A cross-domain fault diagnosis method for gearbox based on cosine-similarity discriminative augmentation under imbalanced samples

The fault diagnosis method based on domain adaptation requires balanced monitoring data across various health conditions. However, in practical industrial scenarios, obtaining specific fault data is challenging. The issue of imbalanced fault data significantly impairs the performance of fault diagnosis methods. To address these challenges, this article proposes a novel approach called the correlation contrastive domain adaptive network with cosine-similarity discriminative Mixup (CCDA-CSDM). First, a data augmentation module based on CSDM is introduced. It selects samples using cosine similarity metrics and generates augmented data through the Mixup strategy. Second, a multilayer dilated convolution is designed for multiscale feature extraction, enabling the acquisition of rich fault information. Third, the correlation contrastive domain adaptation (CCDA) module is employed to minimize global distribution discrepancies between source and target domains while preserving the interclass separation of local samples. Experimental validation is conducted on two gearbox datasets. The proposed CCDA-CSDM achieves diagnostic accuracies of 98.64 and 98.75%, respectively, under imbalanced and unlabeled target domain conditions, demonstrating superior fault diagnosis performance.