SCADA data-driven blade icing detection for wind turbines: an enhanced spatio-temporal feature learning approach

Blade icing is one of the common issues of large-scale wind turbines located in cold regions, which will affect the safety and efficiency of the whole turbine system. Currently, data-driven fault detection has gained increasing interest due to the availability of a large volume of supervisory control and data acquisition (SCADA) data. However, SCADA data has complex time-varying characteristics and strong spatio-temporal correlations among different sensor variables, thus it is still challenging to extract effective fault features for accurate detection. To this end, this paper proposes an enhanced spatio-temporal feature learning approach, called multi-task temporal spatial attention network (MT-STAN). It contains two core modules: a feature extraction module and a multi-task learning module. For better spatio-temporal feature extraction, a spatio-temporal attention block is first developed to extract important variables in the spatial dimension and temporal segments in the temporal dimension via the attention mechanism. Then, we design a multitask learning module, consisting of both deep metric learning and classification learning tasks, to further enhance the discriminative ability of the learned representations and improve the performance of fault detection. The proposed approach is evaluated on a real SCADA dataset, and the results show that our proposed MT-STAN model achieved better detection performance compared with several baseline models.