Condition Monitoring and Anomaly Detection of Wind Turbines Using Temporal Convolutional Informer and Robust Dynamic Mahalanobis Distance

Effective condition monitoring (CM) of wind turbine (WT) is crucial in detecting potential faults and developing preventive maintenance strategies. However, the frequent false alarms and missing alarms decrease the reliability of the WT monitoring system, increasing downtime and replacement costs. Therefore, this article proposes a novel semi-supervised framework for CM and anomaly detection of WT. It only requires the normal data from supervisory control and data acquisition (SCADA) to avoid the negative impact of the imbalanced data. The proposed model is composed of a temporal convolutional informer (TCinformer) and a robust dynamic Mahalanobis distance (RDMD). TCinformer can extract the global long-term features for precise data reconstruction from spatial–temporal features by the TC-based module. RDMD can consider the dynamic correlation and the robustness of the samples to reduce the fluctuations of the conditional indexes (CIs). First, TCinformer is applied to reconstruct the data of the objective variables. Then, RDMD is applied to acquire CIs of WT based on reconstructed errors. Finally, the delay perception (DP) strategy is used to determine the threshold to reduce false alarms and missing alarms based on the initial threshold of RDMD. The experiment results demonstrate the
${\boldsymbol{F}}_{1}$
score and accuracy of the proposed model achieve {0.970, 0.951} and {0.924, 0.921} in two datasets, respectively, which outperforms other state-of-the-art methods in CM and anomaly detection.