An Improved Anomaly Detection Model for Automated Inspection of Power Line Insulators

TL;DR

This work tackles incipient insulator faults under data-scarce and imbalanced conditions by separating detection and anomaly assessment into a two-stage pipeline. It adopts an explainable one-class classifier (FCDD) and introduces a BCE-aligned, modified loss that can be combined with focal loss to enhance pixel-level anomaly localization. Experiments on MVTec-AD and insulator datasets (IDID and SG) show that the modified FCDD improves average and worst-case performance, with semi-supervised training yielding strong gains from a small number of real anomalies. The approach demonstrates practical viability for low-label regimes and offers interpretable heatmaps to aid industry adoption, while highlighting domain-shift and explainability limitations as areas for further work.

Abstract

Inspection of insulators is important to ensure reliable operation of the power system. Deep learning is being increasingly exploited to automate the inspection process by leveraging object detection models to analyse aerial images captured by drones. A purely object detection-based approach, however, suffers from class imbalance-induced poor performance, which can be accentuated for infrequent and hard-to-detect incipient faults. This article proposes the use of anomaly detection along with object detection in a two-stage approach for incipient fault detection in a data-efficient manner. An explainable convolutional one-class classifier is adopted for anomaly detection. The one-class formulation reduces the reliance on plentifully available images of faulty insulators, while the explainability of the model is expected to promote adoption by the industry. A modified loss function is developed that addresses computational and interpretability issues with the existing model, also allowing for the integration of other losses. The superiority of the novel loss function is demonstrated with MVTec-AD dataset. The models are trained for insulator inspection with two datasets -- representing data-abundant and data-scarce scenarios -- in unsupervised and semi-supervised settings. The results suggest that including as few as five real anomalies in the training dataset significantly improves the model's performance and enables reliable detection of rarely occurring incipient faults in insulators.

Figures (6)

• Figure 1: Two-stage insulator fault detection with object detection and anomaly detection. The texts in parentheses represent the model used in this work.
• Figure 2: Aerial images of insulators and extracted images of disks. The extracted images contain very little background compared to aerial images.
• Figure 3: Sample disks (healthy, flashed and broken), and their corresponding ground truth maps for semi-supervised training.
• Figure 4: AUC of FCDD with unsupervised and semi-supervised training for two insulator inspection datasets (data abundance and data scarcity). The proposed semi-supervised model uses the compound modified FCDD-focal loss ($\gamma=0.4$ for data abundance $\gamma=1$ for data scarcity).
• Figure 5: AUC of unsupervised and semi-supervised (modified FCDD + focal loss) models for SG dataset trained with different numbers of training anomalies.