Multi-scale Conditional Domain Adversarial Network with Spectral Penalization for rotating machinery fault diagnosis across varying operating conditions

To address the limitations of conventional convolutional neural network (CNN) in feature extraction of vibration signals and improve fault discriminative capability in adversarial domain adaptation in cross-domain fault diagnosis of rotating machinery, this study proposes a novel fault diagnosis approach based on a Multi-scale Conditional Domain Adversarial Network with Spectral Penalization (MCDANSP). The model directly uses raw vibration signals as input. Its feature extraction module combines multi-scale convolutions and a channel attention mechanism to enhance shared feature extraction between source and target domains. For knowledge transfer, the method uses a conditional domain adversarial network (CDAN) and adds batch spectral penalization (BSP) to preserve feature discriminability during domain alignment. Evaluation on the PU, PHM2009, and SEU datasets yielded accuracies of 82.96%, 84.63%, and 64.03%, respectively. These results surpassed other methods in this research, confirming the method’s superior performance.