Journal of Industrial Information Integration 41 (2024) 100666

Contents lists available at ScienceDirect

Journal of Industrial Information Integration

journal homepage: www.sciencedirect.com/journal/journal-of-industrial-information-integration

Full Length Article

Enhancing SMEs digital transformation through machine learning: A

framework for adaptive quality prediction
Ming-Chuan Chiu * , Yu-Jui Huang , Chia-Jung Wei
Department of Industrial Engineering and Engineering Management, National Tsing Hua University, Hsinchu, Taiwan

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

Keywords: As smart manufacturing expands, businesses see the importance of digital transformation, especially for small
Digital transformation and medium-sized enterprises (SMEs). Unlike larger companies, SMEs face greater challenges when undergoing
Adaptability digital transformation due to technological questions. However, recent advancements in high-performance
Deep learning
computing and reduced hardware costs have made deep learning-based digital transformation more finan-
Quality prediction
cially feasible for SMEs. While previous research utilized machine learning for product quality prediction, there
Small and medium enterprises (SMEs)
remains a lack of comprehensive in adaptive quality prediction specifically designed for SMEs. This study pre-
sents a systematic framework utilizing various machine learning methods and validates research cases using
CRISP-DM (Cross-Industry Standard Process for Data Mining). The first step involves applying XGBoost(eXtreme
Gradient Boosting)for feature selection, the second step utilizes GRU for parameter prediction. Finally, in the
third step, SVM (Support Vector Machine) is employed for quality classification. The integrated framework
achieves high accuracy, with R2 reaching 90 % for predicted parameters and nearly 95 % for classification in-
dicators. Moreover, this research addresses the research gap in quality prediction and adaptability and provide
an effective digital transformation solution for SMEs without substantial investment. The proposed research
framework can be applied SMEs of other domains, such as the machining and traditional manufacturing industry.

1. Introduction while improving product quality and efficiency as well as reducing costs
[6,7]. Furthermore, SMEs account for 99.83 % of total enterprises, 72.7
Over the past few decades, the manufacturing industry has leveraged % of total employment, and 50.6 % of total value-added, providing 55.1
technological innovation to boost economic growth. However, chal- % of exports and holding a crucial position in the economy [8]. Hence,
lenges such as decline of workforce due to an aging population, which SMEs need to actively seek opportunities for digital transformation to
cause difficulties for small and medium-sized manufacturing enterprises enhance SMEs market competitiveness and overcome industry
(SMEs) to recruit and maintain operational quality and viability [1]. challenges.
Additionally, international trade disputes and the COVID-19 pandemic In the past 20 years, digital transformation research has accumulated
have severely impacted the global supply chain and led to material various findings [9] and has proven to promote the efficient develop-
shortages and inflation issues [2]. SMEs, particularly in Asia, are now ment of manufacturing processes and enhance production capacity
facing the challenge of rapid technological transformation and digiti- competitiveness [10]. Quality prediction technology is a vital direction
zation [3]. Consequently, SMEs actively seek digital transformation so- in digital transformation, which helps manufacturers improve product
lutions to enhance SMEs’ market competitiveness [4,5]. Nonetheless, quality and production efficiency. Quality prediction technology facili-
SMEs often encounter more significant competitive pressure and diffi- tates automation and intelligence in the production process, enabling
culties compared to large enterprises due to resource and technological businesses to anticipate product quality in advance and enhance pro-
constraints. duction efficiency and cost reduction [11]. Therefore, the application
Digital transformation holds unique advantages in addressing the and development of quality prediction technology in the digital trans-
issues faced by SMEs. Through smart manufacturing applications, SMEs formation process can assist SMEs in enhancing market competitiveness
can adapt more flexibility to market changes and production challenges, and viability. However, no research currently provides a digital

* Corresponding author at: No. 101, Section 2, Kuang-Fu Road, Hsinchu, 30013, R.O.C, Taiwan.
E-mail address: [email protected] (M.-C. Chiu).

https://doi.org/10.1016/j.jii.2024.100666

Available online 18 July 2024

2452-414X/© 2024 Elsevier Inc. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
M.-C. Chiu et al. Journal of Industrial Information Integration 41 (2024) 100666

transformation solution tailored to SMEs with both quality prediction Industry 4.0 impacts the business models of manufacturing SMEs, of-
and adaptability. fering insights into the qualitative changes and adaptations required for
Taking the example of SMEs manufacturing industries, product these enterprises to thrive in the new digital era [25].
quality is often affected by factors such as temperature fluctuations and In recent years, with the advancement of technology and the rise of
machine equipment depreciation during the manufacturing process. To Industry 4.0, many businesses have begun a digital transformation to
simultaneously achieve capacity and yield, experienced operators play a adapt to the ever-changing market environment. Digital transformation
crucial role [12]. The most relevant factors influencing traditional become an essential part of modern enterprises and societal develop-
manufacturing quality are parameter adjustments and optimization ca- ment, as businesses need to keep pace with technological advancements
pabilities. However, with the rise of the semiconductor industry, fac- to maintain SME’s competitive edge and achieve greater success. Digital
tories often face severe technical personnel shortages, which leads to transformation has multiple definitions [26]. Bondar et al. (2017) [27]
product quality and output issues [13]. If quality prediction can be presented that the broader one defines it as a consistent networking of
accomplished in advance and provide scientific parameter optimization all economic sectors and as an adaption of actors to new circumstances
suggestions, it would not only shorten the trial mold time and reduce of the digital economy [27]. However, in the transformation process,
unnecessary production costs but also address the technical talent especially for SMEs, which are behind other industries. SMEs account for
shortage faced by SMEs [14]. Furthermore, achieving adaptability in 99.83 % of all enterprises and employ 72.7 % of the total workforce,
Industry 4.0, which involves predicting defects and making corre- playing a significant role in the economy. Ulas (2019) [8] presented
sponding parameter adjustments, could result in significant improve- SMEs should start the transformation of SMEs from manufacturing
ments in process capability [15]. technology and then transform the organizational structure and business
In the past, SMEs faced challenges in digital transformation, partic- culture to achieve the digital transformation of productivity [8]. How-
ularly in how to utilize artificial intelligence to predict product quality ever, Vilarinho et al. [28] presented that SMEs lag behind large com-
with limited manpower and budget. The main reason is that achieving panies in terms of digital transformation, which has a negative impact
full-quality inspection in mass-production scenarios is challenging. on the competitiveness of SMEs. Therefore, SMEs integrated artificial
Additionally, adaptive methods have not been widely adopted in SMEs, intelligence with their business operations to enable smart
which highlights a gap in the practical application of advanced tech- manufacturing [7]. Digital transformation can help SMEs maintain
nologies. This study aims to assist SMEs in digital transformation by competitiveness in the era of continuous technological development and
adopting a hybrid machine learning model. Through the utilization of intelligence [29]. The adoption of new digital technology applications
this method and its adaptability, it seeks to propel SMEs towards the has become the norm, and SMEs must adopt digital solutions to remain
final stage of Industry 4.0 maturity. The proposed method architecture competitive [30]. Digital transformation has a positive impact on
integrates XGBoost for feature selection, a GRU time series prediction corporate innovation [31]. Together with the enhancement of regional
model, and an SVM classification model, resulting in promising out- digital industry innovation levels, it can jointly promote corporate
comes in predicting parameters and product quality. innovation. There would be a spatial spillover effect to generate positive
The structure of the article is as follows: Section 1 provides the impacts on the surrounding areas [32]. Many SMEs recognized the
background and a brief introduction to the research. Section 2 conducts importance of digital transformation and started to actively adopt these
a literature review on related studies, starting with an exploration of technologies to improve production efficiency and competitiveness.
Industry 4.0 and digital transformation. It further delves into the stages However, SMEs need to comprehensively consider internal digital
of Industry 4.0 maturity and the concepts and significance of adapt- transformation, especially since SMEs often limited resources for plan-
ability. Additionally, it discusses time series prediction models. Section ning and implementing digital projects [33]. Small and medium-sized
3 elaborates on the proposed method architecture. It includes the use of enterprises often face challenges in digital transformation [34],
hybrid machine learning methods to improve prediction accuracy and including issues related to technology and cybersecurity [35]. Urbach &
forecast yield. Moreover, it employs machine learning techniques to Röglinger (2018) [36] presented that in addition to financial and time
analyze key factors and provide adjustment recommendations, con- resources, SMEs often lack the necessary expertise and skills [36]. The
structing a comprehensive prediction system with adaptive capabilities. success of SMEs in overcoming competency and skill gaps, particularly
Section 4 presents the case study and practical application on the pro- in digitalization, hinges on developing relevant skills through various
duction line to enhance product process capability indicators. Finally, methods such as training, sharing best practices, hiring external experts,
Section 5 discusses the findings and outlines future research directions. reducing resistance, and enhancing the adoption of digital tools [37,38].
Despite the emergence of many digital transformation-related studies, a
2. Literature review comprehensive method combining feature selection, prediction, and
classification as a framework to provide SMEs with a digital trans-
2.1. The development of digital transformation in industrial SMEs formation solution combining quality prediction and adaptability didn’t
develop. The study aims to combine quality prediction with Industry 4.0
With the advent of Industry 4.0, an increasing number of enterprises adaptability, analyzing and grasping the key factors for successful digital
adopted its principles and technologies to enhance performance and transformation and successfully assisting SMEs in completing their
productivity [16,17]. In recent years, numerous studies related to digital digital transformation journey.
transformation have been developed [18–20]. In view of the extremely
high implementation costs of Industry 4.0, these potential benefits 2.2. Adaptability
should be properly assessed [21]. However, during the transformation
process, the manufacturing industry, especially SMEs, found itself Schuh et al. (2020) [39] presented that adaptability in Industry 4.0 is
behind in comparison to other sectors [22]. Chen et al. (2019) [23] defined as the ultimate goal, where enterprises can automatically
presented that Industry 4.0 and digital transformation are crucial ad- respond with the most favorable strategies based on predicted events,
vantages for SMEs in the manufacturing industry, offering opportunities providing solutions to on-site operators [39]. The concept of Industry
to improve future competitiveness and survival through increased pro- 4.0, first introduced at the Hannover Messe in Germany, has gained
ductivity, reduced operational costs, enhanced product quality, and widespread acceptance in business and academic circles. It outlines the
product innovation [23]. The integration of Industry 4.0 and digital future trajectory for the manufacturing sector, emphasizing the inte-
transformation presents both new opportunities and challenges, and is gration of emerging technologies into manufacturing and supply chain
expected to become a key success factor for SMEs in the coming years processes [40]. Javaid et al. (2022) [41] highlighted that the benefits of
[24]. Parra-Sánchez and Talero-Sarmiento (2023) [25] analyzed how adopting this approach include enhanced competitiveness and

