Comparison of Machine Learning Approaches for

Time-series-based Quality Monitoring of Resistance

Spot Welding (RSW)

Baifan Zhou, Tim Pychynski, Markus Reischl and Ralf Mikut

Abstract In automatic manufacturing, enormous amounts of data are generated

every day. However, labeled production data useful for data analysis is difficult
to acquire. Resistance spot welding (RSW), widely applied in automobile
production, is a typical automatic manufacturing process with inhomogeneous
data structures as well as statistical and systematic dynamics. In resistance spot
welding, an electric current flows through electrodes and the materials in between.
The materials are first heated and melted, then congeal, forming what is known
as a weld nugget, joining the materials together. The nugget size is an important
quality indicator, but can only be precisely obtained by using costly destructive
methods. This paper strives to address the issue of the scarcity of labeled
data by using simulation data generated with a verified finite element model.

Baifan Zhou · Tim Pychynski

Robert Bosch GmbH
Robert-Bosch-Campus 1, 71272 Renningen, Germany
[email protected]
[email protected]
Markus Reischl · Ralf Mikut
Karlsruhe Institute of Technology (KIT), Institute for Automation and Applied Informatics (IAI)
Hermann-von-Helmholtz-Platz 1, 76344 Eggenstein-Leopoldshafen, Germany
[email protected]
[email protected]
Archives of Data Science, Series A
(Online First) DOI: 10.5445/KSP/1000087327/13
KIT Scientific Publishing ISSN 2363-9881
Vol. 5, No. 1, 2018
2 Baifan Zhou, Tim Pychynski, Markus Reischl and Ralf Mikut

Physics-based simulation enables large amounts of labeled data to be generated

with fewer limits on sensors and costs. Based on the simulation data, this paper
explores and compares multiple machine learning methods, predicts the nugget
size with a high degree of accuracy, and conducts an analysis of the influence of
feature number and amount of training data on prediction accuracy.

1 Introduction
Technological advances in computing, connecting and sensoring enables a
new wave of productivity improvement. We are currently in the era of Indus-
try 4.0, a term first coined by Kagermann et al (2011), which represents the
fourth technological revolution. In manufacturing, data mining, as one of the
key technologies of Industry 4.0 (Rüßmann et al, 2015), has created new
intelligent tools for automatic information extraction and knowledge discovery
(Wang, 2007; Nagorny et al, 2017). Automated analysis of sensor data from
industrial environments is important for practices such as identifying root causes
of exceptions or making correct decisions (Zhu et al, 2011).
Although enormous amounts of data are generated every day in automatic
manufacturing, the acquired data is not always relevant (such as missing labels)
and thus suitable for data analysis (Wuest et al, 2016). Resistance Spot Weld-
ing (RSW) is a typical fully automated manufacturing process, widely applied
in the automobile industry, with 3000 to 6000 welding spots per car chassis. In
RSW, the two electrode caps of the welding gun (see Figure 1) press two or three
worksheets between the electrodes with force. An electric current then flows
from one electrode, through the worksheets, to the other electrode, generating a
substantial amount of heat as a result of electric resistance. The materials in a
small area between the two worksheets, known as the welding spot, will melt,
and form a weld nugget connecting the worksheets. The electrode caps directly
touching the worksheets wear out easily due to high thermo-mechanical loads
and oxidation, and need to be changed on a regular basis.

Figure 1: An example of a welding

gun (𝐼 = electric current, 𝐷 = welding
cross-section diameter).
Comparison of ML Approaches for Time-series-based Quality Monitoring 3

The quality of the welding spots is typically quantified by the weld nugget
diameter as defined in the international standards (ISO, 2004) and German
standard (DVS, 2016). The welding quality is then determined as accepted
or rejected according to specific tolerance bands. This work strives to predict
the spot diameter as a numeric value rather than a classification as accepted
or rejected, because the exact diameter values are of great interest for process
experts, who can potentially gain better insights by analyzing which input factors
would influence the diameter values to what extent. After predicting the numeric
values of spot diameters, a classification can still be made according to different
welding conditions and user-defined tolerance bands.

