Generalization classification regularization generative adversarial network for machinery fault diagnostics under data imbalance

Recently, generative adversarial networks (GANs) have emerged as powerful solutions for fault diagnosis under data imbalance. However, the challenges of the poor generalization ability of GANs, difficulty in capturing diverse feature representations and reliance on data preprocessing still limit effective applications of GANs-based fault diagnosis methods. To tackle these challenges, an innovative generalization classification regularization generative adversarial network (GCRGAN) is proposed, aimed at machinery fault diagnostics under data imbalance. Firstly, a new generator consisting of a random sampling module, a generating module and a generalization module is designed to improve the generated sample quality under limited data. Meanwhile, original vibration signals serve as the generator input to capture diverse feature representation and accelerate the model convergence. Subsequently, an innovative regularization loss is incorporated into the discriminator loss function, which can improve the generalization and stabilize the learning process under limited data. Then, a classifier module is developed to enforce the generator to generate high-quality samples. Finally, case studies of machinery fault diagnostics under extremely data imbalance have been conducted to demonstrate our proposed approach. Experiment validation results have demonstrated that our GCRGAN approach generates high-quality samples and yields superior diagnostic results under data imbalance, compared with mainstream data augmentation methods.