A novel meta-transfer learning method for few-shot cross-domain fault diagnosis based on parallel synergistic network

Numerous research has confirmed the efficacy of utilizing deep learning for diagnosing faults in rotating machinery. However, this approach often overlooks the practical constraints in industrial scenarios, such as the variable operating environments characteristic of mechanical systems, and the limited availability of fault data. In this article, we propose an improved model-agnostic meta-learning method for few-shot cross-domain fault diagnosis based on parallel synergistic network (IMAML-PSN). This method first utilizes fourier transform-gramian angular summation field (FT-GASF) for data preprocessing to emphasize fault features and ensure that they are not affected by cross-domains. Then, the training moderator variable optimizes the meta-learning training process and improves diagnostic performance according to different meta-tasks. A feature encoder based on parallel synergistic attention mechanism captures the interaction relationships between multidimensional features and learns shared diagnostic knowledge. Comparative cross-domain experiments on bearing fault datasets confirm the method’s diagnostic efficacy and robustness, under the complex operating environments constrained by sample scarcity.