Figure 2: Welded worksheets torn apart.

However, the nugget diameter is difficult to measure, and the volume of data
with nugget diameter (or labeled data in data science) is low. The common
practice is to tear the welded worksheets apart and measure the nugget directly,
for example using a micrometer (see Figure 2). Performing quality control in
the automobile industry usually requires the destruction of several car chassis,
which is extremely expensive.
Many previous studies have adopted the approach of data-driven models.
Leveraging the power of data science, data-driven models can assess the nugget
diameters based on the recorded process data. These models are beneficial as
they can potentially perform quality control for every welding spot reliably,
ensuring process capability and reducing costs for quality control. Many authors
treated the problem as classification (Martín et al, 2007; Sun et al, 2017),
while others evaluated the predicted diameters as a regression problem (Ruisz
et al, 2007; El Ouafi et al, 2010; Wan et al, 2016). Their dataset size ranges
from approximately 10 (Cho and Rhee, 2004) to around 3000 labeled welding
4 Baifan Zhou, Tim Pychynski, Markus Reischl and Ralf Mikut

spots (Boersch et al, 2016). Various methods were explored, including Random
Forests (Sumesh et al, 2015), Neural Networks (Afshari et al, 2014), etc. Most of
them used process curves, such as electrode displacement (Li et al, 2012), while
others also used scalar process parameters, such as coating (Yu, 2015).
Most of the previous studies have mentioned that their data is collected
from laboratory experiments. It is questionable whether models developed from
laboratory data are applicable in real production, as the welding conditions (such
as cooling time and wear) are usually different. Apart from the labeled data
amount problem, other difficulties exist in building successful data-driven models.
This paper summarizes these difficulties systematically as three challenges for
data analysis in RSW and in automatic manufacturing:

• Challenge 1 is the limit on labeled data amount, which results from the
difficulty in collecting labeled data in production or even in the laboratory.
• Challenge 2 is the limit on features. Additional sensors, like electrode
displacement, or extra measurements, such as actual worksheet thickness,
may be necessary for reliable quality prediction in manufacturing pro-
cesses. However, more sensors mean higher costs, and an increased risk of
expensive machine stops due to sensor failures. When deciding whether
to install an extra sensor, its benefit needs to outweigh the disadvantages.
It is, therefore, important to understand which sensors or measurements
(or features in data science) are necessary and to quantify their benefit to
justify the higher costs.
• Challenge 3 is the limit on coverage of relevant situations. Quality failures
are very unusual in manufacturing data, because the welding quality
is good under normal conditions, which is the most frequent case in
production. Moreover, quality failures may be of different types, making it
even more difficult to collect sufficient data for reliable quality prediction.

This work will address the first two challenges. They are interpreted as three
questions in collecting costly labeled data:

• Question 1: “How much labeled data to collect first?” (Challenge 1)

• Question 2: “Which sensors should be installed first?” (Challenge 2)
• Question 3: “Which precision level should the sensors and measurements
of welding spot diameters have?” (Challenge 2)
Comparison of ML Approaches for Time-series-based Quality Monitoring 5

The authors of this paper suggest the use of physics-based simulation data
generated with a verified Finite Element Method (FEM) model (Schif, 2017)
to help overcome these three challenges. An established simulation model has
four advantages:

1. Each simulation run produces one labeled data point. Running a lot of
simulations can thus offer a large amount of labeled data.
2. There is almost no limit for obtaining features that are costly or difficult
to realize through sensor measurements in laboratory or production.
3. Using astutely designed scenarios, rare situations in production can be
studied in detail.
4. There is no measurement sensor precision issue in data generated from a
simulation model.

