Hypertuned-YOLO for interpretable distribution power grid fault location based on EigenCAM

Abstract

Ensuring the reliability of electrical distribution networks is a pressing concern, especially given the power outages due to surface contamination on insulating components. Surface contamination can elevate surface conductivity, thereby resulting in failures that can lead to power shutdowns. Addressing this challenge, this paper proposes an approach for real-time monitoring of electrical distribution grids to prevent such incidents. A hypertuned version of the you only look once (YOLO) model is tailored for this application. We refine the model's hyperparameters by integrating a genetic algorithm to maximize its detection performance. The EigenCAM technique enhances the visual interpretability of the model's outcomes, providing operators with actionable insights for maintenance and monitoring tasks. Benchmark tests reveal that the proposed Hypertuned-YOLO outperforms Detectron (Masked R-CNN), YOLOv5, and YOLOv7 models. The Hypertuned-YOLO achieves an F1-score of 0.867 and a [email protected] of 0.922, validating its robustness and efficacy.

Keywords

• EigenCAM
• Convolutional neural networks
• You only look once
• Power grids