Contrastive Learning for Anomaly Detection in Power Distribution Monitoring Images Using Text- Induced Image Augmentation

Ensuring the reliable operation of power distribution systems is critical, and anomaly detection within these systems plays a vital role in maintaining their stability. However, this task is challenging due to the limited availability of labeled anomaly samples and the significant variability in normal conditions caused by environmental factors such as lighting, weather, and environmental interferences. In this paper, we propose an innovative approach that leverages contrastive learning, augmented with text-induced image generation, to enhance anomaly detection capabilities. Our method integrates traditional image augmentation techniques, like noise addition and rotation, with advanced text-guided editing. Specifically, “assimilation text” is used to generate varied normal images that simulate different benign conditions, while “dissimilation text” creates synthetic anomalies, reflecting potential faults like transformer leakage or arc discharge. This dual augmentation strategy not only balances the dataset by increasing the diversity of both normal and abnormal samples but also improves the model’s robustness in distinguishing between normal and anomalous conditions in complex monitoring environments. The experimental results demonstrate that our approach significantly enhances anomaly detection performance, offering a reliable solution for real-time monitoring and fault detection in power distribution networks.