Deep transfer learning strategy for efficient domain generalisation in machine fault diagnosis.

Area of Science:

• Machine Learning
• Signal Processing
• Mechanical Engineering

Background:

• Automated fault diagnosis relies on vibration sensor data for machine health monitoring.
• Data-driven models require extensive labeled data, and their performance degrades with domain shifts.

Purpose of the Study:

• To develop a novel deep transfer learning strategy for robust machine fault classification.
• To enhance domain generalization in automated fault diagnosis systems.

Main Methods:

• A deep transfer learning approach fine-tuning lower convolutional layers and transferring parameters from deeper dense layers.
• Utilizing time-frequency representations (scalograms) of vibration signals as input.
• Evaluating performance on two distinct target domain datasets with varying sensor precision and data availability.

Main Results:

• The proposed strategy achieves near-perfect accuracy in fault classification.
• Effective performance is demonstrated even with low-precision sensors and limited unlabelled run-to-failure data.
• Sensitivity analysis of individual layer fine-tuning was performed.

Conclusions:

• The novel transfer learning strategy significantly improves domain generalization for machine fault diagnosis.
• This approach offers a reliable solution for real-world applications with limited labeled data and varying sensor conditions.