2
M.-C. Chiu et al. Journal of Industrial Information Integration 41 (2024) 100666

performance, increased versatility and flexibility, and improved profit- techniques to analyze and anticipate product quality, ensuring align-
ability [41]. In the context of Industry 4.0, manufacturing systems aim ment with specific quality standards. Instances of quality issues like
to achieve the direction of smart manufacturing, where Cyber-Physical rework, maintenance, or disposal result in significant time and monetary
Systems (CPS) integration is considered a core technology of Industry inefficiencies during production [51]. Through extensive product data
4.0 [42]. Kumar et al. (2019) [43] presented that CPS includes Internet collection and analysis, predictive models are developed for accurate
of Things, wireless embedded network systems, cloud computing, big anticipation and effective quality control. Bai et al. (2017) [52] pre-
data, and artificial intelligence in manufacturing production systems sented that predicting product quality aids businesses in reducing defect
[43]. Industry 4.0, based on the foundation of CPS integration, enables rates and quality-related concerns, ultimately enhancing production
rapid, flexible, and efficient data collection and analysis across machines efficiency and elevating customer satisfaction, thus enhancing their
[44]. As the intelligent digital transformation in the manufacturing in- competitive advantage [52]. As a central element of smart
dustry is still in its early stages, Rajnai et al. (2018) [45] presented that manufacturing, product quality prediction plays a vital role in process
there is no universally accepted and established method for assessing an analysis. Ren et al. (2020) [53] presented that product quality serves as a
enterprise’s readiness for Industry 4.0 [45]. However, to assist enter- feedback mechanism for adjusting production process parameters,
prises in understanding their level of industrialization and to help en- facilitating evaluation and improvement, thus averting rework occur-
terprises plan the development in Industry 4.0, the National Academy of rences and resource wastage [53]. Consequently, enhancing product
Science and Engineering in Germany (Acatech) defined the Industry 4.0 quality remains a prime focus for contemporary enterprises. With
maturity index, as shown in Fig. 1. ongoing advancements in computer and artificial intelligence technol-
Stich et al. [46] presented that Industry 4.0 is divided into six stages, ogies, integrating progressive concepts into quality management for
namely Computerization, Connectivity, Visibility, Transparency, Pre- automation and intelligence has become a notable trend [54]. Saadallah
dictability, and Adaptability [47]. Each stage builds upon the founda- et al. (2022) [55] presented that the continual evolution of sensors in
tion of the previous one, with detailed explanations and divisions of the smart factories and the industrial Internet of Things (IoT) presents
required capabilities and actual benefits for each stage. As this study enhanced opportunities to apply data-driven methods to machine
focuses on the predictability and adaptability of Industry 4.0, we will learning for quality prediction [55]. As applications expand, meeting
introduce the fifth and sixth stages. The fifth stage, predictability, is quality requirements for predictive machine learning models across
characterized by enterprises being able to collect historical data, extract diverse dimensions becomes a significant challenge in manufacturing
valuable information and knowledge, and adopt accurate data predic- processes. A primary obstacle is improving the precision of quality
tive models to forecast and judge potential events in future monitoring and prediction. Effectively extracting relevant information
manufacturing scenarios, enabling decision-makers to anticipate out- gains importance, prompting the integration of data-driven model
comes in advance [48] and take corresponding decisions. The sixth methods from the field of artificial intelligence. Fertig et al. (2022) [56]
stage, Adaptability, is defined by enterprises automatically responding presented that these methodologies capture interactions, internal data
with the most favorable strategies based on predicted events and correlations, and data relationships to extract knowledge for predicting
providing optimized solutions to on-site operators [39]. The application unfamiliar data [56]. Due to the nature of quality prediction, which
of adaptive technologies for continuous improvement in enterprises is a involves detection or projection, the approach primarily involves su-
critical strategy in Industry 4.0 [47–50]. The goal of the proposed pervised learning methods. Supervised learning aims to estimate values
framework in this study is to achieve the fifth stage of predictability and or categorize targets (regression or classification) based on selected
the sixth stage of adaptability. This will be accomplished by predicting input variables [57]. In this context, machine learning and deep learning
product quality to achieve predictability and analyzing critical factors to techniques can be applied. Common machine learning methods include
provide recommendations for parameter optimization to enhance yield, SVM, ANN, and MLP, while deep learning models encompass CNN,
also achieving adaptability in Industry 4.0. LSTM, and other variations. The strategies for quality prediction exhibit
variation, often contingent on data availability and intended applica-
tion. Msakni et al. (2023) [58] utilized time series data and machine
2.3. Application of deep learning in quality prediction learning algorithms for quality control in manufacturing [58]. Abd
Elnaby et al. [59] combined ML with Six Sigma DMAIC to enhance
Product quality prediction employs statistical and machine learning

Fig. 1. Six stages of industry 4.0 maturity [39].

3
M.-C. Chiu et al. Journal of Industrial Information Integration 41 (2024) 100666

plastic bottle production quality, significantly reducing defects and does not meet the standard, on-site personnel can adjust and optimize
costs, and improving the Sigma level at an Egyptian company [60]. The parameters in real-time based on the key factors identified by machine
convergence of machine and deep learning with quality underscores learning to enhance product yield and reduce production costs. Our
three fundamental tasks: quality factors identification, quality predic- approach also achieves the predictability and adaptability stages of In-
tion, and quality classification. Quality factors identification involves dustry 4.0. The overall method framework is illustrated in Fig. 2.
recognizing, evaluating, and clarifying the connections between process
variables and product quality with an aim to gain comprehensive in- 3.1. Data preprocessing
sights into process interdependencies [61].
From recent research, several important conclusions can be drawn: In the data preprocessing part, although the data provided by the
(1) Predicting quality is a highly diverse and extensively researched field companies has a certain level of completeness, there are still some
within the manufacturing industry. The relevance and popularity of this missing values. Therefore, in this study’s data preprocessing, the mean
field are likely to continue growing in the coming years. (2) General imputation technique is adopted to handle missing values. The data is
methods for quality prediction encompass three main steps: the first step then normalized to fall within the range of 0 and 1. This method elim-
involves data collection and aggregation, the second step is the training inates the comparison limitations caused by different data units,
of predictive models, and the third step involves utilizing the prediction providing a consistent basis for comparative analysis. The formula for
results as references for practical operational decisions. (3) Given the normalization is as follows (Eq. (1)), where X represents the original
increasing availability of data in manufacturing, along with advance- data, Xmin is the minimum value in the time series data, Xmax is the
ments in computational technology and corresponding hardware and maximum value in the time series data, and Xscaled represents the new
software solutions, quality inspection based on predictive models pre- transformed value.
sents a promising and cost-effective approach to inspection processes.
(4) The current state of research underscores the immense potential of X − Xmin
Xscaled = (1)
machine learning and deep learning methods in enhancing quality Xmax − Xmin
assurance and inspection processes.
3.2. Model training
2.4. Discussion
In the model training part, this study will apply XGBoost for feature
From previous literature, the application of Industry 4.0 and digital selection, which will be described in detail in Section 3.2.1. Next, the
transformation in small and medium-sized enterprises (SMEs) is GRU (Gated Recurrent Unit) will be utilized for predicting machine
currently limited [62]. Previous research showed that due to resource parameters, with a comprehensive description and method introduction
and manpower constraints, as well as inadequate technological capa- in Section 3.2.2. Finally, SVM (Support Vector Machine) will be
bilities, few SMEs have been able to simultaneously achieve quality employed to predict the future product quality, with a detailed
prediction and Industry 4.0 adaptability in digital transformation efforts description and method introduction presented in Section 3.2.3.
[63,64]. In addition, the field of quality prediction didn’t offer suitable
measures for addressing defective products, which suggests that the final 3.2.1. Using XGBoost for feature selection
stage of achieving full Industry 4.0 adaptability has not been reached XGBoost, short for eXtreme Gradient Boosting, is one of the boosting
[65]. Therefore, developing a simplified and practical digital trans- algorithms and a common supervised learning technique. The concept of
formation plan that combines adaptability and quality prediction has this algorithm is to combine a large number of classifiers to form one
become the main objective of this research. Furthermore, with the rapid classifier. In other words, it continuously performs feature selection to
development of high-performance computing and the decrease in generate a tree and learn a new function to compensate for the residuals
hardware costs in recent years, deep learning become a financially of the previous prediction [66]. Each tree is correlated with others,
feasible option for SMEs [10]. aiming to correct the mistakes caused by the preceding tree. This
Previous studies have not sufficiently addressed the combination of method combines the advantages of Bagging and Boosting to solve
feature selection and time series prediction models in a manner that classification and regression problems efficiently.
supports the specific needs of SMEs. This study aims to fill this gap by Additionally, XGBoost offers feature selection functionality [67].
proposing a novel method that integrates XGBoost for feature selection Feature selection helps to identify the most useful subset of features in
with a GRU time series model for predicting machine parameters, the original data to enhance the accuracy of classification or model. It is
complemented by an SVM classification model. This hybrid approach is used to quantify the impact between features, eliminate redundant or
designed to facilitate digital transformation, improve adaptability, and irrelevant features in machine learning, create predictive models, and
implement effective quality prediction in SMEs, thereby guiding them find feature combinations related to the predicted class. In XGBoost,
towards the advanced stages of Industry 4.0 maturity. significant features are chosen as nodes in the tree, and the feature
magnitude minimally influences the model’s performance. By extracting
3. Method appropriate features, we can improve the training accuracy [68].
XGBoost can be represented in an additive form, as shown in Eq. (2)
The method framework of this study consists of four steps. In step 1, [69]:
data collection and preprocessing are conducted. In step 2, Machine K
∑
learning techniques are employed to determine the correlation among yi =
̂ fk (xi ), fk ∈ F (2)
various machine data and identify key factors related to product quality. k=1

