Fault Detection of Wind Turbine Converters with Time Sequence Processing and Attention Model

Fault detection methods of converters can improve the stability of wind turbines and reduce the maintenance cost. In order to achieve more accurate converter fault detection, this paper proposes an encoder-decoder structure based on the Long Short-Term Memory (LSTM) network. The LSTM network is considered to extract the sequence information of SCADA data to improve the accuracy of fault detection. Moreover, we introduce attention models into our method. Attention models can enable the network to selectively focus on the most useful variables and data during training and testing period. Specifically, we use the time attention model to calculate the importance of data at different times in the sequence, and the feature attention model to learn the importance of different variables. Experimental results on SCADA data show that our method achieves 7% -14% higher fault detection accuracy than other anomaly detection methods.