Rotating Machinery Fault Diagnosis Method Based on Enhanced Transfer Learning

This chapter offers a novel transfer learning paradigm specifically for defect identification in rotating machinery. The approach tackles a fundamental problem in cross-domain applications: the distributional shift between source and target domains, which usually reduces the efficacy of standard machine learning methods under different operational settings. The suggested method jointly combines geodesic flow kernel (GFK), correlation alignment (CORAL), and a multi-modal feature extraction process to close this gap. To capture both shallow and deep aspects, the later hierarchically combines time-domain statistical characteristics, wavelet scattering coefficients, and stacked auto-encoders (SAE). The Case Western Reserve University (CWRU) bearing dataset provides empirical validation; generalization performance is rigorously assessed on two separate target domains: the Shandong University of Science and Technology (SDUST) and Xi’an-Jiaotong University (XJTU-SY). Classification accuracy of 95.56% (SDUST) and 98.81% (XJTU-SY) revealed by experimental results highlight the flexibility of the framework in heterogeneous data distribution. These results support the possibility of the model as a scalable solution for industrial defect detection, especially in cases with limited labelled target-domain data.