XAI-guided Insulator Anomaly Detection for Imbalanced Datasets

TL;DR

This work tackles insulator defect detection in UAV imagery under severe class imbalance and motion blur. It introduces a two-stage YOLOv8-based pipeline to detect insulators and shells, followed by a reweighted shell classifier trained through undersampling and logistic regression, plus Layer-wise Relevance Propagation for pixel-level damage localization and a sharpness-based image-quality filter. The approach yields per-class improvements (e.g., up to $12\%$ for Broken and $7\%$ for Flash) and high detection accuracy ($mAP_{50}=0.97$ for insulators and $mAP_{50}=0.98$ for shells), with interpretable heatmaps confirming localization quality; macro-averaged accuracy reaches $96.8\%$ at an empirically chosen sharpness threshold. The method addresses class imbalance and provides a transferable framework for explainable, high-precision industrial inspection in power-grid maintenance.

Abstract

Power grids serve as a vital component in numerous industries, seamlessly delivering electrical energy to industrial processes and technologies, making their safe and reliable operation indispensable. However, powerlines can be hard to inspect due to difficult terrain or harsh climatic conditions. Therefore, unmanned aerial vehicles are increasingly deployed to inspect powerlines, resulting in a substantial stream of visual data which requires swift and accurate processing. Deep learning methods have become widely popular for this task, proving to be a valuable asset in fault detection. In particular, the detection of insulator defects is crucial for predicting powerline failures, since their malfunction can lead to transmission disruptions. It is therefore of great interest to continuously maintain and rigorously inspect insulator components. In this work we propose a novel pipeline to tackle this task. We utilize state-of-the-art object detection to detect and subsequently classify individual insulator anomalies. Our approach addresses dataset challenges such as imbalance and motion-blurred images through a fine-tuning methodology which allows us to alter the classification focus of the model by increasing the classification accuracy of anomalous insulators. In addition, we employ explainable-AI tools for precise localization and explanation of anomalies. This proposed method contributes to the field of anomaly detection, particularly vision-based industrial inspection and predictive maintenance. We significantly improve defect detection accuracy by up to 13%, while also offering a detailed analysis of model mis-classifications and localization quality, showcasing the potential of our method on real-world data.

Figures (5)

• Figure 1: Process diagram showing our pipeline operating on an example insulator image. From left to right: A An input insulator image is processed by applying a YOLOv8 network $\mathcal{N}_I$ for whole insulator detection. Subsequently we perform a cropping operation $\mathcal{C}$ based on the predicted bounding box. B The individual shells are detected, via a YOLOv8 model $\mathcal{N}_s$ . The individual shells are cropped based on the predicted boxes via the cropping operator $\mathcal{C}$. C We perform classification using $\mathcal{N}_c$. D Heatmaps are generated for shells classified as "Damaged" using Layerwise Relevance Propagation, where the flow of relevance is indicated by the red arrows.
• Figure 2: Detection results on a sample of insulators from the test dataset using the training YOLOv8 model. One can see that insulators can be detected in a variety of backgrounds. Also multiple detection's are possible, including different colors, brightness and orientation.
• Figure 3: Detection results on a sample of shells from the test dataset. One can see that shells are detected in a variety of environments, orientations and colors.
• Figure 4: The results of our method showing sample heatmaps of shells with different damage types. The first row contains original images of damaged shells. Each row corresponds to a different architecture, comparing their performance. Darker pixels with a shift towards the red color spectrum indicate higher relevance.
• Figure 5: Sharpness versus prediction for the test dataset. We vary the threshold for the sharpness cutoff and plot the accuracy of each individual class accordingly.