Thermal-Aware Generative Adversarial Augmentation for Small-Sample Induction-Motor Fault Detection Using Infrared Thermography

Fault diagnosis in induction motors is challenging when only a limited number of labeled infrared thermal images are available for training. Effective small-sample augmentation requires generated images to preserve class-related appearance patterns as well as visual realism. To address this challenge, a thermal-aware generative adversarial network (TA-GAN) is proposed for induction-motor fault diagnosis using pseudo-colour infrared images. An adaptive attention module combines feature intensity and Sobel-gradient information to emphasize component–state-related thermal patterns. A dual-branch critic jointly evaluates global and attention-weighted features, while an appearance-consistency loss constrains bright-region intensity, gradient structure, and spatial position during generation. The trained generator augments the limited real samples, and the combined dataset is used to train a lightweight classifier. Experiments on acquisition-batch-separated data partitions show improved classification performance over the evaluated real-only baselines under matched classifier-training budgets. The complete framework performs best among the evaluated component and individual-loss configurations. Generated-sample evaluation further indicates improved appearance similarity to real training images compared with a basic conditional GAN, while evaluation on a different motor provides additional evidence of transfer under the tested conditions. Image generation is performed offline, allowing the deployed system to retain the lightweight MobileNetV4 classifier architecture.