Implementing and evaluating the quality 4.0 PMQ framework for process monitoring in automotive manufacturing

Abstract

This study presents an applied integration of machine learning (ML) within the Process Monitoring for Quality (PMQ) framework to address persistent limitations in traditional quality control systems, particularly their inability to manage high-dimensional and real-time manufacturing data. This research enhances the PMQ framework with a novel Validate phase that introduces human oversight and interpretability into the ML decision-making loop. The modified framework has been implemented in a high-precision automotive component facility. The study relied on various ML algorithms, such as Decision Trees (DT), Random Forest (RF), Gradient Boosting Machine (GBM), Logistic Regression (LR), Support Vector Machine (SVM), and Artificial Neural Networks (ANN), to classify and predict defects in engine valves during manufacturing processes. The findings highlighted that GBM and RF provided the best performance, achieving an F1 score of 0.98 and an AUC of 0.99. Feature importance analyzes identified seat height and undercut diameter as key predictors, reinforcing the relevance of interpretable ML in industrial quality management. Beyond technical accuracy, this work demonstrates how structured human-machine collaboration can foster trust in AI-driven quality control, offering a scalable blueprint for Quality 4.0 adoption. The findings contribute to academic literature and industrial practice by bridging conceptual frameworks and real-world implementation strategies for AI-enhanced quality assurance.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-10226-4.

Introduction

The fourth industrial revolution, referred to as Industry 4.0, has fundamentally changed organizational operations^1,43. Smart manufacturing became highly dependent on innovations in automation, digital networking, and real-time data analysis, which allowed systems to run with greater speed, accuracy, and adaptability⁷. Simultaneously, this growing complexity, which shows in higher data volumes, more variables, and tighter tolerances, has started to highlight the shortcomings of traditional quality control systems such as Statistical Process Control (SPC) and Six Sigma⁴². Although quite successful once, these conventional models are not always suited to handle the data-rich nature of modern industrial systems. Therefore, more flexible and insight-driven quality methods are needed².

This demand has led to Quality 4.0, a concept that brings digital technologies into the field of quality control. Originally proposed by Jacob²⁴, and then developed by Radziwill (2020), the concept focuses on enabling teams to use digital technologies to support operational performance and strategic decision-making. Quality 4.0, according to Sader et al.⁴⁰, combines more traditional methods of quality assurance with technologies including artificial intelligence (AI), the Internet of Things (IoT), and big data analytics. However, Antony et al.⁴ noted that the human contribution remains indispensable. Although machines may process information rapidly, it is human judgment that interprets these insights and applies them effectively. Together, these views suggest that Quality 4.0 is not a wholesale replacement of traditional quality systems, but rather a complementary evolution, one that opens up new possibilities for improving efficiency, reliability, and sustainability¹³.

Central to this shift is the growing use of advanced analytics, particularly machine learning (ML) and deep learning (DL), to uncover patterns in manufacturing data and provide actionable feedback¹⁵. These methods have shown the potential to improve predictive accuracy, reduce waste, and identify process inefficiencies before they escalate. However, adoption across the industry remains slow. Case studies documenting successful applications of these technologies are relatively scarce⁹. Beyond that, many engineers and quality practitioners are not yet comfortable working with AI-based tools, often due to a lack of training or exposure^14,19. Furthermore, quality engineers typically refrain from using ML-based predictions due to the lack of trust in AI-generated predictions²³. Although the academic literature has explored the theoretical promise of Industry 4.0 and Quality 4.0 in-depth, practical examples that demonstrate the value of these tools compared to traditional quality techniques are limited. As long as that remains the case, the hesitation toward adopting AI in quality control is likely to persist. In general, this research aims to be among the first studies to empirically demonstrate how Quality 4.0 frameworks can be systematically applied in high-precision manufacturing environments, provide interpretable ML-based predictions, and increase the trust of quality professionals in deploying ML tools in quality monitoring and defect detection and prediction.

Therefore, this study applies ML algorithms within the Process Monitoring for Quality (PMQ) framework, a structured Quality 4.0 approach designed for real-time quality monitoring and continuous improvement (Escobar et al., 2021b; Escobar et al.¹⁶,. The PMQ framework, structured around key phases such as Identify, Acsensorize, Discover, Learn, Predict, Redesign, and Relearn, guides organizations in effectively capturing data, applying advanced analytics, and iteratively improving their processes¹⁷. However, this study suggests a slight modification to the PMQ framework by adding an extra phase after the Predict phase, named Validate, to increase trust in the framework among quality practitioners and ensure the interpretability and reliability of findings. Therefore, this research demonstrates the implementation of the modified PMQ framework in an automotive manufacturing environment, specifically within the context of engine valve production, where six critical dimensional features bottom, stem middle, stem upper, head diameter, seat height, and undercut diameter must be rigorously monitored to ensure product integrity and performance. Unlike other studies focused solely on algorithmic innovation, this research prioritizes applied integration. Consequently, our contribution lies in embedding interpretable ML models and operational validation mechanisms, such as the newly introduced Validate phase within the PMQ framework, to build trust and ensure actionable outcomes in real-world settings. Lastly, this study aims to offer a roadmap for organizations looking to transition from conventional quality management practices to more intelligent, data-driven quality systems, ultimately enhancing their competitiveness in the era of Industry 4.0.

Background on quality 4.0

Quality 4.0 is a modern quality management framework that integrates traditional quality principles with digital transformation technologies embedded in Industry 4.0²⁰. It emphasizes continuous monitoring, predictive analytics, real-time decision-making, and proactive defect prevention, with the aim of improving process robustness, reducing variability, and fostering operational excellence³⁷. For example, ML algorithms offer unique advantages in learning from historical and real-time data, identifying nonlinear relationships, and anticipating quality deviations before they occur^28,46. When deployed correctly, these models can improve defect detection rates, improve root cause analysis, support preventive maintenance, and optimize decision-making³². Furthermore, their ability to process large and multidimensional datasets far surpasses that of traditional SPC tools such as X-bar charts or Hotelling’s T² (Mjimer et al., 2023). Despite these benefits, the successful implementation of Quality 4.0 is still modest, with empirical studies estimating successful adoption rates that range between 13% and 20%¹⁴. This slow uptake reflects a combination of various organizational and technological challenges³¹. First, quality professionals and decision-makers struggle with the interpretability of ML predictions, resulting in a great deal of hesitation in adopting them without human intervention²³. Furthermore, the weak technological infrastructure and the high initial cost hinder the integration of traditional quality management practices with advanced technologies^3,11,47. Resistance to change and cultural shock also significantly hinder the adoption of Quality 4.0^25,39. Therefore, researchers^12,27 suggested that Quality 4.0 adoption must be governed by critical principles such as transparency and interpretability, infrastructure readiness, human involvement and validation, and training and upskilling. Despite these efforts to facilitate the adoption of Quality 4.0, there remains a notable gap in empirical research that demonstrates how organizations can systematically implement advanced Quality 4.0 strategies effectively. Much of the existing literature is conceptual, with limited real-world case studies focusing on critical aspects such as the following.

