above.
dynamic subspace statistic Td2 , static subspace statistics Ts2 and SPEs are
The confidence level is 0.99 to ensure the consistency of the KDE method
all 100 %. This shows that EDLV can successfully separate dynamic and
when setting the control limits. Considering that PCA is used in DPCA,
static subspaces and detect faults. In addition, S2d begins to alarm when
DiCCA and EDLV methods, the variance contribution rate of PCA is set to
95 % in order to select the appropriate principal components. By using the fault happened. After the fault is controlled, S2d also drops below the
the previous method of parameter determination, select μ = 0.74 when γ control limit. In other words, when system starts to control, EDLV can
takes the minimum value in Fig. 1. Fig. 2 shows the autocorrelation and timely detect changes in the normal operating conditions of the system.
Fault 10 is a random fault in which variables change randomly at

8
X. Wang et al. Journal of the Taiwan Institute of Chemical Engineers 156 (2024) 105292

Fig. 8. Autocorrelation and cross-correlation of Xe of three-phase flow facility.

Fig. 9. Fault detection results of EDLV for fault case 3.3 of three-phase flow facility.

Table 2
FDR(%) of DPCA, SFA, DiCCA and EDLV of three-phase flow facility.
DPCA SFA DiCCA EDLV
2 2
Fault T SPE T SPE S2d S2e T2v T2s SPEs S2d T2d Ts2 SPEs

1.1 23.37 50.5 100 76.22 13.14 17.04 76.52 64.52 74.75 10.01 100 79.56 70.38
1.2 23 42.68 71.03 80.42 3.62 5.9 75.65 58.7 77.89 2.92 33.58 80.71 67.96
1.3 19.76 37.29 49.88 50.78 10.36 12.03 50.08 40.15 50.05 6.8 29.8 60.01 48.35
3.1 98.48 100 100 99.38 51.3 78.01 99.03 92.77 99.24 64.42 100 99.14 95.76
3.2 82.24 99.91 43.69 72.36 14.43 42.44 60.05 22.57 66.91 26.07 63.28 63.66 32.42
3.3 97.64 99.68 99.23 99.55 12.82 44.11 99.25 91.41 99.51 28.06 99.58 99.3 94.93
4.1 34.65 92.19 100 89.52 76.19 80.01 88.43 57.56 91.39 66.66 100 91.61 80.83
4.2 10.6 31.12 80.87 86.1 45.82 53.65 83.84 39.92 88.5 34.72 16.89 88.57 76.77
4.3 18.23 95.27 100 85.44 55.81 61.61 88.37 42.25 89.74 42.05 100 89.64 81.07
5.1 71.78 96.91 64.06 31.62 34.91 18.95 57.78 31.62 75.49 36.97 60.45 70.75 46.86
5.2 81.57 88.35 67.33 55.68 41.75 20.16 73.42 58 90.60 43.09 62.75 86.26 62.37

9
X. Wang et al. Journal of the Taiwan Institute of Chemical Engineers 156 (2024) 105292

Table 3
FAR(%) of DPCA, SFA, DiCCA and EDLV of three-phase flow facility.
DPCA SFA DiCCA EDLV
2 2
Fault T SPE T SPE S2d S2e Tv2 Ts2 SPEs S2d T2d Ts2 SPEs

1.1 5.66 14.28 73.12 1.73 2 1.41 1.69 2.47 3.15 1.41 74.04 9.22 1.82
1.2 13.55 37.23 7.8 2.82 8.02 2.45 2.76 2.54 3.73 8.27 7.82 4.17 4.17
1.3 9.57 8.96 5.91 2.8 5.45 2.35 2.28 1.37 2.81 5.7 5.85 4.94 3.49
3.1 16.04 33.8 63.96 1.82 5.23 1.23 2.57 1.02 4.18 5.19 24.01 9.73 2.73
3.2 51.64 88.36 5.59 4.5 7.5 2.59 3.3 1.92 4.4 7.53 5.89 6.16 5.75
3.3 22.07 10.45 5.69 4.46 6.53 2.84 6.04 1.73 5.03 6.42 5.7 5.92 5.48
4.1 3.7 48.31 100 22.56 3.38 2.14 23.05 2.63 23.21 2.11 100 26.19 23.24
4.2 9.54 24.19 6.62 8 5.86 6.48 7.34 5.95 11.14 5.94 5.86 7.39 7.26
4.3 29.53 42.32 65.76 18.18 11.36 5.45 17.1 6.56 17.86 12.33 61.49 19.18 13.24
5.1 69.94 99.04 50.76 27.77 26.5 2.93 41.08 13.82 32.42 27.71 51.66 39.43 15.35
5.2 9.92 19.34 5.07 3.40 10.43 4.11 12.83 7.28 14.20 11.24 4.65 13.62 10.93

