Causal Disentanglement-Based Hidden Markov Model for Cross-Domain Bearing Fault Diagnosis

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Predictive Maintenance

Background:

• Accurate fault diagnosis in industrial machinery is crucial for predictive maintenance.
• Deep learning shows promise but struggles with generalization due to limited industrial failure data.
• Existing methods often overlook signal theory principles in favor of pure data-driven approaches.

Purpose of the Study:

• To develop a robust and generalizable fault diagnosis method for bearings under complex industrial conditions.
• To address the challenge of limited industrial failure data by leveraging causal inference and transfer learning.
• To improve the accuracy and reliability of predictive maintenance systems.

Main Methods:

• Proposed the causal disentanglement-based hidden Markov model (CDHM) to capture underlying causality in vibration signals.
• Constructed a time-series structural causal model (SCM) to represent signal interconnections.
• Designed a hidden Markovian variational autoencoder (VAE) for disentangling fault-relevant and fault-irrelevant signal components.
• Leveraged cross-domain consistency and domain sensitivity for mutually reinforcing optimization of causal disentanglement and transfer learning.

Main Results:

• The CDHM effectively disentangles essential fault patterns from system and environmental interference in bearing vibration signals.
• The model demonstrates robust generalization across diverse operating conditions, validated on CWRU, IMS, and PU datasets.
• Achieved more accurate and generalizable fault representation compared to traditional signal-processing and deep learning methods.

Conclusions:

• The CDHM offers a novel perspective on bearing vibration signal analysis by incorporating causal inference.
• The proposed method significantly enhances the robustness and generalizability of fault diagnosis in industrial settings.
• The CDHM shows strong potential for practical industrial applications in predictive maintenance.