Industrial Process Monitoring Based on Dynamic Overcomplete Broad Learning Network

Area of Science:

• Industrial Process Monitoring
• Machine Learning
• Data Analysis

Background:

• Industrial processes often exhibit complex characteristics like nonlinearity, non-Gaussianity, and time correlation.
• Existing Overcomplete Broad Learning System (OBLS) models handle nonlinearity and non-Gaussianity but neglect time correlation, limiting monitoring accuracy.

Purpose of the Study:

• To develop an effective dynamic overcomplete broad learning system (DOBLS) that accounts for time correlation in industrial process data.
• To enhance fault monitoring accuracy by addressing the combined challenges of nonlinearity, non-Gaussianity, and time correlation.

Main Methods:

• Proposed a dynamic overcomplete broad learning system (DOBLS) using matrix extension to incorporate 'time lag' into raw batch process data.
• Applied the Overcomplete Broad Learning System (OBLS) to analyze the extended dynamic input data for monitoring.
• Established a monitoring model capable of handling nonlinearity, non-Gaussianity, and time correlation simultaneously.

Main Results:

• The DOBLS model demonstrated superior performance on a penicillin fermentation simulation platform.
• The model achieved more comprehensive feature extraction and efficient self-updating of process data.
• Demonstrated an average monitoring accuracy improvement of 3.69% over BLS and 1.26% over OBLS across 26 fault types.

Conclusions:

• The proposed DOBLS model effectively tackles the complexities of industrial process data, including time correlation.
• DOBLS offers improved fault monitoring accuracy and efficiency compared to existing methods like BLS and OBLS.
• The model's ability to handle dynamic data and self-update efficiently makes it a valuable tool for industrial process monitoring.