Fault diagnosis method for unbalanced gearbox bearing under variable working conditions based on generation-enhanced domain adaptation

Abstract

Aiming at the challenge of insufficient minority-class recognition in gearbox bearing fault diagnosis under the combined influence of variable working conditions and class imbalance, this study proposes a fault diagnosis method based on generative enhancement and domain adaptation. Firstly, a Wasserstein generative adversarial network with gradient penalty (WGAN-GP) is employed to generate minority-class fault samples, thereby supplementing the distribution support of scarce categories in the source domain. Then, the generated samples and real source-domain samples are jointly used to train a domain-adversarial diagnostic network composed of a feature extractor, a fault classifier, a domain discriminator and a gradient reversal layer. Through adversarial training, the feature extractor learns domain-invariant representations between the source and target domains, while the fault classifier maintains the discriminative boundary of minority faults. This strategy alleviates the local decision boundary compression caused by clustering of minority samples with majority or similar classes during cross-condition migration. Experiments are conducted on the Southeast University (SEU) Bearingset dataset under a cross-load task (Load20 → Load30).with the source-domain sample counts of Health, Ball, Inner and Outer classes being 8000, 2000, 2000 and 500, respectively. Comparative experiments include convolutional neural network with L2-support vector machine loss (CNN-L2SVM), Class-weighted domain-adversarial neural network (Class-weighted DANN), local maximum mean discrepancy domain-adversarial neural network (LMMD-DANN), Random Oversampling-DANN, CIDSAN-Adapted derived from class-imbalanced deep subdomain adaptive network (CIDSAN), and the proposed method. Results show that the Outer-class recall rates of CNN-L2SVM, Class-weighted DANN and LMMD-DANN are 0.05%, 0.68% and 1.53%, respectively. CIDSAN-Adapted improves the recall to 53.34%, while the proposed method achieves 97.07%. Confusion matrix and three-dimensional t-distributed stochastic neighbor embedding (3D t-SNE) visualizations further demonstrate that generative enhancement increases feature support for the Outer class and improves the discriminative boundary of minority classes in the target domain. The proposed method provides a data-driven diagnostic scheme for prognostics and health management of marine transmission equipment and supports intelligent ship operation and maintenance under limited fault data and variable operating conditions.