Power-Segmenter: A Multi-Head Semantic Segmentation for Power System

This paper addresses the limitations of traditional computer vision technology and single model training methods in power system status detection and defect recognition. While computer vision technology has greatly enhanced equipment inspection and defect detection capabilities, the current approach of employing single model training for singular tasks has been proven to be inefficient and resource-intensive, leading to suboptimal utilization of data, limited adaptability, and high maintenance costs. This paper introduces a novel multi-task semantic segmentation model for power system condition monitoring, known as Power-Segmenter. This model streamlines multiple tasks into one model and effectively utilizes original feature learning, allowing for shared knowledge and information distillation across tasks, thereby improving data utilization and enhancing model generalization capabilities. From a computational perspective, multi-task learning optimizes efficiency by reducing redundant operations and mitigating the problem of training set imbalance in mixed training. Moreover, the application of this model is highly significant to power systems, as it eliminates redundancy in independent training, reducing the difficulty and cost associated with staff training and facilitating widespread implementation across numerous inspection and operational centers. The findings from this study underscore the potential of multi-task learning in improving the performance and efficiency of computer vision technology in power system status detection and defect recognition, marking a significant attempt in power system safety and stability condition monitoring.