After filtering the critical factors for machine parameters, a time series
Where ̂ y i represents the predicted value of the model, K denotes the
prediction model is utilized to forecast the machine parameters.
number of decision trees, fk represents the k-th submodel, xi represents
Simultaneously, an SVM classifier is trained to distinguish between good
the i th input sample, and F represents the set of all decision trees. The
and defective products. Subsequently, the learned SVM classifier is
XGBoost objective function consists of two parts: the loss function and
employed to determine whether the predicted future combination of
the regularization, as shown in Eqs. (3) and (4):
machine parameters will result in defective products, achieving the goal
n
of quality prediction and calculating the yield to know the future batch’s ∑ ( 1)
)
L(φt ) = l yi , ̂
y i(t− + ft (xi ) + Ω(fk ) (3)
yield of products. In step 3, we use various relevant indicators to validate i=1
and assess the feasibility of the model. In step 4, if the predicted yield

4
M.-C. Chiu et al. Journal of Industrial Information Integration 41 (2024) 100666

Fig. 2. Proposed framework.

second derivative of sample xi , ωj represents the output value of the j-th

Ω(fk ) = γT + λ‖ w ‖2 (4)
leaf node, and Ij represents the set of samples in the j-th leaf node. From
Where L(φt ) represents the objective function at the t-th iteration, Eq. (6), it can be observed that the objective function is a convex
yi
̂ (t− 1)
represents the target value at the (t-1)-th iteration, and Ω(fk ) function. By taking the derivative of ωj and equating the differential
represents the regularization at the t-th iteration. The purpose of regu- function to zero, the objective function can achieve its minimum value
larization is to reduce overfitting, and γ and λ are coefficients to control at ωj ,as shown in Eq. (6).
the simplicity of the decision trees. T represents the number of leaf nodes ∑
i∈Ij gi
in the model. The objective function can be expanded using the Taylor ω∗j = − ∑ (6)
expansion as shown in Eq. (5): i∈Ij hi + λ

∑n [
1
]
1 ∑ T Eq. (7) is used to evaluate the quality of the tree model. A smaller
L(∅) ≅ gi ft (xi ) + hi ft2 (xi ) + γT + λ ω2j value indicates a better model. We can obtain the scoring formula for
i=1
2 2 j=1
⎛ ⎞ splitting nodes in the tree as follows:
T ∑
∑ 1 ⎝∑ (∑ )2
[ gi ωj + + λ⎠ω2j + γT (5)
i∈Ij gi
≅ T
2 i∈I ∧ 1∑
j=1 i∈I j j L(∅)min = − ∑ + γT (7)
2 j=1 i∈Ij hi + λ
Where gi represents the first derivative of sample xi , hi represents the

5
M.-C. Chiu et al. Journal of Industrial Information Integration 41 (2024) 100666

Tree-based ensemble learning algorithms such as Random Forest and

ht = (1 − zt ) ⊙ ht− 1 + h̃t ⊙ zt (11)
XGBoost can generate rankings of individual features based on Gini
importance, enabling the integration of these rankings to quickly reduce Two activations, σ (sigmoid) and tanh, are used to model the gates
the search space [70]. In XGBoost, the relative importance of features and normalize the input. rt and zt are the outputs of the reset gate and
can be measured using several metrics, such as split weights [71]. Split update gate, respectively, and used to remember the past weights
weight represents the number of times a feature is used to split the data through two hidden layers h̃t and ht , which are updated after each
in all boosting trees. More important features are utilized more operation. After feature selection of machine parameters, this study
frequently during the construction of boosting trees, while other fea- utilizes the Gated Recurrent Unit to predict future machine parameters.
tures are used to improve the residuals. Gain, on the other hand, mea-
sures the actual reduction of impurity at a node and represents the 3.2.3. SVM classifier determines the classification of good and defective
average gain of all splits using a particular feature. After fitting the products
XGBoost model, feature rankings based on weight and gain can be ob- The SVM (Support Vector Machine) algorithm was initially proposed
tained. The optimal feature subset can be selected based on the trade-off by Vapnik in 1963 [75] as a method for constructing linear classifiers
between learning performance and model simplicity (i.e., fewer fea- [75]. SVM is a statistical-based supervised learning algorithm in ma-
tures). For average precision reduction, a significant drop in accuracy chine learning. Its concept involves finding a hyperplane, also known as
indicates highly correlated and useful features that contribute to the decision boundary, through model training to maximize the margin
learning improvements, and vice versa. In contrast, an irrelevant feature between two classes. A wider margin allows the model to have better
contributes minimally, and a redundant feature contributes limitedly classification performance on new data, distinguishing the two classes
due to its high correlation with other more important features. There- [76]. The points closest to the hyperplane are referred to as support
fore, redundant and irrelevant features can be removed without sacri- vectors. Given a labeled training dataset as follows:
ficing accuracy. The benefits of feature selection include better
interpretability, simplified modeling, shorter learning time, and (x1 , y1 )⋯⋯(xn , yn ), xi ∈ Rd and yi ∈ (− 1, +1) (12)
enhanced generalization [71].
Among them, xi represents the feature vector, and yi is the class label
In this study, XGBoost’s importance scores are used for feature se-
(negative or positive) of the training set. The optimal hyperplane can be
lection. If there is a significant difference between the importance scores
defined as follows:
and the previous ones, the later factors may be considered as irrelevant
or redundant features and will not be included as variables in the sub- wxT + b = 0 (13)
sequent prediction model. Since the machine has numerous parameters
Where w is the weight vector, x is the input feature vector, and b is
and it is difficult to determine the interactions or correlations between
the bias. w and b satisfy the following inequality for all elements in the
them, XGBoost is employed for feature selection in predicting each
training set:
machine parameter.
{
wxTi + b ≥ 1 if yi = 1
3.2.2. Predicting machine parameters using GRU (14)
wxTi + b ≤ 1 if yi = − 1
Time series is a sequence of data points ordered by time with time-
dependent interactions between them. Time series reflects the devel- The objective of training the SVM model is to find w and b in order to
opment process and inertia of socio-economic phenomena and is used to separate the data with a hyperplane and distinguish the classes. Simul-
predict the trend. It is a set of observed results reflecting the develop- taneously, the hyperplane is defined as 2 / || w ||, where || w || is the
ment process of socio-economic phenomena. Recurrent neural networks norm of the vector. The above example represents all data can be
are often utilized for time series prediction with time dependency. perfectly classified, which belongs to the hard-margin SVM. In cases
Traditional recurrent neural networks do not perform well in long-term where perfect separation is not possible, it falls under soft-margin SVM.
memory, meaning recurrent neural networks cannot capture long-term This means that there may be misclassifications within the samples, and
dependencies. Therefore, Long Short-Term Memory (LSTM) networks error must be minimized. Let εi be non-negative slack variables repre-
and Gated Recurrent Unit (GRU) were introduced to address this issue. senting misclassifications within the samples. In most cases, εi = 0,
LSTM resolved the problem of vanishing gradients in recurrent neural indicating correct classification. Let C be the parameter controlling the
network models [72]. However, LSTM has slower training speed, which penalty for misclassifications, where a larger value of C implies a greater
led to the emergence of GRU to accelerate training and reduce memory penalty for misclassification errors. Another criterion for calculating w
while achieving the same accuracy as LSTM [73]. In fact, GRU has been and b in SVM is to minimize all misclassification errors of the training
proven to perform better on some smaller and less complex datasets. The samples [77]. Therefore, the optimization problem for computing w and
difference in structure between GRU and LSTM lies in the number of b can be expressed as follows:
gates used. GRU adopts only two gates, combining the input and forget
M
gates of LSTM into a single update gate. This simplifies the computation 1 ∑
miinimize ‖ w ‖2 + C εi
compared to LSTM and results in faster training. The internal structure 2 i=1
of GRU is similar to that of LSTM, but GRU combines the update and { (15)
yi (w⋅xi + b) ≥ 1 , i = 1, …, n
reset gates into a single update gate. The model has two gates: an update subject to
gate that controls the scope, retaining previous information in the cur- εi , i = 1, …, n
rent state, and a reset gate that determines whether to associate previous
When the input space is non-linear, the soft-margin SVM cannot find
information with the current state. The values of the update gate and
a hyperplane that minimizes the number of misclassified data points and
reset gate are determined by the previous hidden state ht− 1 and the
performs well. To address this, data can be transformed into a higher-
current input xt . The operation process of GRU can be described by the
dimensional space, known as the kernel space, using a kernel function
following four equations [74]:
K(x, x1 ). x and x1 represent two different vectors. Consequently, by
rt = σ (wt ⋅[ht− 1 , xt ]) (8) introducing the kernel function K(x, x1 ) into the solution of the general
optimal separating plane, non-linear transformations can achieve linear
zt = σ (wz ⋅[ht− 1 , xt ]) (9) classification. The following quadratic optimization problems can be
uniformly solved:
h̃t = tanh(wh ⋅[rt ⊙ ht− 1 , Xt ]) (10)

