Feature transfer and fuzzy clustering for bearing degradation prediction under complex operating conditions

• Transfer of bearing features across different environments. • Proposes new degradation features to measure the degree of bearing degradation. • Prediction of bearing degradation using feature transfer and fuzzy clustering. Bearing degradation prediction is essential for ensuring the reliable and efficient operation of rotating machinery, enabling timely fault diagnosis and maintenance scheduling. However, under complex operating conditions, variations in load and environment often lead to significant distribution discrepancies between laboratory and engineering data, thereby degrading prediction accuracy. Therefore, a degradation prediction method integrating transfer learning and fuzzy clustering is proposed. First, kurtosis and waveform index are identified as key degradation indicators through correlation analysis, enhancing feature representativeness while reducing redundancy. Subsequently, two semi-supervised domain adaptation methods, namely Transfer Component Analysis (TCA) and Correlation Alignment (CORAL), are employed to transfer and align features from laboratory bearings to wind turbine bearings. Finally, fuzzy c-means clustering is applied to determine cluster centers corresponding to healthy and faulty states, and the degradation degree is quantified based on the distance between the transferred features and these centers. Experimental results demonstrate that the proposed method achieves an RMSE of 0.0580 and an MAE of 0.0451, outperforming BP, LSTM, TCA, and CORAL methods. In addition, t-SNE visualization and statistical metrics confirm that the proposed approach effectively improves cross-domain feature alignment. Further validation on the XJTU-SY dataset demonstrates its robustness and generalization capability.