Research On Multi-sensor Data Fusion Based On Prediction And D-S Evidence Theory

With the continuous development of modern technology and the increasing integration of various systems,a single source of information can no longer meet the requirements of accuracy and reliability of data information.This requires the use of multi-sensor collaboration to collect data to reduce the uncertainty of measurement and increase the credibility of the data.Multi sensor data fusion technology is a national key research project.Now it has been widely used in military,navigation system and other fields,especially in intelligent greenhouse control.In order to accurately judge the environmental condition of intelligent greenhouse,this dissertation uses multi-sensor information fusion technology to synthesize the environmental information.At the same time,aiming at the problem that the data collected by sensors are interfered by the outside world and contain uncertain information,a two-stage data fusion model for greenhouse environmental monitoring is proposed,and some key technologies are studied.In the fusion process of two-level data fusion model,the first level fusion center first discriminates the validity of the same kind of sensor group data,and corrects the outliers,and then weighted average the corrected similar data to get the fusion result.The environmental parameters are optimized by the local fusion center and then sent to the secondary fusion center for global fusion,so as to accurately judge the greenhouse environment.In the local fusion center,this dissertation uses the box line graph in statistics to divide the outliers of sensor data,and then uses the radial basis function(RBF)neural network prediction model to correct the outliers instead of simply eliminating them.In order to solve the problem that traditional RBF neural network training method is easy to fall into local optimum,an improved cuckoo search(CS)algorithm based on adaptive step size and dynamic discovery probability is designed to optimize the neural network parameters and establish a prediction model.Finally,the local fusion results of homogeneous sensors are calculated by the improved weighted average algorithm.The experimental results show that the improved CS-RBF neural network model is more accurate.In the global fusion center,an improved D-S evidence theory is designed as a decision fusion method to fuse heterogeneous data.The basic probability assignment function is established by using fuzzy rough set,and the conflict degree of the generated evidence is studied.A new method based on angle cosine and reference distance function is designed to redistribute the weight coefficient of focus element and construct confidence matrix to solve the possible evidence conflict problem in the process of fusion.The results show that it can make a good decision and judge the fusion results according to the decision criteria.