The use of artificial intelligence algorithms to achieve automatic control of power operation safety is of great significance for safety production. However, the existing automated detection methods for violations by power operators have low accuracy and cannot accurately identify complex violations. To address the characteristics of violation images in power construction sites, existing algorithms were optimized and an improved YOLOv8 algorithm based violation recognition method was proposed. A channel attention mechanism was added to the Backbone network to increase the ability to filter low-level features, and a CBAM attention mechanism was added to the Head network to increase the ability to filter high-level features. At the same time, the CIoU Loss cost function of YOLOv8 was replaced with EIoU Loss to increase the network’s ability to express violation features. The feasibility of the algorithm improvement was confirmed through experimental verification using illegal image samples collected from the construction site of power transmission and transformation operations. The mAP index increased by up to 4.77%, reaching a practical level.