Multiple classifiers inconsistency-based deep adversarial domain generalization method for cross-condition fault diagnosis in rotating systems

Abstract

Unknown fault operating conditions and the absence of fault data pose significant challenges for real-time fault diagnosis, as the generalization capability of models is heavily reliant on transferable knowledge from a single operating condition. To overcome these limitations, a novel deep adversarial domain generalization framework based on multiple classifiers inconsistency (DADG-MCI) is designed to improve generalized ability without the need for target domain data during training. Initially, unique features of the multiple source domains are captured through the probability output inconsistency of the multiple domain-specific classifiers. Subsequently, adversarial training facilitates finer-grained global feature alignment across multiple source domains, which ensures that the extracted deep features possess strong generalization capabilities. Most importantly, DADG-MCI introduces the multiple classifiers inconsistency to measure multi-domain distributional discrepancy based on Wasserstein distance, which captures feature distribution differences between domains through joint optimization of the multi-classifier module. Finally, two challenging rotating machinery fault datasets are used to evaluate the performance of DADG-MCI for cross-condition fault diagnosis. Compared to several state-of-the-art methods, DADG-MCI achieves the highest average diagnostic accuracies and successfully applies to unseen operating conditions.