High-performance fault diagnosis of smart grid insulators using YOLOv11-based instance segmentation for real-time energy infrastructure monitoring

Ensuring the reliability and stability of smart grids heavily depends on the condition of insulators and their structural bindings. Faults such as cracks, contamination, and binding failures can lead to severe consequences, including power outages, equipment degradation, and increased maintenance costs. Traditional fault detection methods, which rely on manual inspection and basic image-processing techniques, are often inefficient, labor-intensive, and prone to errors. To address these challenges, this study presents an advanced deep-learning-based fault diagnosis system utilizing YOLOv11 with instance segmentation for precise and real-time detection of insulators and binding defects in smart grids. A comprehensive dataset of 1320 high-resolution images of insulators and bindings was created, encompassing diverse environmental conditions and labeled with pixel-wise segmentation masks for precise fault localization. The YOLOv11 model was trained and validated on this dataset using advanced data augmentation techniques such as rotation, flipping, and brightness variations to enhance robustness. The proposed YOLOv11-based system achieved a mean average precision (mAP) of 84.1% at an IoU threshold of 0.5, outperforming YOLOv8 (75.4%), YOLOv5 (72.5%), faster R-CNN (79.5%), and mask R-CNN (82.1%). Additionally, the model operates at an inference speed exceeding 200 frames per second (FPS), making it highly suitable for real-time deployment on edge computing devices, unmanned aerial vehicles, supervisory control and data acquisition systems, and smart grid monitoring stations. Unlike previous deep-learning approaches that primarily rely on bounding box detection, the proposed system integrates instance segmentation, allowing for pixel-level identification of defect regions. This enables more accurate fault assessment, reduces false positives, and enhances predictive maintenance strategies. The high detection speed and accuracy make this model a viable solution for automated smart grid fault diagnostics, reducing reliance on manual inspections and mitigating potential failures before they occur. Future research will focus on optimizing the model for low-power edge computing devices, incorporating multi-sensor fusion techniques (thermal, LiDAR, infrared), and extending its capabilities to detect a broader range of smart grid faults. By bridging the gap between high accuracy, real-time performance, and deployability, this work establishes a new benchmark for AI-driven fault detection in smart grids.