Discriminative Condition-Guided Generative Model for Induction Motor Fault Diagnosis With Limited Data

The scarcity of fault samples degrades the accuracy of data-driven intelligent fault diagnosis (IFD). An auxiliary classifier generative adversarial network (GAN) has therefore emerged as a dominant paradigm for generating multiclass data to mitigate this issue; however, this framework suffers from low intraclass diversity and gradient instability. Specifically, the classifier's strong class-label separability compresses the category's support space, reducing diversity and also causing optimization conflicts. To this end, this article proposes a discriminative condition-guided generative model (DCGM) to synthesize high-fidelity data across categories for induction motor fault diagnosis under small-sample conditions. First, a discriminative classifier is integrated into the conditional GAN to replace the auxiliary classifier, theoretically alleviating diversity constraints and instability. After that, an adaptive feature matching loss based on supervised contrastive learning is proposed to enhance the synthetic quality of the class-label data. Then, three novel evaluation metrics are developed to quantitatively assess the generated data quality. Extensive experiments on induction motor datasets demonstrate that DCGM achieves state-of-the-art evaluation scores compared to diffusion-, transformer-, and GANs-based models. Finally, small-sample fault diagnosis further validates the superiority of the proposed approach, highlighting its potential in engineering applications. To the best of our knowledge, this is the first to introduce the novel discriminative generative framework for conditional data generation in the IFD field.