Hybrid Approach Integrating Deep Learning-Autoencoder With Statistical Process Control Chart For Anomaly Detection Case Study in Injection Molding Process

Received 11 June 2024, accepted 5 July 2024, date of publication 9 July 2024, date of current version 19 July 2024.
Digital Object Identifier 10.1109/ACCESS.2024.3425582
Hybrid Approach Integrating Deep

Learning-Autoencoder With Statistical
Process Control Chart for Anomaly
Detection: Case Study in Injection
Molding Process
FAOUZI TAYALATI , IKHLASS BOUKROUH , LOUBNA BOUHSAIEN ,
ABDELLAH AZMANI, AND MONIR AZMANI
Intelligent Automation and Biomed Genomics Laboratory, FST of Tangier, Abdelmalek Essaâdi University, Tetouan 93000, Morocco
Corresponding author: Faouzi Tayalati (faouzi.tayalati@etu.uae.ac.ma)
This work was supported by the Ministry of Higher Education, Scientific Research and Innovation, the Digital Development Agency
(DDA), and the National Center for Scientific and Technical Research (CNRST) of Morocco (Smart Digital Logistic Services Provider
(DLSP) Project-AL KHAWARIZMI Artificial Intelligence (AI)-PROGRAM).
ABSTRACT Detecting anomalies in the injection molding process remains a challenging task, demanding
significant resources, data, and expertise due to their impact on cost and time reduction. While traditional
methods like statistical process control (SPC) using control charts are widely used for detecting irregularities,
they can catch predefined patterns such as systematic, upward shift, downward shift, cyclic, and mixture
patterns. However, they still have limitations in identifying anomalies beyond theses common patterns.
Numerous unnatural patterns may exist in process data, indicating that the process is out of control. In our
study, we propose an innovative strategy to enhance anomaly detection by integrating Statistical Process
Control (SPC) with a Long Short-Term Memory (LSTM) based Autoencoder (AE). The main objective is
to detect variations in the Melt cushion parameter, a crucial aspect of the injection molding process. The
LSTM-AE model determines optimal threshold levels based on reconstruction loss rates across all time-
series sequences, complementing traditional control charts’ upper and lower limits. With a model achieving a
coefficient of determination (R-squared) of 0.993 and a Mean Absolute Error of 0.0146 through training, and
incorporating multiple process limits—Upper Limit Control (ULC), Lower Limit Control (LLC), Maximum
Threshold, and Minimum Threshold—the enhanced control chart exhibits significant advancements in
anomaly detection. In four distinct scenarios, the integrated model demonstrates its capability to detect
anomalies in the Melt cushion that exceed predefined limits, as well as anomalies showing ascending and
descending trends. This ultimately enhances the robustness and efficiency of anomaly detection in injection
molding processes.
INDEX TERMS Injection molding process, melt cushion parameter, anomaly detection, statistical process
control, LSTM-auto encoder.
I. INTRODUCTION challenge of rising costs and demanding quality standards [2].

Injection molding process remains one of the most used One of the important strategies involves anomaly detection
industry in plastic manufacturing [1]. However, as with many (AD), a promising concept for cost reduction. Its imple-
manufacturing sectors, injection molding process is facing mentation within injection molding remains crucial, enabling
early detection of variations, and consequent reduction in
The associate editor coordinating the review of this manuscript and production expenses. Anomalies within injection molding
approving it for publication was Diego Bellan . processes can manifest in various forms, encompassing
2024 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
95576 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ VOLUME 12, 2024
F. Tayalati et al.: Hybrid Approach Integrating Deep Learning-Autoencoder With SPC Chart
both quality defects and maintenance aspects [3]. Quality their method is to streamline the intricate injection molding
defects in injection molding range from compromising process, facilitating swift identification and response to
part integrity to functionality. Early detection is vital to defects while enhancing the efficiency of the control system
prevent defective parts, reduce scrap rates, and maintain as a whole.
quality standards. Maintenance-related anomalies also pose From the literature review, statistical process control
challenges, affecting productivity and profitability. Common remains one of the most widely used methods for anomaly
maintenance issues include times and cost of break down [4]. detection. However, despite the benefits it offers, Statistical
By identifying and addressing potential equipment failures Process Control (SPC) also comes with limitations, which can
before they occur, manufacturers can minimize downtime, be summarized into four main points:
optimize equipment lifespan, and enhance overall operational • Limitations in complex processes: SPC is most effective
reliability. in processes with stable and predictable behavior.
The need to identify anomalies in industrial processes, In highly complex processes with numerous interacting
especially in injection molding, has driven extensive research variables, it can be challenging to isolate the effects of
into various statistical methodologies. These methods aim individual factors and identify root causes of variation.
to pinpoint deviations from normal operating conditions, This complexity may result in overlooking critical
enabling timely interventions and process optimization. factors or interactions, leading to incomplete insights
Among the tools used for anomaly detection, the main into process behavior.
methodologies include Moving Average, Cumulative Sum, • Assumption of normal distribution [13]: SPC techniques
Deep Learning with Probability, and Statistical Process such as control charts often assume that process
Control (SPC) [5]. data follows a normal distribution. However, in real-
As shown by Cox [6], The Moving Average is foundational world scenarios, process data may exhibit non-normal
statistical technique that calculates the average of a sliding distributions due to various factors such as asymmetry
window of data points, enabling the detection of shifts or or multimodality. When data deviates significantly
trends in process behavior over time. On the other hand, from normality, the effectiveness of SPC methods
Cumulative Sum, as demonstrated by Grunkemeier et al. [7], may be compromised, potentially leading to inaccurate
allows tracking the cumulative deviation from a specified conclusions or decisions.
reference value, facilitating the early detection of persistent • Overreliance on historical data: SPC relies heavily
process shifts or abnormalities. Additionally, Deep Learning on historical data to establish control limits, identify
with Probability, as demonstrated by Ketonen and Blech [8], patterns, and make process adjustments. While historical
leverages advanced neural network architectures. Deep data provides valuable insights into past performance,
learning techniques combined with probabilistic models offer it may not always capture future changes or variations
enhanced capabilities for anomaly detection, particularly in in the process. Unforeseen events, process improve-
complex and high-dimensional datasets. Among the most ments, or changes in operating conditions may not
widely utilized tools within the Statistical Process Control be adequately accounted for in historical data, leading
framework are the control charts. These charts, including the to suboptimal decisions or actions based on outdated
X-bar and R (Range) charts for monitoring process mean and information.
deviation, provide a systematic approach to tracking process • Delayed detection of changes: SPC is designed to
stability and detecting deviations from established control detect significant shifts or trends in process behavior
limits [9]. Integrating these statistical approaches not only over time. However, it may not always promptly
enables the identification of anomalies, but also empowers detect small or rapid changes that occur within short
manufacturers to proactively manage process variability, timeframes. These small changes, if left undetected,
optimize production efficiency, and uphold product quality can accumulate over time and impact product quality
standards. Several studies have explored the application of or process performance. Additionally, SPC methods
the control charts in injection molding processes. Rohani may have limitations in detecting changes in certain
and Chan [10] employed quality tools and control charts types of variation, such as sporadic or systematic
to enhance the quality of plastic injection molded lenses shifts, leading to delayed responses and potential quality
utilized in telecommunication devices. Their study showed issues.
the ability to decrease the defect rate from 13.49% to 10%. In this paper, novel concept is explored to integrate the
Chaciński et al. [11] deployed Statistical Process Control, benefits of statistical control charts with modern machine
integrating process capability indicators and performance learning methods, aiming to overcome limitations. It pro-
indicators into their study. They assessed three product poses a combined approach incorporating control charts,
types, focusing on product mass to improve defect detection, Long Short-Term Memory (LSTM) [14], and Autoencoder
meet customer specifications, and reduce non-compliance. techniques for detecting anomalies in the injection molding
Cao et al. [12] introduced an approach using SPC principles. process. With this triple approach, LSTM captures long-
They utilized data from different parts to create control term dependencies, Autoencoder calculates reconstruction
charts tailored to specific part families. The objective of errors, and Statistical Process Control (SPC) establishes
VOLUME 12, 2024 95577

