Rotating machinery cross-operating condition fault diagnosis based on joint weighted multi-scale migration algorithm

In rotating machinery fault diagnosis, the distribution difference of vibration signals across working conditions and insufficient feature extraction significantly reduce the diagnostic accuracy and are often accompanied by negative feature extraction transfer, making transfer learning from the source domain to the target domain complicated and unstable. In this paper, an innovative cross-working condition transfer learning network, Joint weighted multi-scale network (JWMS-NET), is designed. The JWMS-NET model innovatively combines a multi-scale dynamic convolutional residual network as a feature extractor, a pseudo-label target domain classification loss, and a Jensen–Shannon divergence (JSD) distribution loss. First, a JSD-based distribution loss metric is proposed to enhance the confusion and alignment of cross-domain features, thereby reducing the impact of inconsistent fault feature distributions in the two domains. Second, in response to the negative transfer problem that may occur during feature extraction, this paper designs an improved pseudo-label target domain classification loss mechanism. This mechanism effectively adjusts the loss weights of samples in different target domains through adaptive weight allocation, suppresses the negative transfer effect, and prompts the feature extractor to learn more robust and cross-domain consistent features. Finally, JWMS-NET can further improve diagnostic accuracy by optimizing the feature representation of the target domain and flexibly controlling the decision boundary. The average accuracy of the JWMS-NET model in this paper reached 99.66%, 99.84%, and 99.41% in 26 cross-speed migration tasks on three datasets. Through subsequent experimental comparisons, the JWMS-NET model significantly outperforms the existing domain adaptation model, verifying its superiority and robustness in bearing fault diagnosis.