Adaptive Cointegration Analysis and Modified RPCA With Continual Learning Ability for Monitoring Multimode...

Area of Science:

• Process monitoring and control
• Statistical process control
• Machine learning for industrial applications

Background:

• Current multimode process monitoring often requires data from all modes and prior knowledge.
• Existing recursive methods may forget learned knowledge and struggle with drastic variations.
• Nonstationary processes with frequently varying and emerging modes pose significant monitoring challenges.

Purpose of the Study:

• To develop an advanced monitoring system for nonstationary processes with constantly emerging modes.
• To improve fault detection accuracy and adapt to gradual changes in process dynamics.
• To overcome limitations of existing recursive and traditional multimode monitoring techniques.

Main Methods:

• Proposes adaptive cointegration analysis (CA) to update models with normal samples and adapt to gradual changes.
• Develops a modified recursive principal component analysis (RPCA) incorporating elastic weight consolidation for continual learning.
• Introduces novel statistics and uses recursive kernel density estimation for threshold calculation.

Main Results:

• The proposed adaptive CA distinguishes real faults from normal variations effectively.
• The modified RPCA with continual learning consolidates knowledge, improving model accuracy and reducing performance degradation for similar modes.
• Demonstrates superior performance in accuracy, memory, and computational complexity compared to recursive CA and recursive slow feature analysis.

Conclusions:

• The developed adaptive CA and modified RPCA offer a robust solution for nonstationary process monitoring under evolving conditions.
• The method shows significant effectiveness on both numerical and practical industrial system case studies.
• Highlights the superiority of the proposed approach over state-of-the-art recursive algorithms for complex industrial monitoring tasks.