• The integration of advanced digital technologies into existing quality management systems⁴².

• The role of human factors, including decision-making processes, organizational culture, and workforce training, in facilitating or hindering Quality 4.0 adoption⁹.

In response, Escobar et al.^16,17 proposed the PMQ practical framework that is designed to address real-time process monitoring and continuous improvement challenges in modern manufacturing environments. The PMQ framework follows a cyclical methodology that encompasses the key stages: Identify, Acsensorize, Discover, Learn, Predict, Redesign, and Relearn. This iterative process aligns closely with Quality 4.0 principles, using machine learning algorithms, real-time data collection, and iterative feedback loops to enhance predictive capabilities and proactively mitigate quality issues (Escobar et al., 2021b). Moving beyond traditional SPC, PMQ offers a dynamic approach to quality management, enabling organizations to anticipate defects, optimize resource allocation, and improve overall process efficiency. However, empirical evidence that shows its practical effectiveness and its adoption among quality professionals remains scarce. This can be attributed to their lack of trust in quality predictions generated from ML methods²³.

This lack of case-based research highlights the need for more in-depth studies showcasing how PMQ can be operationalized in industrial settings that are data-intensive, high-variability, and real world. Industries such as automotive manufacturing, with their stringent quality requirements and complex supply chains, provide ideal contexts for testing and refining these advanced frameworks. Therefore, the present study aims to contribute to the growing body of knowledge on Quality 4.0 by:

• Modifying the PMQ framework by adding a validate phase after the prediction phase to incorporate domain expertise and interpretability of ML-based predictions.

• Demonstrating how the PMQ framework can be implemented effectively in a high-precision manufacturing environment.

• Providing empirical insights into the methodological technological integration and organizational adjustments required for successful adoption.

• Offering practical recommendations to address the common barriers to Quality 4.0 implementation, thereby supporting both academic research and industrial best practices³⁸.

By addressing these critical areas, this study aims to bridge the gap between theoretical models and real-world applications, offering valuable contributions to the field of modern quality management in the era of Industry 4.0.

Methodology

Case explanation

This research focuses on a leading Asian-based automotive part manufacturer well-known for manufacturing engine valves used in a variety of vehicle types, from passenger cars to heavy-duty trucks. Engine valves constitute a vital element of the company’s portfolio, with demand expected to increase due to market trends. Therefore, the company has started expanding its manufacturing capacity to meet the expected increase in demand.

However, recent internal evaluations found alarming increases in scrap rates and ongoing quality issues. The company specifically found dimensional discrepancies, most importantly in seat height and crucial diameters, as the main causes of these flaws. Such deviations can affect the operational integrity of engine valves, which are essential for effective combustion and low leakage. For example, the valve seat must seal exactly during operation; even small geometric flaws could lower combustion efficiency and cause mechanical breakdowns. Similarly, variations in head and undercut sizes could alter airflow dynamics, causing early wear or loss of performance. These quality flaws not only raise production costs but also damage reputation. Looking for more proactive data-driven tools for early defect discovery and root cause analysis, corporate leadership responded by initiating initiatives to improve the quality control capacity of the corporation.

Therefore, the objective of this study and the company is to reduce scrap rates, improve cost efficiency, and maintain high standards of product quality amidst expanding operations. To achieve this, the study collected a dataset of 1,000 valves, each described by six critical features: stem bottom, middle, and upper diameters, head diameter, seat height, and undercut diameter. Each valve was labeled based on a binary quality outcome (defective vs. non-defective), forming the basis for supervised learning models. These features were selected based on their direct role in achieving a secure fit, adequate airflow, and thermal stability. Figure 1 provides a technical illustration of an engine valve, highlighting the six critical features under investigation.

Fig. 1.

Technical illustration of an engine valve.

The company initially employed traditional SPC tools, such as univariate X-bar/S control charts and multivariate Hotelling T² control charts, to detect process anomalies. Although these methods provided a basic level of control, they frequently failed to identify certain defects due to the high-dimensional nature of valve measurements and complex interdependencies between critical features. Recognizing the limitations of conventional quality control techniques, the company decided to explore a more robust solution under the umbrella of Quality 4.0, using data-driven methodologies to improve process efficiency and product quality. Therefore,

To this end, the PMQ framework was selected as a robust and advanced approach to enhance the capabilities of defect detection. This framework integrates ML algorithms, real-time data analytics, and continuous improvement principles to uncover hidden patterns in production data, identify root causes of quality deviations, and optimize manufacturing processes.

Data preprocessing and model Preparation

A dataset comprising 1,000 data points was collected from the valve production line, with each data point representing a single product and containing six key dimensional features: (i) Stem bottom diameter, (ii) Stem middle diameter, (iii) Stem Upper diameter, (iv) Head diameter, (v) Seat height, and (vi) Undercut diameter. Although the sample size might appear limited for data-intensive environments, it was deemed sufficient for a pilot study representing typical production conditions within a controlled timeframe. Each data point was labeled as either defective or non-defective, forming the problem as a binary classification task. This dataset facilitated the development and evaluation of machine learning models aimed at predicting product quality outcomes.

Encoding and data cleaning

• Binary Encoding: The quality outcomes were encoded as binary values in accordance with the PMQ framework’s assumption of discrete quality states, where defective items were labeled as 0, and non-defective items were labeled as 1 (refer to Eq. (1)).

