Anomaly detection for automated inspection of power line insulators

Inspection of insulators is important to ensure reliable operation of the power system. Deep learning has recently been explored to automate the inspection process by leveraging aerial images captured by drones along with powerful object detection models. However, a purely object detection-based approach exhibits class imbalance-induced poor detection accuracy for faulty insulators, especially for incipient faults. In order to address this issue in a data-efficient manner, this article proposes a two-stage approach that leverages object detection in conjunction with anomaly detection to reliably detect faults in insulators. The article adopts an explainable deep neural network-based one-class classifier for anomaly detection, that reduces the reliance on plentifully available images of faulty insulators, that might be difficult to obtain in real-life applications. The anomaly detection model is trained with two datasets -- representing data abundant and data scarce scenarios -- in unsupervised and semi-supervised manner. The results suggest that including as few as six real anomalies in the training dataset significantly improves the performance of the model, and enables reliable detection of rarely occurring faults in insulators. An analysis of the explanations provided by the anomaly detection model reveals that the model is able to accurately identify faulty regions on the insulator disks, while also exhibiting some false predictions.