6
M.-C. Chiu et al. Journal of Industrial Information Integration 41 (2024) 100666

n n
∑ 1∑ ( )
TP + TN
maxw(a) = ai − ai aj − yi yj K x, x1 (16) Accuracy = (20)
i=1
2 ij=1 TP + FP + FN + TN
Where ai is the Lagrange multiplier corresponding to each sample, TP
the generalized optimal separating plane decision function can be Precision = (21)
TP + FP
expressed as:
( ) TP
∑ ( ) Recall = (22)
f (x) = sgn yi ai K x, x + b
1
(17) TP + FN
s,v
2⋅precision⋅recall
In this study, the classification has two classes (Good and Defective). F1 score = (23)
precision + recall
Each machine parameter is associated with a class label. SVM classifies
new data points by substituting them into the decision function. If the The evaluation metrics in the second stage are also crucial as the core
result of the calculation is greater than 0, the data point is classified into of this study is to predict product quality in advance. If the predicted
one class; if it is less than 0, it is classified into the other class. machine data can accurately classify the product quality, it demon-
strates the effectiveness of the proposed method in this research.
3.3. Model validation
4. Case study
The performance evaluation of the model in this study is divided into
two parts: prediction and classification. First is the difference between This case study focuses on a small to medium-sized enterprise located
predicted machine data and actual machine data, which is commonly in Taiwan, specializing in the production of IC packaging trays and
known as Mean-Square Error (MSE). It is a measure of accuracy that memory card cases. To ensure the research meets scientific validation,
represents the standard deviation of the differences between predicted we have employed the CRISP-DM (Cross-Industry Standard Process for
and observed values. The definition is given by formula (18), assuming Data Mining) framework [78] for analysis. The study is divided into six
there are n validation data points (i = 1,2,…,n), where ̂ y i represents the main phases: Section 4.1 provides a detailed introduction to Business
predicted machine data and yi represents the actual machine data. Understanding; Section 4.2 elaborates on Data Understanding and Data
Preparation; Section 4.3 explains Modeling, Evaluation, and Deploy-
n
1∑ ment; and finally, a discussion is conducted in Section 4.4.
MSE = y i )2
(yi − ̂ (18)
n i=1

Another indicator to measure the interpretability of the prediction 4.1. Business understanding
model is the coefficient of determination R2 , which measures the pro-
portion of the variance in the dependent variable that is predictable According to, in the Business Understanding phase, we engage in
from the independent variables. The closerR2 is to 1.0, the more discussions with the enterprise to gain a deep understanding of the
explanatory power the model has, and an R2 value of 0.70 indicates a specific problems faced by the business. With outstanding design and
production capabilities in IC packaging trays and tape packaging, the
representative model capability [78]. The formula for calculating R2 is
company has become a well-known IC packaging tray manufacturer in
shown in formula (19), where SSres represents the residual sum of
Taiwan. However, in the past, the company heavily relied on-site ad-
squares, and SStot represents the total sum of squares.
justments for parameters. Due to the need to validate of new process
SSres parameters through repeated Design of Experiments (DOE), a significant
R2 = 1 − (19)
SStot amount of time is required to adjust the parameters [79]. This resulted
The second part involves measuring the classification performance in a waste of production capacity, manpower, resources, and equipment.
using a binary Confusion Matrix, as shown in Table 1. True Positive (TP) To address these challenges, this research proposes a method framework
represents cases where the model correctly identifies the product as for quality prediction. It utilizes the sensors and data collection system
good when it is actually good. False Positive (FP) indicates cases where of the small to medium-sized enterprise to gather data during the pro-
the model incorrectly identifies the product as good when it is actually duction process. Through data analysis and deep learning techniques,
defective. False Negative (FN) represents cases where the model incor- accurate prediction of injection molding machine parameters and
rectly identifies the product as defective when it is actually good, and product yield is achieved. Simultaneously combined with Human Ma-
True Negative (TN) indicates cases where the model correctly identifies chine Interface (HMI), it provides operators with better parameter
the product as defective when it is actually defective. suggestions. This digital transformation process not only reduces labor
Furthermore, the performance is measured using the following costs but also improves process capability indicators, promoting the
evaluation metrics: Precision, Recall, F1-score, and Accuracy. These digital transformation of small and medium-sized enterprises. The
metrics are calculated as shown in Eqs. (20) to (23). Larger values of completion of an Industry 4.0 adaptive system enables adjustment rec-
these metrics indicate better model performance. ommendations based on critical factors if the predicted yield does not
meet expectations, further enhancing process capability and indirectly
reducing costs and manpower. The proposed digital transformation
framework is illustrated in Fig. 3, and the results of the method vali-
dation will be presented in this section.
Table 1
Confusion matrix.
4.2. Data understanding and preparation
Actual

Positive (Good Negative (Defective Subsequently, we proceed with the Data Understanding and Data
product) product) Preparation phases of CRISP-DM. The company provided 17 types of
Predicted Positive (Good TP FP sensor data from the machine. In addition to the sensor data, the com-
product) pany also provided the weight of the product at each mold opening. If
Negative (Defective FN TN the weight exceeds the upper or lower control limit, it is considered a
product)
defective product; otherwise, it is considered acceptable. As mentioned

7
M.-C. Chiu et al. Journal of Industrial Information Integration 41 (2024) 100666

Fig. 3. System architecture diagram.

in Section 3, this research aims to identify key factors related to product adjusted in order of importance. Correctly setting these parameters en-
quality and provide real-time parameter adjustment and optimization sures complete filling and uniform cooling of molded parts, meeting the
based on these factors to improve product yield and reduce production required dimensions and appearance performance. The adjustment
costs. Therefore, this research applied the XGBoost model to determine recommendations will be presented in Section 4. Furthermore, the
the importance of the 17 sensors on the product weight. This allows on- actual sensors and corresponding variables are shown in Fig. 4.
site personnel to identify possible reasons for product defects. The
importance of each key factor related to product quality is shown in
Table 2. 4.3. Modeling, evaluation, and deployment
From Table 2, we can observe a huge gap in importance between "#6
1st stage holding average pressure" and "#7 1st stage injection lowest Next, we proceed with the Modeling and Evaluation phases of CRISP-
position". The top six key factors are crucial references for on-site DM. Section 4.3.1 explains the training process and outcomes to
personnel in adjusting machine settings, and the operator will be examine the importance and impact of each step in the research method.
Section 4.3.2 compares this study with other methods. Finally, we carry
out the Deployment phase of CRISP-DM, offering recommendations for
Table 2 company parameters, with the results detailed in Section 4.3.3.
Evaluation of the importance of product quality.
Item# Variable name Importance 4.3.1. Model training results
In GRU model training, the impact of different hyperparameters on
1 Plasticizing position 0.139
2 1st stage injection maximum pressure 0.133 the model is evaluated through experimental design. A Taguchi method
3 2nd stage injection maximum pressure 0.128 with 3 factors and 4 levels is employed for hyperparameter optimiza-
4 2nd stage injection time 0.114 tion, considering the following hyperparameters: Dropout rate (0.2, 0.3,
5 2nd stage injection screw distance 0.111 0.4, 0.5), Optimizer (Adam, Adagrad, Adadelta, RMSprop), and
6 1st stage holding average pressure 0.095
7 1st stage injection lowest position 0.061
Learning rate (0.01, 0.005, 0.001, 0.0005). The L16 orthogonal table is
8 1st stage injection screw distance 0.049 employed to select the best combination of hyperparameters, and the
9 1st stage injection time 0.037 results can be found in the appendix.
10 1st stage holding screw distance 0.031 Moreover, the GRU gated recurrent neural network is built using the
11 Temperature C 0.030
Keras package in Python. The GRU handles regression problems, with 64
12 Temperature E 0.029
13 Temperature A 0.018 neurons in the hidden layer, 5 key variables in the input layer, and 1
14 Temperature D 0.010 neuron in the output layer for predicting machine parameters. The batch
15 Temperature B 0.003 size is set to 128, trained for 500 epochs, with a learning rate of 0.001,
16 1st stage holding pressure position 0.002 and utilizing the Adaptive Moment Estimation (Adam) algorithm along
17 1st stage holding pressure time 0.000
with the mean squared error (MSE) loss function to evaluate the gap

8
M.-C. Chiu et al. Journal of Industrial Information Integration 41 (2024) 100666

Fig. 4. Actual photos of sensors.

between the computed values and the actual values. Table 3 compares
three types of recurrent neural networks: RNN, LSTM, and GRU. All
three models adopt the same feature selection method for prediction.
The results indicate that LSTM and GRU outperform the traditional RNN
in terms of evaluation metrics.
In Part 3, the SVM classifier determines whether a product is good or
defective. The case data is split into a training set and a test set in an 8:2
ratio, and the SVM model is trained. Fig. 5 shows the confusion matrix of
the SVM classifier for classifying good and defective products, where
0 represents good products, and 1 represents defective products. From
this confusion matrix, the classification model achieves an accuracy of
93 %, precision of 95 %, recall of 96 %, and F1 score of 95 %. In other
words, the proposed method in this study also demonstrates a certain
level of accuracy in predicting yield.

