Multivariate fault isolation of batch processes via variable selection in partial least squares discriminant...

Area of Science:

• Chemical Engineering
• Process Control
• Statistical Process Monitoring

Background:

• Multivariate statistical monitoring of batch processes is crucial for identifying process abnormalities.
• Traditional fault isolation methods like contribution plots suffer from the smearing effect due to variable correlations.
• High autocorrelations in batch processes exacerbate the smearing effect, hindering accurate fault identification.

Purpose of the Study:

• To develop an improved fault isolation method for batch processes that overcomes the limitations of traditional techniques.
• To enhance the identification of faulty variables contributing to process abnormalities.
• To provide process engineers with more reliable information for root-cause diagnosis.

Main Methods:

• A novel variable selection-based fault isolation method is proposed.
• The method transforms fault isolation into a variable selection problem within partial least squares discriminant analysis (PLS-DA).
• A sparse partial least squares model is utilized to solve the variable selection problem.

Main Results:

• The proposed method effectively addresses the smearing effect inherent in correlated variables.
• It emphasizes the relative importance of individual process variables for fault diagnosis.
• This leads to more precise identification of variables contributing to process deviations.

Conclusions:

• The variable selection-based approach offers a significant advancement in multivariate fault isolation for batch processes.
• It provides a more accurate and interpretable alternative to traditional contribution plots.
• The method empowers process engineers with better tools for effective root-cause analysis and process improvement.