IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 13, No. 3, September 2024, pp. 3052∼3060

ISSN: 2252-8938, DOI: 10.11591/ijai.v13.i3.pp3052-3060 ❒ 3052

Machine learning-based decision-making approach for

predicting defects detection: a case study
Elena Barzizza, Nicoló Biasetton, Riccardo Ceccato, Alberto Molena
Department of Management Engineering, Universitá degli Studi di Padova, Vicenza, Italy

Article Info ABSTRACT

Article history: In today’s highly competitive global market, industries must produce faultless
products to achieve profitability. Machine learning (ML) algorithms provide a
Received May 10, 2023
possible method to improve quality standards by enabling the prediction of the
Revised Nov 10, 2023 outcome of quality control processes. This article presents a real case study
Accepted Dec 3, 2023 based on ML algorithms suggested to develop a knowledge-based intelligent
supervisory system to predict defect products in the fashion industry. Defect
Keywords: detection is formulated as a binary classification problem, and several ML algo-
rithms have been compared to determine the most suitable one on the available
Defect production data. The random forest (RF), LightGBM, and C5.0 algorithms exhibit compa-
Fashion industry rable high-end performances on the pre-processed dataset made available by the
Machine learning company. Nevertheless, since the aim of the analysed industry is to reduce the
Predictive quality rate of false negative observations (i.e., the proportion of defected-free products
Quality control wrongly classified), the best method results is RF, as it minimizes this metric.

This is an open access article under the CC BY-SA licence.

Corresponding Author:
Biasetton Nicoló
Department of Management Engineering, Universitá degli Studi di Padova
Stradella S. Nicola, 3 Vicenza 36100, Italy
Email: [email protected]

1. INTRODUCTION
Nowadays, quality represents a strategic factor in all industries that contributes to classifying compa-
nies as profitable and non-profitable. Quality can be considered as an index of internal performance, as it allows
one to develop diversifying product or service differentiation policies. In recent years, the quality of products
and services has even increased in importance due to the massive spread of timely written online reviews by
customers at no cost who want to share their feelings and opinions after use or after experience. Therefore,
the provision of defective or poor quality products can significantly affect both the reputation of the brand and
the loyalty of customers to a certain brand; the aforementioned are some of the reasons that have led numerous
companies to adopt a zero-defect policy.
The definition of a quality management process has changed significantly over the years [1]. The
first quality control (QC) process was developed during the mass production period (1900-1940), and it simply
consisted of the inspection of the final products. However, increasing pressure from the market led to a major
theoretical shift in the first place toward the concept of process quality. The idea underlying this theory was
that looking for errors was much less efficient than finding the source of the errors and removing them. In
the early 60s, quality assurance theory emerged. The process quality, i.e. a posteriori reaction to QC of either
products or processes, was not enough anymore, and industries wanted to prevent risks of failure products
or services identifying their major causes before their appearance. With increasing complexity of products
and interdependencies with the supply chain, ISO 9000 certification has been introduced to define quality

