Unsupervised GAN With Fine-Tuning: A Novel Framework for Induction Motor Fault Diagnosis in Scarcely Labeled Sample Scenarios

Abstract

Induction motor serves as the indispensable component of industrial equipment, where the failure of the key part can lead to significant losses, and its intelligent fault diagnosis (IFD) based on supervised deep learning methods requires extensive labeled data. However, acquiring sufficient labeled samples in industrial equipment usually remains a significant challenge. To tackle this problem, a novel semi-supervised IFD framework based on an improved unsupervised generative adversarial network (GAN) and fine-tuning is proposed. Specifically, an enhanced learning strategy is developed by introducing gradient normalization constraint and a hinge loss function into the modified GAN model to stabilize the adversarial training process and improve the discrimination, which can boost the model’s feature distribution learning ability among various unlabeled data categories. The discriminator trained by unlabeled samples is then employed to identify the fault after being fine-tuned with a limited number of labeled samples. Finally, the effectiveness of the proposed framework is verified by two induction motor cases, including mechanical and electrical failures. The results demonstrate that the proposed method has significant superiorities in training stability and identification accuracy under the condition of very few labeled samples.