This work gives a short introduction to the FEM simulation to cover some points
that the readers may be interested in. The FEM simulation models mechanical
effects (elastic-plastic part deformation and thermal expansion), thermal effects
(temperatures and heat transfer) and electrical effects (electric current density
and electric potential field), taking into account non-linear changes of material
and contact properties (such as electrical and thermal contact conductivity).
Strong interactions between all three fields result in a very dynamic process
behavior and require a multi-field coupled simulation.
The model has been verified using measurements from lab tests under
controlled conditions. Compared quantities were curves of electrical voltage
drops, temperatures and electrode displacements measured during welding as
well as spot diameters and electrode imprints determined by destructive or optical
methods after welding. The main effort of the FEM simulation lies in:

• gathering all relevant information such as worksheet and electrode

geometries, temperature-dependent material data, contact properties etc.
These features are inputs to the simulation,
• setting up a parametric Finite Element Model in an automated simulation
loop, including simulation preparation and data extraction, using com-
mercial and open source tools, such as Python, software packages for the
Finite Element Method, etc.
6 Baifan Zhou, Tim Pychynski, Markus Reischl and Ralf Mikut

2 Data Description
The most normal conditions in production, i.e. only random variation without
spatter or other disturbances, have been selected as the simulation scenario for
the data studied in this paper. This simple scenario consists of one welding
machine, one type of worksheet pair with identical nominal sheet thickness and
material, and three welding programs for three different target spot diameters
(see Figure 3). A total of 13,952 welding spots with diameter measurements
were simulated. Two types of data exist for each welding spot.

• Time series: Process curves are series of values with time stamps,
referred to as time series in data science. Among the extracted 20 time
series, process input curves exist, such as electric current 𝐼, voltage 𝑈,
resistance 𝑅, and process feedback curves, such as force of the electrode 𝐹,
displacement 𝑠 of the electrode, temperature 𝑇 of certain measurement
positions, etc.
• Single features: Numeric values or text strings that remain constant for a
welding process. These are nominal and measured geometry or material
properties of the caps and worksheets, the number of spots welded,
positions of the welding spots, welding programs, etc. The simulation
dataset can contain up to 235 single features.

1
Diameter (normalized)

0
ProgNo : 1 ProgNo : 2 ProgNo : 3

(a) (b)
Figure 3: (a) Process curves for the three welding programs. ProgNo indicates the program numbers.
The welding programs prescribe the way the welding process is performed, by specifying the process
curves and some additional welding parameters. (b) Boxplot of the diameters for the three welding
programs.
Comparison of ML Approaches for Time-series-based Quality Monitoring 7

3 Methods and Experiments

The approach for exploring answers to the three questions proposed above
centers on training machine learning models with different data subsets and
comparing their performance:

• Data splitting into training and test set in this paper is not the conventional
random splitting. In manufacturing, the historical data collected for
training may have slightly different statistical properties than the data in
the later application phase. As mentioned in Section 1, the electrode caps
are regularly changed in RSW. The caps in the collected training dataset
will therefore always be different from the caps in the test set. In this
paper, 14 caps are simulated in total. Data generated with 9 caps (7973
data points) is used for training, while data from the remaining 5 caps
(5979 data points) is used for testing.
• Data is split into subsets of different training data numbers for the
purpose of answering question 1 (“How much labeled data to collect at
the start?”). A series of training subsets is built by randomly selecting a
different number of data points from the training dataset. To allow direct
comparison of the testing results, the test dataset always contains the
same 5979 data points. Machine learning methods are applied to build
models using these subsets with different sizes of training data, and tested
on the same test set (see Table 1).

Table 1: Data splitting to subsets of different training data size.

Set # of Data Points

Training sets 100 250 500 1000 2000 5000 7973

Test sets 5979 5979 5979 5979 5979 5979 5979

• Data is split into subsets of different features for the purpose of answering
question 2 (“Which features are important?”). All available features, i.e.
time series features as well as single features, are divided into four subsets
as in the following enumeration. By comparing the performance of
machine learning models trained using the four different feature subsets,
it is possible to estimate the importance of these features.
8 Baifan Zhou, Tim Pychynski, Markus Reischl and Ralf Mikut

› Production Set: Features that are always available in production, such as current,
voltage, resistance (15 single features and 4 time series).
› LabLowCost Set: Production Set + features available in laboratory with relatively
low cost (16 single features and 10 time series).
› LabHighCost Set: LabLowCost Set + features available in laboratory with higher
cost. (29 single features and 14 time series)
› Complete Set: LabHighCost Set + other features that are difficult to realize or
extremely costly (235 single features and 20 time series).

