Abstract:

Large-scale multi-hop wireless IoT networks are increasingly being deployed to support critical applications with stringent QoS requirements. Robust fault detection in such networks is vital for ensuring normal network operations and enabling proactive maintenance especially in adversarial environments. However, the existing researches primarily focus on data- driven attack detection, leaving natural network fault detection largely unexplored. Due to the absence of dedicated routers, multi-hop IoT networks require network nodes to transmit their own data and also relay data for others, which introduces inherent and cascaded anomalies, making it difficult to locate anomaly sources. Furthermore, the underlying CSMA/CA based communication protocols adopted in IoT networks incur un- predictable delays in transmissions, thereby complicating timely anomaly detection and accurate diagnosis. This paper proposes innovative network fault detection technologies tailored for large- scale multi-hop IoT networks. We introduce a spatio-temporal graph neural network (STGNN) model capable of accurately detecting both node-level and edge-level faults. Our model takes both temporal data and metadata for context-aware forecasting and reconstruction. To address the lack of labeled data, we adopt an unsupervised learning paradigm. We evaluated our model on various large-scale network topologies and transaction log datasets, and results demonstrate that our model consistently outperforms baseline models across most cases.