A cross-domain intelligent fault diagnosis method based on three-branch meta metric learning network for few-shot fault diagnosis

Abstract Despite the advancements made by deep learning methods in rolling bearing fault diagnosis, their reliance on large labeled datasets restricts their practical applicability in real-world scenarios where such data is often limited. This raises the challenge of building models that can function effectively with minimal data and perform well under varying operational conditions. To address this issue, we propose the Three-Branch Prototype Network (TBPN), which adopts a meta-learning strategy, forming tasks by randomly sampling original signals from different operating conditions. At first, we enhance the original signals corresponding to known operating conditions using a denoising strategy based on the Gram matrix in the time and frequency domains. These enhanced time-domain and frequency-domain signals, along with the original signals, are fed into the TBPN model as three branches to extract and fuse fault features. Next, a metric learner is applied to derive prototype representations for each type of fault and calculate the distances between these prototypes and the query fault features, which are then used in a softmax function for multi-class fault classification. The TBPN model demonstrates superior performance in providing rapid and accurate fault classification for rolling bearings, even under unknown operating conditions, by updating its parameters using minimal sample data. To fully assess the performance of our proposed method, we conducted extensive experiments across various industrial settings using the Case Western Reserve University Rolling Bearing Dataset and the Laboratory Rolling Bearing Dataset. The results underscore the effectiveness of TBPN in addressing the challenge of few-shot fault classification in complex environments.