different moments. This fault occurs at the feed temperature of stream 2. separated and returned to their respective storage tanks. More detail can
The detection results of the four methods are shown in Fig. 5. In be found in [32].
Fig. 5(a)-(b), the FDR of T2 and SPE statistics of DPCA are 25.47 % and There were 23 variables collected in this process monitoring, among
61.61 %. Although DPCA describes the general change of the fault, which fault cases 1, 3, 4 and 5 were selected. Because for fault case 2, all
DPCA does not detect a fault in most cases. Fig. 5(c)-(F) shows the methods cannot get good detection results, supervised methods should
detection results of SFA for fault 10. The FDR of T2 and SPE statistics are be adopted for this fault. The 24th variable, the pressure measurement of
65.13 % and 85.38 %, respectively. It can be seen although the fault is PT417, was excluded because the 2-inch riser was only used for fault
detected most of the time in the SPE statistics, T2 are mostly below the case 6. The training data used T2 and T3 combined data sets, with a total
control limit between 380 and 550 sampling points. This may result in of 23025 samples.
missing alarms during fault detection. In Fig. 5(k)-(m), the FDR of Tv2 , Ts2 For the process monitoring of the three-phase flow facility, the above
and SPEs statistics of DiCCA are 34.63 %, 31.88 % and 78.13 %, four methods are used. The lagged numbers are all set to 2. The confi­
respectively. This indicates that DiCCA are not effective for fault 10 and dence level is 0.99 to ensure the consistency of the KDE method when
cannot effectively detect the fault in most of the time. This is caused by setting the control limits. Considering that PCA is used in DPCA, DiCCA
insufficient extraction of dynamic latent variables by DiCCA. In Fig. 5 and EDLV methods, the variance contribution rate of PCA is set to 95 %
(g)-(j), the FDR of Td2 , Ts2 and SPEs statistics of EDLV are 89.84 %, 88.08 in order to select the appropriate principal components. By using the
% and 87.95 % respectively, showing high detection rates in both dy­ previous method of parameter determination, select μ = 0.68 when γ
namic and static subspaces. In addition, the S2d statistic also detects the takes the minimum value. Figs. 7 and 8 show the autocorrelation and
changes of dynamic latent variables, representing the real changes of cross-correlation of tk and Xe. It is successfully separated into dynamic
random faults. and static subspaces.
Table 1 shows the FDR of DPCA, SFA, DiCCA and EDLV for 18 faults Fault case 3.3 is top separator input blockage. The fault was to
in the TE process. The highest FDR of each fault are signed in bold type. gradually close the valve VC404 simulating a blockage that develops
For faults 1 to 7, these fault types are all step faults, and it can be seen over time. In ball valves, when the valve position is close to the full open
from the detection results that the fluctuation of statistics is obvious. For position, the change in pressure drop is almost not significant, but when
faults that are easy to detect, the four methods can obtain satisfactory the valve position is close to the full closed position, the change in
results. For faults that are difficult to detect, such as faults 8 and 10, the pressure drop is much more sensitive. As can be seen from Fig. 9, the Td2 ,
fault type is a random variable, and it will be difficult to detect when the Ts2 and SPEs statistics EDLV increase first, and suddenly decrease when
variable changes slightly. And fault 16 is an unknown fault. The detec­ the valve is fully closed. And S2d describes the change of latent variables,
tion results of each statistic of DPCA, SFA and DiCCA are not all satis­ which first exceeds the control limit because the valve starts to be closed,
factory. For example, for fault 16, the FDR of T2 and SPE statistics of SFA and then drops because the latent variables have not changed much
are 50.5 % and 90.25 %. This shows that SFA is not accurate enough for during this period.
space division. However, EDLV considers different timing information in Tables 2 and 3 shows the FDR and FAR of DPCA, SFA, DiCCA and
the process data, and can divide the process data into dynamic and static EDLV for 11 faults in the three-phase flow facility. SFA has the lowest
subspaces respectively for monitoring. Therefore, for these faults that false alarm rates for most cases while EDLV has the highest FDR for most
are difficult to detect, the Td2 , Ts2 and SPEs statistics of EDLV all achieve cases. The FAR of DiCCA is low, but its FDR is not high either. Moreover,
high detection rates. From the point of varying speed, S2d can timely the detection results of EDLV in static subspace are better. This illus­
monitor the change of dynamic latent variables, thus enabling EDLV to trates the EDLV for dynamic and static subspace segmentation is effec­
detect the occurrence of faults. tive. Both fault cases 1 and 3 progressively close the valve, introducing
fault conditions that simulate the blockage. Fault type 4 is to gradually
5. Three-phase flow facility open the valve to simulate the leakage situation. It can be seen that all of
these faults are temporal related, and EDLV all shows high fault detec­
The three-phase flow facility provides a measured and controlled tion rates.
flow of water, oil and air to a pressurized system. Due to the changing
operational conditions, the large volume of data, and the complexity of 6. Conclusion
the rig, the three-phase flow facility is an ideal benchmark for evaluating
the performance of process monitoring methods [31]. As shown in Considering that SFA and DLV only extract unilateral temporal fea­
Fig. 6, the system consists of the feeding units, a gas and liquid tures, a new dynamic process fault detection method called enhanced
two-phase separator, respective coalescers, and a three-phase separator. dynamic latent variable (EDLV) is proposed with a new measure for
Finally the fluid mixtures are separated in the three-phase separator. The different temporal features. EDLV can extract the features of dynamic
air is returned to the atmosphere and the emulsions of oil and water are latent variables more comprehensively by focusing on the varying speed
and correlation at the same time. Then, the process data is divided into a

