Research On Multi-source Information Fusion Based On Optimized Neural Network

With the popularization of big data and multi-sensor technology,multi-source information fusion technology has developed rapidly.In recent years,this technology has been widely used in many fields.Due to the characteristics of D-S(Dempster-Shafer,D-S)evidence theory and BP(Back Propagation,BP)neural network,they have become key research objects in this field.On the basis of previous studies,this thesis conducts an in-depth study of the problems in the above two methods.The research content is as follows:First of all,this article introduces the relevant knowledge and fusion methods of multi-source information fusion theory,and introduces the D-S evidence theory and BP neural network methods that will be studied in this article in detail.Aiming at the difficulty of the classic D-S evidence theory to solve the conflict problem and the problem of insufficient fusion accuracy,a method of redistributing evidence sources is adopted.The support degree obtained by the similarity matrix constructed by the similarity coefficient is accurately measured through the information entropy,and the initial weight is obtained.The size of the conflict is measured by comparing the initial weight and the average weight,and the initial weight is appropriately disturbed,and then a reasonable weight is obtained to redistribute the evidence sources and combine the new evidence sources.This article verifies its comprehensive performance through simulation and analysis.Secondly,in view of the problem that the BP neural network relies on the principle of gradient descent and is easy to fall into the local minimum,the Glowworm Swarm Optimization(Glowworm Swarm Optimization,GSO)is used to replace the gradient descent to determine its weight and threshold.Aiming at the problem of slow convergence speed and low convergence accuracy of classic GSO,adaptive variable step size and mutation operations are introduced into GSO.If the current firefly individual cannot find a better individual than its own during the optimization process,it will perform mutation operation,otherwise,the position update strategy of variable step size will be implemented to improve its global search ability and accelerate the convergence speed and accuracy.Through simulation and analysis,the comprehensive performance of the improved method in this paper is verified.Finally,in view of the problem of low fusion accuracy and long time spent by a single method,this paper combines the improved D-S evidence theory with the improved GSO optimized BP neural network,and then proposes a multi-source information fusion method based on the optimized neural network.D-S evidence theory has a good performance in dealing with the problem of uncertainty,gradually moving the information closer to the direction of less uncertainty.BP neural network has good processing ability for nonlinear data,but when the amount of information is large,the samples will be contradictions and randomness between them.If the data is directly input to the BP neural network for processing,it will lead to poor training effect,long time,and sometimes even non-convergence.Combining the two methods to achieve complementary advantages,improve the accuracy of the fusion and the convergence time.It is applied to the evaluation of traffic safety level,and the comprehensive performance of the method is verified through simulation and analysis.