Substation Anomaly State Detection Network Based on Asymptotic Feature Fusion

With the continuous development of the power grid, it is becoming increasingly important to enhance the on-site risk management and control capabilities in substation. In order to address the problem of low intelligence in the safety supervision, this paper proposes an end-to-end substation anomaly state detection network called AFFNet (Asymptotic Feature Fusion Network) based on YOLOv7. To address the problem of diverse types and insufficient feature extraction capabilities, the S-ELAN network is designed to learn more features by controlling the shortest and longest gradient paths. In order to avoid the information loss caused from the significant semantic gap between non-adjacent features, a progressive feature fusion strategy is designed to avoid the issue. Furthermore, an adaptive spatial fusion method is utilized to reduce the inconsistency of feature information at each spatial position. In addition, considering the problem of imbalanced positive and negative samples in the dataset, the GHM (Gradient Harmonizing Mechanism) loss function is introduced to adaptively adjust the weight information of different samples with the gradient information. The AFFNet is tested on the substation anomaly state dataset, and the overall accuracy of the proposed model reaches 82.06%, which is significantly better than existing one-stage detection networks. The result demonstrates that the AFFNet model can accurately identify substation anomaly states and improve the risk investigation and protection capabilities of substations. And the ablation results verify that the asymptotic feature fusion is effective for the recognition of the substation anomaly states.