| With the advent of the era of big data,massive amounts of irregular network data with a certain topology,such as social networks,transportation networks,etc.In practical applications,sensors usually cannot work stably due to limited storage space,poor working environment and limited computing power,resulting in abnormal and missing data collected by sensor networks.Moreover,with the wide application of sensors,their data-centered characteristics are becoming increasingly prominent.Whether they can effectively capture useful information from network data is the key factor to evaluate the success of their application.Therefore,anomaly detection of network data is the focus of current research.The existing anomaly detection methods mainly include: network anomaly detection based on manual operation,which is cumbersome and difficult to realize on a large scale;the data anomaly detection method based on digital signal processing(DSP)fails to make effective use of the implied correlation between sensor nodes.Based on the difficulties and shortcomings of existing research,in recent years,a network data anomaly detection method based on graph signal processing(GSP)has been proposed.Compared with the existing anomaly detection methods,graph can better describe the correlation between sensor nodes.Therefore,how to use graph signal processing method to more efficiently realize network data anomaly detection and repair is the focus of current network data research.This paper mainly studies from the following two aspects:(1)Aiming at the problem of network data anomaly detection and location,this paper proposes a dual channel malfunction detection model using graph high pass filter,which uses the corresponding relationship between network topology and network data to detect and locate abnormal data in sensor networks.The proposed algorithm not only improves the detection accuracy compared with the existing frequency domain anomaly monitoring methods based on graph Fourier transform,but also can locate the abnormal nodes location.Firstly,the current data to be detected and the historical data are directly input into the two graph high pass filters in the vertex domain(spatial domain),and the decision threshold is used to detect whether there is a high-frequency component in the filtered output result.If a high-frequency component is detected,it is determined to be abnormal,and the search rules are further used to determine the location of the abnormal value.Simulation results show that compared with the graph frequency domain anomaly detection method,the proposed anomaly detection algorithm not only improves the accuracy of anomaly detection,but also locates the position of anomaly nodes.(2)Aiming at the problem of repairing network abnormal data,a data repair algorithm based on multi shift operator is proposed in this paper.The network data is modeled as a time-varying graph,and the multi shift operator composed of Kronecker product is used to characterize the data relevance,and the data relevance in time and space dimensions is fused,so as to repair the data of abnormal nodes by using the information of network nodes in multiple dimensions.In order to solve the optimization problem of data repair more efficiently,a fusion solution method based on conjugate gradient descent(CGD)is designed.The simulation results show that the CGD fusion method has fewer iterations and higher efficiency than the gradient descent method(GD).Moreover,the abnormal data repair algorithm proposed in this paper has better data repair performance than the existing abnormal data repair algorithm based on Cartesian product graph model. |
