Incremental Pyraformer–Deep Canonical Correlation Analysis: A Novel Framework for Effective Fault Detection in Dynamic Nonlinear Processes

• Published in: Algorithms Vol. 18; no. 3; p. 130
• Main Authors: Ding, Yucheng, Zhang, Yingfeng, Huang, Jianfeng, Peng, Shitong
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 01.03.2025
• Subjects: Analysis; Continuously stirred tank reactors; Correlation analysis; data driven; Data transmission; Datasets; DCCA; Deep learning; dynamic nonlinear process; Effectiveness; False alarms; Fault detection; Fault location (Engineering); incremental learning; Injection molding; Learning; Machine learning; Maintenance; Methods; Neural networks; Nonlinear phenomena; Production management; Real time; Reliability (Engineering); Sparsity; Statistical analysis; Technology application; Variables; China; United States > US; Tennessee
• ISSN: 1999-4893; 1999-4893
• DOI: 10.3390/a18030130

Smart manufacturing systems aim to enhance the efficiency, adaptability, and reliability of industrial operations through advanced data-driven approaches. Achieving these objectives hinges on accurate fault detection and timely maintenance, especially in highly dynamic industrial environments. However, capturing nonlinear and temporal dependencies in dynamic nonlinear industrial processes poses significant challenges for traditional data-driven fault detection methods. To address these limitations, this study presents an Incremental Pyraformer–Deep Canonical Correlation Analysis (DCCA) framework that integrates the Pyramidal Attention Mechanism of the Pyraformer with the Broad Learning System for incremental learning in a DCCA basis. The Pyraformer model effectively captures multi-scale temporal features, while the BLS-based incremental learning mechanism adapts to evolving data without full retraining. The proposed framework enhances both spatial and temporal representation, enabling robust fault detection in high-dimensional and continuously changing industrial environments. Experimental validation on the Tennessee Eastman (TE) process, Continuous Stirred-Tank Reactor (CSTR) system, and injection molding process demonstrated superior detection performance. In the TE scenario, our framework achieved a 100% Fault Detection Rate with a 4.35% False Alarm Rate, surpassing DCCA variants. Similarly, in the CSTR case, the approach reached a perfect 100% Fault Detection Rate (FDR) and 3.48% False Alarm Rate (FAR), while in the injection molding process, it delivered a 97.02% FDR with 0% FAR. The findings underline the framework’s effectiveness in handling complex and dynamic data streams, thereby providing a powerful approach for real-time monitoring and proactive maintenance.