An Intelligent Multi-Local Model Bearing Fault Diagnosis Method Using Small Sample Fusion.

Area of Science:

• Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Accurate bearing fault diagnosis is crucial for industrial safety and preventing motor failures.
• Deep learning methods have advanced motor operation safety but often require substantial monitoring data.
• Harsh industrial conditions limit data collection for bearing sensors, especially for special motor bearings.

Purpose of the Study:

• To develop an effective bearing fault diagnosis method for scenarios with limited monitoring data using transfer learning.
• To address the challenge of small sample fusion in multi-local model bearing fault diagnosis.
• To improve the reliability and intelligence of industrial motor operations through enhanced fault diagnosis.

Main Methods:

• A parallel Bi-LSTM sub-network was constructed to extract features from vibration and current signals.
• Features were serially fused for classification, establishing a source domain fault diagnosis model.
• Maximum Mean Difference algorithm measured data distribution differences; transfer learning fine-tuned the model for the target domain.

Main Results:

• The proposed transfer learning method achieved higher fault diagnosis accuracy with small sample fusion compared to existing methods.
• The method significantly reduced the early training time of the fault diagnosis model.
• Generalization ability of the fault diagnosis model was substantially improved.

Conclusions:

• The developed transfer learning strategy effectively diagnoses bearing faults even with limited data.
• The approach enhances diagnostic accuracy (over 80%) and reduces training time (by 15.3%).
• This method offers a reliable and intelligent solution for industrial motor fault diagnosis, improving operational safety and efficiency.