4.3.2. Comparison with other methods

The comparison between this study and other methods is presented
in Table 3. Previous studies such as Zewdie et al. (2021) [80] proposed
using Random Forest (RF) + LSTM for data-driven prediction of Total
Electron Content (TEC) [80], Adaboost + GRU for crude oil price pre-
diction [81], and XGBoost + GRU without feature selection for predic- Fig. 5. Confusion matrix for quality prediction.
tion. The results of this study are compared with the aforementioned
four methods. Feature selection methods can generally be categorized
into three categories: filter methods, wrapper methods, and embedded Table 4
methods. Filter methods, like statistical methods, only evaluate the Performance comparison of prediction methods.
correlation between variables and the prediction value without XGBoost GRU RF+LSTM Adaboost+GRU XGBoost +GRU
considering the relationship between variables. As a result, filter
MSE 0.0174 0.0138 0.0045 0.0037 0.0025
methods tend to select redundant variables. Therefore, filter methods R2 0.78 0.83 0.868 0.882 0.903
are not considered in this study, and instead, wrapper methods and
embedded methods, such as RF and XGBoost, are compared. Wrapper
methods and embedded methods can detect the mutual impact between 4.3.3. On-site validation of self adaption
variables and are more likely to discover features that improve the From Sections 4.1 to 4.3, it is evident that the proposed method in
model for the machine learning algorithm we intend to utilize. From the this study exhibits a certain level of accuracy in predicting both machine
results in Table 4, it can be observed that the proposed method frame- parameters and product quality. Although the model training and results
work outperforms other feature selection methods and the method meet a certain standard, it means that we can predict future yield.
without feature selection, including commonly employed Adaboost and However, if we only predict yield, it can only serve as a reference for the
RF. This indicates that the proposed method framework exhibits higher company, and the decisions that can be taken are either to stop pro-
accuracy. duction immediately or to continue. It may help avoid unnecessary
waste, but if we can anticipate that future yield may not meet the
standard and know which key factors affect the quality, we no longer
Table 3 rely on the experience or intuition of operators. Instead, we apply data
Comparison of RNN, LSTM, and GRU performances. analysis based on machine learning to identify key factors and provide
RNN LSTM GRU adjustment recommendations to on-site personnel. This not only reduces
MSE 0.0036 0.0027 0.0025 unnecessary waste but also improves process capability, which brings
R2 0.831 0.901 0.903 greater benefits to the enterprise. For the case study in this research, as

9
M.-C. Chiu et al. Journal of Industrial Information Integration 41 (2024) 100666

the predicted yield does not meet the company’s established standard, perfectly align with the machine’s configured settings. Therefore, after
the proposed method utilizes XGBoost to analyze key factors influencing discussing it with the on-site engineers, we propose adjusting the orig-
quality and adjust directions based on the deviation of defective inal machine setting from 92.3 to 95. The adjustment recommendations
parameters. for other key factors follow the same approach as described above. To
Fig. 6 presents the average values of good and defective products for verify whether the yield improved after the adjustments, we conducted
different factors and provides adjustment directions. By utilizing his- an experiment during the first iteration. Before the adjustments, we
torical data of good and defective product machine parameter values, tested 21,024 products, out of which 18,578 were good products,
the company can make directional adjustments. For example, in the " #1 resulting in a yield of 88.37 %. After the adjustments, we tested 22,580
Plastic Position " key factor, the average machine parameter value for products, with 20,366 being good products, resulting in a yield of 90.19
defective products is 93.05, whereas for good products, it is 94.87. %.
Therefore, it is clear that the set value needs to be increased to avoid Finally, we implement the method and parameter recommendations
more defective products. But, machine parameter values may not in the company. To validate the adaptability of the proposed method

Fig. 6. Suggestions for parameter variations.tiff.

10
M.-C. Chiu et al. Journal of Industrial Information Integration 41 (2024) 100666

framework, we requested the company to continue collecting data to issues if the predicted yield does not meet the standard, enabling pro-
verify the continuous improvement of yield. Under the same machine active measures based on future possibilities and achieving the concept
and product conditions, we obtained data on yield improvement after six of Industry 4.0 adaptability.
adjustments from November 2022 to April 2023, as shown in Table 5 Kumar et al. (2020) [85] proposed a network model for establishing
and Fig. 7. The 1st instance demonstrates the results of our proposed upper and lower control limits and detecting process parameter faults,
parameter recommendations. After discussing this method and its out- providing a quality control system based on intelligent data for injection
comes with industry and academic professionals, it was implemented molding products [85]. Farahani et al. (2022) [84] further introduced a
within the company. Fig. 7 showcases the results of the company’s ad- predictive maintenance framework that effectively predicts and man-
justments using our method, which highlights the 2nd to 6th instances ages maintenance operations in plastic injection molding plants and
(implemented by the company). The yield after each of the six adjust- demonstrates the capability to predict cooling issues by utilizing inte-
ments is higher than the yield before adjustments, with the overall yield grated data from injection machines and in-mold sensors [84]. Although
reaching above 90 % after the six adjustments. The overall average yield the system improved injection molding product quality, it could not
achieves 96.72 %, indicating that the proposed method framework can predict product quality in advance to reduce defect rates. Guo et al.
bring continuous improvement to the enterprise. (2019) [86] combined reinforcement learning and artificial neural
We can further see the results of individual parameter adjustments in network models to optimize process parameters, demonstrating fast and
Fig. 8. Take the " #1 Plastic Position " key factor as an example, after the stable convergence in process parameter optimization [86]. Our study
initial adjustment, we observed five consecutive production runs with can predict product quality in advance, reducing unnecessary waste and
no abnormalities. Therefore, we maintained the settings based on the providing key parameter adjustment suggestions as a digital trans-
recommendations from the first adjustment. However, taking the " #4 formation plan for small and SMEs. Dong et al. (2021) [87] proposed a
2nd stage injection time " key factor as an example; after the initial self-learning parameter optimization method that iteratively updates
adjustment, we noticed that the adjusted setting was too high during the parameters in each iteration until the best parameters are found [87].
production process. Therefore, after discussing it with the on-site engi- Wu et al. (2023) [88] presented a generative machine learning-based
neers, we decided to make slight adjustments to the setting value. Due to multi-objective optimization model for injection molding, which can
variations in product types, size and appearance requirements, as well as predict part qualification and optimize processing variables for minimal
differences in raw material batches, on-site personnel may also make energy consumption and weight differences among parts [88].
adjustment based on specific dimensional, appearance, and raw material Dănuţ-Sorin et al. (2020) [89] proposed the concept of adaptability in
batch requirements. intelligent plastic injection molding manufacturing [89]. Gim et al.
(2023) [90] demonstrate the effectiveness of transfer learning in opti-
4.4. Discussion mizing injection molding processes for high-surface quality parts that
reduced the need for costly and time-consuming optimization after mold
To demonstrate the significance and impact of this research in both transfer [90]. However, as revealed by the mentioned studies, product
academic and practical aspects, this section compares with previous quality cannot be known in advance, and adjustments can only be made
studies in the injection molding industry. Table 6 illustrates the com- after defects occur. Previous studies have shown that product quality
parison with related research. In previous studies, Farahani et al. (2021) cannot be known in advance, and adjustments can only be made after
[82] proposed a linear regression model for developing a quality defects occur, despite some research suggesting adaptability. Our study
monitoring system using process data from various mold sensors and addresses these gaps by proposing a method framework incorporating
machine data sources [80]. This research only compared various quality prediction and offering relevant parameter adjustment sugges-
methods for predicting product weight, contributing less to practical tions. This approach aims to achieve digital transformation with In-
applications. Silva & Alenya (2021) [77] used Artificial Neural Network dustry 4.0 adaptability and quality prediction. In contrast, our method
(ANN) and SVM as classifiers to detect part conformity or defects in a framework achieves Industry 4.0 adaptability and demonstrates a sig-
plastic injection molding company in Portugal [77]. However, this nificant improvement in process capability through on-site validation.
proposed framework lacks the ability to identify key parameters related Considering the company falls within the category of SMEs, the
to product quality, providing only the advantage of quality prediction. singularity of its product line, to some extent, limits our ability to
In contrast, our proposed method can overcome this limitation by demonstrate a diverse range of cases. In SMEs, the scope of products and
identifying key parameters and improving process capability indices. services is often more specialized, which restrict new cases from the
Párizs et al. (2022) [83] compared various machine learning methods same enterprise. Therefore, this study has chosen to focus on an in-depth
for predicting quality in multi-cavity co-injection molding [83]. analysis of a single case, aiming to extract valuable insights and un-
Although this study trained classifiers with a small amount of data and derstanding by exploring its details and complexities. Future research
analyzed selective feature selection methods and principal component will endeavor to extend validations to a broader range of industrial
analysis for defining key features from pressure curves for each classi- sectors to enhance the generalizability and practicality of the research
fier, the study focused on quality classification and did not address findings, thereby increasing the universal applicability and impact of the
improving product quality. Similarly, our research also employs feature study.
selection methods for parameter prediction and achieves the goal of
predicting product quality and adaptability. It not only predicts product 5. Conclusion
quality but also enhances product yield. Unlike the mentioned studies,
this research provides relevant adjustment recommendations to address This research developed a hybrid machine learning model to address

Table 5
Yield improvement.
Times Before 1st 2nd 3rd 4th 5th 6th
Product

Good 18,578 20,366 26,254 25,548 17,360 25,455 36,468

Defects 2446 2214 2143 2053 1272 1338 1236
Total 21,024 22,580 28,397 27,601 18,632 26,793 37,704
Yield (%) 88.37 90.19 92.45 92.56 93.17 95.01 96.72

11
M.-C. Chiu et al. Journal of Industrial Information Integration 41 (2024) 100666

Fig. 7. Yield variation chart.