1

• Data Cleaning: To ensure data integrity, a thorough cleaning process was conducted. Outliers resulting from sensor malfunctions were identified using statistical techniques such as the Z-score method and subsequently removed. Missing values were imputed using the mean imputation process for continuous variables, preserving the overall data distribution and mitigating potential biases¹⁴.

Data partitioning and model validation

• Data Split: The dataset was partitioned into training, validation, and testing sets to prevent overfitting and ensure a robust model. Specifically, on the 1,000 data points, 900 (90%) were allocated for training and validation, while 100 (10%) were reserved for final testing to ensure an unbiased performance evaluation of the models.

• Cross-Validation: A 10-fold cross-validation approach was utilized during model training^18,29. This approach balances computational efficiency with the need to minimize variance in model performance estimation. Each fold used 90% of the data for training and 10% for validation, and the process was repeated in all folds. This method provides a robust estimate of model performance, effectively reducing the risk of overfitting³³.

PMQ framework

The PMQ framework, illustrated in Fig. 2, serves as the foundation for integrating advanced analytics into the existing quality system. The original PMQ framework comprises the following key stages: Identify, Acsensorize, Discover, Learn, Predict, Redesign, and Relearn^16,17. However, this research suggests adding an extra phase to address a common issue faced by quality professionals, which is the lack of trust in ML-extracted findings²³. The suggested phase is named Validate and it comes after the Predict phase. This added phase aims to add an extra layer of insurance to quality professionals by explaining the predictions from ML algorithms, verifying the accuracy of the predictions, incorporating domain expertise in the predictions, and preventing any unnecessary changes in the subsequent Redesign phase. The application of each stage in this study is described below:

Fig. 2.

PMQ cycle.

Stage 1: Identify.

A project decision matrix was used to prioritize the valve production line for quality improvements due to its high defect rates and associated cost implications. The availability of historical production data and the feasibility of deploying IoT-based sensors were key factors in this decision¹⁷.

Stage 2: Acsensorize.

Digital sensors were installed throughout the production line to capture real-time dimensional measurements. To ensure data reliability, all sensors underwent rigorous calibration procedures and regular performance checks, minimizing the risk of measurement errors.

Stage 3: Discover.

The raw data collected from the sensors were explored using descriptive statistics and correlation analysis to identify patterns and relationships among variables. Noise reduction techniques and data integration strategies were used to merge multiple sensor streams, improving the quality of the data for subsequent analysis¹⁴.

Stage 4: Learn.

Several machine learning algorithms were applied to the preprocessed dataset, including Decision Trees, Gradient Boosting, Random Forests, SVM, ANN, and Logistic Regression. Hyperparameter tuning was carried out by cross-validation to optimize model performance, balancing predictive accuracy with computational efficiency. Model evaluation metrics included accuracy, precision, recall, F1 score, and Area Under the Receiver Operating Characteristic Curve (AUC-ROC).

Stage 5: Predict.

The best-performing models were deployed to classify items as defective (0) or non-defective (1). This stage involved model optimization techniques and the use of ensemble methods to improve classification accuracy and robustness, particularly in identifying subtle defect patterns.

Stage 6: Validate.

This phase provides interpretability of ML-based predictions to quality professionals, boosting their confidence and understanding of the findings from the Predict phase. Additionally, domain knowledge is used to validate the accuracy of predictions.

Stage 7: Redesign.

This analysis provided critical engineering insights, allowing the identification of process inefficiencies and guiding the development of targeted process improvement strategies to reduce scrap rates.

Stage 8: Relearn.

Recognizing that manufacturing environments are dynamic, a plan for continuous model improvement was established (Escobar et al., 2021b). This involves periodic retraining of the models using new data to account for evolving production conditions, thereby ensuring sustained model accuracy and relevance over time.

Quality control charts

This study focuses on monitoring the quality of six critical features associated with engine valve production in the automotive industry. Given the continuous nature of the data, which can be represented numerically, variable quality control charts are used to ensure effective process monitoring. Specifically, both univariate and multivariate control charts are utilized to capture different dimensions of process performance.

The X-bar/S control chart is implemented to monitor each process separately, enabling the detection of out-of-control signals specific to individual quality characteristics. This SPC tool comprises two components: (i) an X-bar chart, which tracks the sample means, offering insights into shifts or trends in the process average, thereby identifying potential issues related to process centering, and (ii) an S chart, which monitors the sample standard deviations to detect changes in process variability, which may indicate instability or the presence of special causes of variation. The general equations (Eq. 2 to 5) for the X-bar and S-charts are as follows:

2

3

4

5

Where UCL and LCL represent the upper and lower control limits, respectively, is the grand average of all samples, and S is the average sample standard deviation. Where , , and are control chart constants dependent on sample size. This dual-chart approach enables quality practitioners to promptly identify abnormal patterns or trends, facilitating proactive corrective actions to maintain process stability and product quality.

While univariate control charts focus on individual process variables, complex manufacturing environments often involve multiple interrelated characteristics. To address this, the Hotelling T² chart, a multivariate quality control (MQC) tool, is employed to monitor the integrated performance of all six features simultaneously (Mjimer et al., 2023).

Unlike univariate charts, Hotelling’s T² chart captures the covariance structure between variables, allowing for the detection of subtle shifts in the process mean vector that might remain undetected when variables are analyzed independently (Mjimer et al., 2023; Thefeid⁴⁵,. This is particularly valuable in modern manufacturing systems where critical-to-quality (CTQ) characteristics are often correlated. The Hotelling’s T² statistic and upper control limit are calculated by utilizing Eqs. 6 and 7:

6

7

Where is the sample size, is the vector of sample means, µ is the vector of population means, and S is the sample covariance matrix. As for the UCL, is the number of quality characteristics, and is the critical value from the F-distribution at a significance level of α with degrees of freedom and .

By accounting for the interdependencies between variables, Hotelling’s T² chart enhances the sensitivity of quality monitoring, making it indispensable for ensuring process stability and product reliability in multivariate production environments^21,36.

Machine learning tools

Machine Learning represents a transformative approach to data analysis and decision-making, grounded in the principles of AI. Over recent decades, ML has evolved from a theoretical concept into a robust, practical tool, widely adopted across various domains, including quality management and manufacturing systems.

