Adaptive Multimode Process Monitoring Based on Mode-Matching and Similarity-Preserving Dictionary Learning.

Area of Science:

• Industrial Process Monitoring
• Machine Learning
• Data Science

Background:

• Industrial processes evolve, creating new operational modes.
• Existing SCADA systems struggle to adapt models to emerging modes, causing model mismatch.
• Updating models for new modes can lead to catastrophic forgetting of historical data.

Purpose of the Study:

• To develop a method that adapts to new industrial process modes.
• To prevent models from forgetting historical data when learning new modes.
• To improve fault detection and reduce false alarms in multimode industrial processes.

Main Methods:

• Proposed a jointly mode-matching and similarity-preserving dictionary learning (JMSDL) method.
• Learned data from new modes to adaptively match emerging modes.
• Introduced a similarity metric to preserve representation ability for historical data.

Main Results:

• The JMSDL method accurately matches new modes while retaining performance on historical modes.
• Demonstrated effectiveness through numerical simulation, CSTH process, and industrial roasting experiments.
• Significantly outperformed state-of-the-art methods in fault detection and false alarm rate.

Conclusions:

• The JMSDL method effectively addresses model mismatch and catastrophic forgetting in multimode industrial processes.
• The proposed method offers superior performance in fault detection and false alarm reduction.
• JMSDL provides a robust solution for monitoring evolving industrial environments.