Bearing Life Prediction Based on Data Distribution Adaptation Transfer Learning Regressor

To address the challenges of low prediction accuracy caused by large variations in bearing operating conditions, scarce labeled data, and the absence of annotations, this study proposes a bearing remaining useful life (RUL) prediction method based on an improved data distribution adaptation regressor. The proposed approach first extracts multidimensional features from time-domain, frequency-domain, and time-frequency domain representations. It then employs a dual maximum mean discrepancy (MMD) loss mechanism to simultaneously optimize marginal and conditional distribution discrepancies, while utilizing learnable projection matrices for dimensionality reduction and distribution alignment. Furthermore, the Transformer architecture is integrated into an iterative optimization loop, leveraging its self-attention mechanism to capture complex feature dependencies. An adaptive balancing factor is introduced to dynamically adjust the weighting strategy for distribution alignment. Experimental results show that in various conditions and domain scenarios, the proposed method greatly lowers mean squared error and mean absolute error compared to traditional transfer learning, demonstrating better cross-domain prediction.