Robust fault detection approach for wind farms considering missing data tolerance and recovery

Abstract

The advancement in sensing technologies and infrastructure allows real-time condition monitoring on wind turbines (WTs), which helps improve the power generation efficiency, lower the maintenance costs of wind farms (WFs). Practically, the real-time measurements could be unavailable at the Supervisory Control and Data Acquisition end due to unintended events such as sensor faults and communication loss, which significantly depreciates the condition monitoring and fault detection performance. Aiming to mitigate the missing data impact on data-driven WF applications, this study develops a robust anomaly detection approach for WT fault detection using a denoising variational autoencoder. In presence of missing measurements, the proposed approach can not only sustain high fault detection performance but also recover the missing data as an auxiliary function. The proposed approach is tested on a realistic offshore WF and compared with other autoencoder variants and traditional anomaly detection methods. The testing results verify the outstanding robustness of the proposed approach against missing data events and demonstrate its great potential in missing data recovery.