An Intelligent Fault Diagnosis Method for Aviation Precision Bearings based on Transfer Learning

Abstract

To meet the requirements of efficient, accurate, and cross-operating-condition fault diagnosis for aviation bearings, this paper proposes an intelligent fault diagnosis method based on transfer learning. Addressing the data difference between the source domain and target domain, preprocessing methods such as resampling (unifying sampling frequency), filtering denoising, and signal segmentation (improving sample set stability) are adopted to unify data characteristics, laying a foundation for subsequent transfer learning. Class weights are introduced to solve the sample imbalance problem. After multiple model parameter optimizations and comparisons of the recognition effects of various classifiers, it is concluded that Random Forest achieves the best recognition performance. To tackle the decline in diagnostic performance caused by inter-domain distribution differences, a three-layer transfer learning framework featuring feature-model-sample collaboration is proposed. To enhance the credibility and interpretability of the diagnostic process, a full-process interpretability analysis system covering the three-layer transfer learning strategy is constructed, and a visual analysis flow chart showing the joint completion of transfer by the three-layer transfer learning strategy and the original model is presented, thereby improving the transparency and acceptability of diagnostic results. Relevant experiments demonstrate that the proposed intelligent analysis method exhibits excellent performance in practicality, robustness, and interpretability. The research results of this paper can provide important reference for studies in related fields.