Cross-condition fault diagnosis with limited data: a multi-loss adversarial framework and efficient feature learning

Abstract Bearing fault diagnosis based on deep learning methods has been widely used. However, collecting sufficient bearing fault samples in an industrial environment is difficult, significantly limiting the effective training of deep learning models. To address the challenges of feature extraction difficulty and low diagnostic accuracy in deep learning models under small-sample conditions, this paper proposes a small-sample transfer learning diagnostic method called Multi-Loss Domain Adversarial Networks (MLDAN) based on multi-loss domain adversarial techniques. A feature extractor tailored for small-sample issues, named Conv-SCMABlock-MSCABlock Networks (CSMN), is also introduced. Firstly, the vibration signals of rolling bearings from diverse domains are preprocessed using the Fast Fourier Transform (FFT), and subsequently, these processed signals are fed into the CSMN. The gradient reversal layer of MLDAN is then employed to facilitate adversarial learning between the CSMN and the domain discriminators. An adaptive MK-MMD loss function is introduced to reduce the distribution differences between the source and target domains and minimize domain discrepancies further. In contrast, a contrastive loss function with a soft-margin mechanism is applied to enhance feature discriminability. Finally, fault classification was performed using a fault classifier. Leveraging a multi-loss domain adversarial optimization strategy, this study constructed 18 different working conditions across the CWRU and PU datasets under a limited-data scenario, where the target domain contained only 50 samples per fault type. The experimental results demonstrate that the proposed method achieved remarkable performance on both datasets, with average classification accuracies reaching 99.36% and 97.38%, respectively. Moreover, even when the PU dataset was limited to only 10 and 25 samples per class, the method still achieved average fault classification accuracies of 76.36% and 89.76%, respectively. On average, each proposed innovation contributed to an improvement of 2.23% in classification accuracy, further validating the effectiveness and robustness of the approach.