Adaptive pruning techniques in YOLOv5 for industrial training on equipment operation

In the context of power grid operation, the detection of equipment status is of crucial importance for ensuring the stability and safety of the power system. However, the detection of power grid equipment confronts several challenges, such as the similarity in the morphology of electrical equipment, significant variations in the scale of target objects, and the balance between detection speed and accuracy. These issues have always been key research topics in this field. Based on the pruning strategy and dynamic attention mechanism, this paper proposes an efficient and lightweight object detector, named YOLO-AKD. The outcome is a new strategy that can significantly enhance the inference speed of real-time object detectors while maintaining accuracy. To verify the effectiveness of our strategy, we have created a self-made dataset of power grid switch cabinets, which contains the status of 12 types of electrical equipment. At the same time, a network architecture called YOLO-AKD has been established. We trained our YOLO-AKD from scratch on the self-made dataset of power grid switch cabinets without relying on the pre-trained weights of any other large-scale datasets. Through experimental comparisons with the current state-of-the-art real-time object detectors, including YOLOv6 and RTMDet. Taking RTMDet as an example, in the switch cabinet dataset, our method YOLO-AKD improves the [email protected] by 0.2%, and at the same time, the computational cost of FLOPs decreases from 14.8 FLOPs to 7.9 FLOPs, fulfilling the objective of auxiliary teaching in power grid operation training.