Nonlinear quality-related fault detection using combined deep variational information bottleneck and variational...

Area of Science:

• Industrial process monitoring
• Machine learning for fault detection
• Deep learning applications

Background:

• Deep learning is widely used for fault detection in nonlinear industrial processes.
• Existing methods often fail to account for quality-related faults.
• A need exists for methods that can distinguish between quality-related and unrelated fault information.

Purpose of the Study:

• To propose a novel deep VIB-VAE algorithm for industrial fault detection.
• To extract latent variables representing separated quality-related and unrelated information.
• To enhance the monitoring and distinction of quality-related and quality-unrelated faults.

Main Methods:

• Combines Variational Autoencoder (VAE) and deep Variational Information Bottleneck (VIB).
• Deep VIB maximizes mutual information between latent variables and process quality while minimizing it with observation.
• VAE learns quality-unrelated information using quality-related latent variables as auxiliary input.

Main Results:

• Two monitoring statistics are designed based on the two-part latent variables.
• Reconstruction error from VAE is utilized to enhance fault detection performance.
• The algorithm's effectiveness is validated through numerical and real-world hot strip mill process cases.

Conclusions:

• The proposed deep VIB-VAE algorithm effectively monitors and distinguishes quality-related and quality-unrelated faults.
• The method demonstrates superior performance in fault detection for industrial processes.
• This approach offers a significant advancement in handling quality-aware fault detection.