Combinations of possibilities of training data number and feature set

result in 7 × 4 = 28 subsets.

• Extra data sets with manually generated noise are for the purpose
of answering question 3 (“Which precision level should the sensors
have?”). Adding noise to the aforementioned non-noisy 28 subsets results
in another corresponding 28 subsets with noise. The noise levels are
a best engineering guess derived from discussions with process and
measurement experts.

› Sensor data (time series plus single features): Gaussian noise with 2 % standard
deviation (in the time series the noise is added for every single sample point).
› Spot diameter measurement: Gaussian noise with 0.1 mm standard deviation.

In total, 7 × 4 × 2 = 56 subsets are used to train and test machine learning

models and compare their performance. The workflow of experiments on these
subsets is standard (Mikut et al, 2007) and is displayed in Figure 4.

Figure 4: The workflow of data analysis in this paper.

Comparison of ML Approaches for Time-series-based Quality Monitoring 9

First, some simple features are extracted from the time series. The extracted
features are minimum, maximum, minimum position, maximum position,
mean, median, standard deviation, length. We decided to extract only these
features to obtain interpretable features. These extracted features combined
with all other single features are then evaluated and selected using step-forward
selection, starting with the evaluation of each single feature, and incrementally
adding more features. Features that result in models with the best regression
accuracy (RMSE) for the prediction of spot diameters are selected. After feature
selection, three machine learning methods (polynomial regression, neural
networks and k-nearest neighbors) are applied to build different data-driven
models. These models are implemented using the MATLAB toolbox SciXMiner
developed by Mikut et al (2017).
The feature extraction and machine learning models are intentionally kept
simple, for better generalization and automation in other similar applications
in automatic manufacturing. The characteristics of the three machine learning
methods and the reasons for choosing them are explained below.

• Polynomial regression is an established method even in traditional en-

gineering and statistics. The solving of polynomial regression is deter-
ministic. This explains why the performance of polynomial regression
(see Figure 8) is rather stable, after a certain amount of training data
is fed into the model. The features selected by polynomial regression
and the model formula provide transparency for understanding with
engineering knowledge.
• Neural networks are a representative non-linear method. They have
potentially unlimited power to approximate any functions, thus they are
usually powerful enough to solve any supervised learning problem if
adequate data is available. There are, however, two drawbacks. Neural
networks are difficult to interpret, and they often require a large amount
of training data. After a relatively large amount of data has been used
for training, it can be seen in the results that the performance of neural
networks (see Figure 9) even exceeds that of the polynomial model.
• The most substantial characteristic of k-nearest neighbors is that it is
non-parametric, i.e. no parameters are learned through training. Instead,
all training data is stored in the model. The process of prediction is merely
10 Baifan Zhou, Tim Pychynski, Markus Reischl and Ralf Mikut

to use an average of similar points measured by a pre-defined distance

in the training data. K-nearest neighbors would work well if the local
structure of the data were linear, and becomes time-consuming if the
training dataset is too large.

The hyper-parameters of these models are chosen with 5-fold cross validation
with exemplary training datasets. For all machine learning models trained on
subsets without noise, the hyper-parameters are chosen with the non-noisy
subset of all training data (with 7973 training data points) and the Production
Feature Set. For all machine learning models trained with subsets with noise,
this is performed again with the noisy subset of all training data (with 7973
training data points) and the Production Feature Set.
The hyper-parameter selection results in a first order polynomial regression
model without interaction terms (namely a linear regression model, referred to
as Polynomial1), a multi-layer perceptron with one hidden layer of 16 neurons
(referred to as MLP16), and a k-nearest neighbors model (referred as KNN3).
Details are listed in Table 2.

Table 2: Machine learning models and hyper-parameters selected through 5-fold cross validation.

