Multiscale ECA network: a rotation mechanical domain adaptation method with minimal class confusion

Unsupervised rotation mechanical fault diagnosis methods have become popular, but existing unsupervised methods still have some issues. For example, it is challenging to capture vibration signal features at different scales and to address partial class confusion. To improve the diagnostic performance, this study introduces a multiscale, efficient channel attention (ECA) attention mechanism, joint adaptation network (JAN), and minimum class confusion (MCC) for addressing the aforementioned issues. First, the authors design a multiscale fault feature extraction module to capture discriminative information at different scales in vibration signals. Second, the authors introduce the ECA mechanism to weight the extracted features at the channel level, enhancing useful features and suppressing redundant features. Then, the authors employ the JAN method to establish local maximum mean discrepancy, enabling adaptation between corresponding subdomains of the source and target domains, avoiding the problem of being too close. Finally, the authors use MCC as the loss function to reduce prediction confusion between correct and ambiguous categories in target samples, thus improving transfer performance. Experimental results demonstrate that the proposed method exhibits excellent performance in unsupervised rotation mechanical fault diagnosis tasks.