In modern energy systems, substations are the core of electricity transmission and distribution. However, similar appearance and partial occlusion pose significant challenges for automatic identification of electrical devices, and the percentage of oscillation suppression presents a gradient distribution with a range of 30–70%. To address these issues, we collect and annotate the sub-station multi-device dataset (SMDD) in Pascal VOC format with the LabelImg tool, which includes 3101 various devices images from three high-voltage substations. The PyTorch framework in Python is used to construct the main structure of the neural network. Further, feature reassignment and probability weight network (FRPWNet) are constructed based on the classic structure pattern of backbone-neck-head. Specifically, considering the similar appearance of electrical devices, the feature reassignment module (FRA) is designed to enhance the expression of feature in-formation and aggregate the global feature. To address the problem of occlusion between electrical devices, a Probability Weight Assignment (PWA) method is proposed to strengthen the model's focus on the center region of the occluded devices. In the experiment, the hook tool is designed and used for visualizing the internal data of neural networks, and some ablation experiments are conducted on the SMDD dataset. FRPWNet achieves the best detection performance compared with classical object detection networks. Also, it is verified that the global feature from FRA can effectively recognize the similar and small devices. • A context information fusion method is designed to improve device detection performance. • A feature reassignment module was designed to track small-size devices. • A feature weight method was proposed to focus on the margin of occluded devices.
Li et al. (Fri,) studied this question.