Journal homepage: http://ijai.iaescore.com

Int J Artif Intell ISSN: 2252-8938 ❒ 3053

management standards to help industries ensure that they meet the basic requirements of the customer and other
stakeholders related to products or services. In the last ten years, the quality management process shifted from
being implemented due to market pressure to being implemented due to the general importance of delivering
high-quality results.
Improving product and service quality is one of the main goals of the fourth industrial revolution [2],
and the interconnection between automation, machine learning (ML) algorithms, and real-time data is the key to
its completion. From a manufacturing perspective, the ability to efficiently collect and analyse huge amount of
data using powerful ML algorithms can improve QC [3] revealing hidden patterns in the data and thus allowing
for a broader analysis of the impact of all factors that contribute to manufacturing processes. Within the QC
field, ML applications are highly diverse as they range from regular classification based on collected data to
computer vision and image recognition. Similarly, there is no universal ML approach, algorithm, or method to
solve all different QC problems, as the specific features of production systems have a substantial impact on the
algorithms to be implemented [4]. The application of ML algorithms to predict or classify product quality has
a relevant historical background. One of the earliest examples can be traced back to 1993, and concerns the
prediction of the outer diameter of parts produced through injection moulding using backpropagation artificial
neural networks [5].
In recent years, we also assisted to a growing literature on ML algorithms applications to classify or
predict product quality thanks to the exponential increases in computer power, database technologies, optimi-
sation methods, new efficient and robust ML algorithms, and the availability of big data. From the review of
the literature on ML and deep learning (DL) for quality prediction in manufacturing [6], it emerges that more
than 80 papers have been published on this topic between 2012 and 2021, with an exponential increase in the
last few years. The success of ML algorithms can be attributed to their ability to handle high-dimensional,
multivariate data, greatly reducing the need for human effort and improving the quality of the product [7].
These algorithms can provide suitable solutions for collecting fast and reliable information and understanding
the implicit relationships existing within large data sets collected in complex and dynamic environments [8].
An example of an application of ML to optical inspections on finished products adopting a decision tree for the
optimal classification of battery separator defects can be found in [9].
ML algorithms have also been adopted to predict the final quality of products in the early stages of the
manufacturing process see [10]-[13] and to detect defects in mature organisations (i.e. organisations that merge
different tools to improve the quality of manufacturing processes, such as lean production, standards confor-
mity, six sigma, design for six sigma) that only generate a few defects per million opportunities [3]. Several
ML algorithms have been trained and evaluated to predict dimensional defects in a real multistage automotive
assembly line in [14] while [15] analysed a welding process using both the random forest (RF) and the C4.5
decision tree algorithm for classification [16] to correlate the sound of the arc with the quality of the weld.
Schorr et al. [17] predicted the quality of drilled and reamed hydraulic valve bores, by estimating the diameter
and concentricity of re-drilled reamers, starting from torque measurements. Lastly, one of the most intriguing
research areas involves geospatial big data, which comprises vast amounts of data including information about
geographic locations. The utilization of geospatial big data can enhance the quality and reliability of produc-
tion processes by monitoring resources and, most importantly, preventing production anomalies to avoid the
manufacturing of defective products. For a discussion on such topics, you can refer to [18]-[20].
This study presents the learning process and pattern recognition strategy for a knowledge-based intel-
ligent supervisory system, in which the main goal is the prediction of defective products. Defect detection is
formulated as a binary problem, i.e. OK for defect-free products and KO for defect products. To achieve the
aforementioned aim, different supervised ML algorithms are compared, in order to identify the most suitable
one. A real case study of a company working in the fashion industry is presented. The paper is structured as
follows. Section 2 briefly describes the ML algorithms used to predict defective products. In section 3 the case
study is presented, while in section 4 the results are presented and discussed. Section 5 concludes the article
by highlighting the main contributions of this study.

2. METHOD
Classification is the assignment of an object defined by a set of features to one of several predetermined
classes using a learning rule. Thus, it belongs to the macro area of supervised ML algorithms that aims to infer a
function to describe labelled training data (e.g., data with classification class) to predict the output of categorical

Machine learning-based decision-making approach for predicting defects detection ... (Elena Barzizza)
3054 ❒ ISSN: 2252-8938

variable [21]. As stated previously, the goal of this study is to predict the QC outcome, and since this output
is encoded as a binary variable (OK/KO), the following classification ML algorithms have been implemented:
logistic regression; multivariate adaptive regression splines (MARS); generalized additive model (GAM) using
splines; linear discriminant analysis (LDA); quadratic discriminant analysis (QDA); neural network (nnet); RF
(ranger); support vector machines (SVM); light gradient boosting machine (lightGBM); elastic Net regression
(glmnet); and C5.0 decision tree algorithm.
The choice of these algorithms has been driven by the literature review and the experience of other
real case studies gained by the authors. Given that the input variables can encompass both qualitative and
quantitative features, it is imperative that the selected models exhibit the capability to process both types of
variables. The following sections provide a theoretical and concise introduction to the different algorithms that
have been used.
2.1. Logistic regression
Logistic regression is a statistical modelling technique used for classification when the response vari-
able is binary. It is part of the generalised linear model category and applies a logistic function to the linear
combination of predictor variables to estimate the probability of the binary outcome of the response variable
[22]. The model parameters are estimated using maximum likelihood estimation. In logistic regression, the
logit transformation is applied on the odds, which are the probability of success divided by the probability of
failure. So, logistic regression can be represented using (1), where p is the probability of a binary event and
x1 ,x2 ,...,xk are the explanatory variables:

