Design Of Quantizers For Multi-sensor Distributed Estimation Fusion System

Distributed estimation fusion is an important research topic of multiple sensors system. In the distributed estimation fusion system, locally optimal estimation is made on the basis of the observation and then transferred by individual sensor to the fusion center where the fusion is performed. In the practical setting, the observation and estimation from individual sensors must be compressed and coded effectively because of communication bandwidth constraints. Also, due to the fact that the distribution of observation statistics is hard to known, using a training sequence to acquire the final estimation is the only available way. Multi-sensor distributed estimation fusion was extensively introduced to cope with this situation. The study mainly explored the design of optimal quantizers used for multi-sensor distributed estimation fusion system in the case that the joint probability distribution is unknown.To deal with the unknown joint probability distribution, the current work firstly developed a design model and an algorithm for locally optimal quantizers based on hard C-means and fuzzy C-means clustering respectively. The developed model and algorithm integrated the advantages of C-means clustering with simple algorithm, strong local search ability, and powerful ability to process large data set, and direct sum estimation approach proposed by V.Megalooikonomou and Y.Yesha which can greatly reduced storage requirements for the fusion center.Secondly, it is also theoretically necessary to consider a globally optimal estimation model, because the estimation result based on the locally optimal quantizers was local optimization. Hence, this study developed a globally optimal estimation model and proposed a search algorithm on the basis of hard C-means in combination with the design of individual quantizer. Finally, computer simulations were conducted respectively under two methods of designing locally optimal quantizers on the basis of hard C-means and fuzzy C-means clustering. Then the results were compared with the estimation error based on regression trees method for designing quantizers. The results showed that the estimation error based on hard C-means method was lower above 10 percent than the estimation error based on regression trees method, and the fuzzy C-means method can further reduce the estimation error about two percent than hard C-mean method. It proved that the methods for designing locally optimal quantizers on the basis of C-means clustering can improve performance of the system significantly. Meanwhile, computer simulations were also conducted under the globally optimal fusion model. The results indicated that the global optimal model can further improve the fusion performance of the system about two percent than the model based on the locally optimal quantizers.