An improved mixture robust probabilistic linear discriminant analyzer for fault classification.

Area of Science:

• Machine Learning
• Statistical Modeling
• Industrial Process Monitoring

Background:

• Traditional probabilistic models like PPCA have limitations in handling complex industrial data.
• Probabilistic Linear Discriminant Analysis (PLDA) enhances classification by separating within-class and between-class information.
• Outliers and non-Gaussian distributions are common challenges in industrial fault detection.

Purpose of the Study:

• To introduce a novel fault classification method robust to outliers and non-Gaussian data.
• To enhance the classification capability in industrial processes.
• To develop an effective state inference method for fault diagnosis.

Main Methods:

• Development of a mixture robust probabilistic linear discriminant analysis (MRPLDA) model.
• Incorporation of Student's t-priors to handle noise and hidden variables.
• Application of a variational Bayesian expectation-maximization algorithm for parameter estimation.
• Proposal of a novel state inference method considering joint probabilities between test and training samples.

Main Results:

• The MRPLDA model demonstrates robustness against outliers and non-Gaussian variables.
• The proposed state inference method improves fault classification accuracy.
• Validation through a numerical example and the Tennessee Eastman (TE) process.

Conclusions:

• The MRPLDA method offers a powerful tool for fault classification in industrial settings.
• The developed state inference technique enhances diagnostic reliability.
• The approach is effective for complex industrial process monitoring and fault detection.