 
p
logit(p) = ln = β0 + β1 x1 + β2 x2 + · · · + βk xk (1)
1−p

2.1.1. Multivariate adaptive regression splines

MARS is a flexible nonparametric statistical procedure [23]. It approximates the nonlinearity of the
model by using separate regression slopes (splines) in distinct intervals of the independent variable space, and
so the gradient of the regression line may change in different intervals, which are defined by the so-called
knots. MARS uses a fast but intensive search procedure to determine the variables to use and the end points
of the intervals for each variable. It also allows for searching for interactions between variables. The selection
of the optimal MARS model is composed of two phases: first, it constructs a large number of basis functions
(BFs), which are selected to overfit data; then, in the so called backward phase, the redundant BFs (the ones
which made the least contributions) are deleted. It can be considered as a white-box model, as its structure is
interpretable, meaning that the relationships between variables and the response can be understood through the
model’s node and spline structure.
2.1.2. Generalized additive model using splines
GAM using splines is a general additive model that uses splines functions to model nonlinear relation-
ships [24]. Each predictor variable is modelled using a separate spline function, in which a spline is a piecewise
polynomial function defined over two knots. These functions smoothly connect polynomial function between
the various knots, allowing GAM to model non-linear relationships between predictor variables and the re-
sponse variable without assuming any particular shape of the relationship. The parameters of this model are
estimated using maximum likelihood estimation. The key difference between this algorithm and the aforemen-
tioned MARS algorithm is the method used to model the nonlinearity. Although MARS uses separate slopes
at different intervals in the independent variable space, GAM uses smooth functions, such as cubic splines, to
model the nonlinearity.
2.1.3. Linear discriminant analysis
LDA [25], is a statistical technique used to predict the class membership of an observation. It is
generally used to classify observations between two classes, but it can also be extended to handle multiple
classes. LDA assumes that all classes can be separated by linear boundaries, so it constructs multiple linear
discriminant functions, each representing a hyperplane in the feature space, to differentiate classes. In case
of two classes, as in the case study presented here, LDA draws a hyperplane and projects the data onto this
hyperplane, in order to maximize the separation of the two categories. This hyperplane is created considering
two criteria simultaneously: maximizing the distance between the means of two classes and minimizing the

Int J Artif Intell, Vol. 13, No. 3, September 2024: 3052–3060

Int J Artif Intell ISSN: 2252-8938 ❒ 3055

variation within each category. It is also important to underline that LDA is based on the assumption that the
covariance matrices are equal for all classes.

2.1.4. Quadratic discriminant analysis

QDA is a statistical technique used for classification, very similar to LDA, with the difference that
QDA relaxes the hypothesis that all covariance matrices are equals; as a result this technique is capable of cap-
turing more complex relationships between predictor variables and response variables. It works by estimating
the parameters of a quadratic discriminant function for each class, which is a quadratic equation involving the
predictor variables. When a new observation is presented, QDA computes its discriminant scores based on
these quadratic functions and assigns it to the class with the highest score.

2.1.5. Neural network

Neural networks draw inspiration from the architecture and functioning of biological neural networks,
and possess the capacity to model intricate relationships between inputs and outputs [26]. They consist of a
series of interconnected nodes that are arranged in layers. This multitude of nodes is known as perceptrons or
neurons, which are responsible for transmitting signals to other neurons located in the subsequent layer. The
signal is transmitted when a threshold value is exceeded, so the neurons have to compute an activation function
to send the output to the subsequent neurons. The training phase of these complex structures involves forward
and backward propagation: once the network was fed with values in a forward direction, the backward step
optimizes the weights, to enhance the accuracy of the model.

2.1.6. Random forest

RF [27], is a class of statistical algorithms that can be applied for both regression and classification.
It consists of n decision trees that are typically grown on an individual training data set extracted from the
original data set. To address the issue of correlated trees, a random sub-sample of m predictors, from the full
set of p available predictors, is selected as split candidates at each split point of a tree. This technique helps to
ensure that each tree is built using a slightly different set of predictors, thus increasing the diversity among the
trees in the forest, allowing a greater reduction in the prediction variability.

2.1.7. Support vector machines

SVM is a supervised ML algorithm [28]. The fundamental principle of SVM involves identifying the
appropriate hyperplane for classification using support vectors, which are the points of each class located at
the margin. This method is effective for both linearly separable and linearly non-separable data: indeed, in the
case of linearly separable data, SVM finds a hyperplane which separates the data with the largest margin, while
for non-separable data it maps the input data into a high dimensional space with a kernel function to find the
separating hyperplane.

