Wind turbine fault detection based on deep residual networks

Abstract

Condition monitoring and fault detection for wind turbines (WTs) can effectively lower the effect of failures. A large amount of data would be generated during the operation of WTs, and these data have the following characteristics: (1) The data volumes are too large, and it is difficult to extract fault features. (2) Signals are correlated, and a fault may lead to multiple sensors to alarm. (3) Dimension and orders of magnitude are different between parameters. (4) The supervisory control and data acquisition (SCADA) data are fluctuating, which can easily lead to false alarms and missed alarms with certain one-sidedness. (5) Fault alarms possess a lag. To solve the problems of inaccurate and untimely fault detection (FD) caused by these data characteristics, a new deep network called deep residual network (DRN) is proposed in this paper for WTs’ detection. In the proposed method, the raw data collected by SCADA system are directly applied as the inputs of the DRN. Then, a convolutional residual building block (CRBB) is established by using convolutional layers, squeeze and excitation units. Meanwhile, the improved meta-ACON (active or not) is introduced to replace of rectified linear unit (ReLU). The high-level features are extracted from the raw data by stacking multiple CRBBs. Finally, the FD results are obtained by feeding the extracted features to the softmax classifier. The proposed DRN is validated by using the data from the SCADA system. The results indicate that the proposed DRN achieves better performance, and outperforms some published fault detection methods.