The Research Of Missing Data Reconstruction Methods In Wireless Sensor Networks

Wireless sensor networks have been widely used in various fields,such as environmental monitoring,agricultural monitoring and aviation and so on.Sensor nodes are usually deployed in harsh and directly exposed environment,due to the hardware failure,the limited energy,channel interference,or human caused factors,the receiving sensor node is unable to obtain accurate data of other sensor nodes,those sensory data are often abnormal,or even missing,and the larger WSN scale is,the more data may be lost.So,how to reconstruct the large number of loss data is the key research problem.Therefore,it’s a significant meaningful problem to construct themissing data reconstruction model,especially the large-scale missing data reconstruction model to ensure the minimum error,and to meet the characteristics of sensory data.This paper is around the abnormal data detection,especially in-depth study on how to reconstruct the missing data in wireless sensor networks,the main research contents are as follows:The characteristics of sensory data are analyzed firstly,including:（1）The spatio-temporal correlation model is constructed.The sensory data correlations of different time intervals of single node are analyzed,and the spatial correlation is also analyzed of spatially adjacent nodes,that is the similarity of sensory data between different sensor nodes.（2）The sparse characteristic of sensory data is analyzed.According to the definition of sparse,a measurement matrix is constructed of n nodes in the sensory data obtained by t time periods,the matrix is not strictly sparse,but the majority of sensory data has the smooth change characteristic,so an appropriate sparse matrix is selected to transform the measurement matrix to obtain the sparsity.（3）The low rank characteristic of sensory data is analyzed.After the measurement matrix is singular value decomposited,the main components of the measurement matrix are concentrated on a small number of singular values,other singular values can be approximated to 0,which proves the measurement matrix having low rank characteristic.Based on the spatial-temporal correlation,the random missing data reconstruction method based on fuzzy distance is put forward to reconstruct the random lack of sensory data effectively,the main research contents are as follows:（1）A multiple linear regression missing value reconstruction method based on time correlation is studied,which uses historical data to estimate the current data of one node.The method is only suitable for smoothly changing sensory data within a certain time period,but not suitable for the sudden jump sensory data influenced by some special factors or the data acquisition time interval is too long.（2）A random point missing values reconstruction method based on fuzzy distance is proposed,which calculates the Euclidean distance between the node to be reconstructed and its neighbor nodes to construct a T-S fuzzy logic model.And the correlation coefficient is calculated according to the Euclidean distance and the spatial correlation,then the estimation value is to be reconstructed.（3）A point random missing value reconstruction method based on spatio-tempral correlation is put forward,which combines the above two methods,not only can handle smoothly changing data,and can also handle non-stationary changing data.In order to solve the case of large-scale random missing in measurement environment,a reconstruction model based on singular value decomposition is proposed.The main contents are as follows:（1）A perception environment modeling method is proposed.Environmental matrix,missing rate matrix,measurement matrix,reconstruction matrix,and the equality relationships between them are defined.（2）A reconstruction algorithm based on singular value decomposition is proposed.Due to the environmental matrix and the reconstruction matrix has low rank characteristic,so the proposed algorithm is to solve reconstruction matrix with a low rank in the case of a known measurement matrix and different missing rate,which has a better performance especially in large scale missing of sensor data.（3）In order to improve the accuracy of the reconstruction matrix,a new reconstruction algorithm for large scale missing data based on the spatio-temporal correlation is put forward.Experiments show that the reconstruction error of the algorithm is reduced.If all the sensory data collected from the monitored area in the same time is built into a vector data,that is data frames,the sensor data collected at different time frames can be combined into a measurement matrix.In order to study how to use the history frames to reconstruct the current frame,a missing data reconstruction method based on dictionary learning is proposed,the main contents of the method are as follows:（1）A mathematical representation of continuous data frames is proposed.Time correlation between different data frames is calculated to prove that the time interval becomes smaller,the temporal correlation between data frames is larger,that is more closer the current data frame and its historical data frame is,more similar is.So,it’s feasible to choose the appropriate historical data frame to estimate the current frame.（2）The sensor data reconstruction method based on dictionary learning is put forward.The method adopts the K-SVD dictionary learning algorithm to train the data within a certain time period to get the basis function dictionary,and based on the dictionary,the ₁l minimum norm method is used to sparse coding,finally,according to the dictionary and sparse coefficient,the reconstructed sensory data is obtained.（3）An adaptive dictionary update algorithm is put forward.The sensory data of wireless sensor networks is real-time,continuous data stream,if the missing data is reconstructed based on the initial fixed dictionary,which doesn’t meet the stream characteristic of sensory data,at the same time if each reconstructed data frame is added to the dictionary as training data set,the algorithm performance is too low.So,the adaptive dictionary update algorithm is presented.When the correlation between the current reconstructed data frame and the greatest weight data frame of the dictionary training set is less than a pre-set threshold,the training data set is updated,so as to update the dictionary.The missing data reconstructed based on updated dictionary obtained higher accuracy and smaller error.At the last,the data quality of the wireless sensor networks is studied.The two most important dimensions of data quality,i.e.the completeness and correctness are analyzed by collecting the sensory data of the actual lab environment.At the same time,the algorithms presented in this study are verified and analyzed use these sensory data.