Multi-scale physics-informed reinforcement learning method for bearing fault diagnosis under data scarcity

Identifying mechanical faults is crucial for maintaining stability in industrial systems. However, the lack of labeled fault data significantly undermines the accuracy and generalization of diagnostic methods in practical applications. To address this issue, this paper presents a novel multi-scale physics-informed neural network with enhanced reinforcement learning (MPINN-ERL) for bearing fault diagnosis under data scarcity. First, a multi-scale convolutional bi-directional gated recurrent unit with physics-inspired residual regularization module is developed to effectively integrate physics-inspired dynamic residual regularization into the system, thereby improving the stability of diagnostic performance. Furthermore, a conditional channel-spatial Wasserstein generative adversarial network is designed to generate augmented fault samples for alleviating the limited-sample problem. Finally, a novel loss function and an ERL algorithm are introduced to dynamically adjust the weight of physics-inspired dynamic residual regularization, thereby improving the model’s diagnostic stability when labeled data are scarce. Experimental results on the evaluated bearing fault datasets show that the proposed method achieves competitive diagnostic accuracy. Compared with several representative state-of-the-art deep learning methods, MPINN-ERL also exhibits relatively stable diagnostic performance.