Zero-Shot Bearing Fault Diagnosis Under Varying Conditions Based on Prototypical Vision Transformers

Bearings are critical components susceptible to various faults, making accurate diagnosis essential. While data-driven methods have significantly improved fault diagnosis, zero-shot fault diagnosis remains challenging. This is primarily due to constantly varying working conditions and the emergence of novel fault types, which typically necessitate costly model fine-tuning or retraining. To overcome these limitations, we propose a novel Prototypical Vision Transformer (PViT) for zero-shot bearing fault diagnosis. PViT innovatively combines the powerful feature representation capabilities of Vision Transformer (ViT), especially with limited data, with the strong representation learning of prototypical networks. This synergy eliminates the need for model fine-tuning or retraining. Our method first transforms raw vibration signals into time-frequency images, enabling cross-modal representation. Subsequently, a ViT model is developed using a meta-learning scheme to extract generalized, fault-related features from these images. A prototypical network is then employed for fault classification, allowing the recognition of unseen fault types without modifying the ViT parameters. Extensive experiments on public datasets validate the proposed PViT. Results demonstrate high diagnostic accuracy, even with limited training data, under varying working conditions, and in the absence of certain fault types during training, showcasing its robustness and practical applicability.