upper and lower control limits (UCL and LCL). The case plastic material is fed into a heated cylinder via a hopper and
study presented in this study focuses on the ‘‘Melt cushion processed by a rotating screw. Friction and heating units melt
parameter,’’ a critical aspect of the injection molding process the material, accumulating it at the screw tip until pressure
[15]. The melt cushion refers to the distance between the prompts the screw to retract, creating a melt reservoir. The
end of the screw and the nozzle tip when the injection stroke injection phase starts when the mold closes, and the molten
ends. This parameter plays a pivotal role in ensuring both the material fills it. Following this, packing, cooling, and ejection
stability of the process and the mechanical integrity of the occur. Once solidified, the part is ejected as the mold opens,
machine. A stable melt cushion is essential for maintaining completing the cycle [18]. This process is facilitated by an
both process and product quality, as well as ensuring the injection molding machine.
mechanical stability of the machine. By monitoring the melt Figure 1 illustrates the principal components of the injec-
cushion parameter for anomalies, issues related to material tion equipment. The injection machine is a complex system
quality, temperature fluctuations, and melt uniformity can be with several essential components, including the injection
identified early. This early detection helps prevent defects system and mold clamping mechanism. The injection system
in product quality and damage to critical machine compo- comprises the hopper, where raw material is stored, the
nents like the non-return valve and screw, as outlined by barrel where it’s melted and homogenized, the screw system
Womer, Zhao and Schiffers [16]. for injecting material into the mold, and the nozzle for the
The general contributions of this article can be outlined as injection point [19]. Conversely, the clamping unit consists
follows: of the mold, which shapes the final product, the clamping
• To introduce a novel method for enhancing the identifi- mechanism that securely holds the mold halves together
cation of process irregularities, our aim is to merge the during injection, and the ejector mechanism, which opens
Statistical Process Control (SPC) method with LSTM- the mold and ejects the finished part once it’s cooled and
AE. We seek to enhance conventional control charts solidified.
by incorporating detectability thresholds to the existing
upper and lower limits. B. MELT CUSHION PARAMETER
• Develop a robust tool for quality optimization and The Melt cushion parameter, also referred to as decom-
maintenance. This approach enables manufacturers to pression distance, signifies the distance the screw moves
improve anomaly detection, ensuring early identifica- backward after the injection phase, preceding mold opening,
tion of process variations, leading to better product Figure 2. This action forms a cushion of material in front of
quality and reduced defective parts. It also minimizes the screw, compensating for variations in material viscosity
unplanned downtime and maximizes productivity by and mold temperature [15]. It’s pivotal in injection molding,
promptly resolving issues related to equipment failures serving as a buffer to ensure proper material packing in the
or process abnormalities. Overall, this leads to sig- injection cavity. Detecting cushion issues can be challenging
nificant cost savings by reducing scrap, rework, and and time-consuming to resolve. The material left at the
maintenance expenses. screw’s front during its forward position is termed the
• To implement the developed approach in real injection melt cushion. In conventional injection molding, the melt
process as case study, with a particular focus on the cushion should never reach zero to maintain part quality and
‘‘Melt cushion parameter’’ to validate the effectiveness dimensional consistency. Moreover, the cushion value plays
of the developed approach. a critical role in preserving resin integrity before injection.
The article is divided into four main sections. The first Adjusting the cushion value isn’t directly possible on molding
section outlines the study’s scope, focusing on analyzing machines, but it’s crucial for transmitting pressure into
the injection molding process and introducing the ‘‘Melt the cavity. Without a proper cushion, achieving consistent
cushion’’ parameter. In the subsequent section, the literature cavity pressure becomes challenging, leading to fluctuations
review examines previous research on anomaly detection in part dimensions and difficulties in managing viscosity
in injection molding, covering the use of control chart shifts.
tools and machine learning methods. The third section The irregularities observed on the Melt cushion have
introduces the study’s innovative approach, which combines multiple impacts on process stability. Firstly, the cushion
control charts with LSTM-AE to improve anomaly detection parameter compensates the shrinkage during cooling phase
in injection molding. Finally, the fourth section presents [20], ensuring the final part retains its desired dimensions and
the research results and discusses their implications for geometry. Secondly, it prevents short shots by optimizing the
enhancing injection molding processes. injection process to adequately fill the mold cavity, reducing
the risk of incomplete parts [21]. Additionally, adjusting the
II. SCOPE OF STUDY cushion parameter helps reduce part stress by minimizing
A. INJECTION MOLDING PROCESS injection pressure at the end of the process, thereby improving
Injection molding, a widely used plastic manufacturing overall part quality. Moreover, it optimizes material flow
process, comprises five key steps: Plasticizing, injection, within the mold cavity [22], leading to enhanced the quality of
packing, cooling, and ejection [17]. During plasticizing, the molded part. Lastly, the cushion parameter influences gate
95578 VOLUME 12, 2024

FIGURE 1. The standard structure of the injection molding machine.
a framework that utilizes sensor data obtained from plastic

injection machines along with deep learning models to detect
anomalies and securely cluster outliers. This framework aims
to provide effective solution tailored for factories transition-
ing to smart factory systems, with three main contributions.
Firstly, it adopts a suitable approach for manufacturing sites.
Secondly, it secures outliers using Auto Label with pseudo
labeling. Thirdly, it enables decision-makers to identify
potential defect causes. The benefits of this approach are
incorporated into the Two Phases Anomaly Detection system
architecture, which utilizes an LSTM classifier alongside
FIGURE 2. a. The Melt cushion position before injection step. b. The Melt the pseudo labeling technique, achieving a high reliability
cushion position after injection step. rate surpassing 90% accuracy. Additionally, clustering visu-
alization of defective data is facilitated through the Self
Organizing Map (SOM) algorithm. Nunes et al. [25] highlight
sealing, crucial for preventing material leakage and ensuring the importance of predictive maintenance in industry for
proper mold cavity filling. optimizing maintenance schedules and prolonging equipment
lifespan. They introduce a data-driven methodology called
III. LITERATURE REVIEW generalized fault trees to assess equipment reliability in real-
Many research has been explored into anomaly detection time. Applied to predictive maintenance of injection molds,
aspects, this review focuses on three main areas of anomaly their approach, combining a new training algorithm with
detection: injection molding processes, Statistical Process isolation forest anomaly detection, demonstrates significant
Control (SPC) in injection molding, and Long Short-Term cost savings of 27.05% compared to preventive mainte-
Memory Autoencoder (LSTM-AE) techniques for general nance and 63.43% compared to corrective maintenance.
anomaly detection. Silva et al. [26] discuss the implementation of Artificial
At the level of anomaly detection, several approaches Intelligence (AI) in plastic injection molding processes to
have been explored. Lughofer and Pichler [23] propose an enhance competitiveness in Industry 4.0. They highlight the
automated method for predicting quality issues in injection importance of real-time Quality Prediction methodologies in
molding machines using only regular production data. Their reducing costs for companies dealing with complex industrial
approach analyzes process variables to detect anomalies and processes. The paper delineates the implementation process,
predicts quality criteria directly from time series trends. encompassing data collection, real-time classification, and
By dynamically updating models, they successfully identified methodologies such as Data Augmentation and Human-
anomalies and production changes during real processes, in-the-Loop labeling. The AI-driven approach enables pre-
achieving high correlations up to 0.98, with observed quality diction and alerting of process quality issues, leading to
trends and low error rates below 5%. Lee et al. [24] explore increased productivity and reduced non-compliant parts
the growing significance of smart factories, particularly in production. Testing across various processes and materials
sensor data-based anomaly detection research. They present shows significant improvements, including up to a 12%
VOLUME 12, 2024 95579

increase in Overall Equipment Effectiveness (OEE) and a in measurement accuracy and the necessity for repeated
notable decrease in non-conforming parts production. measurements. The implementation of statistical regulation
At the level of SPC application, significant research has in production management is highlighted as a means to
been conducted to detect anomalies. RC Kanu [5] applies reduce costs and enhance efficiency. Future research sugges-
Statistical Process Control (SPC) using x-bar and range tions include evaluating production process capability using
control charts to monitor process stability in the injection process capability indices. The study underscores the effec-
molding of plastic parts. Control chart analysis revealed tiveness of Shewhart control charts in improving quality and
variability in process parameters such as plasticizing time, efficiency in manufacturing processes, resulting in reduced
cooling time, cushion final position, and screw position at non-conforming products and downtime. The application of
change-over. The study aims to integrate materials from these control charts is recognized as a growing trend in
courses on SPC and plastics injection molding to enhance manufacturing, aiming to increase productivity and decrease
students’ understanding of both subjects. Students produced costs. Further studies are proposed to explore the use of
ASTM (American Society for Testing and Materials) tensile modern indices for assessing process capability and achieving
and impact specimens while examining process variability continuous improvement in results. The study conducted by
affecting parts quality. Polycarbonate (PC) and acrylonitrile- Chaciński et al. [11] investigates process stability in injection
butadiene-styrene (ABS) blend resin from Bayer were used, molding, specifically focusing on product weight, which
dried at 200 ◦ F before processing. A 60-ton Sandretto holds significant importance for plastic molding quality.
Injection Molding Machine was employed, with specific They delve into the implementation of SPC techniques,
processing parameters. Data collection involved printing which offer a systematic approach for evaluating process
parameters after each molding cycle and weighing parts capability and pinpointing areas for enhancement. Utilizing
after 48 hours. Control charts showed out-of-control process metrics such as Cp and Cpk [29] The study centers on three
parameters, indicating potential production of poor quality distinct production processes, each conducted on separate
products. Variability sources were identified as cushion final production nests, with a primary emphasis on product weight
position, screw change-over position, actual cooling time, as a pivotal quality parameter in plastic molding. Evaluation
and plasticizing time. The study conducted by Uluksan [27] criteria encompass the calculation of potential and actual
aims to address water leakage issues in condenser dryer process capability indices (Cp and Cpk), providing precise
water tanks produced through plastic injection molding insights into process performance and stability. To ensure
and vibration welding, employing statistical process control accuracy and repeatability, measurement data are gathered
(SPC) techniques. Initially, the study assessed the cost using an AXIS B6M industrial platform scale. The authors
and criticality of the problem and outlined the manufac- identify specific areas for improvement within each process,
turing workflow. A cause-and-effect diagram was created underscoring the significance of reducing variability and
to identify potential causes of leakage, with a focus on enhancing alignment to bolster stability and overall quality.
the injection molding and vibration welding processes. In addition, the study contributes significantly to existing
Control charts were utilized to evaluate process stability, literature by showcasing the tangible application of SPC
revealing poor capability initially. Improvement efforts were methods in assessing and refining injection molding pro-
then directed towards mold maintenance, custom fixtures cesses. Through the strategic utilization of process capability
for welding, and standardized controls for rubber band indicators, manufacturers can optimize operations, curtail
usage. Following these improvements, process capability defects, and elevate product quality.
significantly improved, as confirmed through hypothesis At the application level of deep learning methods,
testing. Additionally, an automated leakage testing system significant advancements have been achieved in anomaly
was implemented to further reduce defective products. The detection using the LSTM-Autoencoder (LSTM-AE) model.
study concludes that ongoing improvement efforts, including Muaz et al. [30] propose a novel approach for predicting
the use of design of experiments, can effectively enhance product quality in injection molding, integrating both time
plastic injection molding and vibration welding processes series and non-time series parameters using deep learning.
across various products and industries. The paper authored Their model, trained on data from in-mold pressure and tem-
by Malindzakova et al. [28] focuses on evaluating the quality perature sensors, achieves superior performance compared to
characteristics of plastic moldings for automotive industry benchmark models, with significantly lower error rates. This
requirements using Shewhart control charts. These charts automated system, coupled with modern containerization
are utilized to assess the measured width and length of practices, enhances the quality control process for injection
the plastic moldings through statistical analysis. The study molding operations. Jung et al. [31] investigate the applicabil-
reveals an initial non-compliance with required standards, ity of machine learning in quality prediction for the injection
followed by observed improvements in the production molding industry, aligning with the pursuit of sustainable
process over time. Emphasizing the importance of setting growth in Industry 4.0. They assess various machine learning
control charts to monitor variability and maintain desired algorithms, finding that autoencoders outperform others in
production parameters, the research also addresses challenges accuracy, precision, recall, and F1-score metrics. Feature
95580 VOLUME 12, 2024