Fig. 8. Changes in output.tiff.

practical issues and academic contributions in small and medium-sized the research gap in virtual measurement and adaptability fields and
manufacturing industries. Additionally, it rigorously validated the provides an effective digital transformation solution for these industries.
effectiveness of this approach using the CRISP-DM methodology. It fills The proposed method framework demonstrates insert an average of 95

12
M.-C. Chiu et al. Journal of Industrial Information Integration 41 (2024) 100666

Table 6 prediction and provides practical assistance to enterprises in predicting

Comparison with other studies. yield in advance and offering adjustment directions to improve process
Research Yield Parameters Industry 4.0 capability. Furthermore, it complements academic gaps and resolves
prediction adjustment adaptability practical problems.
Farahani et al. (2021) V This method can be readily applied in various manufacturing do-
[82] mains as long as sensors are installed and capable of collecting param-
Párizs et al. (2022) V eters. In the future, its applicability can extend to diverse applications
[83] such as healthcare and agriculture. We can adapt and modify the pro-
Silva& Alenya. V
(2021) [77]
posed method framework to find the most suitable machine learning and
Farahani et al. (2022) V deep learning models for different situations and achieve yield predic-
[84] tion as well as parameter optimization suggestions tailored to different
Kumar et al. (2020) V industries. By improving the accuracy of predictions and classifications,
[85]
we can enhance decision support systems, thereby increasing overall
Guo et al. (2019) [86] V
Dong et al. (2021) V operational efficiency and reducing costs. With technological advance-
[87] ments, we can further enhance the model’s performance, including
Wu et al. (2023) [88] V reducing training time, improving big data processing capabilities, and
Dănuţ-Sorin et al. V enhancing adaptability to new situations. Additionally, the model can be
(2020) [89]
Gim et al. (2023) V V
expanded to real-time data analysis, enabling businesses to respond
[90] quickly to market changes and internal operations and helping them
Proposed Method V V V take firm steps on their development path.

% accuracy in prediction and classification models. Utilizing feature Funding

selection and adjustment recommendations, the process capability index
is improved by 41 %. In academic contributions, previous studies have National Science and Technology Council of Taiwan partially and
yet to present digital transformation strategies developed for SMEs. This financially support this research under Contract no. NSTC 112-2221-E-
study’s approach, incorporating machine learning and deep learning 007 -091 -MY2
techniques, represents the first application of the framework in the
quality prediction field. Simultaneously aiding SMEs in achieving In- CRediT authorship contribution statement
dustry 4.0 adaptability. Compared to large enterprises, the method
proposed in this study is suitable for SMEs because it only requires deep Ming-Chuan Chiu: Writing – review & editing, Writing – original
learning methods and sensors to achieve digital transformation. SMEs do draft, Supervision, Project administration, Methodology, Funding
not need to invest significant amounts of capital and resources in acquisition, Conceptualization. Yu-Jui Huang: Writing – original draft,
building facilities like fully automated factories, as large enterprises Software, Methodology, Investigation, Formal analysis, Data curation,
often do. Moreover, this method can be applied not only in this industry Conceptualization. Chia-Jung Wei: Writing – review & editing, Visu-
but also in other domains with enough samples, showcasing its adapt- alization, Validation, Software.
ability concept to automatically respond with the most favorable stra-
tegies based on predicted events and providing optimization suggestions
to supervisors. Declaration of competing interest
In practical contributions, it allowed early quality prediction, saving
labor and material costs, and significantly improving product yield. The authors declare that they have no known competing financial
According to the company’s feedback, the increased yield from applying interests or personal relationships that could have appeared to influence
this method leads to an average monthly revenue increase of 20 % for a the work reported in this paper.
specific product, accumulating substantial gains in the long term.
Additionally, this method resolved the reliance on experienced opera- Data availability
tors for parameter adjustments, elevating the process capability signif-
icantly. Thus, this research addresses the shortcomings in quality The data that has been used is confidential.

Appendix

Table A1, Table A2

Table A1
Experimental design.

Dropout rate Momentum optimizer Learning rate

Level 1 0.2 Adam 0.01

Level 2 0.3 Adagrad 0.005
Level 3 0.4 Adadelta 0.001
Level 4 0.5 RMSprop 0.0005

13
M.-C. Chiu et al. Journal of Industrial Information Integration 41 (2024) 100666

Table A2
Hyperparameter optimization results.

Dropout rate Momentum optimizer Learning rate R2

1 0.2 Adam 0.01 0.877

2 0.2 Adagrad 0.005 0.86
3 0.2 Adadelta 0.001 0.878
4 0.2 RMSprop 0.0005 0.865
5 0.3 Adam 0.005 0.889
6 0.3 Adagrad 0.01 0.892
7 0.3 Adadelta 0.0005 0.9
8 0.3 RMSprop 0.001 0.874
9 0.4 Adam 0.001 0.903
10 0.4 Adagrad 0.0005 0.887
11 0.4 Adadelta 0.01 0.895
12 0.4 RMSprop 0.005 0.871
13 0.5 Adam 0.0005 0.881
14 0.5 Adagrad 0.001 0.883
15 0.5 Adadelta 0.005 0.876
16 0.5 RMSprop 0.01 0.854

Reference application, J. Ind. Inf. Integr. 31 (2023) 100413, https://doi.org/10.1016/j.

