Fault Detection and Diagnosis in Industrial Processes with Variational Autoencoder: A Comprehensive Study

Area of Science:

• Artificial Intelligence
• Machine Learning
• Process Engineering

Background:

• Variational Autoencoders (VAEs) are powerful deep generative models increasingly used for industrial process monitoring.
• The fault diagnosis capabilities of VAEs, crucial for industrial applications, require further investigation.
• Existing studies have not comprehensively evaluated VAE variants for their process monitoring and fault diagnosis performance.

Purpose of the Study:

• To comprehensively study the process modeling and monitoring capabilities of several Variational Autoencoder (VAE) variants.
• To investigate and define effective fault detection schemes across latent, residual, and combined domains.
• To develop and evaluate a deep reconstruction-based contribution diagram for fault diagnosis under fault propagation.

Main Methods:

• Defined three distinct fault detection schemes: latent domain, residual domain, and combined domain.
• Introduced a deep contribution plot and a deep reconstruction-based contribution diagram for fault diagnosis.
• Comparatively evaluated four deep VAE models (static VAE, dynamic VAE, LSTM-VAE, GRU-VAE) on the Tennessee Eastman process benchmark.

Main Results:

• Recurrent VAE models (LSTM-VAE and GRU-VAE) demonstrated strong performance in process monitoring.
• The deep reconstruction-based diagnosis mechanism proved effective for fault diagnosis in deep domains.
• Comparative analysis highlighted the advantages of recurrent VAEs for complex industrial monitoring tasks.

Conclusions:

• Recurrent VAEs are highly recommended for industrial process monitoring due to their enhanced modeling and diagnostic capabilities.
• The proposed deep reconstruction-based diagnosis mechanism significantly improves fault diagnosis accuracy.
• This research provides valuable insights into selecting and applying VAE variants for effective industrial process monitoring and fault diagnosis.