Research On Data Anomaly Detection Method Of Sensor Network Based On Graph Neural Networks

With the progress of microcomputer and communication technology,low-cost and highly sensitive sensor networks have been widely used in a variety of fields,such as meteorological monitoring,forest-fire prevention and military reconnaissance.Nevertheless,since sensors are usually deployed in places that are difficult for humans to supervise,they are vulnerable to both environmental factors and malicious human attacks.Damaged sensors may transmit abnormal data to the control center,resulting in the decision-making of control center being affected and leading to property damage and even casualties.In order to reduce the impact of abnormal data in sensor network,the research work of this paper mainly focuses on the problem of abnormal data detection in sensor networks.The specific work is summarized as follows:(1)As the sensor network is easily affected by natural factors,this paper addresses the problem of fault sensor detection in wireless sensor network(WSN)by proposing a fault detection model named GCN-GRU,which hybridizes graph convolutional network(GCN)and gate recurrent unit(GRU).The model consists of three layers: input layer,spatiotemporal processing layer and output layer.And the input layer receives the sensor network data and the graph model constructed by WSN and transmits them to the spatiotemporal processing layer.In the spatiotemporal processing layer,the spatial distribution features of WSN and the characteristics of faults in high-dimensional space are extracted by GCN,and they are constructed as the high-dimensional data of time series as the input of GRU.Then the temporal evolution features of sensor network data and the temporal and spatial evolution characteristics are extracted and fused by GRU.Finally,the fault detection results are obtained in the output layer.To evaluate the performance of the GCN-GRU,this paper compares GCN-GRU model with fault detection algorithms for WSN.Numerical experiments show that the GCN-GRU model has better fault detection performance.(2)Sensor networks in smart grid are not only affected by natural factors,but also suffer false data injection attacks(FDIA)by criminals.Different from abnormal data caused by natural factors,FDIA has the characteristics of strong concealment and interference,which poses a serious threat to the security and stability of the smart power grid.In order to detect false data injection attacks in smart grids,a graph attention network(GAT)based FDIA detection model is proposed in this paper.Firstly,we construct a graph model for the sensor network in smart grid,in which the sensors are modeled as nodes on the graph,sensor connections as edges between graph nodes,and sensor data as graph signals.Secondly,we feed the adjacency matrix and graph signals into the graph attention layer.Specifically,the graph attention layer assigns different weights to the neighbors based on their importance and updates the features of each node in the graph based on neighbor information and assigned weights.Finally,we fit and output FDIA detection results through a fully connected layer with the features extracted by the graph attention layer.According to numerical experiment results,our proposed model outperforms comparative algorithms with same condition,which demonstrates its effectiveness.