In this study, several ML algorithms are used to analyze quality-related data, each offering distinct advantages in pattern recognition, anomaly detection, and predictive analytics. The following algorithms are considered: Decision Tree (DT), Gradient Boosting Machine (GBM), Random Forests (RF), Support Vector Machines (SVM), Artificial Neural Networks (ANN), and Logistic Regression (LR). The mathematical formulation for each of the mentioned algorithms is shown in Supplementary Appendix A.

Each algorithm has unique strengths in handling structured data, non-linear relationships, and complex interactions between variables. Table 1 provides an overview of these ML techniques, highlighting their typical applications in quality control and process optimization.

Table 1.

Description of ML algorithms and their application in quality.

Algorithm	Brief Description	Application in Quality Management	Reference
Decision Trees	A simple and interpretable model for classifying data based on feature splits.	Classifying data and identifying key factors influencing product quality.	Wuest et al.49
Gradient Boosting Machines	An ensemble method that builds trees sequentially to improve prediction accuracy.	Enhancing fault detection by improving prediction accuracy in complex processes.	Said et al.41
Random Forests	An ensemble learning method that averages multiple decision trees for robust predictions.	Classifying data and identifying key factors influencing product quality.	Wright and König48
Support Vector Machines	A model that finds the hyperplane best-separating classes in a high-dimensional space.	Monitoring process stability by identifying critical boundaries between different quality states.	Escobar and Morales-Menendez (2019)
Artificial Neural Networks	A model inspired by the human brain and is effective for capturing nonlinear relationships.	Predictive maintenance and defect detection.	Cai et al.8
Logistic Regression	A statistical model predicting binary outcomes based on input features.	Assessing probabilities of quality outcomes based on process data.	Choudhary et al.10

Multiple metrics are used to ensure a robust assessment of the ML models implemented. First, a confusion matrix is established to highlight the model’s classification performance using four key metrics, including the true positive (TP), true negative (TN), false positive (FP), and false negative (FN) values. False positive, also known as Type I Error (α), occurs when a non-defective item is incorrectly classified as defective. Where False negative, Known as Type II Error (β), occurs when a defective item is mistakenly classified as non-defective. Additionally, other metrics such as the F1 score and the ROC curve are also used to evaluate ML models. By aggregating two important factors, precision, and recall, into a single metric via their harmonic mean, the F1 score provides a balanced assessment of the performance of a model (see Eq. 8). This guarantees that the classification accuracy evaluation considers false positives as well as false negatives. On the contrary, the Receiver Operating Characteristic (ROC) curve offers a graphical representation of the trade-off between accurately spotting positive cases (true positives) and the risk of misclassifying negatives as positives (false positives).

8

.

By integrating these ML techniques with traditional quality control tools, this study aims to develop a comprehensive framework for proactive quality management, capable of identifying patterns and predicting potential process deviations in real time⁴⁴.

Implementation of the PMQ quality 4.0 framework

Identify

With the expansion of the company’s production, the quality of the engine valve deteriorated, manifesting itself in increased scrap rates, uneven seat diameters, and increased customer complaints. Given that engine valves are critical components that influence engine efficiency, emissions, and overall vehicle performance²⁶, addressing these defects became a strategic priority. A project decision matrix was used to assess the feasibility and importance of investigating valve quality. Through a structured problem definition process, the team identified key quality concerns and narrowed the scope to engine valves, ensuring a targeted approach to diagnosing the factors that contribute to quality degradation. This step established a strong foundation for subsequent data-driven analysis and improvement initiatives.

Acsensorize

To capture real-time production data and monitor key quality parameters, the company deployed a large array of high-precision sensors on the valve manufacturing line. These sensors recorded essential dimensional measurements, including stem diameters, head diameter, seat height, and undercut diameter, all of which are critical to ensure proper valve function. Consistent with the data collection and cleaning procedures outlined in Sect."Data preprocessing and model Preparation", this stage prioritized accuracy and reliability by filtering out noisy or incomplete data points. The integration of sensor-based monitoring not only improved the granularity of quality tracking but also ensured that key production variables were systematically recorded. By the end of this stage, the team had compiled a robust high-integrity dataset suitable for in-depth defect analysis, laying the groundwork for advanced statistical and AI-driven diagnostics.

Discover

The Discover stage emphasizes the exploration, analysis, and interpretation of cleaned data to find important trends, patterns, and possible causes of variance influencing engine valve quality. This stage seeks to identify important process elements affecting quality results and evaluate their conformity to given criteria. Six main characteristics, inlet valve stem size at its lower, middle, and upper parts, valve head diameter, seat height, and valve undercut diameter, that considerably affect engine valve quality were identified. These parameters are essential for ensuring optimal performance, durability, and efficiency of the engine valve system.

The inlet valve slips through the valve guide. Determining its diameter at several points, bottom, middle, and top, guarantees appropriate fit, best seal, and efficient heat dissipation. Any variation from the prescribed specifications may compromise engine performance, wear resistance, and sealing capacity. Whereas the head diameter controls the flow area of the air-fuel mixture (intake valve) and exhaust gases, which is vital for engine power output and combustion efficiency. The distance between the valve head and the valve seat in the cylinder head is expressed by the seat height. A correct sealing effect, compression ratio optimization, and combustion efficiency depend on maintaining ideal seat height. Furthermore, the undercut, a reduced-diameter area below the valve head, improves gas flow dynamics, reduces turbulence, and helps to avoid carbon accumulation.

Exploratory analysis is conducted to gain a deeper understanding of the variations within these key features. The analysis encompassed establishing distribution plots to examine normality and outliers, assessing process capability indices (Cp and Cpk) to evaluate the manufacturing processes’ consistency with design specifications, and constructing univariate X-bar/S and multivariate Hotelling T² quality control charts for monitoring variation trends. Figure 3 shows histograms, and statistical summaries indicate that each feature follows a normal distribution without notable anomalies. However, a detailed process capability assessment reveals both strengths and areas requiring improvement.

Fig. 3.

Histogram plot for each key feature.

• High Process Capability: The lower and upper inlet valve stem diameters demonstrate strong manufacturing consistency, with Cp values exceeding 1.33, indicating a well-controlled process. However, a minor deviation in Cpk for the upper stem diameter suggests a shift from the optimal process centering.

• Moderate Process Capability: The middle stem diameter exhibits moderate performance (Cp > 1.33), but the lower Cpk value highlights a need for further process optimization.

