Jiang Guoqian, Zhou Junchao, Wu Xin, Xu Xiangdong, He Qun, Xie Ping
Accurate and reliable anomaly detection is of great significance to ensure the safe and efficient operation of wind turbines. However, due to the complex structure and variable operation conditions of wind turbines, the measure SCADA data usually present complex nonlinear and strongly correlated and coupling characteristics. To better capture spatial correlations among different sensor variables, a new wind turbine anomaly detection approach based on dilated causal convolution network is proposed. Specifically, Focal Loss is introduced to improve the traditional loss function to address the data imbalance issue. The proposed approach can extract effective multiscale spatially correlated features with different receptive field sizes and effectively model the hidden spatial causality among different sensor variables. Furthermore, it can provide an end-to-end anomaly detection solution for wind turbines, which can directly learn useful spatial features from raw SCADA data and build the nonlinear mapping relationship between original data space and condition label space, thus finally outputting the corresponding detection results. A real case study with SCADA data from a wind farm is used to verify the feasibility and effectiveness of the proposed approach.