importance analysis identifies temperature and time as Given the challenges of adapting model-based methods to
significant factors affecting quality. These findings under- changing environments, they propose an Autoencoder-LSTM
score the potential of artificial intelligence techniques for method for remaining useful life prediction. This approach
sustainable management in Industry 4.0, offering insights utilizes Autoencoder to extract features from original bearing
for businesses seeking to optimize their manufacturing data and LSTM for life prediction. Experiments demonstrate
processes. Huang et al. [32] address the challenges of the effectiveness of the extracted features and show the
predicting polymer melt flow behavior in injection molding method’s superiority compared to other approaches. Further
due to discrepancies between simulation and actual results. experiments confirm higher prediction accuracy compared to
Their proposal involves utilizing a machine learning approach traditional machine learning and related methods.
to establish a virtual-actual correction model, focusing on Upon reviewing the literature, several gaps and opportu-
improving the accuracy of mold cavity pressure. Unlike nities have emerged, which serve as the primary motivation
traditional methods, this approach doesn’t require extensive for the development of this paper. Specifically, the literature
data for training, enhancing its practicality for specific gap in this area can be addressed through three main
settings. Experimental verification using injection molding points:
machines, molds, and materials validates the method’s effec- • Integration of Control Chart with LSTM in Injection
tiveness. By employing an autoencoder to extract features Molding: While existing research, such as the work by
and a multilayer perceptron to establish relationships between Essien and Giannetti [36], has explored LSTM applica-
simulated and actual data, the proposed method significantly tions in manufacturing for predictive maintenance and
improves the relationship between simulated and actual quality prediction, a gap remains in integrating control
cavity pressures, demonstrating its potential for enhancing chart techniques with LSTM specifically for injection
injection molding process optimization. The correlation molding. Control charts are standard for quality control
between simulated and actual cavity pressures significantly in manufacturing, but their combination with LSTM
improved, rising from 81% to 98%. Ramezankhani et al. [33] for real-time anomaly detection or quality prediction
tackled the challenges in advanced manufacturing stemming in injection molding is unexplored. Bridging this gap
from sparse and heterogeneous data. They introduced a could lead to innovative methods for enhancing process
sim-to-real transfer-learning framework aimed at mitigating monitoring and quality assurance in injection molding
data scarcity and heterogeneity in smart manufacturing operations
applications. Their methodology integrates wide-and-deep • Combining LSTM and Autoencoder: While some works
learning to handle structured sensory data and thermal have explored the use of LSTM or Autoencoder
images separately, with a primary focus on predicting individually, few have combined these two methods in
product quality metrics. The authors employed a convolu- the field of the injection molding process. This presents
tional variational autoencoder (ConvVAE) to extract main an opportunity to investigate the potential synergy
representations of thermal images in an unsupervised manner. between LSTM, Autoencoder, and other deep learning
Furthermore, they integrated sim-to-real transfer learning by techniques, potentially leading to more effective predic-
leveraging theory-based heat transfer simulations. Through tive models.
evaluation in an industrial thermoforming process case • The literature reveals a lack of focus on monitoring the
study, ConvVAE demonstrated superior performance over cushion parameter in injection molding studies. Existing
conventional methods, even with limited data. Additionally, research mainly concentrates on post-defect indicators
the authors conducted model explainability analysis using like weight and final dimensions of the product. Yet,
SHAP (SHapley Additive exPlanations) values, confirming monitoring the cushion parameter is vital as it acts
alignment between predictions, theoretical expectations, and as a pre-defect indicator, aiding early detection and
data correlation statistics. Po and Kim [34] presented an identification of potential issues.
approach to detect anomalies within smart factories by These gaps motivate exploring a triple combination of
utilizing a Deep Autoencoder Neural Network. This method Statistical Process Control (SPC), LSTM, and Autoencoders
capitalizes on the advancements of Industry 4.0 to improve for anomaly detection in injection molding processes. This
operational efficiency while addressing the challenge of integrated approach is offering comprehensive real-time
promptly identifying and handling abnormal events. The monitoring and anomaly identification by leveraging statisti-
Deep Autoencoder Neural Network is enhanced with sup- cal insights, temporal learning, and feature extraction power,
plementary processing stages aimed at refining the residual enhancing accuracy and efficiency in quality assurance and
image, thereby enhancing the precision of anomaly detection. process optimization.
Given their proficiency in reconstructing input data, autoen-
coders are particularly suitable for such applications, making IV. IMPLEMENTATION
them highly effective for detecting anomalies in smart factory A. METHODOLOGY
environments. Yang et al. [35] address the importance of The proposed approach aims to develop a model with the
accurately assessing the degradation of bearings in injection ability to identify anomalies by combining the LSTM-Auto
molding machines and predicting their remaining useful life. Encoder and Statistical Process Control (SPC) chart
VOLUME 12, 2024 95581

FIGURE 3. The proposed method combining LSTM-Autoencoder and control chart for anomaly detection.
methodologies. The methodology is delineated into sev- (SPC) framework. The establishment of the control chart aids
eral stages. Initially, data collection involves assembling in deriving the Upper Limit Control and the Lower Control
a dataset containing 10160 observations related to the of the Process, which serve as two critical indicators utilized
cushion parameter. This dataset is segregated into two for anomaly detection. Secondly, it facilitates the subsequent
groups: normal instances, regrouping 8934 observations, and training of the LSTM-AE model, enabling the determination
anomalous instances, comprising 1226 observations. The of the maximum residual, which is utilized as the threshold
dataset representing normal situations depicts the machine’s for anomaly detection. While constructing the control chart,
operation under optimal conditions, signifying a production a new dataset is created, centered around the sample means.
phase devoid of quality defects or process variations. Within On the other hand, the anomaly dataset is exclusively
this study, the dataset serves as dual purposes. Firstly, allocated to validate the efficiency of anomaly detection.
it is employed to construct the control chart and execute a The model is subjected to validation using this dataset to
capability study adhering to the Statistical Process Control pinpoint any deviations from the normal situation, effectively
95582 VOLUME 12, 2024

FIGURE 4. The original and the trend of the dataset - Normal observations.
establishing a template for detecting abnormal situations.

After defining these thresholds, testing is performed on
abnormal situations to validate the model. Following this, the
integration of the control chart and the established thresholds
forms the basis for a new control chart that incorporates
both SPC detection and anomaly detection. The ultimate
objective is to develop a unified SPC control chart featuring
upper and lower limits, enabling the dual prediction of
anomalies. Figure 3 illustrates the flowchart of the proposed
methodology.
B. DATA ANALYSIS
1) DATASET OF THE NORMAL SITUATION
The data was collected from a workshop specializing in
producing ‘‘Connector cables’’ for the automotive sector in
northern Morocco’s injection molding industry, where an
ARBUG-Gmbh injection molding machine, with a capacity
of 150 tons, was in operation. The dataset encompasses
observations of cushion melt parameters gathered across
10,160 production cycles, amounting to over 100 hours FIGURE 5. Normality analysis of the melt cushion distribution.
of production time. The dataset is categorized into two

groups, normal situations, which include 8,934 observations,
and anomaly situations, which comprise 1,226 observations. evaluation of process stability [37]. Various techniques,
Figure 4 displays the graphical representation of the data such as constructing histograms [38] and box-and-whisker
collected during normal situation. plots [39], are commonly employed in this analysis. These
The dataset representing normal situations reflects the methods help verify whether the data conforms to a normal
typical production conditions. According to the production distribution, which is fundamental for reliable SPC outcomes.
datasheet, the cushion parameter is expected to be 6.61 cm3 , In Figure 5, the observations demonstrate a normal
with an engineering tolerance of +/- 0.5 cm3 , Lower Spec- distribution, indicating that the dataset adheres closely to
ification Limit (LSL) as 6.11 cm3 and Upper Specification the Gaussian model. Supporting this observation, Table 1
Limit (USL) as 7.11 cm3 . Any value within this tolerance provides statistical information, with an average cushion
range is considered acceptable, while values falling outside parameter value of 6.06 cm3 and a standard deviation
are identified as anomalies. of 0.07 cm3 . Additionally, the histogram illustrates the
distribution’s symmetry around the mean, reinforcing the nor-
2) NORMALITY ANALYSIS mality of the dataset. Notably, approximately 90% of the
Assessing the normality of a dataset is a crucial step, observations fall within the range of 6.5 cm3 to 6.57 cm3 ,
particularly in Statistical Process Control (SPC), where further affirming the concentration of data points around
adherence to a Gaussian distribution is often essential the central tendency. This thorough analysis underscores the
for the accurate establishment of control limits and the dataset’s compliance with normality assumptions, bolstering
VOLUME 12, 2024 95583

TABLE 1. Main statistical outcomes from the normal dataset.
FIGURE 7. Analysis of capability using the provided normal dataset.
of the process relative to the target or nominal value.

It evaluates both the spread of the process distribution
and its alignment with the target value. Cpk is calculated
as the minimum of two indices, the minimum distance
between the process mean and the nearest specification
FIGURE 6. Illustrative gaussian curve showing defect limits. limit, expressed as follows:
Cpk = min((USL − µ)/3σ, (µ − LSL)/3σ )) (2)
the reliability of subsequent SPC analyses and decision- where µ is the mean of the values, USL represents the upper
making processes. specification limit, LSL stands for the lower specification
limit, and σ denotes the standard deviation of the process.
C. SPC FRAMEWORK Cp assesses if the distribution can fit within the specifi-
1) PROCESS CAPABILITY cation limits, while Cpk determines if the overall average
As defined by Oberoi et al. [40], Process capability is is centered. When the overall average aligns with the
a statistical metric that gauges the capacity of a process specification center, Cp and Cpk values match. If they differ,
to consistently generate output that aligns with predefined it indicates the overall average is not centered. The greater the
criteria or customer expectations. In other words, it assesses gap between the values, the more the overall average is offset.
the extent to which a process can meet quality standards Palmer and Tsui [42] explained that the interpretation of
within the tolerance limits set for the product. The capability the process can be done according to the value obtained of
of a process is typically evaluated in terms of its ability each capability index:
to perform within specified upper and lower limits or • If (Cp > 1 and Cpk > 1): The process is capable, and the
tolerances. These tolerances represent the acceptable range of process mean is well-centered within the specification
variation for key characteristics or dimensions of the product limits. This is the ideal scenario, and it signifies that the
or service. Process capability analysis aims to determine process has a low probability of producing defects.
whether the inherent variability of the process is small • If (Cp > 1 and Cpk < 1): The process is capable, but
enough to ensure that the majority of output falls within the process mean is not well-centered. This suggests that
these tolerance limits. The most commonly used indices for although the process can produce within specifications,
evaluating process capability are Cp and Cpk there is a risk of producing a higher proportion of defects
• Cp (Process Capability Index): Process capability due to the lack of centering.
assesses a process’s ability to meet specifications by • If (Cp < 1 and Cpk < 1): The process is not capable,
evaluating the spread of the process distribution relative and the process mean is not well-centered. This indicates
to the width of the specification limits. It’s calculated as that the process is producing a substantial number of
the ratio of the tolerance width to the process variation, defects, and immediate actions are required to improve
expressed by the following formula: the process
Regarding process capability, the application of the Sta-
Cp = (USL - LSL)/6σ (1)
tistical Process Control (SPC) framework has demonstrated
where USL is the upper specification limit, LSL is the lower promising outcomes, as presented in Table 2. With a Cp of
specification limit and σ is the standard deviation of the 2.37 and a Cpk of 2.36, the process epitomizes outstanding
process. performance. These metrics show precise alignment within
• Cpk (Process Capability Index, adjusted for centering): tolerance boundaries, highlighting the process’s reliability
Is similar to Cp but also takes into account the centering and robustness in consistently delivering high quality parts
95584 VOLUME 12, 2024

