Realization And Optimization Of Graph Learning In Hardware Sensor Network

With the advancement of sensor technology and signal processing,the applications of sensor network in social production and life have developed rapidly.The data analysis and processing tasks in sensor network usually involve a large amount of structured data,where the structure carries key information about the nature of the data.In practical applications,the topology of sensor network is significantly different from the structure constructed by the geometric distance between the hardware nodes.Therefore,in order to effectively represent,process and analyze the data in the sensor network,the potential structure between the nodes must be estimated properly.In this case,a key task is to infer a graph topology that can describe the observed characteristics of the data in the sensor network,thereby capturing the potential relationships between the nodes in the network.This master thesis studies the graph learning in sensor networks,and its application in data reconstruction and recovery.Due to the failure of the sensor node and the interference in signal transmission,the node data will be contaminated by noise or even lost.In view of this situation,this paper studies the problem of graph representation learning of sensor networks under the condition of incomplete sampling with the help of tools and methods in graph signal processing.Based on the global smoothness of graph signals in topology,this thesis proposes two different synchronization frameworks for joint sensor network data repair and graph representation learning for two different graph signal smooth representation models.These two methods can complete the inference of the topology structure when the node data is incomplete,and the results obtained by the topology structure inference can capture the hidden patterns between the sensor node data.In addition,this method can synchronously recover the missing data in the node in the process of completing the graph representation learning task,thereby achieving joint optimization of node data recovery and graph representation learning.Finally,simulations and experiments verify the effectiveness of the proposed method in processing sensor network data.In addition,the proposed graph learning framework is applied to the sensor network monitoring system,where we reconstruct the data of unsampled monitoring points while capturing the topological relationship between sensor nodes.