In order to solve the problem of low efficiency and accuracy in the detection of non-standard safety helmet worn by the existing substation patrol personnel, a lightweight substation personnel non-standard behavior detection model based on improved YOLO is proposed. The model consists of a feature extraction network, an ECA-SPP network, an ECA-PANet network and a prediction network; MobileNet V3 is used in the feature extraction network; feature maps of four scales are extracted and input into the SPP and PANet networks, and are optimized based on an attention mechanism; The effectiveness of the proposed model is verified by the data set of the detection of non-standard wearing of safety helmets in substations. The experiment results show that the proposed model mAP is a 0.8244 and FPS is a 38.06, which is obviously better than other models such as Faster RCNN, YOLOv4 and YOLOx, and has higher accuracy and faster detection speed. It can provide a reference for real-time detection of substation personnel wearing non-standard safety helmet.