GTAINet: Graph neural network-based two-stage anomaly identification for locking wire point clouds using hierarchical attentive edge convolution

With the application of 3D sensors, studies on various vision tasks based on point clouds have been explored in different fields. In the field of the high-speed train safety inspection, the vision-based artificial intelligence inspection technology has received greater attention in recent years. In this paper, we introduce a graph neural network-based (GNN) two-stage anomaly identification method (GTAINet) for locking wire point clouds. GTAINet consists of two sub-networks, a segmentation network and a classification network. The segmentation network splits bolt, wire and background points, then the segmented bolt and wire points are fed into the classification network to identify whether the sample is broken. However, as an essential type of geometric data structure, how to extract the rich geometric representation from point clouds is the key to the above vision tasks. To this end, we proposed a hierarchical attentive edge convolution (HAEConv) to establish a GNN. HAEConv is able to recover hierarchical topological information from point clouds and attentively recalibrate each edge's response, which allows the model to better capture more useful information of local and global geometric structure. Another critical point that limits the segmentation and classification performance is the lack of training data, in particular the lack of anomaly locking wire samples. To address this challenge, we propose a synthetic algorithm that can synthesize massive fake anomaly locking wire data using parameterized Be ' zier curves. Experiments demonstrate that the proposed networks based on HAEConv outperform popular existing methods on both segmentation and classification tasks. In addition, the synthetic method presented allows us to pre-train a model with very strong generalization ability, which can significantly improve model performance.