Global Attention-Based Approach for Substation Devices Classification and Localization
Zhimin Guo, Yalin Li, Yangyang Tian, Hao Liu, Shaoguang Yuan, Chunyu Hou · 2023
Accurately identifying various electrical devices is crucial for the safe and stable operation of substations. However, the size of devices in substations varies, and some devices have similar appearances, which often leads to inspection errors. To address these issues, we designed the path aggregation network with global attention mechanism (GAM-PAN). Considering the scale variation of electrical devices in images, the path aggregation module is introduced to fuse the multi-scale features. To address the problem of recognizing similar devices, global attention mechanism (GAM) is s embedded into the feature aggregation module. The GAM enhances the fusion features of similar devices and reconstructs the fused feature maps. In addition, a transfer learning strategy is adapted to train GAM-PAN, which speeds up the convergence rate of the network. GAM-PAN is tested on a substation multi-device dataset (SMDD) and achieves a detection accuracy of 84.32%, compared to four classical object detection methods. Furthermore, it was verified that global attention can effectively solve the problem of recognizing similar targets.