Violation behaviors of substation operators remain obstacle to power safety production. Previous work mostly relies on detecting objects such as helmets in the image to judge the behavior of substation operators, rather than extracting characteristics of substation operators’ behavior. In this work，violation behaviors are divided into two categories. One can be characterized by absence of tools, such as not wearing safety helmets and not wearing working clothes. The other violation behaviors such as falling on the ground, climbing or crossing do not have specific tools features. Thus, a violation behaviors detection strategy which can accurately identify two kinds of violation behaviors is developed by combining object detection model based on YOLOv5, pose estimation model based on HRNet and skeleton-based action recognition model based on ST-GCN. The results of experimental verification on data from a substation prove the effectiveness of the proposed strategy.