Wind Turbine Fault Detection With Multimodule Feature Extraction Network and Adaptive Strategy

Unexpected failures in wind turbines (WTs) lead to tremendous economic losses and safety hazards to wind farms. Intelligent condition monitoring and fault detection can prevent unscheduled downtime and improve reliability and efficiency. However, the variational external environments and operation states result in time-varying correlations and fluctuations of variables, which pose challenges to precise modeling and detection for WTs. This article proposes a dual-stage temperature monitoring and fault detection method, which can be applied to multiple key components in WTs. First, a novel multimodule feature extraction network (MMFENet) is designed to model the normal behavior of temperature data in the monitoring component. The proposed model decomposes the correlations into distinct patterns and devises different modules for quantification. Specifically, the static module aims to characterize the stationary correlations resulting from the inherent structure to provide reference expectations of fluctuations in the dynamic module and accelerate the model training. The dynamic module is devoted to depicting time-varying interdependence caused by variational operation states to pursue precise modeling. In addition, the proposed network also uses a heterogeneous data fusion module to use external information. Second, an adaptive detection strategy is devised for converting the prediction errors into detection results. Based on the evaluation results in the training procedure, parameter ranges during online detection can be adaptively determined to contain multiple rational values, which can enhance robustness and scalability. Experimental results using field data from real wind farms demonstrate the effectiveness and superiority of the proposed method.