Incremental Learning Method for Wind Turbine Fault Detection Models Considering False Negatives

Fault detection (FD) algorithms based on supervisory control and data acquisition (SCADA) data have been widely used in the operation and maintenance (O&M) of wind turbines (WTs). However, the performance of FD models will degrade due to the time-varying operating conditions (TVOC). Incremental learning (IL) methods can be employed to update the models online to adapt to TVOC, but the error accumulation caused by false negatives will lead to a continuous decrease in the FD rate (FDR). A novel IL method considering false negatives is proposed to improve the performance of WT FD models. First, a data buffer is built to cache some new normal data used in model updating. Second, a processing strategy for false negatives is proposed to block some high-risk data from being added to the buffer, thereby weakening the error accumulation. Third, when the buffer is full, update the FD model using the data in it. A real-world SCADA dataset with gearbox faults and four different FD algorithms with various model updating strategies are used in the experiments. The results demonstrate that the proposed method can lower the false alarm rate (FAR) of all four FD algorithms. After processing false negatives, for example, for the FD model based on Gaussian kernel regression, its FDR increased by about 11%, and for the FD model based on multivariate state estimation technique, its FDR increased by about 14%. The results of hyperparameter experiments and early stopping experiments show that the proposed method has good potential for practical applications.

I. Introduction

The demand for new technologies in the operation and maintenance (O&M) of wind turbines (WTs) has grown in recent years. It is crucial to monitor the conditions of WT by analyzing the operational data to detect incipient faults and reduce O&M costs [1]. Methods based on supervisory control and data acquisition (SCADA) data have received a lot of attention for condition monitoring of WT [2], [3], [4]. However, there is an imbalance problem with SCADA data [5], meaning that there is much less fault data than normal data. The reasons for this include WT not being allowed to operate for a long time under fault conditions [6], inefficient manual labeling, and so on. Therefore, fault detection (FD) algorithms based on normal behavior modeling [7], [8] have been widely used. These FD models are trained using only normal data and can detect faults by analyzing the differences between current and normal conditions.