A Novel Distributed Fault Detection Approach Based on the Variational Autoencoder Model.

Area of Science:

• Industrial Process Monitoring
• Machine Learning for Engineering
• Chemical Engineering Systems

Background:

• Traditional distributed fault detection models treat units independently, ignoring critical inter-unit relationships and leading to incomplete information.
• Missing data significantly degrades the performance of industrial fault detection systems.
• Variational Autoencoders (VAE) offer powerful nonlinear feature extraction capabilities essential for complex industrial data.

Purpose of the Study:

• To propose a distributed fault detection method (DVAE) that incorporates local and neighboring unit information.
• To develop a method capable of reconstructing missing data in industrial processes.
• To enhance the accuracy and robustness of fault detection in large-scale industrial systems.

Main Methods:

• System variables were divided into local and neighboring units based on system mechanisms.
• A distributed Variational Autoencoder (DVAE) model was established for each local unit to map multivariable data to a latent variable space.
• Latent variables capturing local and neighboring unit information were utilized for fault detection using Euclidean distance.

Main Results:

• The proposed DVAE method effectively describes local and neighboring unit information, capturing complex inter-unit relationships.
• The method demonstrated strong performance in reconstructing missing data, a common challenge in industrial settings.
• Validation on the Tennessee Eastman process confirmed the DVAE method's efficacy in fault detection and data imputation.

Conclusions:

• The DVAE method provides a more comprehensive approach to distributed fault detection by considering unit interdependencies.
• The integration of VAEs enables robust fault detection even with missing sensor data.
• This approach offers significant potential for improving the safety and efficiency of industrial operations.