• Low Process Capability and Process Instability: The head diameter, seat height, and undercut diameter show significant variability and off-centeredness, as indicated by low Cp and Cpk values, underscoring the need for stricter quality control measures.

Figure 4 summarizes the Cp and Cpk values for each key feature, providing a comparative analysis of process stability and capability.

Fig. 4.

Process capability for each key process.

To further analyze process stability, variation pattern analysis using statistical process control charts such as X-bar/S charts and Hotelling T² charts are employed to monitor real-time variations.

• X-bar and S Charts: These charts illustrate that while the stem diameters exhibit general stability, an out-of-control signal in sample 21 (middle stem diameter) suggests an assignable cause, requiring root-cause analysis and corrective action.

• Head Diameter and Seat Height Instability: The head diameter chart reveals instances where values hit the lower control limit (sample 26), raising concerns about dimensional consistency. Additionally, seat height exhibits unusual variance patterns, indicating potential inconsistencies in the machining or assembly processes.

• Hotelling T² Analysis: This multivariate approach further supports the identification of anomalies and process drifts, strengthening the case for targeted process refinements.

Figures 5 and 6 visually represent the results of this statistical monitoring, highlighting key areas where interventions are necessary.

Fig. 5.

X-bar/S control chart for each key process.

Fig. 6.

Hotelling T² control chart.

The analysis successfully identifies critical dimensions affecting engine valve quality, ensuring a data-driven approach to quality control. While some processes exhibit high consistency, others (e.g., head diameter, seat height, and undercut) require corrective measures to minimize variability. Statistical control charts provide early warning signals for deviations, enabling proactive quality assurance. Further exploration using predictive analytics and AI-based optimization techniques in subsequent phases will enhance precision and defect reduction.

By leveraging the insights gained in the Discover phase, the next stages of the PMQ Quality 4.0 framework will focus on refining process parameters, optimizing quality control mechanisms, and integrating advanced digital manufacturing solutions for continuous quality improvement.

Learn

The preceding analysis highlights the limitations of traditional control charts in effectively identifying defective items within complex manufacturing processes. Escobar¹⁵a) further elaborates on this issue, emphasizing that conventional control charts are inadequate when dealing with hyperdimensional data spaces. Given these constraints, the PMQ Quality 4.0 framework advocates leveraging ML and DL algorithms to detect out-of-control conditions and defective items more effectively.

To establish a robust baseline, multiple classification models, Support Vector Machine (SVM), Logistic Regression (LR), and Decision Tree (DT), were deployed. These models were selected due to their proven ability to handle high-dimensional tabular datasets, efficiently capture complex, non-linear relationships, and provide interpretable results through feature importance analysis³⁰. Furthermore, these models demonstrate strong generalization capability and robustness against overfitting, making them suitable for real-time quality monitoring tasks in manufacturing environments³⁴.

The comparative evaluation of baseline models revealed that DT outperformed the others, achieving an F₁ score of 0.93 and an AUC of 0.96. Meanwhile, SVM and LR models demonstrated comparable performance, each achieving an identical F₁ score of 0.73. The AUC scores for SVM and LR were 0.8 and 0.78, respectively, indicating similar discriminatory power between defective and non-defective items. These initial findings establish a strong foundation for further optimization in the next phase.

Predict

This phase aims to enhance quality classification performance by employing advanced ensemble learning models such as Random Forest (RF), Gradient Boosting Machine (GBM), and Artificial Neural Network (ANN). The objective is to systematically assess their performance against the baseline models established in the Learn phase.

Table 2 presents the confusion matrices for all six models. The results indicate that GBM and RF emerged as the best-performing models, each achieving zero false positives and one false negative, along with high counts of true positives (26) and true negatives (73).

Table 2.

Confusion matrix for all six models.

Confusion matrix for SVM model
	Predicted Negative	Predicted Positive
Actual Negative	55	9
Actual Positive	10	26
Confusion matrix for LR model
	Predicted Negative	Predicted Positive
Actual Negative	55	9
Actual Positive	10	26
Confusion matrix for DT model
	Predicted Negative	Predicted Positive
Actual Negative	70	3
Actual Positive	1	26
Confusion matrix for RF model
	Predicted Negative	Predicted Positive
Actual Negative	73	0
Actual Positive	1	26
Confusion matrix for the GBM model
	Predicted Negative	Predicted Positive
Actual Negative	73	0
Actual Positive	1	26
Confusion matrix for ANN model
	Predicted Negative	Predicted Positive
Actual Negative	69	4
Actual Positive	8	19

Conversely, SVM and LR exhibited weak performance, each generating 9 false positives and 10 false negatives. The ANN model, while an improvement over SVM and LR, still exhibited 4 false positives and 8 false negatives.

In addition to confusion matrices, the F1 score, and AUC were calculated for all models to provide a more comprehensive evaluation. The performance rankings based on these metrics are as follows: (i) Best performers: GBM and RF (F₁ = 0.98, AUC = 0.99), (ii) Third-best: DT (F₁ = 0.93, AUC = 0.96), (iii) Moderate performers: ANN (F₁ = 0.76, AUC = 0.83), and (iv) Lowest-performers: SVM and LR (F₁ = 0.73, AUC = 0.80 and F₁ = 0.73, AUC = 0.78, respectively).

These results underscore that tree-based algorithms (GBM, RF, and DT) are the most effective for classifying engine valves. Their superiority is attributed to their ability to: (i) Capture non-linear relationships within high-dimensional manufacturing data, (ii) Effectively partition the feature space, improving classification accuracy, and (iii) Handle structured tabular data efficiently, making them more suitable than deep learning-based models in this context³⁵. In addition, the findings are consistent with findings that tree-based methods often outperform neural networks in smaller, structured datasets^6,22. Figures 7 and 8 provide a comparative visualization of the model’s performance, illustrating the ROC curve and a side-by-side evaluation of F₁ scores and AUC values.

Fig. 7.

ROC curve.

Fig. 8.

Performance comparison of employed models using F₁ score and AUC.

These insights validate the effectiveness of advanced ML-based classification techniques in Quality 4.0-driven manufacturing environments, offering a robust alternative to traditional quality control methods and enabling a more data-driven approach to defect detection and process optimization.

