A discriminator-free adversarial network for bearing fault diagnosis under unseen operating conditions

Abstract

Rolling bearings are critical to rotating machinery, contributing to nearly 40% of equipment failures and resulting in substantial economic losses. Domain adaptation methods have shown promise in rolling bearing fault diagnosis. However, their reliance on prior target-domain data limits effectiveness under variable, real-world operating conditions. To overcome these challenges, this paper proposes a novel domain adversarial generalization network. First, a novel data augmentation method is introduced, which generates new training samples from various operating conditions. This enhances data diversity and smooths the data distribution. Second, a multi-channel feature extractor is designed to capture domain-invariant features from multiple perspectives, improving the model's robustness. Additionally, a label smoothing regularization strategy is incorporated to mitigate overfitting, and a new $L1,2$-norm is applied for domain discrepancy calculation. This optimizes the efficiency of feature alignment and eliminates the need for additional discriminators. Extensive experiments are conducted on three datasets, and the results demonstrate that the proposed model achieves superior performance under various operating conditions, especially in scenarios with significant data distribution differences. These findings highlight the model's potential for industrial applications.