Machine Learning Model Hyper-parameter Value

Multivariate polynomial regression Order of the polynomial 1

Multi-layer perceptron (one hidden layer) Number of neurons in the hidden layer 16
K-nearest neighbor k 3

The experiments show that many features are highly correlated and cause
problems in feature selection due to random effects. For example, some features
are selected for better performance in the feature selection stage, but could show
better or worse performance in a later stage of performance comparison, or even in
another repetition of training and testing. To circumvent this problem, for training
machine learning models on subsets with more features, the selected features
from smaller feature sets will always be used, and the models have the opportunity
to add more features. It is worthy to note this issue of highly correlated features,
for this phenomenon may be prevalent in manufacturing data analysis.
Comparison of ML Approaches for Time-series-based Quality Monitoring 11

4 Results
To answer the three questions proposed in Section 1, it is necessary to compare
the performance of the various models built with different methods and trained
with different subsets. In order to compare the performance, an appropriate
measure is important. Various performance measures were proposed in the
literature. This section first discusses the various performance measures, then
uses the selected measure to compare the models and provides answers to the
proposed questions. Diameter prediction is treated as a regression problem in
this work, as process experts can obtain more information from a predicted
diameter than a mere good/bad classification.

0.12 1

Correlation Coefficient
0.99
0.1
RMSE (mm)

0.98
0.08 Production
Production
LabLowCost
LabLowCost 0.97
LabHighCost
0.06 LabHighCost
All
All
100 250 500 1000 2000 5000 7973 100 250 500 1000 2000 5000 7973
Training data number Training data number

(a) (b)
1 1
Error Within 10 Percent
Error Within 5 Percent

0.95

0.995
0.9
Production Production
LabLowCost LabLowCost
0.85 LabHighCost LabHighCost
All All
0.99
100 250 500 1000 2000 5000 7973 100 250 500 1000 2000 50007973
Training data number Training data number

(c) (d)

Figure 5: Comparison of performance measures by testing Polynomial1 on non-noisy datasets.

RMSE (or MSE; see Figure 5 a) stands for root mean squared error (or mean
squared error) between the reference diameters and predicted diameters. It
is a classic performance measure for regression tasks, and has been used
12 Baifan Zhou, Tim Pychynski, Markus Reischl and Ralf Mikut

in Martín et al (2009) and Boersch et al (2016), etc. RMSE is also used

in this paper for feature selection. The disadvantages of RMSE are that
it is scale-dependent (Hyndman and Koehler, 2006), and not particularly
intuitive compared to other measures.
Correlation Coefficient (or R-squared; (see Figure 5 b), used in Muhammad
and Manurung (2012), is good for judging the models’ ability to predict
the general trend of the training data. Since one of the goals of diameter
prediction is to minimize the number of product failures, only knowing the trend
is not sufficient for this goal.
The most intuitive measures for process experts are perhaps User-Defined
Errors (UDE, Boersch et al (2016); Figure 5 c, d). These are more suitable
for helping to set the factor of safety, thus closer to the goal of minimizing
product failures. Two User Defined Errors are adopted in this paper: Error
Within 5 % and Error Within 10 %. Error Within 5 % means the percentage of
predictions with relative errors smaller than 5 % of the reference value. Error
Within 10 % is likewise.
Comparing models trained on different feature sets (see Figure 5) shows that,
as expected, when more features are available in training, the models also have
better performance in testing: All › LabHighCost › LabLowCost › Production.
Comparing models trained on different training data subsets shows that,
when the training data number is greater than 250, the performance improves
insignificantly as the training data number further increases.
From results of testing the first-order polynomial model (see Figure 5), trained
on different numbers of training data and features, it can be seen that RMSE,
Correlation Coefficient and Error Within 5 % have similar trends. This is also
the case for the multi-layer perceptron model and k-nearest neighbor model.
Error Within 10 % cannot differentiate the models trained with different data
number and feature sets, since they always show very good results. Therefore,
in the following discussions, only Error Within 5 % will be compared.
Comparing models built using different machine learning methods can be
made by comparing Figure 5 c, Figure 6, and 7, or in Figure 8. The comparison
indicates the performance of the models built with three machine learning
methods is: Polynomial ≈ MLP > KNN. Figure 6 shows that MLP16 tends
to overfit for the Feature Set LabLowCost and LabHighCost. In this regard,
Polynomial1 is more advantageous.
Comparison of ML Approaches for Time-series-based Quality Monitoring 13

