MPINet: Multiscale Physics-Informed Network for Bearing Fault Diagnosis With Small Samples

Deep learning is increasingly prevalent in the bearing fault diagnosis, while the deficiency of fault samples could diminish the diagnostic efficacy of data-driven models that depend on extensive training data. For that, a novel multiscale physics-informed network (MPINet) is proposed for bearing fault diagnosis with small samples. Our fundamental idea is incorporating physical knowledge into the training process for enabling the model could better learn the fault features. To pursue this goal, a physics-informed block (PIB) is developed to extract fault features, which is customized for each failure mode. By this process, multiple independently trained PIBs encode the physical knowledge of their corresponding failure mode into the model, and thus yield multiscale fault features. Finally, the diagnosis result is obtained by using a new classifier head to merge these multiscale features. Extensive experimental results show that our MPINet can obtain superior diagnosis performance with small samples.