jii.2022.100413.
[19] Y.C. Tsao, C.W. Ho, C.C. Wu, The role of digital transformation in improving
[1] W.H. Huang, Y.J. Lin, H.F. Lee, Impact of population and workforce aging on
collaborative planning to address unexpected crisis, J. Ind. Prod. Eng. 40 (2023).
economic growth: case study of Taiwan, Sustain 11 (2019) 1–13, https://doi.org/
[20] K. AlOmoush, C. Lassala, S. Ribeiro-Navarrete, The role of digital business
10.3390/su11226301.
transformation in frugal innovation and SMEs’ resilience in emerging markets, Int.
[2] A. Raj, A.A. Mukherjee, A.B.L. deSousa Jabbour, S.K. Srivastava, Supply chain
J. Emerg. Mark. (2023), https://doi.org/10.1108/IJOEM-12-2022-1937.
management during and post-COVID-19 pandemic: mitigation strategies and
[21] V. Fernandez-Viagas, J.M. Framinan, Exploring the benefits of scheduling with
practical lessons learned, J. Bus. Res. 142 (2022) 1125–1139, https://doi.org/
advanced and real-time information integration in Industry 4.0: a computational
10.1016/j.jbusres.2022.01.037.
study, J. Ind. Inf. Integr. 4 (2022) 100281, https://doi.org/10.1016/j.
[3] M. Koumas, P.E. Dossou, J.Y. Didier, Digital transformation of small and medium
jii.2021.100281.
sized enterprises production manufacturing, J. Softw. Eng. Appl. 14 (2021)
[22] G.C. Kane, D. Palmer, A.N. Phillips, D. Kiron, N. Buckley, Achieving Digital
607–630.
Maturity, MIT Sloan Management Review and Deloitte University Press, 2017.
[4] Q. Chen, W. Zhang, N. Jin, X. Wang, P. Dai, Digital transformation evaluation for
[23] C.-L. Chen, Value creation by SMEs participating in global value chains under
small- and medium-sized manufacturing enterprises using the fuzzy synthetic
industry 4.0 trend: case study of textile industry in Taiwan, J. Glob. Inf. Technol.
method DEMATEL-ANP, Sustain 14 (2022), https://doi.org/10.3390/
Manag. 22 (2019) 120–145.
su142013038.
[24] M.Z. Yaqub, A. Alsabban, Industry-4.0-enabled digital transformation: prospects,
[5] Z. Yang, J. Chang, L. Huang, A. Mardani, Digital transformation solutions of
instruments, challenges, and implications for business strategies, Sustain 15
entrepreneurial SMEs based on an information error-driven T-spherical fuzzy cloud
(2023), https://doi.org/10.3390/su15118553.
algorithm, Int. J. Inf. Manage. 69 (2023) 102384.
[25] D.T. Parra-Sánchez, L.H. Talero-Sarmiento, Digital transformation in small and
[6] R. Krishnan, Challenges and benefits for small and medium enterprises in the
medium enterprises: a scientometric analysis, Digit. Transform. Soc. (2023).
transformation to smart manufacturing: a systematic literature review and
[26] Y. Peng, C. Tao, Can digital transformation promote enterprise performance?
framework, J. Manuf. Technol. Manag. (2024), https://doi.org/10.1108/JMTM-
—From the perspective of public policy and innovation, J. Innov. Knowl. 7 (2022)
07-2022-0255.
100198, https://doi.org/10.1016/j.jik.2022.100198.
[7] M. Ghobakhloo, N.T. Ching, Adoption of digital technologies of smart
[27] S. Bondar, J.C. Hsu, A. Pfouga, J. Stjepandić, Agile digital transformation of
manufacturing in SMEs, J. Ind. Inf. Integr. 16 (2019) 100107, https://doi.org/
system-of-systems architecture models using Zachman framework, J. Ind. Inf.
10.1016/j.jii.2019.100107.
Integr. 7 (2017) 33–43, https://doi.org/10.1016/j.jii.2017.03.001.
[8] D. Ulas, Digital transformation process and SMEs, Procedia Comput. Sci. 158
[28] S. Vilarinho, I. Lopes, S. Sousa, Developing dashboards for SMEs to improve
(2019) 662–671, https://doi.org/10.1016/j.procs.2019.09.101.
performance of productive equipment and processes, J. Ind. Inf. Integr. 12, (2018)
[9] X. Zhu, S. Ge, N. Wang, Digital transformation: a systematic literature review,
13–22, https://doi.org/10.1016/j.jii.2018.02.003.
Comput. Ind. Eng. 162 (2021) 107774, https://doi.org/10.1016/j.
[29] A. Garzoni, I. DeTuri, G. Secundo, P. DelVecchio, Fostering digital transformation
cie.2021.107774.
of SMEs: a four levels approach, Manag. Decis. 58 (2020) 1543–1562, https://doi.
[10] M.C. Chiu, H.Y. Tsai, J.E. Chiu, A novel directional object detection method for
org/10.1108/MD-07-2019-0939.
piled objects using a hybrid region-based convolutional neural network, Adv. Eng.
[30] H. Gil-gomez, Sustainable technology and entrepreneurship, Sustain. Technol.
Informatics 51 (2022) 101448, https://doi.org/10.1016/j.aei.2021.101448.
Entrep. 2 (2023) 100030, https://doi.org/10.1016/j.stae.2022.100030.
[11] G. Dutta, R. Kumar, R. Sindhwani, R.K. Singh, Digitalization priorities of quality
[31] M. Liu, C. Li, S. Wang, Q. Li, Digital transformation, risk-taking, and innovation:
control processes for SMEs: a conceptual study in perspective of Industry 4.0
evidence from data on listed enterprises in China, J. Innov. Knowl. 8 (2023)
adoption, J. Intell. Manuf. 32 (2021) 1679–1698, https://doi.org/10.1007/
100332, https://doi.org/10.1016/j.jik.2023.100332.
s10845-021-01783-2.
[32] S. Li, L. Gao, C. Han, B. Gupta, W. Alhalabi, S. Almakdi, Exploring the effect of
[12] A.M.R. Tortora, V. DiPasquale, C. Franciosi, S. Miranda, R. Iannone, The role of
digital transformation on firms’ innovation performance, J. Innov. Knowl. 8 (2023)
maintenance operator in industrial manufacturing systems: research topics and
100317, https://doi.org/10.1016/j.jik.2023.100317.
trends, Appl. Sci. 11 (2021), https://doi.org/10.3390/app11073193.
[33] S. Peillon, N. Dubruc, Barriers to digital servitization in French manufacturing
[13] T.H. Sam, P.S. Fernandez, Y.J. Nan, W. Yan, W. Jie, B. Zou, Factors influencing
SMEs, Procedia CIRP. 83 (2019) 146–150, https://doi.org/10.1016/j.
company performance during COVID-19 : case study of a semiconductor company
procir.2019.04.008.
in Malaysia, (2023). https://doi.org/10.17605/OSF.IO/RA5DS.
[34] C. Buck, J. Clarke, R. Torres de Oliveira, K.C. Desouza, P. Maroufkhani, Digital
[14] V. DeSimone, V. DiPasquale, S. Miranda, An overview on the use of AI/ML in
transformation in asset-intensive organisations: the light and the dark side,
Manufacturing MSMEs: solved issues, limits, and challenges, Procedia Comput. Sci.
J. Innov. Knowl. 8 (2023), https://doi.org/10.1016/j.jik.2023.100335.
217 (2022) 1820–1829, https://doi.org/10.1016/j.procs.2022.12.382.
[35] D. Rani, N.S. Gill, P. Gulia, A forensic framework to improve digital image evidence
[15] M. Pejić Bach, A. Topalović, Ž. Krstić, A. Ivec, Predictive maintenance in industry
administration in IIoT✰, J. Ind. Inf. Integr. 38 (2024) 100568 https://doi.org/
4.0 for the SMEs: a decision support system case study using open-source software,
10.1016/j.jii.2024.100568.
Designs 7 (2023), https://doi.org/10.3390/designs7040098.
[36] N. Urbach, M. Röglinger, Introduction to digitalization cases: how organizations
[16] L. Barreto, A. Amaral, T. Pereira, Industry 4.0 implications in logistics: an
rethink their business for the digital age, Digit. Cases How Organ. Rethink Their
overview, Procedia Manuf. 13 (2017) 1245–1252, https://doi.org/10.1016/j.
Bus. Digit. Age (2018) 1–12.
promfg.2017.09.045.
[37] L. Dörr, K. Fliege, C. Lehmann, D.K. Kanbach, S. Kraus, A taxonomy on influencing
[17] N. Chen, D. Sun, J. Chen, Digital transformation, labour share, and industrial
factors towards digital transformation in SMEs, J. Small Bus. Strateg. 33 (2023)
heterogeneity, J. Innov. Knowl. 7 (2022) 100173, https://doi.org/10.1016/j.
53–69.
jik.2022.100173.
[38] B. Nyagadza, Sustainable digital transformation for ambidextrous digital firms:
[18] S. Perera, X. Jin, P. Das, K. Gunasekara, M. Samaratunga, A strategic framework for
systematic literature review, meta-analysis and agenda for future research
digital maturity of design and construction through a systematic review and

