Conflict Evidence Fusion Method Of D-S And Its Recognition Application

The application of space infrared target recognition,due to the environmental disturbance,perceptual noise,transmission error caused by multi-sensor in the front-end detection stage,and the data conversion,denoising and normalization operations involved in the data preprocessing stage.When the decision-making level is fused,the multi-source information to be fused often has a certain degree of incompleteness and uncertainty.How to comprehensively analyze and process this kind of information and then implement efficient fusion,so as to give play to the advantages of information fusion,and obtain more accurate and robust target recognition results is a very challenging task.Dempster-Shafer(D-S)evidence theory has powerful uncertain information analysis and fusion capabilities,and can meet the specific needs of target recognition,so it is considered to be an ideal theoretical tool.However,when there is a high degree of conflict between the evidence,you will get results that are contrary to common sense.In response to this problem,this paper takes the modification of the evidence source as an entry point,focuses on the weighted average solution,and has made the following progress in traditional and deep learning methods:In using the traditional method of belief entropy to determine the weight of evidence,the conflict evidence fusion method based on Pignistic probability transformation(PPT)entropy and entropy distance is studied,and conflict is comprehensively measured from the perspectives of self-confidence entropy and mutual confidence entropy through PPT entropy and entropy distance.The weight of evidence is reasonably determined,so as to effectively integrate conflicting evidence.The current conflict evidence fusion method based on belief entropy only unilaterally uses the belief entropy itself to measure the uncertainty of evidence in order to determine the weight of evidence,thereby modifying the source of evidence.However,the effect of mutual belief entropy is not considered,which leads to unreasonable results after fusion.For this problem,we introduce pignistic probability transform into belief entropy,and propose a conflict evidence fusion method based on PPT entropy and entropy distance.The PPT entropy measures the conflict between evidences from the perspective of selfbelief entropy.Compared with other belief entropies,it can make full use of the information of the intersection subset and more accurately measure the uncertainty of evidence.Entropy distance is a new distance measurement method,which measures the conflict between evidence from the perspective of mutual belief entropy.The complementary combination of two measures can determine the weight of evidence more comprehensively.Compared with the typical method based on belief entropy,the results of two simulation experiments show that the proposed method improves the reliability of the correct target by 3.8% and 1.6%,respectively,and has a faster convergence speed;the experimental results of target recognition application show that the reliability of the correct target proposed method is increased by 0.6%.In the traditional method of using evidence distance to determine the weight of evidence,the conflict evidence fusion method based on Pignistic probability transformation divergence is proposed.By introducing Pignistic probability transformation divergence,it overcomes the problems of current divergence-based conflict evidence fusion methods that multi-element subsets are treated as singleelement subsets and the time complexity is high,and the conflict evidences are efficiently merged.Divergence is a way to describe the difference of evidence,and essentially an effective tool to measure the distance of evidence.However,the conflict evidence fusion method based on Belief Jensen–Shannon divergence uses multi-element subsets as single-element subsets to measure the distance between evidences,which can not reflect the interaction between single element and multi-element subsets.Although the improved method solves this problem,the time complexity of the improved method is relatively high.For this problem,we combine pignistic probability transform with Belief Jensen–Shannon divergence to construct pignistic probability transform divergence,which can reflect the interaction between single element and multi-element subsets,and satisfy boundedness,Non-degeneration and symmetry.On this basis,the conflict evidence fusion method based on pignistic probability transform divergence is proposed.This method uses pignistic probability transform divergence to measure the distance between evidences to obtain credibility weights.Deng entropy is used to measure the uncertainty of evidence to obtain the information weight.Then the two types of weights are combined as the final weight of evidence to modify the source of evidence.Target and fault recognition simulation experiment results show that compared with other four typical divergence methods,both the proposed method and Wang’s method obtain the highest correct target credibility,and the time complexity is one third of Wang’s method.In the aspect of using the deep learning method to determine the weight of evidence,the conflict evidence fusion method based on the weight-determining neural network is studied,and the optimal value of the evidence weight is adaptively obtained through the weight-determining neural network,so as to effectively solve the problem of evidence conflict.At present,the method of determining the weight of evidence based on belief entropy or distance of evidence has a single factor that affects the weight.Even if the two are combined,it cannot ensure that the weight is the most reasonable and optimal.In view of the strong adaptive learning and information mining capabilities of convolutional neural networks,we introduce deep learning methods into conflict evidence fusion,and use the weight-determining neural network to determine the weight of evidence.This method uses evidence as the network input and the corresponding weight as the output.Through backpropagation,the network parameters are updated to mine potentially useful information that affects the weight value,such as importance,reliability,and unknown information hidden in or between evidences.In this way,the weight of evidence is comprehensively determined and continuously optimized.When the custom weight loss function obtains the minimum value,the network searches for the optimal solution of the weight.The experimental results show that on the simulation data set,Yan and Han’s method and the proposed method achieve the highest recognition accuracy at the same time;on the UCR data set,compared to the other four typical traditional methods,the recognition accuracy of the proposed method is different Increased 4.06%,10.11%,10.11% and 16.19%.