Nonlinear Chemical Process Fault Diagnosis Using Ensemble Deep Support Vector Data Description.

Area of Science:

• Chemical Engineering
• Machine Learning
• Process Monitoring

Background:

• Traditional Support Vector Data Description (SVDD) has limitations in complex nonlinear chemical process monitoring due to its shallow learning structure.
• Existing methods struggle with satisfactory fault detection performance in intricate industrial scenarios.

Purpose of the Study:

• To develop an advanced anomaly detection technique for more effective chemical process fault monitoring.
• To improve fault detection accuracy and enable robust fault cause identification.

Main Methods:

• Proposed an Ensemble Deep Support Vector Data Description (EDeSVDD) model incorporating deep feature extraction.
• Utilized an ensemble learning strategy based on Bayesian inference to generate diverse DeSVDD sub-models.
• Developed a fault isolation scheme using distance correlation coefficients to identify root cause variables.

Main Results:

• The EDeSVDD model demonstrated superior fault detection performance compared to traditional SVDD and the basic Deep SVDD (DeSVDD) model.
• The proposed method effectively identified the cause variables responsible for detected process faults.
• Validated performance on the Tennessee Eastman process, confirming enhanced monitoring capabilities.

Conclusions:

• EDeSVDD offers a significant advancement in nonlinear chemical process monitoring and fault analysis.
• The integration of deep learning and ensemble methods provides a more robust and accurate anomaly detection solution.
• The fault isolation capability of EDeSVDD is crucial for practical industrial applications.