14
M.-C. Chiu et al. Journal of Industrial Information Integration 41 (2024) 100666

directions, Sustain. Technol. Entrep. 1 (2022), https://doi.org/10.1016/j. [65] D. Horváth, R.Z. Szabó, Driving forces and barriers of Industry 4.0: do
stae.2022.100020. multinational and small and medium-sized companies have equal opportunities?
[39] G. Schuh, R. Anderl, R. Dumitrescu, A. Krüger, M. tenHompel, Industrie 4.0 Technol. Forecast. Soc. Change 146 (2019) 119–132, https://doi.org/10.1016/j.
maturity index. Managing the digital transformation of companies–update 2020, techfore.2019.05.021.
Acatech Study 64 (2020). [66] T. Chen, C. Guestrin, XGBoost: a scalable tree boosting system, in: Proc. ACM
[40] I. Castelo-Branco, T. Oliveira, P. Simões-Coelho, J. Portugal, I. Filipe, Measuring SIGKDD Int. Conf. Knowl. Discov. Data Min, 2016, pp. 785–794, https://doi.org/
the fourth industrial revolution through the Industry 4.0 lens: the relevance of 10.1145/2939672.2939785, 13-17-Augu.
resources, capabilities and the value chain, Comput. Ind. 138 (2022) 103639, [67] A. Ogunleye, Q.G. Wang, XGBoost model for chronic kidney disease diagnosis,
https://doi.org/10.1016/j.compind.2022.103639. IEEE/ACM Trans. Comput. Biol. Bioinforma. 17 (2020) 2131–2140, https://doi.
[41] M. Javaid, A. Haleem, R.P. Singh, R. Suman, Artificial intelligence applications for org/10.1109/TCBB.2019.2911071.
industry 4.0: a literature-based study, J. Ind. Integr. Manag. 7 (2022) 83–111. [68] C.P. Hsieh, Y.T. Chen, W.K. Beh, A.Y.A. Wu, Feature selection framework for
[42] Y. Lu, Cyber physical system (CPS)-based industry 4.0: a survey, J. Ind. Integr. XGBoost based on electrodermal activity in stress detection, IEEE Work. Signal
Manag. 2 (2017) 1750014. Process. Syst. SiPS Des. Implement. (2019) 330–335, https://doi.org/10.1109/
[43] N. Kumar, J. Kumar, Efficiency 4.0 for Industry 4.0, 15, 2019. SiPS47522.2019.9020321.
[44] E. Taddei, C. Sassanelli, P. Rosa, S. Terzi, Circular supply chains in the era of [69] W. Dong, Y. Huang, B. Lehane, G. Ma, XGBoost algorithm-based prediction of
industry 4.0: a systematic literature review, Comput. Ind. Eng. 170 (2022) 108268, concrete electrical resistivity for structural health monitoring, Autom. Constr. 114
https://doi.org/10.1016/j.cie.2022.108268. (2020) 103155, https://doi.org/10.1016/j.autcon.2020.103155.
[45] Z. Rajnai, I. Kocsis, Assessing industry 4.0 readiness of enterprises, in: 2018 IEEE [70] M.C. Chiu, C.Y. Wen, H.W. Hsu, W.C. Wang, Key wastes selection and prediction
16th World Symp. Appl. Mach. Intell. Informatics, 2018, pp. 000225–000230. improvement for biogas production through hybrid machine learning methods,
[46] V. Stich, G. Gudergan, V. Zeller, Need and Solution to Transform the Sustain. Energy Technol. Assessments 52 (2022) 102223, https://doi.org/
Manufacturing Industry in the Age of Industry 4.0 – A Capability Maturity Index 10.1016/j.seta.2022.102223.
Approach, in: L. Camarinha-Matos, H. Afsarmanesh, Y. Rezgui (Eds.), Collaborative [71] X. Shi, Y.D. Wong, M.Z.F. Li, C. Palanisamy, C. Chai, A feature learning approach
Networks of Cognitive Systems, PRO-VE 2018. IFIP Advances in Information and based on XGBoost for driving assessment and risk prediction, Accid. Anal. Prev.
Communication Technology, 534, Springer, Cham, 2018, https://doi.org/ 129 (2019) 170–179, https://doi.org/10.1016/j.aap.2019.05.005.
10.1007/978-3-319-99127-6_3. [72] R. Fu, Z. Zhang, L. Li, Using LSTM and GRU neural network methods for traffic flow
[47] M. Holm, O. Danielsson, A. Syberfeldt, P. Moore, L. Wang, Adaptive instructions to prediction, in: Proc. - 2016 31st Youth Acad. Annu. Conf. Chinese Assoc. Autom.
novice shop-floor operators using Augmented Reality, J. Ind. Prod. Eng. 34 (2017) YAC 2016, 2016, pp. 324–328, https://doi.org/10.1109/YAC.2016.7804912.
362–374, https://doi.org/10.1080/21681015.2017.1320592. [73] Z. Zhang, D. Robinson, J. Tepper, Detecting hate speech on twitter using a
[48] N. Rane, Transformers in Industry 4.0, Industry 5.0, and Society 5.0: Roles and convolution-gru based deep neural network, in: Proc. ACM Web Conf. (WWW
Challenges, 2023. 2018) Proceeding, 2018, pp. 745–760, 10.475/123_4.
[49] N. Rane, ChatGPT and similar generative artificial intelligence (AI) for smart [74] R. Dey, F.M. Salemt, Gate-variants of gated recurrent unit (GRU) neural networks,
industry: role, challenges and opportunities for industry 4.0, industry 5.0 and in: 2017 IEEE 60th Int. Midwest Symp. Circuits Syst., 2017, pp. 1597–1600,
society 5.0, SSRN Electron. J. (2023), https://doi.org/10.2139/ssrn.4603234. https://doi.org/10.1109/MWSCAS.2017.8053243.
[50] S. Ahmadi, Elastic data warehousing: adapting to fluctuating workloads with [75] V.N. Vapnik, Pattern recognition using generalized portrait method, Autom.
cloud-native technologies, J. Knowl. Learn. Sci. Technol. 2 (2023) 282–301, Remote Control 24 (1963) 774–780.
https://doi.org/10.60087/jklst.vol2.n3.p301. ISSN 2959-6386. [76] S. Huang, C.A.I. Nianguang, P. Penzuti Pacheco, S. Narandes, Y. Wang, X.
[51] X. Yan, G. Duan, The real-time prediction of product quality based on the U. Wayne, Applications of support vector machine (SVM) learning in cancer
equipment parameters in a smart factory, Processes 10 (2022) 967, https://doi. genomics, Cancer Genom. Proteom.. 15 (2018) 41–51, https://doi.org/10.21873/
org/10.3390/pr10050967. cgp.20063.
[52] Y. Bai, C. Li, Z. Sun, H. Chen, Deep neural network for manufacturing quality [77] B. Silva, J. Sousa, G. Alenya, Machine learning methods for quality prediction in
prediction, in: 2017 Progn. Syst. Heal. Manag. Conf. PHM-Harbin 2017 - Proc, thermoplastics injection molding, Int. Conf. Electr. Comput. Energy Technol.
2017, https://doi.org/10.1109/PHM.2017.8079165. ICECET 2021 (2021) 1–6, https://doi.org/10.1109/ICECET52533.2021.9698455.
[53] L. Ren, Z. Meng, X. Wang, R. Lu, L.T. Yang, A wide-deep-sequence model-based [78] D.J. Leinweber, Stupid data miner tricks, J. Invest. 16 (2007) 15, https://doi.org/
quality prediction method in industrial process analysis, IEEE Trans. Neural Netw. 10.3905/joi.2007.681820.
Learn. Syst. 31 (2020) 3721–3731, https://doi.org/10.1109/ [79] M.R. Khosravani, S. Nasiri, T. Reinicke, Intelligent knowledge-based system to
TNNLS.2020.3001602. improve injection molding process, J. Ind. Inf. Integr. 25 (2022) 100275, https://
[54] S. Yingying, H. Lianjuan, W. Jianan, W. Huimin, Quantum-behaved RS-PSO-LSSVM doi.org/10.1016/j.jii.2021.100275.
method for quality prediction in parts production processes, Concurr. Comput. [80] G.K. Zewdie, C. Valladares, M.B. Cohen, D.J. Lary, D. Ramani, G.M. Tsidu, Data-
Pract. Exp. 34 (2022) e5522, https://doi.org/10.1002/cpe.5522. driven forecasting of low-latitude ionospheric total electron content using the
[55] A. Saadallah, F. Finkeldey, J. Buß, K. Morik, P. Wiederkehr, W. Rhode, Simulation random forest and LSTM machine learning methods, Sp. Weather 19 (2021),
and sensor data fusion for machine learning application, Adv. Eng. Informatics 52 https://doi.org/10.1029/2020SW002639 e2020SW002639.
(2022) 101600, https://doi.org/10.1016/j.aei.2022.101600. [81] G.A. Busari, D.H. Lim, Crude oil price prediction: a comparison between AdaBoost-
[56] A. Fertig, M. Weigold, Y. Chen, Machine learning based quality prediction for LSTM and AdaBoost-GRU for improving forecasting performance, Comput. Chem.
milling processes using internal machine tool data, Adv. Ind. Manuf. Eng. 4 (2022) Eng. 155 (2021) 107513, https://doi.org/10.1016/j.compchemeng.2021.107513.
100074, https://doi.org/10.1016/j.aime.2022.100074. [82] S. Farahani, B. Xu, Z. Filipi, S. Pilla, A machine learning approach to quality
[57] H. Tercan, T. Meisen, Machine learning and deep learning based predictive quality monitoring of injection molding process using regression models, Int. J. Comput.
in manufacturing: a systematic review, J. Intell. Manuf. 33 (2022) 1879–1905, Integr. Manuf. 34 (2021) 1223–1236.
https://doi.org/10.1007/s10845-022-01963-8. [83] R.D. Párizs, D. Török, T. Ageyeva, J.G. Kovács, Machine learning in injection
[58] M.K. Msakni, A. Risan, P. Schütz, Using machine learning prediction models for molding: an industry 4.0 method of quality prediction, Sensors 22 (2022) 2704.
quality control: a case study from the automotive industry, Comput. Manag. Sci. 20 [84] S. Farahani, V. Khade, S. Basu, S. Pilla, A data-driven predictive maintenance
(2023) 1–28, https://doi.org/10.1007/s10287-023-00448-0. framework for injection molding process, J. Manuf. Process. 80 (2022) 887–897,
[59] Z. Abd Elnaby, A. Zaher, R.K. Abdel-Magied, H.I. Elkhouly, Improving plastic https://doi.org/10.1016/j.jmapro.2022.06.013.
manufacturing processes with the integration of Six Sigma and machine learning [85] S. Kumar, H.S. Park, C.M. Lee, Data-driven smart control of injection molding
techniques: a case study, J. Ind. Prod. Eng. 41 (1) (2023) 1–18, https://doi.org/ process, CIRP J. Manuf. Sci. Technol. 31 (2020) 439–449, https://doi.org/
10.1080/21681015.2023.2260384. 10.1016/j.cirpj.2020.07.006.
[60] Z. Abd Elnaby, A. Zaher, R.K. Abdel-Magied, H.I. Elkhouly, Improving plastic [86] F. Guo, X. Zhou, J. Liu, Y. Zhang, D. Li, H. Zhou, A reinforcement learning decision
manufacturing processes with the integration of six sigma and machine learning model for online process parameters optimization from offline data in injection
techniques: a case study, J. Ind. Prod. Eng. 41 (2024) 1–18, https://doi.org/ molding, Appl. Soft Comput. J. 85 (2019) 105828, https://doi.org/10.1016/j.
10.1080/21681015.2023.2260384. asoc.2019.105828.
[61] F.M. Abdullah, A.M. Al-Ahmari, S. Anwar, Analyzing interdependencies among [87] Z. Dong, P. Zhao, J. Zheng, K. Ji, Y. Chen, J. Fu, Intelligent injection molding:
influencing factors in smart manufacturing, Sustain 15 (2023) 3864, https://doi. parameters self-learning optimization using iterative gradient-approximation
org/10.3390/su15043864. adaptive method, J. Appl. Polym. Sci. 138 (2021) 50681, https://doi.org/10.1002/
[62] D.T. Matt, V. Modrák, H. Zsifkovits, Industry 4.0 for smes: challenges, app.50687.
opportunities and requirements, 2020. https://doi.org/10.1007/978-3-030 [88] F. Wu, Y.S.Y. Peng, Z. Mao, B. Chen, Generative machine learning-based multi-
-25425-4. objective process parameter optimization towards energy and quality of injection
[63] H. Khusna Mustafa, S. Yaakub, U. Malaysia, O. Yeop Abdullah, Innovation and molding, Environ. Sci. Pollut. Res. 30 (2023) 51518–51530.
technology adoption challenges: impact on SMEs’ company performance, Int. J. [89] I.R. Dănuţ-Sorin, C.G. Opran, G. Lamanna, Lean manufacturing 4.0 of polymeric
Accounting, Financ. Bus. (2018) 57–65. www.ijafb.com. injection molding products, Macromol. Symp. 389 (2020) 10–12, https://doi.org/
[64] K. Karuppiah, B. Sankaranarayanan, I. D’Adamo, S.M. Ali, Evaluation of key factors 10.1002/masy.201900109.
for industry 4.0 technologies adoption in small and medium enterprises (SMEs): an [90] J. Gim, H. Yang, L.S. Turng, Transfer learning of machine learning models for
emerging economy context, J. Asia Bus. Stud. 17 (2023) 347–370, https://doi.org/ multi-objective process optimization of a transferred mold to ensure efficient and
10.1108/JABS-05-2021-0202. robust injection molding of high surface quality parts, J. Manuf. Process. 87 (2023)
11–24, https://doi.org/10.1016/j.jmapro.2022.12.055.

15