10
X. Wang et al. Journal of the Taiwan Institute of Chemical Engineers 156 (2024) 105292

dynamic subspace and a static subspace, and PCA algorithm is used in [9] Lan T, Tong C, Shi X, Luo L. Dynamic statistical process monitoring based on
generalized canonical variate analysis. J Taiwan Inst Chem Eng 2020;112:78–86.
the static subspace to distinguish different types of faults. By setting
[10] Ku W, Storer RH, Georgakis C. Disturbance detection and isolation by dynamic
statistic to monitor the varying speed of dynamic latent variables, EDLV principal component analysis. Chemometr Intell Lab Syst 1995;30:179–96.
can distinguish between changes in normal operating conditions and the [11] Yang J, Zhang Y, Zhou L, Zhang F, Jing Y, Huang M, Liu H. Quality-related
occurrence of faults. Finally, EDLV is applied to TE process and three- monitoring of papermaking wastewater treatment processes using dynamic
multiblock partial least squares. J Bioresour Bioprod 2022;7:73–82.
phase flow facility. By comparing DPCA, DiCCA and SFA algorithms, it [12] Huang J, Yan X. Dynamic process fault detection and diagnosis based on dynamic
is proved that EDLV can divide dynamic process more reasonably and principal component analysis, dynamic independent component analysis and
obtain better detection results. bayesian inference. Chemometr Intell Lab Syst 2015;148:115–27.
[13] Zheng J, Zhao C. Enhanced canonical variate analysis with slow feature for
dynamic process status analytics. J Process Control 2020;95:10–31.
Declaration of competing interest [14] Li G, Qin SJ, Zhou D. A new method of dynamic latent-variable modeling for
process monitoring. IEEE Transactions on Industrial Electronics 2014;61:6438–45.
[15] Dong Y, Qin SJ. A novel dynamic pca algorithm for dynamic data modeling and
The authors declare that they have no known competing financial process monitoring. J Process Control 2018;67:1–11.
interests or personal relationships that could have appeared to influence [16] Dong Y, Qin SJ. Dynamic-inner canonical correlation and causality analysis for
the work reported in this paper. high dimensional time series data. IFAC-PapersOnLine 2018;51:476–81.
[17] Fernandes NC, Rato TJ, Reis MS. Modeling in the observable or latent space? a
comparison of dynamic latent variable based monitoring methods for sensor fault
Acknowledgments detection. Chemometr Intell Lab Syst 2022;231:104684.
[18] Chen X, Zhao C. Linear and nonlinear hierarchical multivariate time delay
analytics for dynamic modeling and process monitoring. J Process Control 2021;
This work was supported by the National Natural Science Foundation
107:83–93.
of China (No. 62073141), National Key Research and Development [19] Dong Y, Liu Y, Qin SJ. Efficient dynamic latent variable analysis for high-
Program of China (2020YFC1522502, 2020YFC1522505), Shanghai dimensional time series data. IEEE Trans Industr Inform 2019;16:4068–76.
Rising-Star Program (No. 21QA1401800). The Shanghai Natural Science [20] Xu X, Ding J. Decentralized dynamic process monitoring based on manifold
regularized slow feature analysis. J Process Control 2021;98:79–91.
Foundation under Grant 22ZR1417000. [21] Shang C, Yang F, Gao X, Huang X, Suykens JA, Huang D. Concurrent monitoring of
operating condition deviations and process dynamics anomalies with slow feature
