A multi-source domain adaptive bearing fault diagnosis method based on progressive multi-stage alignment guidance

Abstract Multi-source domain adaptation effectively addresses challenges such as limited target domain labeling and insufficient single-source domain information in intelligent fault diagnosis. While multi-source domain adaptation offers advantages over traditional methods, it becomes more complex due to greater inter-domain differences. To address the aforementioned challenge, a progressive multi-stage aligned multi-source domain adaptive fault diagnosis method (PMAMDA) is proposed. This network combines a shared feature extractor, several domain-specific feature extractors, domain discriminators, and classifiers, utilizing a multi-stage distributional adaptation strategy. The introduction of hybrid metric, combining variance discrepancy representation (VDR) and Wasserstein distance for adversarial training, enhances domain adaptation performance. Additionally, divergence loss is used to align the outputs of multiple specific classifiers. Through extensive experiments on a bearing dataset, PMAMDA shows exceptional fault diagnosis performance across different speed and load conditions, attaining an average accuracy of 99.18%, which notably surpasses the performance of other methods.