FIGURE 8. Illustrative control chart displaying Upper Control Limit (UCL) and Lower Control Limit (LCL).
TABLE 2. Calculation of the process capability. 3) X-BAR AND R CHART

These control charts are essential for assessing the mean and
variability of a process when measuring continuous data.
X-bar charts are utilized to monitor the central tendency of
a process, while R charts are used to track its variations.
The X-bar chart displays the average of a subgroup of
measurements, and the R chart shows the range of the same
subgroup. X-bar and R charts are ideal for processes with
relatively small subgroup sizes (typically fewer than 10) and
normally distributed data [47].
a: CALCULATION OF THE MEAN CHAR

The mean chart is constructed using the following formulas:
that meet desired criteria. Figure 7 illustrates the representa-
(X 1,X 2, . . . .Xn)
P
tions of process capability. X̄ = (3)
n
where n is the number of observations
2) THEORY OF CONTROL CHART
(X¯1, X¯2, . . . .Xk)
¯
P
Control charts are essential tools in statistical process control X= (4)
(SPC) for monitoring and controlling processes in real time k
[43]. Figure 8 shows visual representation of typical control where k is the number of subgroups
chart. The selection of a control chart type depends on the
characteristics of the data being monitored and the specific UCL = X + A2∗ R̄ (5)
goals of the analysis [44]. where A2 is a constant depending on the number of
Figure 9 illustrates the common types of control charts, observations, and R̄ is the mean of the range
which are classified into two main types: attributes, as dis-
cussed by Chiu and Kuo [45], and variables, as elaborated LCL = X − A2∗ R̄ (6)
by Gibra [46]. Attribute control can be subdivided into charts where A2 is a constant depending on the number of
for percentage defectives and charts for the number of defects observations, and R̄ is the mean of the range
per unit. In variables control, the main emphasis is placed
on monitoring fluctuations in the average and the range
of measurements. Regardless of the specific type, control b: CALCULATION OF THE RANGE CHART
charts generally adhere to the same fundamental structure, The range chart is constructed using the following formulas:
where the mean value is enclosed by upper and lower (R1, R2, . . . .Rk)
P
control limits. The calculation of these control limits varies R̄ = (7)
k
depending on the type of chart being used. In this paper, based where k is the number of subgroups
on the measurable value and sample size, we utilized the
X-Bar and Range control chart concept to define the Upper UCL = D4∗ R̄ (8)
Limit Control (ULC) and Lower Limit Control (LCL) of the
where D4 is a constant depending on the number of
process. This approach was also used to generate the new
observations (Table3.), and R̄ is the mean of the range
dataset for the LSTM-Auto encoder, with samples consisting
of 3 observations each. LCL = D3∗ R̄ (9)
VOLUME 12, 2024 95585

FIGURE 9. The main type of the statistical process control chart.
TABLE 3. Constant variables for X-bar and range control charts. of 10 randomized cycles of the distribution. There are no
instances of production falling outside of tolerance limits, and
there is no discernible ascending, descending, or cyclic trend
of deviation during this example.
D. LSTM AUTOENCODER
Our research involves collecting substantial amounts of
unlabeled time series data representing both normal and
abnormal operational states. To analyze this data, we employ
unsupervised learning techniques, specifically Autoencoders.
TABLE 4. Upper limit control and lower limit control calculation.
However, conventional Autoencoders struggle to effectively
capture the temporal patterns inherent in time series data [48].
To address this issue, we enhance our approach by inte-
grating long short-term memory (LSTM) neural networks as
sequential layers within the Autoencoder (AE) framework,
renowned for its ability to capture long-term memory in time
series data [49].
1) LSTM CELL STRUCTURE

LSTM (Long Short-Term Memory) is proposed to enhance
the short-term memory capacity of recurrent neural networks
where D3 is a constant depending on the number of (RNNs) by incorporating long-term memory states, com-
observations and R̄ is the mean of the range. monly referred to as cell states [50]. Figure 12 illustrates the
Upon applying the aforementioned formulas, the results for internal structure of an LSTM cell, which comprises input,
the mean and the range are presented in Table 4. Additionally, output, and forget gates [51].
the calculation of the process limits is illustrated in the same Initially, the LSTM cell initiates with a forget gate,
table. tasked with either retaining or discarding the prior cell state
The final result is presented in the figure 10 and 11, information, ct −1 . he decision to forget or retain information
displaying the control chart for all observations of the melt is determined by processing the input data, xt , and the
cushion parameter. The X-bar control chart indicates that the previous hidden state, ht−1 , through a sigmoid activation
center is at 6.61 cm3 , with an upper control limit (UCL) function, yielding an output value, ft , is between [1, 0],
of 6.67 cm3 and a lower control limit (LCL) of 6.54 cm3 . as in (10) and (11)
On the other hand, the Range chart shows a center of
0.107 cm3 , with a lower limit of 0 and an upper limit ft = σ (Wfx xt + Wfh ht−1 + bf ) (10)
of 0.276 cm3 . Furthermore, the illustrative example in the ex
σ (x) = x (11)
Figure 11 demonstrates the good stability of production e +1
95586 VOLUME 12, 2024

FIGURE 10. Mean and range control chart with calculated LCL, UCL, and Mean.
FIGURE 11. Randomized data points from 10 production cycles for analysis of mean and range control chart.
Following this, the input gate initiates the generation of

a new memory state, gt, by feeding xt and ht-1 into the
tanh activation function, as in (12) and (13). Simultaneously,
the input gate determines which portions of the candidate
memory state will be retained, creating an input state, it,
as in (14). Next, the updated state of the memory cell, ct,
is archived as indicated in equation (15)
gt = tanh(Wgx xt + Wgh ht−1 + bg ) (12)

ex − e−x
tanh (x) = x −x (13)
e +e
it = σ (Wix xt + Wih ht−1 + bi ) (14)
FIGURE 12. Inner structure of LSTM cell. g (t) +f (t) (15)
where W stands for the weight matrices of the gates, and b Ultimately, the updated hidden state is generated by the
represents the bias vectors. output gate. ht, derived from the newly updated memory cell
VOLUME 12, 2024 95587

FIGURE 13. Architecture of the proposed LSTM-Autoencoder.
state and the output state, ot, as in (16) and (17) we obtained a final count of 2974 sequences. Subsequently,
these sequences were reshaped into a 2-dimensional array to
ot = σ (Wox xt + Woh ht−1 + bo ) (16)
be fed into the Autoencoder, representing samples and time
ht = tanh(ct ) ⊙ ot (17) steps. In our case, the final size of the dataset is [2974 × 5].
2) AUTOENCODER b: LSTM ENCODER

Autoencoder is unsupervised learning technique used to The primary function of the LSTM encoder is to serve as a
build data-driven models when labeled data is unavailable. sequence folding layer, transforming features into a batch of
The learning process of the Autoencoder model involves feature sequences organized by time. This process resembles
predicting input data and minimizing the difference, known applying convolution operations separately to each timestep
as the reconstruction loss (RL), from the input to the output. of the feature sequences. Figure 14 illustrates how the AE
Figure 13 illustrates the architecture of the proposed AE, encoder collaborates with a sequence of LSTM unit cells,
comprising encoder, latent, and decoder layers arranged which have been trained to discern the crucial features within
sequentially. In this study, the encoder is composed of two the input sequence.
LSTM layers, one with 64 units and the other with 32 units. Each time series of Xi comprises 5 samples gathered over
The latent layer functions as a repeat vector, duplicating the 5 timesteps. This data is then inputted into the encoder.
output of the encoder to match the input requirements of The first layer receives as input sequences with dimension
the decoder [52]. The decoder layer is composed of two according to the features and timesteps, on our case the
LSTM layers, one with 32 units and the other with 64 units. feature is 1 and the timesteps is 5, the input is [5 × 1], the
Subsequently, the decoder’s output is transformed into a process is obtained on parallel mechanism between each cells
time series distribution that reflects the characteristics of and each timesteps of the sequences, the output of this layer
the input data. The LSTM-AE model acquires knowledge is sequence with dimensional of [5∗ 64]. The second layer
from the training dataset by reducing the loss associated with consists of an LSTM network containing 32 LSTM cells.
reconstruction with both input data (xt) and output data (yt). Each LSTM cell unit is responsible for processing a single
In this paper, the reconstruction loss function is defined as the sample. A sequence of 5 LSTM cells operates sequentially,
mean absolute error (MAE), as in (18). where the first LSTM unit transfers the outcome of the sample
to the second LSTM unit. Following this, the second LSTM
a: MODEL INPUT DATA unit evaluates whether to preserve or discard the preceding
We addressed the anomaly detection challenge by capturing sample from the first LSTM. If it opts to preserve it, the
normal occurrences through time-based sequences derived information is stored in long-term memory. Then, the second
from the newly generated dataset originating from the SPC LSTM cell transmits the information of the sample from the
chart constructed in section III. Initially, the new dataset first LSTM cell, along with the feature information processed
comprised 2978 observations. In this phase, we prepared from the current sample, to the third LSTM cell, and so on.
sequences for the LSTM-AE, each sequence containing a The final LSTM cell, which is the fifth in the sequence,
fixed size of 5 observations for training the model, from integrates all the samples considered valuable, as processed
Sequence 1(X1 = {x1, x2, x3, x4, x5}), Sequence 2 (X2 = by the preceding four LSTM cells. The final LSTM cell
{x2, x3, x4; x5; x6}) until sequence 2974(X2974 = {x2984, outputs the information pertaining to all relevant samples.
x2985, x2986, x2987, x2978}). After sequence preparation, The coveted output is now represented as a [1 × 32] vector,
95588 VOLUME 12, 2024

