The Research And Application Of The Fuzzy Neural Network

With the development of computer science and technology, artificialintelligence shows strong vitality in the field of target recognition. As a branch ofartificial intelligence, fuzzy neural network (FNN) has both the advantages of fuzzysystems and neural networks, and is capable of expressing and processingdeterministic information as well as fuzzy information, also has a good ability oflearning. Therefore fuzzy neural network has been widely used in target recognition.First，the thesis introduced the fuzzy logic theory and fuzzy neural networktheory. Then this thesis used sound characteristics to perform the simulation basedon the Takagi-Sugeno fuzzy neural network. And a group of higher ambiguity datawhich is caused by close class-distance was found during the analysis of errorrecognition data. So this paper proposed a universal learning algorithm of fuzzyneural network by improving the update processing of learning rate and parameter.And it has performed an accurate recognition on the data of close distances betweenclusters. Comparing with traditional way, this algorithm increased the recognitionrate by10%by simulating on sound characteristics data and tumor cells data.Second, this paper analyzed the non-fuzzy output of original fuzzy neuralnetwork and made further research on the data of close distances between clusters.And then it proposed a method to fuzzily the output. Base on this, the paperestablished a new fuzzy neural network model: multi-fuzzy input and multi-fuzzyoutput (MFIMFO) model. In comparison with the original network model, therecognition rate improved significantly. Especially when using sound characteristics,the average recognition rate of the new network increased to around90%from70%of the original network.At last, in the practical application of fuzzy neural network, when using themoment invariants, gray scale and texture feature vectors to recognize the SARimage target in MSTAR, because of the existence of some data whose distancesbetween classes are similar in the feature vector, the traditional fuzzy neural network can not achieve good recognition results. Therefore, based on the new fuzzy neuralnetwork, this paper adopted the improved learning algorithm to construct thenetwork of the recognition of SAR image, and analyzed how to choose anappropriate membership function to perform the simulation. When analyzing thesimulation results of new network, we found that the data processing capability ofthe proposed fuzzy neural network is significantly higher than the traditional one.The overall recognition rate increased about7%, which proved the effectiveness ofthe proposed network model in SAR.