2.1.8. Light gradient boosting machine

LightGBM is a recent ML algorithm [29]. It is built on the gradient boosting framework, which utilizes
the combination of several weak models to create a robust model capable of capturing complex nonlinear
relationships between features and target variables. LightGBM’s innovative tree-building approach uses a
split-point algorithm to efficiently select the optimal split point for each feature. This, along with its design for
fast and efficient training and prediction, enables LightGBM to handle large data sets and excel in scenarios
where training time is constrained.

2.1.9. Elastic net regression

Elastic net regression is a regularization regression model [30]. This model is a combination of lasso
regression (L1) and ridge regression (L2) penalties to overcome the limitations of both methods, especially
when data are affected by multicollinearity (two or more input variables are highly correlated). Similarly
to lasso regression, this method can automatically perform variable selection and shrink coefficients. This
regression model can predict both quantitative and binary output.

2.1.10. C5.0 Decision tree algorithm

C5.0 is an algorithm that aims to build a decision tree capable of classifying observations based on
some input variables. This algorithm works by splitting the sample on the basis of the feature with the highest
information gain; then, every subsample as specified by the first split is split recursively until the subsamples

Machine learning-based decision-making approach for predicting defects detection ... (Elena Barzizza)
3056 ❒ ISSN: 2252-8938

become indivisible. At the lowest level, the splits are evaluated, and those that do not significantly contribute
to the model are pruned until the desired level of accuracy is achieved [31].

2.1.11. Algorithms evaluation and comparison

The input variables will be used by the ML algorithms mentioned above that will return an estimate of
the probability value of belonging to the class of defective products. Subsequently, the classification of observa-
tions into the OK/KO classes is obtained setting a threshold value: if the probability predicted by the algorithm
is higher than the threshold, the observation will be classified as KO, otherwise it will be classified as OK. The
classification success of the different models is measured using the receiver operating characteristics (ROC)
curve, the area under the curve (AUC), and the following indices based on the confusion matrix: total accuracy,
sensitivity, precision, recall, specificity, false negative rate, and false positive rate. It is also fundamental to
underline that the tuning parameters of the ML algorithms have been optimized using a grid search to minimize
the false negative rate (i.e. the proportion of defected products wrongly classified as defected-free ones). The
reason for this strategy lies in the fact that, from an industry perspective, it is worse to classify a product as free
from defects when it is not than to misclassify a defect-free product and reprocess it. Furthermore, for each
ML algorithm, different threshold values were tested to find the one that minimized the false negative rate.

3. DESCRIPTION OF THE CASE STUDY

The case study we are going to present in this paper concerns the tumbling of a fashion accessory
made of acetate, and more specifically, the unit of analysis we are going to investigate will be a batch of
identical products. To produce this product, the manufacturing process is characterised by five consecutive
steps. At the end of each step, an operator evaluates the final output and decides whether the batch of products
can move on to the next step or must redo the step. Therefore, each batch of products can run each step of
the manufacturing process multiple times until the operator is satisfied with the final result. Consequently, the
amount of time necessary to complete a process varies according to several subjective factors. At the end of
the manufacturing process, an operator visually inspects 10% of the batch of products and decides, according
to a series of internal quality standard requirements, whether the batch must be classified as a defected or a
defected-free one. Speaking about data, it is important to underline that the source of the data we analyzed in
this case study is the manufacturing execution system (MES) of the company we worked with, that is a software
designed to optimize the manufacturing process by tracking and documenting the entire production cycle. From
the MES, we obtained 38,743 observations, representing different batches, and after a careful data cleaning
phase, performed in compliance with the company and consisting in the elimination of the observations that
had a null processing or waiting time, mostly due to some technical issues, we ended up with a final data set
consisting of 28,009 observations. In Figure 1, we report the boxplot of the waiting times for the five steps that
compose the productive process before the data cleaning phase: from this figure. It is possible to observe that
there are some negative values.

Figure 1. The waiting times for the five steps that constitute the production process

Int J Artif Intell, Vol. 13, No. 3, September 2024: 3052–3060

Int J Artif Intell ISSN: 2252-8938 ❒ 3057

In this case study we used a total of 22 predictive variables, which can be divided into two main
categories: product features and process characteristics. More in details, among the former, which are a total of
11 out of 22 predictive variables, we can identify features such as color, model, and some typical measurements
of the product under study, such as the design of certain parts of the product itself. On the other hand, as far
as variables related to the production process are concerned, we can include, for example, the processing time
for each of the five steps already mentioned, the percentage of time spent in a certain machinery positioned in
a certain area of the production chain and also the wear of the abrasive paste, being one of the most important
variables in the process and thus in determining the outcome of QC. The values of mean, Q1, median, Q3 and
standard deviation for the last mentioned variable are reported in Table 1. Clearly, the target variable, on the
other hand, is represented by the outcome of QC downstream of each processing step.

