Variational Attention-Based Interpretable Transformer Network for Rotary Machine Fault Diagnosis

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning models, particularly those using vibration signals, are promising for rotary machine fault diagnosis (RMFD).
• Existing methods often struggle to capture complex association relationships within vibration signals.
• Transformer networks excel at identifying signal associations but lack interpretability regarding fault causation.

Purpose of the Study:

• To propose an interpretable deep learning model, the Variational Attention-based Transformer Network (VATN), for RMFD.
• To enhance the mining of association relationships within vibration signals for improved fault diagnosis.
• To establish a clear causal association between signal patterns and specific fault types.

Main Methods:

• Developed VATN by modifying transformer encoders to effectively mine signal associations.
• Incorporated a sparse constraint on attention weights, utilizing variational inference and Laplace approximation for reparameterization.
• Applied Dirichlet distributions to attention weights to embed prior fault-type knowledge.

Main Results:

• VATN demonstrated superior effectiveness compared to other methods on bevel gear and bearing datasets.
• The model successfully mined association relationships within vibration signals.
• Attention weight heatmaps provided interpretable visualizations of causal associations between signal patterns and fault types.

Conclusions:

• The proposed VATN model offers an effective and interpretable solution for rotary machine fault diagnosis.
• VATN successfully addresses limitations of existing deep learning and transformer-based approaches in RMFD.
• The interpretability feature allows for a deeper understanding of fault mechanisms based on vibration signal analysis.