Dual-module multi-head spatiotemporal joint network with SACGA for wind turbines fault detection

Fault detection in wind turbines (WTs) was commonly characterized by an imbalance of fault class data, which could lead to a degradation of fault detection performance. In addition, temporal and spatial interaction information is not considered in the fault detection process, which weakens the model performance. Based on the above problems, this study proposes a novel dual-module multi-head spatiotemporal joint network with slidingwindow auxiliary classifier generating adversary (DMSJN-SACGA). The proposed DMSJN-SACGA in this study consists of four parts: data generation, dual-module feature encoder, multi-head spatiotemporal joint representation, and fault classification decoder. Firstly, the designed SACGA module, which utilizes the labeled fault data of WTs, generates high-quality fault class data to alleviate the problem of imbalanced fault class data of WTs. Secondly, the designed dual-module spatiotemporal joint representation framework learns the interactions between spatial attribute representation and time sequence representation to realize spatiotemporal joint representation. Compared to training with real data only, the key metrics of macro-F1 are 0.23 higher and g-meanF1 are 0.332 higher for the proposed DMSJN-SACGA trained with the addition of generative data. Compared to the other baseline models, the proposed DMSJN-SACGA has a superior performance in realizing the effective classification of WTs fault detection.