Table 1. Mean, Q1, Median, Q3, and standard deviation values for the variable related to wear
of the abrasive paste
Productive step Mean Q1 Median Q3 StDev
Step 1 174.5607 83.40968 172.2402 259.9114 106.5447
Step 2 249.0394 119.9143 241.9183 369.3094 150.1812
Step 3 515.7192 238.9415 543.8706 773.4081 300.9195
Step 4 441.7456 243.1367 399.2906 600.9616 246.1685

In this case study we used R, and we went to test the different ML models described in section 2.
Regarding the settings we used the common training and testing approach, where training data set are used to
train a set of candidate algorithms trying different tuning parameters and testing data are used to evaluate the
performances of the models on never-seen-before data. So, adopting this approach, we divided the observations
and used 75%, so 21,007 observations in the training phase and 25%, so the remaining 7,002 observations in
the testing phase. It is also important to underline that we have used the 10-fold cross validation approach,
that is used to evaluate the generalisation ability of each candidate algorithm and select the best, according to a
relevant model selection criterion, which in our case was the false negative rate.
Moreover, as already mentioned, we evaluated different threshold values, in order to find the one that
minimizes the false negative rate. So, it is possible to affirm that a two-step approach was implemented: in the
first step, various ML models were tested while keeping the threshold fixed—starting, for example, with a value
of 0.5. After testing these ML models, we also explored four different thresholds values (0.5, 0.4, 0.3, 0.2) to
find the ML model-threshold combination that minimized the false negative rate. It’s crucial to emphasize one
more time that minimizing this rate is highly important, indeed, for the company, a false negative (a product
classified as OK when it’s not) is much more it’s much more problematic than a false positive (a product
classified as KO when it’s not). Lastly, an issues that arises is the imbalance in the data. Indeed, as one might
expect, the number of batches not conforming to QC is much lower than those that are compliant. In our
specific case, the total number of observations is 28,009, but only 21.2%, so 5,936 are classified as KO.
In such a situation, the classification might be influenced in favour of the most represented class, i.e.
the OK ones [32]. Therefore, the ML algorithms described in the previous subsections have been implemented
twice: on the original data set provided by the company and on a balanced data set extracted from the original
one using the random undersampling technique that results on a data set made up of 11,545 observations, of
which 5,936 KO and 5,518 OK. Lastly, we rated the different ML algorithms accordingly to their capacity of
minimizing the false negative rate.

4. RESULTS AND DISCUSSION

From the performances we obtained by testing the ten ML models, and considering the false negative
rate as a primary importance index, for the reasons previously discussed, we found that, regardless of the
chosen threshold and the type of dataset used (balanced or unbalanced), three models always emerged as the
best among the ten tested. These three models are: RF, LightGBM, and C5.0. As for the choice of the threshold,
we tested four different levels: 0.5, 0.4, 0.3, and finally 0.2. Regardless of the model and the type of dataset
used, the analysis revealed that the threshold minimizing the false negative rate was 0.2. So, we decided to set
that threshold for a final comparison among the various ML models. We compared RF, LightGBM, and C5.0,
using both datasets and fixing the threshold at 0.2. Table 2 shows the performance of various tested indices for
these three models. The left side of the table represents the performances obtained in the unbalanced dataset,

Machine learning-based decision-making approach for predicting defects detection ... (Elena Barzizza)
3058 ❒ ISSN: 2252-8938

while the right side represents the performances obtained on the balanced dataset. Looking at the table, it can
be noted that RF always stands out as the best model since it minimizes the false negative rate in both cases.