Error Within 5 Percent 1 1

Error Within 5 Percent

0.95 0.95

0.9 0.9
Production
Production
LabLowCost
LabLowCost
0.85 LabHighCost 0.85
LabHighCost
All
All
100 250 500 1000 2000 5000 7973 100 250 500 1000 2000 5000 7973
Training data number Training data number

Figure 6: Testing results MLP16 on non- Figure 7: Testing results KNN3 on non-noisy
noisy datasets. datasets.

1 1
Production
Error Within 5 Percent

Error Within 5 Percent

LabLowCost
0.95 0.95 LabHighCost
All

0.9 0.9
Polynomial1
MLP16
0.85 0.85
KNN3
100 250 500 1000 2000 5000 7973 100 250 500 1000 2000 5000 7973
Training data number Training data number

Figure 8: Testing models trained on non- Figure 9: Testing Polynomial1 on noisy

noisy Production Set. datasets.

Comparing models trained with non-noisy data and noisy data can be made by
contrasting Figure 5 c and Figure 9. The comparison suggests the performance
deteriorates strongly with noisy data. Generally speaking, the difference between
models trained on these three feature subsets, Production, LabLowCost, and
LabHighCost is insignificant.

5 Conclusion and Outlook

Simulation data can provide large amounts of economically obtained labeled
data and features that would be otherwise unavailable. This data contains no or
low measurement error. As a result, it is possible to hold a collection of valuable
14 Baifan Zhou, Tim Pychynski, Markus Reischl and Ralf Mikut

data from production or laboratory. Reviewing the starting point of the three
questions, the following conclusions are made based on the datasets generated
by a simulation model in the selected simple scenario (see Section 2), serving
as a guidance for a first plan of data collection:

• Answer to Question 1: Starting with a collection of 250 labeled data

points.
• Answer to Question 2: Sensors available in Production can already be
useful.
• Answer to Question 3: Measurement uncertainty can lead to significant de-
terioration of the prediction performance. In our case, precision better than
Gaussian errors of 2 % standard deviation for sensors and 0.1 mm standard
deviation for diameter measurements respectively is recommended.

It is important to note that the conclusions regarding the necessary number of

data points and features are highly dependent on the dataset, and the simulation
scenario from which the data is collected. It is therefore safe to say that the
results can be generalized for the specific situation simulated. They can serve
as a first guidance for data collection, but should not be assumed as valid in
general. Since the selected simulation scenario of one welding machine and
three welding programs tends to be oversimplified, these conclusions may not
hold for other scenarios or for complex production data. The optimal data
collection plan may vary depending on the complexity of the actual situations
compared to the simulated situation.
In future research, several topics can be explored. We have plans for combining
the simulated data and laboratory or production data.

• The first option is to use laboratory data or production data to fur-

ther improve and fine-tune the simulation model so that realistic phys-
ical effects and authentic variance of influencing factors can be re-
built in the simulation. This realistic simulation data can be used for
further data analysis.
• After the first point has been realized, a further step is to apply the model
trained on simulated data, and test on the laboratory or production data
to see the extent to which the trained model can be generalized.
Comparison of ML Approaches for Time-series-based Quality Monitoring 15

• A third option is to apply transfer learning, i.e. to pre-train a model

on simulated data, then to fine-tune the model on laboratory data or
production data, and to see how much improvement in performance
is gained.
• The improved simulation model and generated data can be used for
production condition evaluation before a real production process begins,
so that the influence of change of welding parameters, such as welding
gun properties, material properties, distance between the worksheets, can
be systematically evaluated.
• We will also analyze alternative feature extraction methods (such as
features based on Fourier Transformation, Wavelet Transformation or
Principal Component Analysis) to evaluate the difference of prediction
accuracy using different feature extraction methods.
• As actual production conditions may change due to a change in supplier,
new engineering design, etc., it is always questionable to what extent a
model trained on historic or lab data can be applied to production data in
the future, which constitutes a general problem in manufacturing. This is
also called model monitoring and remains an open question.

