Research On Perceived Target Classification Based On Decision Fusion In Wireless Sensor Networks

As an important part of the Internet of Things(IoT),Wireless Sensor Networks(WSNs)have been widely used in industrial,commercial,medical and military fields.As the sensing data in the network continuously increases,the problem of limited energy supply and communication bandwidth in wireless sensor networks has become the biggest challenge.Decision fusion technology can reduce the amount of data transmission in the network reasonably,decrease the bandwidth demand and energy consumption of data communication,improve the accuracy of data processing,and extend the working life of the network,so it is one of the research hotspots in the field of wireless sensor networks.In complicated application scenarios,the sensory data of sensor nodes often contains some uncertain information due to the influence caused by various interference and sensor performance.It is generally considered that the uncertain information is mainly derived from imprecise data from the sensor's incomplete observations of the targets,as well as incomplete data obtained by the partial observations of the targets.It is very difficult for the traditional decision fusion technology based on probability framework to describe and process these uncertain information effectively,while evidence theory can carry out reasonable modeling and reasoning on various uncertain information through popularizing the frame of discernment to power set.Therefore,based on the evidence theory and focusing on the target classification problem of wireless sensor networks,this paper will carry out the following research on the local node decision and multi-sensor decision fusion in the decision fusion process:1.For incomplete data with missing partial attribute,it is easy to obtain wrong classification results if the mode classification of observation targets is carried out according to the single estimation of the missing attribute.Therefore,based on the thought of multple imputation estimation,an incomplete data classification based on extreme learning machine imputation and evidential reasoning(IDCEE)was put forward in this paper.This method firstly uses the extreme learning machine to carry out multiple estimation imputation on the missing attribute information and calculate the reliability of different estimated values.Then a classification method based on evidential reasoning is designed for the relatively credible eigenvectors with estimated values and multiple sub-classification results are obtained.Finally,multiple sub-classification results are fused under the framework of evidence theory,and the thought of creedal classification is adopted to identify and judge the fusion results,so that the observed objects can belong to corresponding single class and composite class with different mass values.The simulation results indicated that the proposed method can efficiently perform the target classification task of incomplete data and obtain more significant performance improvement than other classification.algorithms based on incomplete data imputation methods.2.It is difficult for inaccurate data with a larger inaccuracy to classify such sample data accurately by using the acquired attribute information.Although the traditional Evidential k-Nearest Neighbor(EkNN)algorithm can effectively process uncertain data,it is easy to misjudge the target data to the wrong category when the categories of different Nearest Neighbor samples are greatly different,namely,when the observed sample data is located in the feature overlapping region of training samples of different categories.Therefore,based on the evidential editing method,a new Evidential k-Nearest Neighbor(EkNN)algorithm was put forward in this paper.The main idea of NEkNN algorithm is to consider the expected value and standard deviation of various training sample data sets,and use normalized Euclidean distance to assign category labels with basic belief assignment(BBA)structure to each training sample,so that training samples in overlapping region can offer more abundant and diverse category information.Further,EkNN classification of the observation sample data was carried out in the training sample sets of various categories,and mass functions of the target to be tested under this category was obtained,and Redistribute Conflicting Mass Proportionally Rule 5(PCR5)combination rules are used to conduct global fusion,thus obtaining the global fusion results of the targets.The simulation results indicated that this algorithm can obtain better classification results than other classification methods based on k-nearest neighbor.3.When local soft decisions of multiple sensors are fused,especially when high-conflict evidence is fused,the fusion results will be inconsistent with the facts.This paper analyzed the existing evidence conflict measurement methods,and found that the commonly used conflict calculation methods would fail in specific use scenarios.Therefore,this paper also analyzed the nature of evidence conflicts and proposed the Weighted Combination Method Based on Dissimilarity Measure(WCMDM).WCMDM fully considers the relationship between the evidences and the characteristics of the evidence body itself,and acquires the credibility of the evidences by using the inconsistency between the evidences.Based on this,the credibility of the evidences is corrected by the internal inconsistency of the evidence body,and the final weight coefficient of the evidence is acquired.In the meantime,based on the idea of weighted average,WCMDM makes full use of the weight coefficient of the evidence to suppress the impact of high-conflict evidence on the fusion results in the final evidence fusion process.Therefore,the simulation results indicate that,after comparing with the existing high-conflict evidence fusion methods,WCMDM not only has a faster convergence rate,but also improves the rationality of fusion results more effectively,so it is conducive for the fusion center to make a decision and judgment rapidly.4.For the decision fusion problem of local hard decision of multiple sensors,the existing hard decision fusion methods do not consider the d:istance difference between the target to be tested and various training samples as well as the conflicts between different hard decisions,so that the accuracy of fusion results isn't high when the number of sensors is relatively small.Therefore,a decision fusion method based on decision reliability and relative reliability(DFDR)was proposed in this paper.DFDR adopted the prior knowledge of the target to be tested to put forward a conversion method which transforms the sensor's local hard decision into soft decision,and then the relative reliability of the sensor's local decision was obtained according to the reliability and information amount of the converted soft decision.Later,the local decisions provided by multiple sensors were modified and fused by means of evidence discount,and the category with the reliability in the fusionresults was selected as the classification result of the target.Therefore,the simulation results indicate that DFDR can achieve a better classification accuracy than traditional weighted voting rules,Naive Bayesian rules and reliability-probability decision fusion method.