Validate

This phase is integrated into the typical PMQ framework to bridge the gap between model predictions and process redesign. It aims to increase trust and understanding among quality professionals in ML-based predictions before applying any engineering interventions. Therefore, this phase incorporates various quantitative methods such as feature importance analysis to enhance interpretability, and qualitative approaches like asking experts to ensure that the predictions are contextually meaningful to quality professionals.

Feature importance analysis is performed with the aim of understanding which features contribute the most to the quality of the engine valve. Root node analysis is implemented to examine the initial decision points (root nodes) across multiple decision trees within the ensemble models to determine which variables corresponding to specific manufacturing processes are most frequently selected as primary discriminators in quality classification. The results identified seat height, undercut diameter, and head diameter as the three most critical contributors to defect classification. Notably, seat height surfaced as the primary split in 340 RF and 403 GBM decision trees (see Table 3), underscoring its significant impact on quality performance. This suggests that even minor deviations in this metric can lead to a disproportionate number of quality failures.

Table 3.

Top 3 RF and GBM features and their occurrence as the root node.

Top 3 RF features and their occurrence as the root node
Feature	Occurrence as root node
Seat height	340
Undercut diameter	243
Head diameter	207
Top 3 GBM features and their occurrence as the root node
Feature	Occurrence as root node
Seat height	403
Undercut diameter	326
Head diameter	140

To further validate these findings, permutation feature importance analysis (Figs. 9 and 10) was conducted, which confirmed that seat height had the most significant influence, followed by undercut diameter and head diameter. This method effectively captured nonlinear relationships between features and classification outcomes while also accounting for feature interactions.

Fig. 9.

Feature importance bar chart for RF model.

Fig. 10.

Feature importance Bar chart for GBM model.

The agreement between both methods reinforces the reliability and validity of the model predictions. The explainability provided by this analysis reduces skepticism among quality professionals and increases their trust in the use of ML predictions in quality-related decision-making. Furthermore, these findings are cross-checked against the engineers’ understanding of manufacturing processes. For example, seat height is widely known to affect valve sealing and combustion performance; therefore, its emergence as the most prominent feature contributing to quality classification validates the model and process knowledge. This alignment with engineering concepts builds credibility and establishes a feedback loop where domain knowledge supports data-driven approaches. Lastly, validating the model predictions helps engineers avoid unnecessary and expensive process adjustments and ensures that redesign efforts are allocated properly.

Redesign

Building on the insights from the Learn, Predict, and Validate phases, the Redesign phase aims to identify and implement engineering improvements targeting the processes and features most responsible for quality deviations. From an operational standpoint, focusing on fewer but more influential sub-processes facilitated efficient resource allocation, optimized machine calibration, and enhanced worker training. Based on these findings, several strategic engineering recommendations were proposed and subsequently tested in early trials, leading to improved valve-to-seat fit quality, reduced scrap, and more consistent dimensional adherence:

• Seat Height: Even small inconsistencies can compromise the valve’s sealing function, resulting in leakage, reduced compression, and increased emissions.

Proposed solutions include precision machining, real-time measurement using non-contact laser micrometers, and high-precision CNC equipment to tighten tolerances.

• Undercut Diameter: Variations in the undercut profile can disrupt airflow or increase the risk of carbon buildup.

Proposed solutions included the integration of advanced tooling systems with force sensors to ensure uniformity, which is essential for maintaining consistent flow dynamics.

• Head Diameter: Precise dimensional control is crucial to ensure stable airflow and prevent misalignment during engine operation.

The proposed solution suggested the adoption of diamond-tipped cutting tools and adaptive control systems to compensate for tool wear and maintain machining precision.

Relearn

The final phase of the PMQ Quality 4.0 cycle, the Relearn stage, focuses on continuous improvement through model refinement and adaptive learning. Following the implementation of redesign interventions, updated data, particularly on seat height, undercut diameter, and head diameter, will be systematically collected to assess the impact of the proposed enhancements. Advanced IoT-enabled sensors will be included in the production line to offer a continuous stream of real-time process information. This will allow operators to act quickly to remedy minor deviations before they cause severe quality problems.

Over time, these updated observations will be reintroduced into the ML models, utilizing either incremental learning or periodic retraining techniques (Escobar et al., 2021b). This allows the system to constantly adjust to changing factor conditions, raw material changes, and production needs, therefore maintaining the relevance and efficacy of quality control policies. Furthermore, additional features can be incorporated as the manufacturing ecosystem matures, such as material batch properties (accounting for natural variability in raw materials) and (ii) ambient production floor conditions (including temperature, humidity, and other environmental factors impacting machines and operators). This would significantly improve predictive accuracy and locate previously unnoticed relationships.

Ultimately, this cyclical Relearn approach exemplifies the core philosophy of the PMQ framework, leveraging data-driven, iterative refinements that seamlessly integrate engineering expertise with advanced analytics. This fosters an adaptive, proactive quality culture that remains resilient and effective despite shifting production requirements and technological advancements.

Discussions and implications

This study highlights how an ML-driven implementation of the modified PMQ framework effectively addresses the limitations of traditional quality management approaches. By integrating Industry 4.0 technologies such as IoT-enabled sensor data, real-time analytics, and advanced ML models, this research bridges the persistent gap between theoretical quality frameworks and their practical execution⁹. Specifically, the application of the modified PMQ in engine valve production illustrates the shortcomings of conventional SPC methods, where both univariate (X-bar and S) and multivariate (Hotelling T²) control charts often struggle to detect subtle yet critical variances in high-dimensional, precision-driven manufacturing environments. In contrast, ML-based models effectively capture complex, multifactorial interactions, enabling early and precise defect detection. Such granularity of insights is particularly critical in the automotive sector, where microscopic deviations in key dimensional features such as seat height, undercut diameter, and head diameter can significantly impact engine performance, emissions, and reliability²⁶.

A key contribution of this study is the demonstrated capability of tree-based ensemble models, particularly GBM and RF, to uncover hidden defect patterns that remain undetected by traditional techniques. Beyond achieving superior predictive accuracy, these models generate interpretable feature-importance metrics, offering targeted insights for process improvement. This interpretability is often undervalued in discussions on AI-driven quality management, yet it is essential for quality managers and shop-floor operators who must make rapid, high-stakes decisions based on ML recommendations (Aslam, 2024). This aligns with contemporary discussions on Quality 4.0, emphasizing that data-driven insights should complement, not replace, human expertise, fostering a symbiotic relationship between analytics and domain knowledge⁴.