This paper tries to use data analysis as a guidance for data collection. The iterative
process data-collection → data analysis → data-analysis guided data collection
→ data analysis is an attempt to combine data collection and analysis. A similar
practice would be particularly necessary in fields where a limit on the amount
of labeled data, features, and covering situations is a problem of interest.

Acknowledgements The authors would like to thank Simon Waczowicz for his help using the
SciXMiner toolbox, Alexander Schif for providing the data, Friedhelm Günter, Florian Schmid, Martin
Dieterle for their valuable input on improving the quality of the paper, and Katherine Quinlan-Flatter
for proofreading this article.

References
Afshari D, Sedighi M, Karimi MR, Barsoum Z (2014) Prediction of the Nugget Size in
Resistance Spot Welding with a Combination of a Finite-element Analysis and an
Artificial Neural Network. Materiali in tehnologije 48(1):33–38. DOI: 10.26634/jms.
3.1.3366.
16 Baifan Zhou, Tim Pychynski, Markus Reischl and Ralf Mikut

Boersch I, Füssel U, Gresch C, Großmann C, Hoffmann B (2016) Data Mining in

Resistance Spot Welding. The International Journal of Advanced Manufacturing
Technology, pp. 1–15, Springer. DOI: 10.1007/s00170-016-9847-y.
Cho Y, Rhee S (2004) Quality Estimation of Resistance Spot Welding by Using Pattern
Recognition with Neural Networks. IEEE Transactions on Instrumentation and
Measurement 53(2):330–334. DOI: 10.1109/TIM.2003.822713.
DVS (2016) DVS 2902–3: Widerstandspunktschweißen von Stählen bis 3 mm
Einzeldicke – Konstruktion und Berechnung. Standard, Deutscher Verband für
Schweißen und verwandte Verfahren e.V., Düsseldorf (Germany).
El Ouafi A, Bélanger R, Méthot JF (2010) An On-line ANN-based Approach for Quality
Estimation in Resistance Spot Welding. In: Advanced Materials Research, Vol. 112,
pp. 141–148. DOI: 10.4028/www.scientific.net/AMR.112.141.
Hyndman RJ, Koehler AB (2006) Another Look at Measures of Forecast Accuracy.
International Journal of Forecasting 22(4):679–688. DOI: 10.1016/j.ijforecast.2006.
03.001.
ISO (2004) ISO 14327:2004: Resistance Welding – Procedures for Determining the
Weldability Lobe for Resistance Spot, Projection and Seam Welding. Standard,
International Organization for Standardization, Geneva (Switzerland).
Kagermann H, Lukas WD, Wahlster W (2011) Industrie 4.0: Mit dem Internet der
Dinge auf dem Weg zur 4. industriellen Revolution. VDI Nachrichten 13.
Li Y, Zhao W, Xue H, Ding J (2012) Defect Recognition of Resistance Spot Welding
based on Artificial Neural Network. Advances in Intelligent and Soft Comput-
ing 115(2):423–430. DOI: 10.1007/978-3-642-25349-2_56.
Martín Ó, López M, Martin F (2007) Artificial Neural Networks for Quality Control
by Ultrasonic Testing in Resistance Spot Welding. Journal of Materials Processing
Technology 183(2-3):226–233, Elsevier. DOI: 10.1016/j.jmatprotec.2006.10.011.
Martín O, Tiedra PD, López M, San-Juan M, García C, Martín F, Blanco Y (2009)
Quality Prediction of Resistance Spot Welding Joints of 304 Austenitic Stainless
Steel. Materials and Design 30(1):68–77. DOI: 10.1016/j.matdes.2008.04.050.
Mikut R, Burmeister O, Grube M, Reischl M, Bretthauer G (2007) Interaktive Auswer-
tung von aufgezeichneten Zeitreihen für Fehlerdiagnosen und Mensch-Maschine-
Interfaces. atp - Automatisierungstechnische Praxis 49(8):30–34. ISBN: 978-3-
180919-80-5.
Mikut R, Bartschat A, Doneit W, Ordiano JÁG, Schott B, Stegmaier J, Waczowicz
S, Reischl M (2017) The MATLAB Toolbox SciXMiner: User’s Manual and
Programmer’s Guide. arXiv:1704.03298.
Muhammad N, Manurung YHP (2012) Design Parameters Selection and Optimization
of Weld Zone Development in Resistance Spot Welding. In: Proceedings of World
Academy of Science, Engineering and Technology, 71, p. 1220.
Comparison of ML Approaches for Time-series-based Quality Monitoring 17

