To address the challenges of significant feature distribution discrepancies across different working conditions and the difficulty in distinguishing fine-grained faults in rolling bearings, a dynamic weighted joint distribution adaptation (DW-JDA) network is proposed. The method integrates a continuous wavelet transform (CWT) preprocessing module, a depthwise separable convolution (DSConv) based feature extractor, and a novel entropy-weighted discriminative joint distribution alignment (EW-DJDA) mechanism. First, one-dimensional vibration signals are converted into two-dimensional time–frequency images via CWT to fully capture the time–frequency characteristics of nonstationary signals. Then, a hybrid backbone network based on DSConv is constructed to efficiently extract spatial features. Simultaneously, the EW-DJDA module is introduced, which dynamically adjusts the alignment weights of marginal and conditional distributions according to the prediction uncertainty in the target domain, effectively mitigating negative transfer. Furthermore, to tackle the issue of ambiguous decision boundaries, a Hard Negative Margin Softmax loss function is proposed, which explicitly increases the distance between the target class and the strongest interfering class, thereby enhancing the model’s discriminability for hard-to-distinguish samples. Experiments on the Central South University, Case Western Reserve University, and Beijing Jiaotong University datasets demonstrate that DW-JDA achieves average diagnostic accuracies of 99.55, 99.92, and 99.91%, respectively, showing competitive or best average performance among the compared transfer learning methods. Particularly under strong noise conditions with signal-to-noise ratio = −2 dB, the model maintains an average accuracy of 97.71%, exhibiting excellent robustness and generalization capability.
Xie et al. (Sat,) studied this question.