Fault Detection of Wind Turbine Generator Bearing Using Attention-Based Neural Networks and Voting-Based Strategy

Wind turbines are usually exposed to dynamic and adverse environmental conditions, and faults on mechanical parts are major threats to the economic efficiency of wind farms. Condition monitoring and fault detection play important roles in the reliability of wind turbines. However, due to the dynamic characteristics and fluctuations of variables, precise modeling for representing the complex operation state of key components is still challenging. This study presents a data-driven fault detection method for generator bearing of wind turbines. A dual-stage attention-based recurrent neural network is first trained to model the normal behavior of generator bearing temperature. Based on long short-term memory, an input attention mechanism is specially designed to allocate the variable importances specific to each timeslot. Since the variable importance is distinct at different timeslots, the time-varying correlations among variables in wind turbines can be captured, and hence, the performance of variable estimation can be remarkably improved. In addition, this article focuses on the challenges faced by the practice of wind power industries, where the evolution rates of impending faults are usually unknown and distinct, and proposes a novel voting-based detection strategy. The proposed strategy conducts multiple processes to evaluate a potential fault from different perspectives and, hence, can guarantee the robustness and performance of fault detecting. Eventually, field supervisory control and data acquisition data from real wind farms are utilized for validation and extensive comparison studies verify the effectiveness of this proposed scheme.