Research on State Prediction of Wind Turbine Gearbox Based on Monitoring Data Distribution Similarity and Feature-Based Transfer Learning

To make full use of the limited monitoring data with fault information in the case of inconsistent distribution of the monitoring data among the wind turbines, a combined operational state monitoring method based on spatiotemporal self-coding, exponential weighted moving average, and feature-based transfer learning is proposed. Firstly, a spatiotemporal self-coding network is used to extract the common feature information of similar wind turbines, and then an operational state monitoring network based on group similarity and parameter migration, as well as a wind turbine gearbox (WTGB) operational state prediction model based on deep transfer learning is constructed. Case applications are performed using the actual monitoring data of a wind farm in northern China. The results show that the combined method can effectively detect the fault information, and the classification accuracy can almost above 0.9. The supervisory control and data acquisition (SCADA) monitoring data with fault information can be used to detect the potential failures of other WTGBs, enabling condition-based maintenance, cost savings and improved reliability.