The Research On Data Fusion Algorithms Based On The Prediction Model In Wireless Sensor Network

Today, the wireless sensor network technology not only has been applied to themilitary field for a long time, also has been widely applied to the civil field. A wirelesssensor network is composed of many sensor nodes, it perceives, monitors and gathersinformation from the surrounding environment, and through certain local processing,transfers sense’s data to the user. But most of the sensor nodes are limited by computingand storage resources, as most of WSN applications are deployed in the harsh environment,the sensor nodes’ battery energy is also very limited. Therefore, we need to develop allkinds of technology to save the sensor nodes’ various resources, especially at nodes’ energyconsumption. Data fusion technology, through series of processing to the multisourceoriginal data, such as association, merger and compression, etc., can be better at finishingreasoning, evaluation tasks, etc., than using a single sensor source. In data fusionprocessing, it can reduce the redundant information of data, and reduce the wireless sensornetwork data transmission in quantity. So it has an important influence to wireless sensornetwork’s energy saving, and the whole network life’s extending.This paper is basis of the learning and study in wireless sensor network and data fusiontechnology, focus on the dynamic copy probability model, which is Ken, Disjoint-CliquesξDJCο at Ken architecture and particle swarm optimization. By combining the twoalgorithms disjoint-cliques model and support vector machine, and according to the sensornetwork data fusion characteristics, we put forward a new data fusion model，which isSVR-DJC algorithm, and do some simulation experiments.Ken is a data fusion technology based on replicated dynamic probabilistic models. Thebasic idea is to maintain a pair of dynamic probabilistic models over the sensors’ data, withone copy distributed in the sensor network and another at a PC base station. At every timeinstance, the base station simply computes the expected values of the sensors attributesaccording to the model and uses it as the answers to the SELECT~*FREQ f query, thisprocess is not to need internal communication between base station and sensor nodes.These perception nodes always are monitoring the truth data from surrounding environment,and once they monitoring irregular data, i.e., data that was not accurately predicted by themodel within the required error bound, prior to the routing to base station.In order to using the perception data’s time and space correlation before forecasting,the authors put forward Disjoint-Cliques Models in wireless sensor network. The models utilize the spatial correlation between the sensor nodes, the wireless sensor network isdivided into several or dozens of parts, and at a part we choose a representative node, thisnode maintain a dynamic probabilistic model, and the base station runs a copy of this model.The representative node gathers all the data of its Clique, then calculates the data generatedat next time instance, if the forecasting data is at error range compared to real data, it do notneed to upload the data to the base station, or find the minimum data set meeting the errorrange, and upload the data set to the base station.Support Vector Machine ξSVMο is very popular machine learning algorithm; it cansolve the classification and regression problems. Because the Support Vector Machine isbased on Statistical Learning Theory and Structural Risk, it is very suitable for solvingsmall sample problems. And it is effective to avoid the high dimension problem whichproducing too high computational complexity. Therefore, support vector regressionmachine is very suitable for applications in Ken system, such as DJC models. As there aremany sensors’ attributes, we put forward Disjoint-Cliques Models based on support vectorregression machine ξSVR-DJCο to solve the Ken model prediction problem. Theparameters selection is good or not, it has a great influence on Support vector regressionmachine’s prediction performance. The traditional SVR parameters selection is using thealgorithm of trial and error, but this algorithm has a longer time cost, and need some luck.In this paper, we use particle swarm algorithm to optimize SVR parameters selectionproblem. Particle swarm algorithm using particle swarm intelligence is the search for theoptimal solution. It is the intelligent computing algorithm. Particle swarm algorithm issimple, strong global search ability and fast convergence.This paper gives a simulation experiment of SVR-DJC model. In order to make thesimulation experiment comparable and common, we chose the Intel Research Lab’s indoortemperature and humidity, light and voltage data for experiments. The simulationexperiments compared the particle swarm support vector regression machine ξPSO-SVRοand the trial, error support vector regression machine ξT-SVRο and back-propagationneural network the predicting performance of sensor data. The experimental results showthat the PSO-SVR spent less time than T-SVR in parameter selection, and forecasting databy PSO-SVR is much closer to the real data than T-SVR and BPNN. And the experimentsshow that the DJC model based on the PSO-SVR can be very good at solving wirelesssensor network’s "select*" problem. We also do some experiments in different size of trainset, the experiments show that, the more the train records the better the prediction is. Whilethe sensor nodes have limited resources, it’s better to keep the train records small. So wesuggest that take24hours records as the train set.