Semi-supervised bearing fault diagnosis based on metric learning and domain adversarial learning

Abstract As critical components in rotating machinery, bearing health status directly affects the operational safety of equipment and personnel. Therefore, effective bearing condition monitoring is of great importance. However, challenges such as scarce labeled samples, varying operating conditions, and strong noise interference severely restrict the performance of high-precision fault diagnosis. To address these issues, this paper proposes a fault diagnosis method that combines maximum mean discrepancy and a domain-adversarial neural network to reduce feature discrepancies under variable speed and load conditions. High-confidence pseudo-labels are adopted to improve diagnostic performance under semi-supervised settings. Based on the above strategies, a unified loss function is constructed and a high-precision bearing-fault diagnosis framework adaptable to varying working conditions is established. Meanwhile, a tailored convolutional auto-encoder (CAE) is designed for strong-noise environments to denoise vibration signals and provide high-quality inputs for the network. Experimental results demonstrate that, with ResNet14 as the backbone, the proposed feature alignment method achieves diagnostic accuracy exceeding 98% on both the Case Western Reserve University dataset and the self-built test dataset, even when only 5% of target-domain data are labeled. The constructed CAE module maintains reliable denoising performance even under a high noise level of −10 dB. The entire model contains 1531 973 parameters and requires only 3 s per training epoch. Accordingly, the proposed method is well-suited for semi-supervised domain-adaptive bearing fault diagnosis under real industrial scenarios.