Table 2. Performance of RF, LightGBM and C5.0, threshold equal to 0.2, unbalanced and balanced data
Unbalanced data Balanced data
ML algorithm RF LightGBM C5.0 RF LightGBM C5.0
AUC 0.7447 0.7209 0.7110 0.7294 0.6979 0.6875
Accuracy 0.6659 0.6662 0.6807 0.5377 0.5733 0.5584
Recall 0.7076 0.6490 0.6029 0.9929 0.9572 0.9608
Precision 0.3552 0.3465 0.3520 0.5288 0.5508 0.5420
Sensitivity 0.7076 0.6490 0.6029 0.9929 0.9572 0.9608
Specificity 0.6547 0.6709 0.7014 0.0480 0.1602 0.1259
FNR 0.2924 0.3510 0.3971 0.0071 0.0428 0.0392
FPR 0.3453 0.3291 0.2986 0.9520 0.8398 0.8741

5. CONCLUSION
In this study, conducted in collaboration with a company operating in the fashion industry, we com-
pared various ML models and different settings to identify the one that allows a reliable prediction of QC
outcomes. Specifically, we evaluated the eleven models presented in section 2 to identify the one that, based
on some predictive variables related to product features or process characteristics, minimizes the false negative
rate. Since the data was unbalanced, with approximately 80% of observations being OK (having passed QC),
the study was conducted on both the original data set, consisting of 28,009 observations, and a balanced data
set obtained through undersampling. In our analysis, we also experimented various threshold values to en-
hance performance on the false negative rate. Considering both data sets, our analysis revealed that the optimal
threshold value to minimize the false negative rate is 0.2. Regarding ML models, our analysis showed that three
models outperform the others: RF, LightGBM, and C5.0. Ultimately, the best ML-threshold combination was
found to be RF with a threshold set at 0.2, holding true for both the unbalanced and balanced datasets. Through
this analysis, we obtained a model that accurately predicts QC outcomes while minimizing the risk of false
negatives observations, i.e. the number of products that are deemed defect-free but are defective. While the
findings cannot be broadly generalized due to the case study nature, we can assert that, much like in previous
cases, RF demonstrates satisfactory performance. Future research will be devoted to comparing our results with
others in terms of concrete data for better research integrative value. The implementation of ML algorithms
that are effective in predicting the outcome of QC based on product or process variables can help companies
in all fields understand what to expect from the production process regarding product quality. This subject
(and, therefore, this study) merely represents the initial stage of a possible prospective improvement journey
that could aspire to establish production lines that culminate in zero defective products by understanding, for
instance, whether production defects are associated with specific product or production line factors, through
further analysis.

ACKNOWLEDGEMENTS
Authors would like to thank the leading multinational company in the fashion industry that funded this
project and provided the data analysed in this project. In particular, the authors thank Barbieri Pietro, Bortolan
Luca, Canale Matteo, Giada Marco, Longi Davide, Napol Fabio, Paganin Danny for their invaluable sugges-
tions. This study was carried out within the MICS (Made in Italy—Circular and Sustainable) Extended Partner-
ship and received funding from the European Union Next-GenerationEU (PIANO NAZIONALE DI RIPRESA
E RESILIENZA (PNRR)—MISSIONE 4 COMPONENTE 2, INVESTIMENTO 1.3—D.D. 1551.11-10-2022,
PE00000004). This study was carried out within the MOST—Sustainable Mobility National Research Centre
and received funding from the European Union Next-GenerationEU (PIANO NAZIONALE DI RIPRESA E
RESILIENZA (PNRR)—MISSIONE 4 COMPONENTE 2, INVESTIMENTO 1.4—D.D. 1033 17 June 2022,
CN00000023). This manuscript reflects only the authors’ views and opinions, neither the European Union nor
the European Commission can be considered responsible for them.