Furthermore, the empirical results underscore that traditional SPC tools need not be entirely discarded. Instead, they can serve as an initial diagnostic layer to detect broad process instabilities, while ML models refine these insights to pinpoint nuanced sources of variation. This integrative approach directly addresses the growing need for hybrid quality management solutions that effectively blend Industry 4.0 technologies with established quality principles^24,40. Additionally, the iterative structure of PMQ, specifically its Relearn stage, ensures continuous improvement, even as production conditions evolve. The ability to retrain models periodically and integrate sensor-driven feedback loops mitigates the limitations of static control limits, which often become obsolete in dynamic manufacturing environments prone to frequent changeovers and material variability⁴³.

Beyond technical integration, the study highlights critical infrastructural and cultural challenges associated with adopting a Quality 4.0 framework. On the infrastructural front, implementing IoT sensors, data pipelines, and ML-driven quality control demands substantial investments, robust data governance policies, and specialized technical expertise⁵⁰. On the cultural side, transitioning from conventional SPC methods to AI-powered quality monitoring can generate resistance among employees, who may perceive algorithms as opaque or question the reliability of automated alerts. Consequently, successful implementation necessitates targeted training programs, clear model explanations, and incremental pilot deployments that build confidence by demonstrating tangible process improvements before full-scale adoption¹⁹. This reinforces a critical, yet often underexplored aspect of Quality 4.0; the interdependence between technological advancements and human adaptability in sustaining long-term, data-driven quality enhancements⁹.

From a broader perspective, the findings of this study extend beyond the automotive sector. Industries such as aerospace, healthcare, and electronics face similarly complex quality challenges, requiring strict adherence to precision standards and proactive defect mitigation strategies^1,5. While specific quality variables differ across domains (e.g. aerodynamic contour features in aerospace versus dimensional tolerances in engine components), the core principle of data-rich, real-time monitoring coupled with an iterative, learning-based improvement cycle remains universally relevant. The structured approach of the PMQ framework offers a scalable and adaptable model for organizations transitioning from reactive, manual quality control to proactive, ML-driven systems. Furthermore, the cyclical nature of PMQ allows continuous refinement of predictive models, ensuring that quality improvement efforts evolve alongside shifting production demands and emerging technological advancements.

Theoretical implications

This study makes several theoretical contributions to the evolving body of knowledge on Quality 4.0. It reinforces the empirical foundation of the PMQ framework, demonstrating its practical applicability in high-precision manufacturing, an area where much of the existing literature remains conceptual^9,14. This research extends the PMQ framework by adding an extra phase called Validate. This phase is specially designed to address the common issue of lack of trust among quality professionals in predictions based on ML²³. Moreover, the validate phase offers a structured point of human oversight, allowing engineers to assess and approve model predictions before they influence operational decisions. This way, the modified PMQ framework ensures that ML methods complement human judgment, a critical aspect of Quality 4.0⁴. By showing that traditional univariate and multivariate control charts have limited effectiveness in managing high-dimensional quality data, this research substantiates the role of ML models, particularly tree-based algorithms (GBM, RF, and Decision Trees), in capturing intricate interactions among process variables.

Previous studies on Quality 4.0 have emphasized the importance of real-time analytics and big data, but few have provided empirical validation within specific industrial contexts^4,43. This research addresses that gap by systematically applying the structured stages of the modified PMQ (Identify, Acsensorize, Discover, Learn, Predict, Validate, Redesign, Relearn), demonstrating its effectiveness in the automotive sector. The findings also respond to calls for industry-specific case studies that illustrate how data-driven quality management can accommodate stringent precision requirements and manage interdependent process parameters.

A particularly novel insight from this study is the emphasis on interpretability in ML-driven quality management. By leveraging feature importance metrics from GBM and RF models, the study achieves superior predictive performance and provides actionable intelligence, enabling process engineers to implement targeted interventions on critical defect-prone variables (e.g., seat height, undercut, and head diameter).). This insight aligns with the growing focus on actionable AI, where ML tools are not merely used for automation but serve as enablers of informed and data-driven decision-making in complex manufacturing environments⁴.

Additionally, by embedding iterative ML retraining cycles within the modified PMQ framework, this research complements and extends traditional Six Sigma (SS) and Lean Six Sigma (LSS) methodologies, which often rely on static statistical models that may not adapt effectively to rapidly evolving production conditions¹⁴. The findings reinforce the argument that sensor-driven feedback loops and continuous ML retraining are crucial for sustaining long-term quality improvements in Industry 4.0 settings².

Collectively, these theoretical contributions provide a strong foundation for future research. Researchers are encouraged to replicate and scale the modified PMQ framework across diverse high-precision industries (e.g., aerospace, medical device manufacturing, semiconductor production) to further validate its effectiveness. Additionally, future studies should explore the human factors influencing the adoption of ML-driven quality systems such as operator training, digital literacy, and change management strategies. Another promising avenue for research is the economic feasibility of transitioning from traditional SPC to ML-based quality control, assessing cost-benefit trade-offs and return on investment for organizations considering Quality 4.0 implementation.

Practical implications

The implementation of the PMQ framework underscores several key practical implications, particularly its ability to provide precise and actionable engineering insights. One of the most significant findings is that seat height emerged as the most critical feature influencing defect classification. Even minor deviations in this parameter can significantly impact valve performance, emphasizing the necessity for precision machining, non-contact metrology, and real-time dimensional tracking. Maintaining stringent tolerances in these areas is crucial for reducing emissions and extending engine longevity. Furthermore, underlined by the significance of undercut and head diameter in engine quality, it becomes imperative to have specialized inserts, force sensors, and adaptive machining controls (e.g., diamond-tipped cutters, real-time speed adjustments) to guarantee consistent and accurate manufacturing.

By including ML models in production quality control systems, defect rates may be greatly lowered, material waste could be cut, and rework expenses could be minimized. Early defect identification allows producers to respond before major manufacturing errors happen, therefore saving significant costs, especially in high-volume production settings. The shift from traditional manual and SPC-driven inspections to data-driven quality monitoring necessitates a structural transformation in production lines. This may involve the integration of IoT sensors at key process points to enable centralized dashboards with real-time data collection and processing. Such developments would give line operators instant feedback, therefore enabling faster decision-making and corrective action.

