Improving the Stability of Multi-hop IoT Communication Paths Using Graph Neural Networks

This paper aims to solve the problem of path stability degradation caused by topological dynamics and link quality fluctuations in multi-hop IoT (Internet of Things) communications. It introduces Graphormer's global self-attention and structural encoding methods to efficiently model global and local information of the network and improve the stability of the communication path. It then constructs node features and edge features, designs the Graphormer module, introduces the shortest path distance bias, degree centrality encoding, and channel quality bias to enhance global topological perception, and uses sliding time window dynamic graph input to capture network evolution. The experiments are conducted on the OMNeT++ platform, covering 50–400 node scales and multiple mobility models. The results show that when the number of nodes increases from 50 to 400, the link availability decreases and the packet loss rate increases to a higher level. In terms of path life, Graphormer reaches 29.5 seconds in the Random Direction scenario, which is significantly better than DGI (Deep Graph Infomax) (22.8 seconds) and GAT (Graph Attention Network) (18.9 seconds), and reaches 37.8 seconds in Reference Point Group Mobility, which is better than AODV (Ad hoc On-Demand Distance Vector) (19.5 seconds) and RPL (Routing Protocol for Low-Power and Lossy Networks) (21.0 seconds). In the delay jitter comparison, Graphormer's end-to-end delay fluctuation range is smaller, significantly lower than other models. Graphormer effectively improves path stability through global attention and dynamic graph modeling.