Supplementary materials analysis. AIChE Journal 2015;61:3666–82.
[22] Zhong W, Jiang C, Peng X, Li Z, Qian F. Online quality prediction of industrial
terephthalic acid hydropurification process using modified regularized slow-
Supplementary material associated with this article can be found, in feature analysis. Ind Eng Chem Res 2018;57:9604–14.
the online version, at doi:10.1016/j.jtice.2023.105292. [23] Saafan H, Zhu Q. Improved manifold sparse slow feature analysis for process
monitoring. Comput Chem Eng 2022;164:107905.
[24] Zhao C, Huang B. A full-condition monitoring method for nonstationary dynamic
References chemical processes with cointegration and slow feature analysis. AIChE J 2018;64:
1662–81.
[1] Nawaz M, Maulud AS, Zabiri H, Suleman H. Review of multiscale methods for [25] Wang J, Zhao Z, Liu F. Robust slow feature analysis for statistical process
process monitoring, with an emphasis on applications in chemical process systems. monitoring. Ind Eng Chem Res 2020;59:12504–13.
IEEE Access 2022;10:49708–24. [26] Li Q, Zhao Z, Wang L. Slowness or autocorrelation? a serial correlation feature
[2] Zhu J, Ge Z, Song Z, Gao F. Review and big data perspectives on robust data mining analysis method and its application in process monitoring. J Process Control 2023;
approaches for industrial process modeling with outliers and missing data. Annu 121:1–12.
Rev Control 2018;46:107–33. [27] Zheng J, Zhao C, Gao F. Retrospective comparison of several typical linear dynamic
[3] He Y, Shi H, Tan S, Song B, Zhu J. Multiblock temporal convolution network-based latent variable models for industrial process monitoring. Comput Chem Eng 2022;
temporal-correlated feature learning for fault diagnosis of multivariate processes. 157:107587.
J Taiwan Inst Chem Eng 2021;122:78–84. [28] Tan S, Zhou X, Shi H, Song B. Adaptive slow feature analysis-sparse autoencoder
[4] Hsu C-C, Shih P-C, Tien F-C. Integrate weighted dependence and skewness based based fault detection for time-varying processes. J Taiwan Inst Chem Eng 2023;
multiblock principal component analysis with bayesian inference for large-scale 142:104599.
process monitoring. J Taiwan Inst Chem Eng 2021;119:6–22. [29] Downs JJ, Vogel EF. A plant-wide industrial process control problem. Comput
[5] Wang G, Jiao J, Yin S. Quality-related fault detection approaches based on data Chem Eng 1993;17:245–55.
preprocessing. IFAC-PapersOnLine 2017;50:15740–7. [30] Bathelt A, Ricker NL, Jelali M. Revision of the tennessee eastman process model.
[6] Hardoon DR, Szedmak S, Shawe-Taylor J. Canonical correlation analysis: an IFAC-PapersOnLine 2015;48:309–14.
overview with application to learning methods. Neural Comput 2004;16:2639–64. [31] Zhu J, Shi H, Song B, Tao Y, Tan S. Convolutional neural network based feature
[7] Tian L, Li Z, Yan X. A novel quality-relevant fault detection method based on mica- learning for large-scale quality-related process monitoring. IEEE Trans Industr
som multi-subspace partitioning for non-gaussian industrial processes. J Taiwan Inform 2021;18:4555–65.
Inst Chem Eng 2023;143:104687. [32] Ruiz-Cárcel C, Cao Y, Mba D, Lao L, Samuel R. Statistical process monitoring of a
[8] Huang P, Tao Y, Song B, Shi H, Tan S. Tensor sequence component analysis for multiphase flow facility. Control Eng Pract 2015;42:74–88.
fault detection in dynamic process. Can J Chem Eng 2020;98:225–36.

11