HSPAN-GNN-based fault detection for power transmission lines

Abstract To supply uninterrupted electric power is extremely crucial for the economy and daily life. The traditional manual inspection method for power transmission line fault detection has many disadvantages. Although deep learning has been applied in power transmission line detection, existing schemes have deficiencies in complex environment with low-resolution images and small target. To solve this problem, this paper proposes a visual feature modeling method based on graph convolution and constructs a graph neural network model, HSPAN-GNN (High-priority Subsampling with SPD and Attention Normalization-Graph Neural Network), which combines graph convolution and convolution operations. We build a novel HSPAN-GNN model includes a high-degree priority subsampling module to balance computational efficiency, memory overhead, and accelerate inference. The proposed Space-to-Depth Convolution (SPD-Conv) solves the detection problems in small target and low-resolution scenarios, and the Normalization-based Attention Module (NAM) enhances the detection performance. Experiments have shown that the HSPAN-GNN model can achieve efficient and accurate target detection of transmission lines.