Small satellite networks (SSNs), which are constructed by large number of small satellites in low earth orbits (LEO), are considered as promising ways to provide ubiquitous Internet access. To handle stochastic Internet traffic, on-board routing is necessary in SSNs. However, large-scale, high dynamic SSN topologies and limited resources make on-board routing in SSNs face great challenges. To address this issue, we turn to graph neural network (GNN), a deep learning network inherently designed for graph data, motivated by the fact that SSNs can be naturally modeled as graphs. By exploiting GNN's topology extraction capabilities, we propose a GNN-based learning routing approach (GLR) to achieve near-optimal on-board routing with low complexity. We design high-order and low-order feature extractor and cross process to deal with high dynamic topologies of SSNs, even those topologies that have never been seen in training. Simulation results demonstrate that GLR results in a significant reduction in routing computation cost while achieves near-optimal routing performance in SSNs with different scales compared with typical existing satellite routing algorithms.

翻译：小型卫星网络是由低地球轨道上大量小型卫星建造的,被认为是提供无处不在的互联网接入的有希望的方法。为了处理随机互联网流量,SNSNS系统需要使用机载线路。但是,大规模、高动态的SNSN地形和有限的资源使SSNSS在机载路线上出现巨大的挑战。为了解决这一问题,我们转向图形神经网络(GNNN),这是一个为图表数据而设计的深层次的学习网络,这是一个深层次的网络,其动机是SNSN可以自然地以图表为模型。通过利用GNNN的地形提取能力,我们建议采用基于GNNN的学习路由方法(GLR),以实现低复杂性的近最佳的机载路由。我们设计高顺序和低序地貌提取器和交叉程序,以处理高动态的SNSNNE系统地形,甚至是在培训中从未见过的那些地形。模拟结果表明,GLRM的结果是,通过利用GNNE的地形提取能力,大大降低与SNSMAL系统相比,同时取得不同程度的卫星路路路程计算结果。