Temporal Convolution Network Based on Attention for Intelligent Anomaly Detection of Wind Turbine Blades

Abstract

With the concept of “carbon neutral”, wind power industry ushers in a golden period of development. Wind turbine blades are the key components of wind turbine, their normal operation directly affects the operation of the whole wind power generation. The traditional diagnosis method of wind turbine blades needs to disassemble the wind turbine blades. However, due to the potential of long-short-term memory network model and feedforward network based on attention mechanism in anomaly detection and diagnosis, sequence model has become the mainstream of wind turbine blades anomaly detection. Unfortunately, RNN needs to wait for the end of the forward transmission of the previous time step before the forward transmission of the next time step. It can only input all the historical information by default, and can not achieve the fine control of the input information. Therefore, this paper applies TCN-ATT, a temporal convolution network based on attention mechanism, to intelligent anomaly detection of wind turbine blades by combining dilation convolution, causal convolution, the skip connection of residual blocks and attention module. In this method, causal convolution and dilation convolution are used to adapt to the frame of multi-dimensional time series data of wind turbine blades, which provides more flexibility for changing the size of receptive field. At the same time, attention mechanism is added to capture the relevant features in the sequence, so as to improve the classification accuracy of anomaly detection of wind turbine blades. To verify the effectiveness of the system, this paper compares the performance of TCN-ATT with TCN, LSTM, GRU and other methods. The comparative test shows that TCN-ATT has better performance of feature extraction and anomaly detection. In addition, TCN-ATT is applied to multi fault anomaly detection, which also achieves high classification accuracy.