Nagorny K, Lima-Monteiro P, Barata J, Colombo AW (2017) Big Data Analysis in

Smart Manufacturing: A Review. International Journal of Communications, Network
and System Sciences 10(03):31, Scientific Research Publishing. DOI: 10.4236/ijcns.
2017.103003.
Ruisz J, Biber J, Loipetsberger M (2007) Quality Evaluation in Resistance Spot Welding
by Analysing the Weld Fingerprint on Metal Bands by Computer Vision. International
Journal of Advanced Manufacturing Technology 33(9-10):952–960. DOI: 10.1007/
s00170-006-0522-6.
Rüßmann M, Lorenz M, Gerbert P, Waldner M, Justus J, Engel P, Harnisch M (2015)
Industry 4.0: The Future of Productivity and Growth in Manufacturing Industries.
Boston Consulting Group.
Schif A (2017) Examination of Machine Learning Methods for the Evaluation of
Resistance Spot Welds based on Simulated Process Scenarios. Master’s thesis,
University of Stuttgart.
Sumesh A, Rameshkumar K, Mohandas K, Babu RS (2015) Use of Machine Learn-
ing Algorithms for Weld Quality Monitoring using Acoustic Signature. Procedia
Computer Science 50:316–322. DOI: 10.1016/j.procs.2015.04.042.
Sun H, Yang J, Wang L (2017) Resistance Spot Welding Quality Identification with
Particle Swarm Optimization and a Kernel Extreme Learning Machine Model. The
International Journal of Advanced Manufacturing Technology 91(5-8):1879–1887,
Springer. DOI: 10.1007/s00170-016-9944-y.
Wan X, Wang Y, Zhao D (2016) Quality Monitoring based on Dynamic Resistance and
Principal Component Analysis in Small Scale Resistance Spot Welding Process. The
International Journal of Advanced Manufacturing Technology 86(9-12):3443–3451,
Springer. DOI: 10.1007/s00170-016-8374-1.
Wang K (2007) Applying Data Mining to Manufacturing: The Nature and Implications.
Journal of Intelligent Manufacturing 18(4):487–495. DOI: 10.1007/s10845-007-
0053-5.
Wuest T, Weimer D, Irgens C, Thoben KD (2016) Machine Learning in Manufac-
turing: Advantages, Challenges, and Applications. Production & Manufacturing
Research 4(1):23–45. DOI: 10.1080/21693277.2016.1192517.
Yu J (2015) Quality Estimation of Resistance Spot Weld based on Logistic Regression
Analysis of Welding Power Signal. International Journal of Precision Engineering
and Manufacturing 16(13):2655–2663, Springer. DOI: 10.1007/s12541-015-0340-6.
Zhu S, Kong L, Chen L (2011) Data Mining of Sensor Monitoring Time Series and
Knowledge Discovery. In: Proceedings of the 2nd International Conference on
Artificial Intelligence, Management Science and Electronic Commerce (AIMSEC),
pp. 2914–2917. DOI: 10.1109/AIMSEC.2011.6010471.