Int J Artif Intell, Vol. 13, No. 3, September 2024: 3052–3060

Int J Artif Intell ISSN: 2252-8938 ❒ 3059

REFERENCES
[1] A. Weckenmann, G. Akkasoglu, and T. Werner, “Quality management - history and trends,” TQM Journal, vol. 27, no. 3, pp.
281–293, 2015, doi: 10.1108/TQM-11-2013-0125.
[2] R. Y. Zhong, X. Xu, E. Klotz, and S. T. Newman, “Intelligent manufacturing in the context of industry 4.0: A review,” Engineering,
vol. 3, no. 5, pp. 616–630, 2017, doi: 10.1016/J.ENG.2017.05.015.
[3] C. A. Escobar and R. M. -Menendez, “Machine learning techniques for quality control in high conformance manufacturing envi-
ronment,” Advances in Mechanical Engineering, vol. 10, no. 2, 2018, doi: 10.1177/1687814018755519.
[4] A. Frankó and P. Varga, “A survey on machine learning based smart maintenance and quality control solutions,” Infocommunications
Journal, vol. 13, no. 4, pp. 28–35, 2021, doi: 10.36244/ICJ.2021.4.4.
[5] A. E. Smith, “Predicting product quality with backpropagation: A thermoplastic injection moulding case study,” The International
Journal of Advanced Manufacturing Technology, vol. 8, no. 4, pp. 252–257, 1993, doi: 10.1007/BF01748635.
[6] H. Tercan and T. Meisen, “Machine learning and deep learning based predictive quality in manufacturing: a systematic review,”
Journal of Intelligent Manufacturing, vol. 33, no. 7, pp. 1879–1905, 2022, doi: 10.1007/s10845-022-01963-8.
[7] H. S. Deshmukh, V. S. Jadhav, S. P. Deshmukh, and K. K. Rane, “Towards applicability of machine learning in quality control–a
review,” ECS Transactions, vol. 107, no. 1, pp. 14729–14737, 2022, doi: 10.1149/10701.14729ecst.
[8] G. Köksal, I. Batmaz, and M. C. Testik, “A review of data mining applications for quality improvement in manufacturing industry,”
Expert Systems with Applications, vol. 38, no. 10, pp. 13448–13467, 2011, doi: 10.1016/j.eswa.2011.04.063.
[9] J. Huber et al., “Method for classification of battery separator defects using optical inspection,” Procedia CIRP, vol. 57, pp. 585–590,
2016, doi: 10.1016/j.procir.2016.11.101.
[10] R. Wagner, B. Haefner, and G. Lanza, “Function-oriented quality control strategies for high precision products,” Procedia CIRP,
vol. 75, pp. 57–62, 2018, doi: 10.1016/j.procir.2018.04.069.
[11] F. C. Neto, T. M. Gerônimo, C. E. D. Cruz, P. R. Aguiar, and E. E. C. Bianchi, “Neural models for predicting hole diameters in
drilling processes,” Procedia CIRP, vol. 12, pp. 49–54, 2013, doi: 10.1016/j.procir.2013.09.010.
[12] P. G. Mathews and M. S. Shunmugam, “Neural-network approach for predicting hole quality in reaming,” International Journal of
Machine Tools and Manufacture, vol. 39, no. 5, pp. 723–730, 1999, doi: 10.1016/S0890-6955(98)00061-3.
[13] R. Wagner, J. Fischer, D. Gauder, B. Haefner, and G. Lanza, “Virtual in-line inspection for function verification in serial production
by means of artificial intelligence,” Procedia CIRP, vol. 92, pp. 63–68, 2020, doi: 10.1016/j.procir.2020.03.126.
[14] R. S. Peres, J. Barata, P. Leitao, and G. Garcia, “Multistage quality control using machine learning in the automotive industry,”
IEEE Access, vol. 7, pp. 79908–79916, 2019, doi: 10.1109/ACCESS.2019.2923405.
[15] A. Sumesh, K. Rameshkumar, K. Mohandas, and R. S. Babu, “Use of machine learning algorithms for weld quality monitoring
using acoustic signature,” Procedia Computer Science, vol. 50, pp. 316–322, 2015, doi: 10.1016/j.procs.2015.04.042.
[16] J. R. Quinlan, C4. 5: programs for machine learning. San Mateo, California: Morgan Kaufmann Publishers, 2014.
[17] S. Schorr, M. Möller, J. Heib, and D. Bähre, “Quality prediction of drilled and reamed bores based on torque mea-
surements and the machine learning method of random forest,” Procedia Manufacturing, vol. 48, pp. 894–901, 2020, doi:
10.1016/j.promfg.2020.05.127.
[18] E. Clayton, “Predictive modeling and visual perception algorithms, ambient sound recognition and processing tools, and deep
neural network and vision sensing technologies in the decentralized and interconnected metaverse,” Linguistic and Philosophical
Investigations, vol. 22, pp. 145–161, 2023, doi: 10.22381/lpi2220239.
[19] G. Lăzăroiu, M. Andronie, M. Iatagan, M. Geamănu, R. S, tefănescu, and I. Dijmărescu, “Deep learning-assisted smart process
planning, robotic wireless sensor networks, and geospatial big data management algorithms in the internet of manufacturing things,”
ISPRS International Journal of Geo-Information, vol. 11, no. 5, 2022, doi: 10.3390/ijgi11050277.
[20] G. Duncan, “Motion planning and remote sensing algorithms, predictive geospatial modeling and deep learning artificial intelli-
gence tools, and machine perception and image recognition technologies in the blockchain-based virtual economy,” Analysis and
Metaphysics, vol. 21, pp. 193–209, 2022, doi: 10.22381/AM21202212.
[21] M. Mohri, A. Rostamizadeh, and A. Talwalkar, Foundations of machine learning. Cambridge, Massachusetts: The MIT Press, 2018.
[22] J. A. Bunge and D. H. Judson, “Data mining,” in Encyclopedia of Social Measurement, Elsevier, 2005, pp. 617–624 , doi:
10.1016/B0-12-369398-5/00159-6.
[23] J. H. Friedman, “Multivariate adaptive regression splines,” The Annals of Statistics, vol. 19, no. 1, 2007, doi:
10.1214/aos/1176347963.
[24] S. N. Wood, Generalized additive models: an introduction with R. Boca Raton, Florida: Chapman and Hall/CRC, 2017 , doi:
10.1201/9781315370279.
[25] R. A. Fisher, “The use of multiple measurements in taxonomic problems,” Annals of Eugenics, vol. 7, no. 2, pp. 179–188, 1936,
doi: 10.1111/j.1469-1809.1936.tb02137.x.
[26] C. M. Bishop, Neural networks for pattern recognition. Oxford, United Kingdom: Clarendon Press, 1995.
[27] L. Breiman, “Random forests,” Machine Learning, vol. 45, pp. 5–32, 2001.
[28] B. E. Boser, I. M. Guyon, and V. N. Vapnik, “A training algorithm for optimal margin classifiers,” in Proceedings of the fifth annual
workshop on Computational learning theory, 1992, pp. 144–152 , doi: 10.1145/130385.130401.
[29] G. Ke et al., “LightGBM: A highly efficient gradient boosting decision tree,” Advances in Neural Information Processing Systems,
vol. 30, pp. 3147–3155, 2017.
[30] H. Zou and T. Hastie, “Regularization and variable selection via the elastic net,” Journal of the Royal Statistical Society. Series B:
Statistical Methodology, vol. 67, no. 2, pp. 301–320, 2005, doi: 10.1111/j.1467-9868.2005.00503.x.
[31] S. Rajeswari and K. Suthendran, “C5.0: advanced decision tree (ADT) classification model for agricultural data analysis on cloud,”
Computers and Electronics in Agriculture, vol. 156, pp. 530–539, 2019, doi: 10.1016/j.compag.2018.12.013.
[32] N. Japkowicz and S. Stephen, “The class imbalance problem: A systematic study,” Intelligent Data Analysis, vol. 6, no. 5, pp.
429–449, 2002, doi: 10.3233/ida-2002-6504.