FIGURE 14. Architecture of the proposed encoder and repetition layers.
FIGURE 15. Architecture of the proposed decoder.
serving as encoded features. The third layer is designed to The fourth layer wich is the first step of the encoder
replicate the vector based on the defined timesteps in the is constituted by 32 LSTM units wich it dealing with
model. It can create copies of the [1 × 32] vector equal to the the concerted vectore of the [5∗ 32] size from the third
number of timesteps specified. In our case, with a timestep layer of the encoder. Each [1 × 32] set is now utilized
size of 5, Layer 3 produces 5 duplicates of encoded features, as an input to the decoder, forming a Layer 4 network
resulting in a two-dimensional vector equivalent to [5× 32]. comprising 32 LSTM cell units. Each LSTM cell unit
The RepeatVector layer acts as a bridge between the encoder handles one [1 × 32] encoded feature. Consequently, the
and decoder modules. output of this layer is a matrix with dimensions [5∗ 64].
The layer 5, wich is containing 64 LSTM units, the output
c: LSTM DECODER of each LSTM unit reflects the learning outcome from the
The primary role of the LSTM decoder is to function as encoded feature, where this output is multiplied with the
a sequence unfolding layer, reconstructing the sequence [1 × 64] vector generated by the supplementary TimeDistri-
structure of the compressed input data across timesteps. bution layer. Simultaneously, each LSTM cell unit produces
Figure 15 demonstrates the decoder’s coordination with a secondary output, encapsulating the processed information
LSTM cells to regenerate the outputs. Layer 4, LSTM (32), state. This output is relayed to the subsequent LSTM cell,
mirrors Layer 2, while Layer 5, LSTM (64), corresponds to except for the last LSTM unit. Additionally, performing
Layer 1. matrix multiplication between the output of each LSTM layer
VOLUME 12, 2024 95589

FIGURE 16. a. Loss Curve for Hyperparameter Fine-Tuning 1 and 2. b. Loss Curve for Hyperparameter Fine-Tuning 3 and 4.
(size 5∗ 64) and the TimeDistribution layer (size 64 × 1) TABLE 5. Data split for training and validation.
results in a vector sized [64 × 1], aligning with the input size.
d: THRESHOLD DEFINITION
In neural networks, a ‘‘threshold’’ is a pivotal decision point
for predictions, while an ‘‘anomaly’’ refers to an observation 3) AUTOENCODER IMPLEMENTATION
significantly deviating beyond it. Instances surpassing this In this section, we detail the methodology for implementing
point are classified as anomalies. In our study, the model anomaly detection for the injection molding machine, with a
is trained on a dataset comprising melt cushion parameter specific emphasis on the Melt cushion parameter. The process
values within a typical range, enabling the computation of begins with training and fine-tuning the hyperparameters
reconstruction error rates for normal cushion value data of the LSTM-AE model using the training data. Following
points. Following training and error computation across this, the trained LSTM-AE model is deployed and assessed
all samples, the maximum error rate is designated as the using the testing data.
threshold. Subsequently, this threshold is applied to the
testing melt cushion dataset, now inclusive of anomalies a: TRAINING PHASE
encountered during production. Each sample undergoes The developed model has been trained using data extracted
computation of a reconstruction error rate for every value, from the generated X-Bar control chart, encompassing
and samples are flagged as anomalies if their error exceeds 2978 new observations clustered around the chart’s mean.
the threshold. Subsequent sections of the study are exclusively focusing
1 X on analyzing the dataset derived from the X-Bar control
RL = MAE = |x t − yt | (18)
N chart, with no further treatment of the range data planned
Trainingdata
in this research. The dataset is split into a training set
where xt is the actual measure, the t is the predicted value and a validation set, with 80% allocated for training and
95590 VOLUME 12, 2024

TABLE 6. Hyperparameters fine-tuning.
20% for validation (Table 5). Subsequently, the training

data undergo normalization using the maximum-minimum
scaler. Following this preprocessing step, the LSTM-AE
model learns through minimizing the residual error between
the input data (xt) and the reconstructed output data
(yt). Afterward, the threshold value is determined as the
maximum residual error. Then, the trained LSTM-AE model
is employed to detect anomalies by reconstructing the input
data. Anomalies are detected through a comparison between
the residual error and the threshold value.
Several iterations of training loops were conducted to
enhance the effectiveness of the model. This involved
adjusting key hyperparameters to optimize performance. The
comparison of model performance was based on metrics
such as coefficient of determination named R-squared (R2)
as in (19) and Mean Absolute Error (MAE) between the
validation and training sets. Table 6 outlines the primary FIGURE 17. Reconstruction residual error based on training data (Normal
training configurations used to optimize the training results. situation).
The main hyperparameters and parameters tuned included:
Number of layers [2 or 4], number of LSTM cells per layer of each tuning. Table 7 provides a performance comparison
[32 or 64], Optimizer: [Adam or RMSprop], number of of Autoencoder applications in various contexts, confirming
epochs [ 50 or 100], batch size [32 or 64], learning rate the proposed model’s effectiveness.
[0.01 or 0.001], dropout rate [0.1 or 0.2]
Pn b: PREDICTION OF THE TRAINING DATA AND
(xi − yi )2
R2 = 1 − Pi=1 n 2
(19) RECONSTRUCTION OF THE ERROR
i=1 (xi − ȳ) To evaluate the proficiency of the learned LSTM-AE model in
where xi is the actual value of observation i, the yi is the reconstructing the input data, the trained model is utilized to
predicted value and ȳ is the mean of the observed va predict the training data. Figure 18 illustrates the assessment
The result obtained from the fine tuning of the hyper- between the predicted data (blue line) and the measured
maparemetrs shows that the optimum hyper parameters data (yellow line). The LSTM-AE model effectively captures
are according to the tuning number 4. With a training the time series variations of the Melt cushion parameter,
R-squared of 0.994, the model effectively explains nearly all as evident from the comparison. Despite being trained
the variability in the training data, while the slightly lower and validated with abundant normal data, a residual error
validation R-squared of 0.992 on training data and 0. 991 on persists between the predicted and actual data within the
validation data, it indicates robust generalization to unseen trained LSTM-AE model. The residual error mentioned is
data. Additionally, the low Mean Absolute Errors (MAE) for the absolute error. Additionally, the histogram in Figure 17
both training 0.0206 and validation 0.0146 sets suggest that illustrates the residual errors for the Melt cushion after
the model’s predictions are consistently close to the actual reconstruction using the trained model. The threshold limit
values. Overall, these results underscore the model’s strong for anomaly detection in the Melt cushion is set at the
fit to the data and its ability to generalize well, affirming maximum residual error, which is below 0.04. The histogram
its reliability for making accurate predictions. Figure 16 indicates that the majority of residual errors fall below this
shows the learning curve of the training and the validation threshold.
VOLUME 12, 2024 95591

FIGURE 18. Reconstruction prediction based on normal dataset.
FIGURE 19. The four main types of anomalies observed in the testing data.
TABLE 7. Performance comparison. extent of deviation exhibited by an observation. Observations

surpassing the threshold are categorized as anomalies.
To effectively illustrate the functionality of reconstruction-
based time series anomaly detection, the LSTM-AE model
can be employed to detect anomalies within each input
sequence. This approach enables us to calculate the recon-
struction error rates associated with the normal samples
within the Melt cushion dataset. Once training is completed
and the reconstruction error is computed for all samples,
the LSTM-AE model aims to attain a low mean absolute
error (MAE) and sets the maximum reconstruction error as
a threshold. if the behavior of the testing set diverges from
the training process, leading to a mean absolute error (MAE)
surpassing the threshold (0.04 in our study), it can be flagged
as an anomaly.
c: TESTING THE TRAINED LSTM-AE ON ANOMALY

An anomaly can be described as an observation that signif- SITUATION
icantly deviates from the majority of the data. Alternatively, In this section, we apply the trained LSTM-AE model to
a threshold serves as a reference point for assessing the forecast the testing dataset, encompassing anomaly scenarios.
95592 VOLUME 12, 2024

FIGURE 20. Reconstruction prediction based on testing dataset (Abnormal situation).
Then, we evaluate the precision of these forecasts by Figure 21 visually illustrates the newly established limits
computing the residual errors, which illustrate the differences derived from the maximum and minimum residuals of the
between the predicted and observed data points. Afterwards, reconstructed model. Initially, we set the trial limit at 0.02916.
we compare these residual errors with the predefined Additionally, we introduced a novel limit based on the min-
threshold values outlined in section 3.2. If a residual error imum residual, set at 0.0188, to improve anomaly detection
surpasses this threshold, indicating a deviation from normal by identifying previously unnoticed anomalies, specifically
behavior, the model identifies and raises an alert for the type 4 anomalies. These anomalies, with errors not exceeding
presence of anomalous instances. Figure 19 illustrates the the threshold of 0.02916, were considered acceptable by the
primary anomalies outlined in this paper within the trend of initial model. This approach strengthens anomaly detection,
the testing dataset. These anomalies are categorized into four uncovering overlooked anomalies and contributing to a more
types: thorough data understanding. In Figure 21, both limits are
• Anomaly of type 1: The values surpass the tolerance depicted based on the reconstruction error: the upper limit at
threshold of the Upper Limit Control (ULC) computed 0.0243 and the lower limit at 0.0188, validating the model’s
on the Statistical Process Control (SPC). capability to detect the four outlined anomaly types.
• Anomaly of type 2: The values fall below the lower limit Figure 22 illustrates the model’s capacity to identify
of the Lower Limit Control (LLC) determined on the anomalies through training on normal scenarios. Anomalies
SPC chart. are flagged in red when the reconstruction error surpasses
• Anomaly of type 3: Ascendant or descendant trend in the 0.0243 threshold, whereas those falling below the
the values, signifying a significant increase or decrease 0.0188 minimum threshold are depicted in green. By adjust-
in the mean value of the observations. ing these thresholds’ upper and lower bounds, the model
• Anomaly of type 4: The mean deviates from the effectively pinpoints all four anomaly types. Furthermore,
principal mean and exceeds the tolerance range it not only identifies instances breaching these thresholds
but also recognizes trends in anomalies, as exemplified by
The process of constructing the testing data (Figure 20), anomaly type 3. Figure 23 illustrates two initial detections:
reveals less alignment between the generated values (high- one indicating a negative trend (left figure) and the other a
lighted in yellow) and the original values (highlighted in positive trend (right figure). In the positive trend, a distinct
blue), indicating some residual discrepancies. anomaly is noticeable with three successive points, while
Figure 20 illustrates the generation of testing data, in the negative trend, observations surpass the tolerance
showcasing a notable difference between the generated values threshold, signaling an anomaly. This characteristic greatly
(highlighted in blue) and the original values (highlighted enhances anomaly detection and aids in early anomaly
in yellow). The observed variance during construction, recognition.
quantified at 0.04, is attributed to errors in the process. The identification of anomalies within observations is
Notably, as the threshold approaches zero, the alignment significantly influenced by the specific thresholds set.
between the constructed data and the original data becomes Enhancing the ability to detect anomalies involves carefully
more pronounced. adjusting these thresholds to align with the desired sensitivity
VOLUME 12, 2024 95593

