Image processing-based anomaly detection in electricity distribution systems: An application of SAM and deep learning

The rapid and accurate detection of anomalies, such as fractures, cracks, arc burns, and other defects in components like insulators, towers, and spacers within electrical distribution networks, is of critical importance for the safety of power transmission systems and the efficiency of maintenance processes. Although various methods have been developed in the literature for detecting such anomalies, existing studies are often limited in scope and primarily focus on a single component. In this study, a comprehensive analysis was conducted using the Segment Anything Model (SAM) and four different deep learning-based classifiers to identify anomalies in various components of the distribution network. The open-source dataset obtained from the Roboflow platform consists of a total of 1,539 images, including labeled data for broken insulators, arc-burned insulators, intact insulators, towers, and spacers. In the preprocessing stage, the dataset was re-annotated, expanded, and data augmentation techniques were applied. The analyses were performed using the Python programming language on the Google Colab platform, utilizing open-source libraries such as pandas, numpy, and torch. The study consists of two main stages: segmentation and classification. In the first stage, automatic segmentation of the components in the images was performed using the SAM algorithm, which achieved a segmentation accuracy of 88.55%. In the second stage, the segmentation masks generated by SAM were classified using four deep learning models: MobileNetV2, ResNet50, Vision Transformer (ViT), and VGG-16. The performance of these models was evaluated using accuracy, precision, recall, and F1-score metrics. The comparative results revealed that the MobileNetV2 model achieved the highest accuracy (84,21%) among the classifiers and demonstrated the best performance in anomaly detection within the distribution network. The findings indicate that the Segment Anything Model (SAM) provides a successful approach for anomaly detection in electrical distribution lines. Furthermore, integrating SAM with different deep learning models can enhance its generalization capability, thereby contributing to faster and more efficient maintenance processes in power distribution networks.