A novel wind turbine fault diagnosis based on deep transfer learning of improved residual network and multi-target data

Abstract In industrial production, problems such as lack of data, complex fault types, and low generalizability of deep learning models seriously affect the fault diagnosis of wind turbines. Therefore, we have developed a fault diagnosis model for wind turbines under harsh conditions to address the above problems. First, the collected one-dimensional vibration data is reshaped into two-dimensional form by using the Gramian Angular Field. The two-dimensional form not only extends the spatial structure of the data, but also effectively improves the information expression of the data. In addition, the data is classified into large-scale data, medium-scale data, small-scale data, class-imbalanced data, and heterogeneous data based on the data type. Then, the deep residual network structure is redesigned to improve the diagnostic performance of the model based on the sensitivity of the reshaped data to the size of the convolutional kernel, and the new structure of the network is employed to implement transfer learning. Finally, we adopt the developed fault diagnosis model to achieve the fault diagnosis of bearings and gears in the wind turbine gearbox. Meanwhile, an automatic hyperparameter search mechanism was added to improve the partial hyperparameter optimization in this study. It is demonstrated that the model proposed in this study has excellent diagnostic performance with multi-target data for wind turbines, and has excellent generalizability and reliability.