Machine learning-based decision-making approach for predicting defects detection ... (Elena Barzizza)
3060 ❒ ISSN: 2252-8938

BIOGRAPHIES OF AUTHORS

Elena Barzizza has a Master’s degree in Management engineering at the University of

Padua, Department of Technique and Management of Industrial systems. She is research and teaching
assistant, and second year Ph.D. student in Management Engineering at the University of Padua. Her
research interests include big data analysis, machine learning, and non parametric statistics. She can
be contacted at email: [email protected].

Nicolò Biasetton holds a Master’s degree in Management engineering at the University of

Padua, Department of Technique and Management of Industrial systems. He is research and teaching
assistant, and final year Ph.D. student in Management Engineering. His research interests include
supervised and unsupervised machine learning algorithms, sentiment and emotional analysis. He can
be contacted at email: [email protected].

Riccardo Ceccato holds a Ph.D. in Management Engineering at the University of

Padua, Department of Technique and Management of Industrial systems. He is a lecturer in Ma-
chine Learning and Big Data, a member of the editorial board of the Journal of Applied Statistics
and a reviewer for several international scientific journals. His research interests include machine
learning algorithms and multivariate nonparametric techniques. He can be contacted at email: ric-
[email protected].

Alberto Molena holds a Master’s degree in Management Engineering at the Univer-

sity of Padua, Department of Technique and Management of Industrial systems. He is research
and teaching assistant, and first year Ph.D. student in Management Engineering. His research in-
terests include industrial statistics and design of experiment. He can be contacted at email: al-
[email protected].

Int J Artif Intell, Vol. 13, No. 3, September 2024: 3052–3060