Probabilistic Residual Modeling for Sensor-Based Process-Quality Fault Detection in Industrial Systems.

Area of Science:

• Industrial Process Monitoring
• Fault Detection and Diagnosis
• Statistical Process Control

Background:

• Traditional sensor-based process monitoring uses process and quality variables, often relying on fixed residual structures.
• Existing methods struggle with nonlinear industrial processes where residual variances vary and correlations exist between variables.
• This limitation necessitates advanced methods for accurate fault detection and impact assessment.

Purpose of the Study:

• To propose a novel probabilistic residual modeling method for robust process-quality fault detection.
• To address the limitations of fixed residual structures in nonlinear industrial processes.
• To improve the detection and discrimination of process-related versus quality-related faults.

Main Methods:

• Developed a probabilistic residual modeling approach building on deep variational canonical correlation analysis (DVCCA).
• Introduced conditional residual distributions parameterized by latent operating states inferred from sensor data.
• Utilized residual negative log-likelihoods as monitoring statistics for evaluating abnormality under current operating conditions.

Main Results:

• The proposed method demonstrated superior performance in detecting process-side disturbances and discriminating between process and quality faults compared to PLS, CCA, DCCA, and DVCCA.
• Achieved high true positive rates (e.g., 82.02% for Jx in TPFF air line blockage) with low false alarm rates (e.g., 1.52%).
• Successfully identified specific faults like air line blockage and sensor bias, showcasing effective process-quality fault discrimination.

Conclusions:

• The probabilistic residual modeling method offers a significant advancement in process-quality fault detection for nonlinear industrial systems.
• Conditional residual distributions effectively account for varying operating conditions, enhancing monitoring accuracy.
• The method provides a more reliable tool for maintaining process integrity and product quality.