A lightweight fault diagnosis method based on knowledge distillation under time-varying rotational speeds (June 2024)

Intelligent fault diagnosis models for engineering must meet real-time requirements while satisfying high diagnosis rates. However, collecting a large amount of fault information of different working conditions in the engineering environment is difficult, and the lack of fault data is a problem that has been difficult to solve. This paper presents a lightweight fault diagnosis model for wind turbine gearboxes based on transfer learning and knowledge distillation. The model is designed to be lightweight and can be used under time-varying rotational speeds. First, TL-ResPConv based on transfer learning and partial convolution was designed as a teacher network model to train a lightweight student network model for fault diagnosis based on TL-ResPConv-KD. Then, the student model obtained through knowledge distillation training is used as the distilled fault diagnosis model. Finally, the performance of the different models is validated and compared on a laboratory simulated planetary gearbox dataset. The experiments prove that the proposed model outperforms other models in fault recognition, achieving a balance between lightweight and high diagnostic accuracy. This makes it a practical network with higher diagnostic accuracy for fault diagnosis under time-varying rotational speeds, meeting the demand for lightweight deployment of edge equipment.