Sensor Data Estimation Method For Wireless Sensor Networks

Recently, with the rapid development of sensor technology, wireless communication andcomputer techniques, the Wireless Sensor Networks (WSN) have become the evolution trendand hot research spot in sensor and control field. Great attention has been paid by the military,industry and academe field all over the world. Many researchers gave abundant findings aboutwireless sensor from both fundamental theory and application. At the same time, WSN isbeing widely applied in many fields like military, industry, agriculture, environmentmonitoring, smart traffic, smart home etc. Among these wireless sensor networks applicationsystems, the resource of low-cost wireless sensor node is constrained, i.e., capacity ofcomputing, storage, wireless communication distance and energy is very limited, and also thewireless sensor nodes are easily affected by noise, interference and surrounding environment.So during the wireless transmission, the sensor data missing often occurs, and this phenomenais worse in some special environments. This problem has become a big challenge for dataprocessing methods. And the sensor data estimation is effective way to solve this problem.Also it is a powerful tool to support data inquiry, data aggregating, energy-savingtransmission and early warning mechanism.For the estimation of missing sensor data in wireless sensor networks, many researchershome and abroad have been doing a lot of research works and getting certian research results.But there still exist some important problems needed to be resolved. For example, thecharacteristics of the sensor data are not fully investigated and made use to estimate thesensor data, which lead to high computing complexity; The estimation accuracy is very lowwith high complexity; The estimation problem of the uncertain sensor data in a local field isnot considered; The dynamic data module of sensor data stream is not fully considered. Forthese problems, special research works are done in this paper, and the research results aregotten as follow:(1) For the problem of most of sensor data estimation methods did not consider thecharacteristics of sensor data, whichi lead to high computational complexity, we propose acorrelation analysis-based estimation framework of sensor data. For the problem of highcomputational complexirity of SVR (Support Vector Regression)–based estimation method, based on the framework, we propose correlation analysis-based LS-SVR (Least SquareSupport Vector Regression) sensor data estimation method called CALS-SVR (CorrelationAnalysis-based LS-SVR). In this method, we consider the characteristics of the sensor data ofwireless sensor networks, and extract the most correlated sensor variable to be used as theinput of modeling and estimation. And also we adopt the LS-SVR with low computationalcomplexity to estimate the sensor data. So the estimation efficiency can be improved largely.The experiments results show that the proposed CALS-SVR has better estimation efficiencyand higher estimation accuracy compared to present sensor estimation method based SVR andLS-SVR.(2) For the problem of high computational complexity and low estimation accuracy inexisting sensor estimation method, and also considering the implementation problem ofsensor data estimation method on the wireless sensor node of WSN. We propose thecorrelation analysis-based multiple linear regression sensor data estimation method calledCA-MLR (Correlation Analysis-based Multiple Linear Regression). In this method, weexplore the characteristics of sensor data of wireless sensor networks, and take correlationanalysis on them. And then combine with multiple regression, which has low computationalcomplexity. So the estimation efficiency can be improved largely. The experiments resultsshow that the proposed method has better estimation efficiency and accuracy, so it is verysuitable for applying on the wireless sensor node.(3) For the problem of uncertain sensor data processing, we propose an uncertain sensordata processing framework based on MVPCA (Multiple variable Principle ComponentAnalysis). And for the problem of uncertain sensor data estimation, we propose an uncertaindata estimation method based the framework. In this method, we use the MVPCA to extractthe intrinsic feature of the uncertain sensor data. in this way, most of the uncertainty can beeliminated. Then we adopt the multiple regression based on correlation analysis to estimationthe sensor data. So the uncertain sensor data can be estimated. The experiments results showthat the proposed method can estimate the uncertain sensor data efficiently with highefficiency.(4) For the problem of estimation mode updated not in time and big estimation error ofdynamic stream sensor data, we propose a sensor data stream estimation method, which iscalled KF-CAMLR (Kalman Filter-Correlation Analysis-based Multiple Linear Regression). In this method, the Kalman Filter is combined with multiple regression to estimate dynamicsensor data stream. The Kalman filter adjusts its working states according the estimation error,and at the same time adjusts the model parameters of the multiple regression, so theestimation model adjusts efficiently according to data model in the sensor stream. Theestimation accuracy can be improved largely. The experiments results show that the proposedsensor data stream estimation method based on Kalman can estimate dynamic sensor data insensor data stream efficiently with high estimation accuracy.Finally， the existing problems are alalyzed and research plan for future work aredescribed.