关键词: 风电机组, 时变性能, LSTM-AE神经网络, 自适应阈值, 互信息, 故障诊断

Abstract: Performance assessment and anomaly detection of wind turbine are important for product health monitoring.Previous performance assessment models seldom considered the sequential characteristic of data and complex operation conditions in performance monitoring,which led to low accuracy,and the function modules related to the abnormality are not identified based on the performance assessment,resulting in high maintenance cost.For these problems,a new method was proposed to assess performance and detect anomaly points of wind turbine based on Long Short Time Memory-Auto Encoder (LSTM-AE) neural network.A performance assessment model was constructed based on LSTM units and AE to compute the index that described the abnormality of wind turbine performance.By comparing with the adaptive threshold which was calculated through Support Vector Regression (SVR),the anomaly points were identified.For different performance parameters,Gaussian Copula entropy was employed to estimate the mutual information.By sorting the mutual information,the critical performance parameters were finally identified,and then mapped to function models.A case study results showed that the proposed method could effectively process the data with sequential characteristic and improve the accuracy of anomaly detection.

Key words: wind turbine, time-varying performance, LSTM-AE neural network, adaptive threshold, mutual information, fault diagnosis