Furthermore, Quality 4.0 adoption presents a valuable opportunity for organizations to upskill their workforce, ensuring that engineers and operators can effectively interpret ML-driven insights and manage sensor-based data. This change emphasizes the growing necessity of cross-functional expertise through cooperation among quality engineers, data scientists, and IT experts. Companies without internal analytical capacity should think about creating strategic alliances with outside AI solution providers or funding internal training initiatives to fully utilize advanced ML techniques.

This study primarily focuses on the automotive sector. However, the modified PMQ approach is equally applicable to other industries where strict tolerances and complex assemblies are important. For example, in the aerospace manufacturing industry real-time defect detection is crucial for flight-critical components, where even the slightest deviations can pose significant risks. Similar strict quality criteria are required in the healthcare industry to guarantee patient safety, especially in the manufacturing of medical implants and devices⁵. These results can be used by policymakers and industry stakeholders to establish Quality 4.0 standards, therefore promoting broad acceptance and best practices in data governance, sensor deployment, and defect prediction-driven AI tools.

Moreover, by utilizing Quality 4.0’s digital solutions, organizations become more equipped to handle the increasingly stringent global environmental regulations. For instance, improved defect detection and enhanced process control contribute directly to sustainability goals by reducing waste, optimizing resource usage, and lowering energy consumption associated with rework (Qudus et al., 2025). Such promotion of sustainable manufacturing practices not only enhances operational efficiency but also strengthens brand reputation and customer trust.

Limitations and future directions

First, the focus of this research was on the automotive industry and engine valve production. This may restrict the generalizability of the findings to other industries with different operational characteristics. Therefore, future research should assess the adaptability of the modified framework across different sectors. A comparative analysis across multiple industries would provide deeper insights into the scalability and effectiveness of the modified PMQ framework. Second, this study excluded external factors such as environmental conditions (e.g., ambient temperature, humidity), material batch properties, and operator-induced variability. This highlights another potential area for future research to provide a more holistic understanding of quality variations and defect causality. Lastly, a multiclass quality classification approach instead of the binary classification method of the modified PMQ framework can be explored. This can equip quality professionals with additional flexibility to predict the quality of products, resulting in more accurate predictions.

Advanced ML and IoT technologies, which demand significant infrastructure, labor up-skilling, and data governance, formulate the basis of the modified PMQ framework. Financial restrictions and privacy issues may prevent small and medium-sized enterprises (SMEs) from using such systems. Therefore, future research may explore affordable solutions, cloud-based analytics, and joint learning initiatives to encourage wider implementation of Quality 4.0 approaches.

Moreover, the modern manufacturing industry landscape is characterized by continuous change and dynamic nature. This underlines the need for integrating self-learning models that continuously adapt to evolving process conditions could significantly enhance defect prevention and process optimization. Therefore, future research should focus on embedding reinforcement learning or edge computing frameworks to enable real-time decision-making with minimal delay.

Lastly, investigating the regulatory and policy implications of AI-driven quality management is crucial. This highlights a vital area for future research. For example, researchers can develop standardized industry-wide guidelines for AI adoption in quality management to ensure a consistent and ethical implementation across various sectors. Furthermore, exploring collaborations between academia, industry, and policymakers could further facilitate the establishment of best practices.

Conclusions

This study demonstrates the transformative potential of the Quality 4.0 PMQ framework, leveraging ML techniques, particularly tree-based models such as GBM and RF, to enhance defect prediction and root cause analysis in a real-world, data-intensive automotive manufacturing environment. By integrating systematic sensor data collection with the iterative modified PMQ stages (Identify, Acsensorize, Discover, Learn, Predict, Validate, Redesign, Relearn), the framework not only achieved high classification accuracy (F1 ≥ 0.98 for the best-performing models) but also provided actionable insights into critical quality determinants in the automotive industry, such as seat height, undercut diameter, and head diameter. This shows how well AI-driven approaches can handle complex industrial interactions that lead to reduced scrap rates and enhanced process efficiency. In addition, the modified PMQ framework aids quality professionals in overcoming the lack of trust in ML-based predictions and integrating human expertise with advanced analytics. This integration represents a vital pillar of Quality 4.0 implementation.

The findings of this study reveal important prospects for the automotive industry as well as other industries, especially in sectors like aerospace, healthcare, and electronics, where strict precision and high-dimensional process data call for advanced quality control methods. In addition, crucial features of Quality 4.0, such as real-time monitoring and predictive maintenance, highlight the vital role of the modified PMQ framework in driving proactive, rather than reactive, quality control strategies. Additionally, the effective use of feature importance analysis during the validate phase supports the growing trend of human-focused AI solutions by ensuring that automated decision-making systems are transparent and understandable.

From a managerial point of view, the results highlight the importance of cross-functional collaboration and workforce training to support the smooth integration of sensor-based monitoring, ML analytics, and real-time quality control. Furthermore, adopting Quality 4.0 can support the transformation of the organization to more sustainable practices, as evident in the resultant reduction in waste and improved resource allocation. However, achieving the full potential of Quality 4.0 is also tied to investing in data literacy, algorithm interpretability, and IoT-enabled infrastructure.

Overall, this study bridges the gap between theoretical quality management concepts and real-world implementation, reaffirming the need for adaptive, iterative, and data-driven quality control systems. The findings express strong applicability; however, future research can build on these insights by extending the modified PMQ framework to other industrial sectors with different operational conditions and improving real-time adaptability by integrating reinforcement learning and edge computing for instantaneous defect prediction. By addressing these areas, Quality 4.0 can continue to evolve as a key enabler of operational excellence, resilience, and sustainable quality management in modern smart manufacturing settings.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Author contributions

Author 1 (Fathy Alkhatib) wrote the manuscript, Author 2 (Mohamed Allam) performed analysis, Author 3 (Vikas Swarnakar), collected the data from the industry, reviewed, and revised the whole draft and edited the article, Author 4 (Juman Alsadi) prepared all Figures, Author 5 (Maher Maalouf) validated the data analysis, and reviewed and revised the draft.

Data availability

Data is, however, available from the corresponding author upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.