FIGURE 21. Maximum and minimum threshold limits of the reconstructed anomalous dataset.
FIGURE 22. Detecting the four types of anomalies using minimum and maximum thresholds.
FIGURE 23. Detecting of the anomaly of type 3 ‘‘Descendant trend’’ (left figure ) and ‘‘Ascendant trend’’ (Right
figure.
levels. Table 8 provides a visual representation of how In our study, we conducted four separate trials, systematically
altering these limits can impact the identification process. adjusting the upper limit from 95%, 90%, 85%, down to 80%
95594 VOLUME 12, 2024

TABLE 8. Sensitivity analysis.
of the maximum threshold, and simultaneously adjusting trial 4, with 0.243 as the maximum and 0.0188 as the
the lower limit from 10%, 25%, to 64% of the minimum minimum, identified all four types.
threshold. • Anomaly detection: Through the combination of SPC
In the initial trial, employing thresholds of 0.0277 (95% and LSTM-AE by leveraging data gathered from the
of the maximum threshold) and 0.0029 for the minimum control chart of the mean as training data for LSTM-
limit, only 7 points were detected, making up just 1.71% AEutoencoder (LSTM-AE), the model undergoes effec-
of all observed anomaly situations. Anomaly type 1 was the tive training and exhibits a remarkable proficiency in
only one detected in this setup. In the second trial, with the anomaly detection. In this study, the model attained an
maximum limit lowered to 0.0262 (90% of the maximum) impressive MAPE of 0.0146 and an R-squared value
and the minimum limit raised to 0.0072, the number of of 0.992. Notably, the model successfully identified
detections increased by approximately 2% compared to the various types of anomalies, totaling four distinct types
first trial, with 13 observations identified as anomalies. Here, • Early detectio n: The proposed model demonstrates its
anomalies of types 1, 2, and 4 were detected. Transitioning to capability to achieve early detection of certain types of
the third trial, the detection rate notably increased to 43.6% anomalies, particularly type 3. This solution highlights
(178 anomalies), with types 1, 2, and 4 being identified. that the model not only identifies anomalies beyond
Finally, in the last trial, using thresholds of 0.0243 (80% of the predefined limits but also discerns the ascending and
maximum) and 0.0188 as minimum thresholds, detectability descending trends of anomalies. Such early detection
reached 57.5%, with 235 anomalies detected out of total of anomalies is crucial for initiating timely actions
observations, and all four types of anomalies were identified. to prevent process deviations and make adjustments
These findings underscore the sensitivity of the models to before defective parts are produced or any machine
threshold adjustments. The identification of anomalies within malfunctions occur.
observations is highly affected by the precise thresholds set. This study presents two main limitations that can be
addressed separately to further enhance the quality of the
V. DISCUSSION
results.
• Firstly, the quantity of data is crucial for deep learning
The current paper introduces a fusion of the Statistical
models to achieve high performance in detecting a
Process Mean Control Chart with LSTM Autoencoder
diverse range of anomalies and making accurate pre-
to improve the detection of anomalies in the injection
dictions. In this paper, the trained model was based
molding process, focusing specifically on the ‘‘Melt Injection
on data initially derived from 10160 values, which
Parameter’’ as a case study. The key findings of this study can
were divided into three samples according to the mean
be outlined as follows:
control chart, yielding 3389 values. Since the model
• Smart control charts: The control chart of the mean heavily relies on error reconstruction, having more data
are enhanced by introducing two additional limits, would improve precision in error reconstruction and
the minimum and maximum thresholds, alongside the subsequently enhance detectability.
existing lower and upper control limits (LLC and • Secondly, a limitation lies in the model’s dependency on
ULC). This paper affirms that such augmentation sig- training with normal situations. Ideally, achieving such
nificantly boosts anomaly detection capability. Unlike normal conditions is complex in real-world scenarios
the static LLC and ULC, the min and max limits are due to the degradation and life cycle of machines. Condi-
dynamic, allowing for adjustments based on application tions such as machine state, materials used, maintenance
requirements and the severity of anomaly detection. protocols, and labor factors all need to align optimally,
The determination of these limits often relies on expert which can be challenging to replicate simultaneously
analysis. Through experimentation with different min in experimental settings. Addressing these limitations
and max limit values, the type of anomalies detected can could lead to more robust and applicable results in
vary. In our study, four types were identified. The initial anomaly detection and prediction.
trial, utilizing 0.0277 as the maximum and 0.0277 as • Finally, the real-world conditions frequently involve
the minimum, only detected one type of anomaly, while complex interactions among variables that evolve over
VOLUME 12, 2024 95595

time, a challenge that static models may not effectively hybrid methodologies to enhance efficiency, reduce costs,
address. There is a necessity for more adaptable systems and ensure product quality in industrial operations.
capable of dynamically adjusting to changing conditions The article suggests extending the study to include root
in the injection molding process. A promising solution cause analysis, exploring additional types of anomalies,
to address this challenge is the development of a digital and validating the methodology across various control chart
twin for the injection process [57], this technology formats. Complementing the anomaly detection methodology
creates a virtual model mirroring the manufacturing with root because analysis would provide a more compre-
environment, enabling real-time monitoring, analysis, hensive understanding of the underlying issues. This could
and optimization [58]. involve integrating approaches such as Bayesian networks,
fuzzy logic, or failure tree analysis to pinpoint the root causes
For futures researches, the study can be extended in of each anomaly detected.
several ways to further enhance its scope and applicability, By addressing the identified limitations and pursuing
we propose expanding upon our current methodology by future perspectives, further advancements can be made
integrating supplementary sensor technologies into our towards enhancing the accuracy and applicability of anomaly
data collection process. While our study has leveraged detection methodologies in industrial settings.
data from the dependable Arburg Allrounder 150T and
utilized its internal calculations to monitor the ‘cushion
REFERENCES
parameter,’ we acknowledge that this dataset alone may not
[1] Z. Chen and L. Turng, ‘‘A review of current developments in process and
offer comprehensive insights into all potential equipment quality control for injection molding,’’ Adv. Polym. Technol., vol. 24, no. 3,
anomalies. To enhance the effectiveness of our model, future pp. 165–182, Sep. 2005, doi: 10.1002/adv.20046.
studies could explore the integration of additional sensors. [2] F. Tayalati, M. Azmani, and A. Azmani, ‘‘Artificial intelligence based
plastic injection process for initial parameters setting and process
For example, the inclusion of accelerometers on the motor monitoring-review,’’ in Smart Applications and Data Analysis (Commu-
and hydraulic pump could enable the detection of subtle nications in Computer and Information Science), vol. 1677, M. Hamlich,
vibrations, providing early indicators of underlying issues L. Bellatreche, A. Siadat, and S. Ventura, Eds., Cham, Switzerland:
Springer, 2022, pp. 294–307, doi: 10.1007/978-3-031-20490-6_24.
or equipment wear. Additionally, integrating cavity pressure [3] M. Shamsuzzaman, S. Haridy, A. Maged, and I. Alsyouf, ‘‘Design
sensors to monitor the stability of injection filling within the and application of dual-EWMA scheme for anomaly detection in
mold could offer valuable insights into the manufacturing injection moulding process,’’ Comput. Ind. Eng., vol. 138, Dec. 2019,
Art. no. 106132, doi: 10.1016/j.cie.2019.106132.
process’s performance and identify deviations from expected [4] A. Salonen and M. Deleryd, ‘‘Cost of poor maintenance: A concept for
norms. By incorporating these supplementary sensors into maintenance performance improvement,’’ J. Quality Maintenance Eng.,
our data collection framework, future research efforts can aim vol. 17, no. 1, pp. 63–73, Mar. 2011, doi: 10.1108/13552511111116259.
[5] R. Kanu, ‘‘A study of process variability of the injection molding of
to achieve a more comprehensive understanding of equipment plastics parts using statistical process control (SPC),’’ in Proc. ASEE Annu.
behavior and performance. This enhanced visibility can Conf. Expo., Atlanta, GA, USA, Jun. 2013, pp. 23.110.–23.110.9, doi:
facilitate more proactive anomaly detection and mitigation 10.18260/1-2–19124.
[6] D. R. Cox, ‘‘Prediction by exponentially weighted moving averages and
strategies, ultimately contributing to improved manufacturing related methods,’’ J. Roy. Stat. Soc. B, Stat. Methodol., vol. 23, no. 2,
efficiency and product quality. pp. 414–422, Jul. 1961, doi: 10.1111/j.2517-6161.1961.tb00424.x.
[7] G. L. Grunkemeier, R. Jin, and Y. Wu, ‘‘Cumulative sum curves and
their prediction limits,’’ Ann. Thoracic Surg., vol. 87, no. 2, pp. 361–364,
VI. CONCLUSION Feb. 2009, doi: 10.1016/j.athoracsur.2008.10.053.
The present paper introduces a novel approach that combines [8] V. Ketonen and J. O. Blech, ‘‘Anomaly detection for injection molding
using probabilistic deep learning,’’ in Proc. 4th IEEE Int. Conf. Ind. Cyber-
Statistical Process Control (SPC) with LSTM-Autoencoder Phys. Syst. (ICPS), Victoria, BC, Canada, May 2021, pp. 70–77, doi:
(LSTM-AE) to enhance anomaly detection in injection 10.1109/ICPS49255.2021.9468190.
molding processes, focusing on the critical parameter of [9] A. E. Smith, ‘‘X -bar and R control chart interpretation using neural
computing,’’ Int. J. Prod. Res., vol. 32, no. 2, pp. 309–320, Feb. 1994, doi:
‘‘Melt Injection.’’ The study demonstrates the efficacy of 10.1080/00207549408956935.
this hybrid methodology in detecting anomalies, including [10] J. Mohd. Rohani and K. T. Chan, ‘‘Improving quality with basic statistical
early detection and trend analysis of deviations in the process control (SPC) tools: A case study,’’ Jurnal Teknologi, vol. 35,
pp. 21–34, Jan. 2012, doi: 10.11113/jt.v35.589.
manufacturing process. Moreover, the research highlights [11] T. Chaciński, P. Jaskólski, and M. Pałubicki, ‘‘Study of process stability in
the importance of integrating advanced machine learning injection molding based on product weight,’’ J. Mech. Energy Eng., vol. 6,
techniques with traditional process control methods to no. 1, pp. 21–26, Jul. 2022, doi: 10.30464/jmee.2022.6.1.21.
[12] J. Cao, Y. S. Wong, and K. S. Lee, ‘‘Application of statistical process
improve quality assurance and prevent defects in industrial
control in injection mould manufacturing,’’ Int. J. Comput. Integr. Manuf.,
settings. By leveraging the strengths of both SPC and LSTM- vol. 20, no. 5, pp. 436–451, Jul. 2007, doi: 10.1080/09511920601160148.
AE, the proposed approach offers a promising solution [13] R. Godina, C. Pimentel, F. J. G. Silva, and J. C. O. Matias, ‘‘Improvement
for real-time monitoring and proactive management of of the statistical process control certainty in an automotive manu-
facturing unit,’’ Proc. Manuf., vol. 17, pp. 729–736, Jan. 2018, doi:
production processes. 10.1016/j.promfg.2018.10.123.
Overall, this study contributes valuable insights to the [14] H. D. Nguyen, K. P. Tran, S. Thomassey, and M. Hamad, ‘‘Fore-
field of anomaly detection in manufacturing, paving the casting and anomaly detection approaches using LSTM and LSTM
autoencoder techniques with the applications in supply chain manage-
way for further advancements in process optimization and ment,’’ Int. J. Inf. Manage., vol. 57, Apr. 2021, Art. no. 102282, doi:
quality control. The findings underscore the potential of 10.1016/j.ijinfomgt.2020.102282.
95596 VOLUME 12, 2024

