Research On WSN Abnormal Node Detection Method Based On Graph Neural Network

Wireless Sensor Networks(WSNs)are the wireless ad hoc networks densely deployed in the monitoring area with a collection of multi-hop connected sensor nodes,however,attacks from the outside or the unreliability of the sensor itself may cause the sensor to produce false readings and lead to abnormalities,so anomaly detection technology is one of the key technologies to ensure the normal and smooth operation of wireless sensor networks.Wireless sensor networks not only have massive multimodal observations of nodes,but also have topological information reflected by spatial locations,so they can be modeled into many graph structure data.Graph neural network is a popular deep learning technology that widely used in graph structure applications,it models graph from both network topology and node attributes,and extracts interesting features from node specific attributes based on the edges in the graph structure to perform graph representation learning or downstream tasks.The WSN data features consists of correlation information between node locations,correlation information between modes and correlation information between different moments,but the existing methods do not comprehensively consider these information during anomaly detection,e.g.,some of them only consider multi-node single-modal scenarios,while the rest of them only consider single-node multimodal scenarios,which limits the performance of anomaly detection.At the same time,with the development of wireless communication technology and hardware equipment,the number of nodes and observations of WSN continues to grow,and due to the complex dependence between nodes,the computing cost and hardware requirements of large-scale WSN increase exponentially with the increase of the number of layers of graph neural networks.More importantly,the output of the existing neural network model based on unsupervised anomaly detection methods is often not directly related to the anomaly detection judgment results,and the anomaly score is calculated by reconstructing the error and prediction error to determine the system state.In view of the above problems and challenges,this paper studies the WSN anomaly node detection method based on graph neural network,the main work and innovations are as follows:1)Compared with the existing anomaly detection methods that only consider the multimodal time series data of a single node or the multinode time series data of a single mode,which fail to comprehensively consider the spatial and temporal features of multiple sensor nodes,the anomaly detection framework designed in this paper considers the temporal features of multimodal multinode data flows and the spatial features between multinode locations.The proposed anomaly detection method for a WSN data flow based on a dynamic GNN is organised with temporal features on a single-mode extraction module,a multimodal correlation feature extraction module,and spatial correlation features between the sensor node extraction modules.Graph attention mechanisms are used in all modules,finally,the above features are combined to predict the time series data of WSN nodes and identify abnormal states.2)In the case when a large-scale WSN faces a large number of sensor nodes,this paper divides all nodes into multiple clusters,the data flows of different clusters are trained and inferred on different anchor node devices,and dimensionality reduction is also carried out to reduce the model operation frequency,which can effectively reduce the hardware overhead and time consumption of the network.At the same time,compared with the existing anomaly detection framework based on autoencoder reconstruction,this paper combines the efficient and regular data feature capture capacity of the reconstruction-based model and the graph representation generalization ability improvement brought by selfsupervised learning,proposes a novel anomaly detection framework based on the adaptive fusion of node spatial features and the data flow temporal features of the nodes in the Global and Local spaces via Self-supervised Learning.