Data-Driven Integrated Analysis in Monitoring Manufacturing Processes Using Principal Component Analysis and Multivariate Control Charts (Case Study: Mobarakeh Steel Company, Isfahan)

Introduction and Objectives. In advanced manufacturing industries, especially in industries such as steel that are faced with complex and multi-stage processes, optimizing the monitoring of quality and performance of production lines is of high importance. Traditional univariate monitoring methods, due to their inability to analyze the relationships among different variables, are inefficient in the timely identification of anomalies. Therefore, the combination of multivariate statistical process control (MSPC) methods and Principal Component Analysis (PCA) as an innovative solution can be an efficient tool for reducing data complexity, increasing monitoring accuracy, and improving the final quality of products. The objective of this research is to design and validate a hybrid model based on PCA and multivariate control charts in order to optimize monitoring in the slab production process at Mobarakeh Steel Company.
Methods. The present study was conducted with a quantitative approach and based on the analysis of real data from the production lines of Mobarakeh Steel. First, control data related to the casting process were collected and outlier data were removed using boxplot charts. Then, by performing the KMO and Bartlett tests, the adequacy of the data and the correlation of variables for implementing PCA analysis were confirmed KMO = 0.78، p < 0.001 PCA analysis led to the extraction of three principal components that identified the most important influential variables in the process. These components were given as inputs to the multivariate control models MEWMA and MCUSUM. In specific cases, the Hotelling T² chart was used. The developed model, after implementation, was validated by industry experts.
Findings. The results of the analysis showed that by reducing the data dimensions from 9 variables to 3 key components (PC1, PC2, and PC3), more accurate process monitoring was made possible. In the casting stage, the use of the MEWMA chart led to faster identification of fluctuations and deviations, while the EWMA chart was used for univariate monitoring in order to compare performance. The comparison between these two models showed that MEWMA has higher sensitivity to small changes in the process. Also, the implementation of the hybrid model resulted in 95% of out-of-control points being identified using only four charts, whereas previously they were scattered across more than 45 charts. Regression analysis of the components also confirmed that the principal components well describe the changes of the key variables.
Conclusion. This research shows that the combination of PCA methods and multivariate quality control not only increased the accuracy in monitoring the production process but also significantly reduced the complexity of control models and monitoring costs. The application of this approach leads to the reduction of false alarms, increased speed of response to changes, and overall improvement of the quality system. Also, by appropriate adjustment of the lambda (λ) parameter in the MEWMA model, the overlap of detecting deviation points with univariate methods reaches more than 80%. Based on expert confirmation, the proposed model has high practical capability and is proposed as a practical solution for improving process monitoring in large industries such as steelmaking.