[15] A. R. Agrawal, I. O. Pandelidis, and M. Pecht, ‘‘Injection-molding process [35] J. Yang, Y. Peng, J. Xie, and P. Wang, ‘‘Remaining useful life prediction
control—A review,’’ Polym. Eng. Sci., vol. 27, no. 18, pp. 1345–1357, method for bearings based on LSTM with uncertainty quantification,’’
Oct. 1987, doi: 10.1002/pen.760271802. Sensors, vol. 22, no. 12, p. 4549, Jun. 2022, doi: 10.3390/s22124549.
[16] C.-L. Zhao and R. Schiffers, ‘‘Condition monitoring of non-return valves [36] A. Essien and C. Giannetti, ‘‘A deep learning model for smart manufac-
in injection molding machines using available process and machine data,’’ turing using convolutional LSTM neural network autoencoders,’’ IEEE
in Proc. 35th Int. Conf. Polym. Process. Soc., Çeşme, Turkey, 2020, Trans. Ind. Informat., vol. 16, no. 9, pp. 6069–6078, Sep. 2020, doi:
Art. no. 020054, doi: 10.1063/1.5142969. 10.1109/TII.2020.2967556.
[17] F. Tayalati, A. Azmani, and M. Azmani, ‘‘Application of supervised [37] P. Mach and H. Hochlova, ‘‘Testing of normality of data files for
machine learning methods in injection molding process for initial application of SPC tools,’’ in Proc. 27th Int. Spring Seminar Electron.
parameters setting: Prediction of the cooling time parameter,’’ Prog. Artif. Technol., Meeting Challenges Electron. Technol. Prog., Bankya, Bulgaria,
Intell., pp. 1–17, Apr. 2024, doi: 10.1007/s13748-024-00318-z. 2004, pp. 318–321, doi: 10.1109/ISSE.2004.1490443.
[18] J. Jung, K. Park, B. Cho, J. Park, and S. Ryu, ‘‘Optimization of [38] G. Hatem, J. Zeidan, M. Goossens, and C. Moreira, ‘‘Normality testing
injection molding process using multi-objective Bayesian optimization and methods and the importance of skewness and kurtosis in statistical
constrained generative inverse design networks,’’ J. Intell. Manuf., vol. 34, analysis,’’ BAU J., Sci. Technol., vol. 3, no. 2, p. 7, Jun. 2022, doi:
no. 8, pp. 3623–3636, Dec. 2023, doi: 10.1007/s10845-022-02018-8. 10.54729/ktpe9512.
[19] F. Tayalati, I. Boukrouh, L. Bouhsaien, A. Azmani, and M. Azmani, [39] J. E. Jarrett, ‘‘Analyzing data utilized in process control and continuous
‘‘Design of combined AHP-TOPSIS model for optimizing the selection improvement,’’ Int. J. Econ. Manage. Sci., vol. 5, no. 2, 2016. [Online].
of injection molding machines,’’ Int. J. Eng., 2024. [Online]. Available: Available: https://digitalcommons.uri.edu/cba_facpubs/61/
https://www.ije.ir/article_196709.html
[40] E. H. S. Oberoi, M. Parmar, H. Kaur, and R. Mehra, ‘‘SPC (statistical
[20] D. Annicchiarico and J. R. Alcock, ‘‘Review of factors that affect shrinkage
process control): A quality control technique for confirmation to ability
of molded part in injection molding,’’ Mater. Manuf. Processes, vol. 29,
of process,’’ Int. Res. J. Eng. Technol., vol. 3, no. 6, pp. 666–672,
no. 6, pp. 662–682, Jun. 2014, doi: 10.1080/10426914.2014.880467.
2016.
[21] M. Moayyedian, K. Abhary, and R. Marian, ‘‘The analysis of short shot
possibility in injection molding process,’’ Int. J. Adv. Manuf. Technol., [41] Creating a Capable Process Using Process Capability. Accessed:
vol. 91, nos. 9–12, pp. 3977–3989, Aug. 2017, doi: 10.1007/s00170-017- Jun. 10, 2024. [Online]. Available: https://www.ptonline.com/articles/
0055-1. creating-a-capable-process-using-process-capability
[22] P. Guerrier, G. Tosello, and J. H. Hattel, ‘‘Flow visualization and simulation [42] K. Palmer and K.-L. Tsui, ‘‘A review and interpretations of process
of the filling process during injection molding,’’ CIRP J. Manuf. Sci. capability indices,’’ Ann. Oper. Res., vol. 87, pp. 31–47, Apr. 1999, doi:
Technol., vol. 16, pp. 12–20, Jan. 2017, doi: 10.1016/j.cirpj.2016.08.002. 10.1023/a:1018993221702.
[23] E. Lughofer and K. Pichler, ‘‘Data-driven prediction of possible quality [43] G. S. Wasserman, ‘‘Short run spc using dynamic control chart,’’ Comput.
deterioration in injection molding processes,’’ Appl. Soft Comput., vol. 150, Ind. Eng., vol. 27, nos. 1–4, pp. 353–356, Sep. 1994, doi: 10.1016/0360-
Jan. 2024, Art. no. 111029, doi: 10.1016/j.asoc.2023.111029. 8352(94)90307-7.
[24] S. Lee, Y. Yun, S. Park, S. Oh, C. Lee, and J. Jeong, ‘‘Two phases anomaly [44] R. P. Anjard, ‘‘SPC chart selection process,’’ Microelectron. Rel., vol. 35,
detection based on clustering and visualization for plastic injection no. 11, pp. 1445–1447, Nov. 1995.
molding data,’’ Proc. Comput. Sci., vol. 201, pp. 519–526, Jan. 2022, doi: [45] J.-E. Chiu and T.-I. Kuo, ‘‘Attribute control chart for multivariate
10.1016/j.procs.2022.03.067. Poisson distribution,’’ Commun. Statist., Theory Methods, vol. 37, no. 1,
[25] P. Nunes, E. Rocha, J. Santos, and R. Antunes, ‘‘Predictive main- pp. 146–158, Dec. 2007, doi: 10.1080/03610920701648771.
tenance on injection molds by generalized fault trees and anomaly [46] I. N. Gibra, ‘‘Recent developments in control chart techniques,’’
detection,’’ Proc. Comput. Sci., vol. 217, pp. 1038–1047, Jan. 2023, doi: J. Qual. Technol., vol. 7, no. 4, pp. 183–192, Oct. 1975, doi:
10.1016/j.procs.2022.12.302. 10.1080/00224065.1975.11980695.
[26] B. Silva, R. Marques, D. Faustino, P. Ilheu, T. Santos, J. Sousa, and [47] Z. Wu, M. Yang, M. B. C. Khoo, and P. Castagliola, ‘‘What are the
A. D. Rocha, ‘‘Enhance the injection molding quality prediction with best sample sizes for the Xbar and CUSUM charts?’’ Int. J. Prod. Econ.,
artificial intelligence to reach zero-defect manufacturing,’’ Processes, vol. 131, no. 2, pp. 650–662, Jun. 2011, doi: 10.1016/j.ijpe.2011.02.010.
vol. 11, no. 1, p. 62, Dec. 2022, doi: 10.3390/pr11010062. [48] M. H. Qais, S. Kewat, K. H. Loo, C.-M. Lai, and A. Leung, ‘‘LSTM-
[27] M. Uluskan, ‘‘Decreasing defects in plastic injection molding and vibration based stacked autoencoders for early anomaly detection in induction
welding processes through statistical process control,’’ Open J. Nano, heating systems,’’ Mathematics, vol. 11, no. 15, p. 3319, Jul. 2023, doi:
vol. 6, no. 2, pp. 7–18, 2021. 10.3390/math11153319.
[28] M. Malindzakova, K. Čulková, and J. Trpčevská, ‘‘Shewhart control charts [49] S. Hochreiter and J. Schmidhuber, ‘‘Long short-term memory,’’
implementation for quality and production management,’’ Processes, Neural Comput., vol. 9, no. 8, pp. 1735–1780, Nov. 1997, doi:
vol. 11, no. 4, p. 1246, Apr. 2023, doi: 10.3390/pr11041246. 10.1162/neco.1997.9.8.1735.
[29] V. Moradi and B. S. Gildeh, ‘‘Fuzzy process capability plots for families
[50] Y. Yu, X. Si, C. Hu, and J. Zhang, ‘‘A review of recurrent neural networks:
of one-sided specification limits,’’ Int. J. Adv. Manuf. Technol., vol. 64,
LSTM cells and network architectures,’’ Neural Comput., vol. 31, no. 7,
nos. 1–4, pp. 357–367, Jan. 2013, doi: 10.1007/s00170-012-4003-9.
pp. 1235–1270, Jul. 2019, doi: 10.1162/neco_a_01199.
[30] M. Muaz, H. Yu, W. L. Sung, C. Liu, and B. Drescher, ‘‘A
[51] L. Medsker. (1201). Recurrent Neural Networks | Design and
multitask encoder–decoder model for quality prediction in injection
Application. Accessed: Jun. 10, 2024. [Online]. Available: https://www.
moulding,’’ J. Manuf. Processes, vol. 103, pp. 238–247, Oct. 2023, doi:
taylorfrancis.com/books/edit/10.1201/9781003040620/recurrent-neural-
10.1016/j.jmapro.2023.08.030.
networks-lakhmi-jain-larry-medsker
[31] H. Jung, J. Jeon, D. Choi, and J.-Y. Park, ‘‘Application of machine
learning techniques in injection molding quality prediction: Implications [52] B. M. Dillon, L. Favaro, T. Plehn, P. Sorrenson, and M. Krämer,
on sustainable manufacturing industry,’’ Sustainability, vol. 13, no. 8, ‘‘A normalized autoencoder for LHC triggers,’’ SciPost Phys. Core, vol. 6,
p. 4120, Apr. 2021, doi: 10.3390/su13084120. no. 4, p. 74, Nov. 2023, doi: 10.21468/scipostphyscore.6.4.074.
[32] M. Huang, C. Liu, and K. Ke, ‘‘Calibration of cavity pressure simu- [53] G. Narkhede, A. Hiwale, B. Tidke, and C. Khadse, ‘‘Novel MIA-
lation using autoencoder and multilayer perceptron neural networks,’’ LSTM deep learning hybrid model with data preprocessing for fore-
Polym. Eng. Sci., vol. 61, no. 10, pp. 2511–2521, Oct. 2021, doi: casting of PM2.5 ,’’ Algorithms, vol. 16, no. 1, p. 52, Jan. 2023, doi:
10.1002/pen.25777. 10.3390/a16010052.
[33] M. Ramezankhani, M. Harandi, R. Seethaler, and A. S. Milani, ‘‘Smart [54] H. Zhu, F. Meng, S. Rho, M. Li, J. Wang, S. Liu, and F. Jiang,
manufacturing under limited and heterogeneous data: A sim-to-real ‘‘Long short term memory networks based anomaly detection for
transfer learning with convolutional variational autoencoder in thermo- KPIs,’’ Comput., Mater. Continua, vol. 61, no. 2, pp. 829–847, 2019.
forming,’’ Int. J. Comput. Integr. Manuf., vol. 37, nos. 1–2, pp. 18–36, Accessed: Jun. 11, 2024. [Online]. Available: https://cdn.techscience.
Feb. 2024, doi: 10.1080/0951192x.2023.2257623. cn/files/cmc/2019/v61n2/20191016011452_87994.pdf
[34] P. Po and H. Kim, ‘‘Smart factory mold injection anomaly detection using [55] S. Seshan, D. Vries, J. Immink, A. van der Helm, and J. Poinapen, ‘‘LSTM-
deep autoencoder neural network,’’ in Proc. IEEE Int. Conf. Consum. based autoencoder models for real-time quality control of wastewater
Electron.-Asia (ICCE-Asia), Busan, South Korea, Oct. 2023, pp. 1–3, doi: treatment sensor data,’’ J. Hydroinformat., vol. 26, no. 2, pp. 441–458,
10.1109/ICCE-Asia59966.2023.10326430. Feb. 2024, doi: 10.2166/hydro.2024.167.
VOLUME 12, 2024 95597

