Adaptive fault detection in wind turbine via RF and CUSUM

Abstract

Rapid developments of wind industry arise the issue of heavy monitoring tasks. The residual monitoring based on normal behaviour modelling is a highly recommended method when fault record information is missing. However, it is difficult to achieve efficient normal behaviour modelling and dynamic residual monitoring simultaneously. To this end, a novel adaptive fault detection scheme, which merges random forest (RF) with adaptive cumulative sum (CUSUM), is proposed. The authors exploit RF to explore the non-linear mechanism between features and the target variable robustly, and obtain the residuals quickly. Then, they design the adaptive CUSUM control chart of time-varying shift to sensitively detect the changes of residuals. For illustration, they apply the proposed scheme to the supervisory control and data acquisition data acquired from a wind farm in China. The empirical results demonstrate that the proposed scheme is superior to several competing methods in capturing faults and reducing false alarms. Meanwhile, the authors find it can detect anomaly quickly, automatically and robustly under different signal-to-noise ratios. These provide operators sufficient time to adopt an effective maintenance strategy.