A novel wavelet constrained physics-informed neural network for bearing fault diagnosis

Fault diagnosis models based on data-driven method have always been a hot topic of research, promoting the intelligent development of rolling bearing fault diagnosis methods. However, data-driven fault diagnosis methods fail to directly account for fault mechanisms, thereby diminishing the credibility of their results. Therefore, a novel wavelet constrained physics-informed neural network (WCPINN) is proposed for bearing fault diagnosis. Firstly, an ensemble wavelet convolutional layer is proposed to extract physical information related to the health condition of bearings. Secondly, a novel physics-constrained loss is designed into the training stage to guide the parameters updating direction of the model. Two bearing datasets are utilized to validate the effectiveness of the proposed method. The results suggest that the is capable of fully focusing on the fault-related physical information within the original vibration signals, thereby endowing the model with a certain degree of feature credibility. The addition of physical constraints makes the training of the neural network more targeted, enhancing the stability and accuracy of the diagnosis results.