[56] N. Khan, S. U. Khan, and S. W. Baik, ‘‘Deep autoencoder-based hybrid LOUBNA BOUHSAIEN received the degree from
network for building energy consumption forecasting,’’ Comput. Syst. Sci. the Faculty of Sciences and Technologies of
Eng., vol. 48, no. 1, pp. 153–173, 2024, doi: 10.32604/csse.2023.039407. Tangier, Abdelmalek Essaâdi University, in 2022.
[57] P. Bibow, M. Dalibor, C. Hopmann, B. Mainz, B. Rumpe, D. Schmalzing, She is currently pursuing the Ph.D. degree in
M. Schmitz, and A. Wortmann, ‘‘Model-driven development of a artificial intelligence and data science, applied to
digital twin for injection molding,’’ in Advanced Information Systems human resources management. She is a Statistical
Engineering (Lecture Notes in Computer Science), vol. 12127, S. Dustdar, Engineer. With a strong interest in using data
E. Yu, C. Salinesi, D. Rieu, and V. Pant, Eds., Cham, Switzerland: Springer,
to find useful answers to business problems.
2020, pp. 85–100, doi: 10.1007/978-3-030-49435-3_6.
Her research interests include machine learning,
[58] A. Rehmer, M. Klute, H.-P. Heim, and A. Kroll, ‘‘A digital twin for
predictive analytics, employee behavior modeling,
part quality prediction and control in plastic injection molding,’’ in
Modeling, Identification, and Control for Cyber-Physical Systems Towards and risk assessment.
Industry 4.0. Amsterdam, The Netherlands: Elsevier, 2024, pp. 79–109,
doi: 10.1016/B978-0-32-395207-1.00014-7.
ABDELLAH AZMANI received the Ph.D. degree

FAOUZI TAYALATI received the degree from the in industrial computing from the University of
National School of Mechanics and Microtechnics, Science and Technology, Lille, in 1991. He was a
Besançon, in 2011, and the master’s degree in Professor with the École Centrale de Lille and the
electrical engineering from the Faculty of Applied Institute of Computer and Industrial Engineering,
Sciences, Béthune, France. He is currently an Lens. He is with Intelligent Automation and
Engineer. With more than twelve years of expe- Biomed Genomics Laboratory. He is a Professor
rience in industrial engineering, predominantly with the Faculty of Sciences and Technology of
in the automotive industry. His work explores Tangier, Morocco. He has contributed to many
the integration of artificial intelligence techniques scientific research.
into manufacturing processes, optimization of
production systems, and application of AI in maintenance management,
drawing insights from his extensive industry experience.
MONIR AZMANI received the Ph.D. degree

in computer engineering, automation, and signal
IKHLASS BOUKROUH received the degree from processing from the University of the Littoral Opal
the Faculty of Sciences and Technologies of Coast. Currently, he is an authorized Professor
Tangier, Abdelmalek Essaâ University, in 2021. with the Faculty of Science and Technology
She is currently pursuing the Ph.D. degree in of Tangier and attached to the Laboratory of
artificial intelligence and data science, applied Computer Science, Systems, and Telecommunica-
to the e-commerce sector. She is a Statistical tions, with more than 12 years of experience in
Engineer. With a strong interest in using data teaching and pedagogical supervision. Specializes
to find useful answers to business problems. in computer science, industrial computing, and
Her research interests include machine learning, signal processing, also in the field of artificial intelligence. He has experience
predictive analytics, customer behavior modeling, in teaching, training, intervention, and studies and research.
and risk assessment.
95598 VOLUME 12, 2024

Hybrid Approach Integrating Deep Learning-Autoencoder With Statistical Process Control Chart For Anomaly Detection Case Study in Injection Molding Process

Uploaded by

Hybrid Approach Integrating Deep Learning-Autoencoder With Statistical Process Control Chart For Anomaly Detection Case Study in Injection Molding Process

Uploaded by

Received 11 June 2024, accepted 5 July 2024, date of publication 9 July 2024, date of current version 19 July 2024.

Digital Object Identifier 10.1109/ACCESS.2024.3425582

Hybrid Approach Integrating Deep

I. INTRODUCTION challenge of rising costs and demanding quality standards [2].

VOLUME 12, 2024 95577

95578 VOLUME 12, 2024

FIGURE 1. The standard structure of the injection molding machine.

a framework that utilizes sensor data obtained from plastic

VOLUME 12, 2024 95579

95580 VOLUME 12, 2024

VOLUME 12, 2024 95581

95582 VOLUME 12, 2024

establishing a template for detecting abnormal situations.

of production time. The dataset is categorized into two

VOLUME 12, 2024 95583

TABLE 1. Main statistical outcomes from the normal dataset.

FIGURE 7. Analysis of capability using the provided normal dataset.

of the process relative to the target or nominal value.

95584 VOLUME 12, 2024

TABLE 2. Calculation of the process capability. 3) X-BAR AND R CHART

a: CALCULATION OF THE MEAN CHAR

VOLUME 12, 2024 95585

FIGURE 9. The main type of the statistical process control chart.

1) LSTM CELL STRUCTURE

95586 VOLUME 12, 2024

Following this, the input gate initiates the generation of

gt = tanh(Wgx xt + Wgh ht−1 + bg ) (12)

VOLUME 12, 2024 95587

FIGURE 13. Architecture of the proposed LSTM-Autoencoder.

2) AUTOENCODER b: LSTM ENCODER

95588 VOLUME 12, 2024

FIGURE 14. Architecture of the proposed encoder and repetition layers.

FIGURE 15. Architecture of the proposed decoder.

VOLUME 12, 2024 95589

95590 VOLUME 12, 2024

TABLE 6. Hyperparameters fine-tuning.

20% for validation (Table 5). Subsequently, the training

VOLUME 12, 2024 95591

FIGURE 18. Reconstruction prediction based on normal dataset.

TABLE 7. Performance comparison. extent of deviation exhibited by an observation. Observations

c: TESTING THE TRAINED LSTM-AE ON ANOMALY

95592 VOLUME 12, 2024

FIGURE 20. Reconstruction prediction based on testing dataset (Abnormal situation).

VOLUME 12, 2024 95593

95594 VOLUME 12, 2024

TABLE 8. Sensitivity analysis.

VOLUME 12, 2024 95595

95596 VOLUME 12, 2024

VOLUME 12, 2024 95597

ABDELLAH AZMANI received the Ph.D. degree

MONIR AZMANI received the Ph.D. degree

95598 VOLUME 12, 2024

You might also like