Research on fault diagnosis of wind power gearbox bearings based on transfer learning

With the increasing demands for the reliability and stability of wind turbines, predictive fault diagnosis of wind turbines has become increasingly necessary. Among all components, bearings have the highest failure rate in wind turbines, and their operating condition is crucial for the normal operation of wind turbines. Currently, fault classification is mainly conducted through experimental simulation, and fault signal characteristics are extracted from the data collected from the fault test bench using big data technology for fault diagnosis. Traditional vibration signal analysis requires expert experience to extract signal characteristics, which is difficult and complex, and is heavily influenced by human subjective factors. However, deep learning utilizes artificial neural networks to automatically extract features, resulting in a more comprehensive and objective representation of faults. As a major research area in current fault diagnosis, deep learning has been widely applied. Combining different neural network models with various data processing methods can improve the performance of fault diagnosis and continuously enhance its accuracy. As one of the most powerful theories in deep learning, transfer learning applies learning from known tasks to new tasks. It can enhance the performance of deep neural network models by learning from the data characteristics of known tasks, thereby obtaining accurate fault classification results. In this paper, a bearing fault diagnosis method based on transfer learning utilizes the EfficientNet-b0 network and employs a small sample size of actual operating data from wind turbines for fault diagnosis. The results show that this method achieves accurate fault diagnosis, with a classification accuracy of 100% for four types of faults with small sample sizes.