Research On Key Techniques Of Target Recognition Based On Concept Reasoning

Making full use of prior information to build intelligent reasoning systems is an effective way to enhance the perfomance of target recognition.In order to break through the recognition limitation of single information source by using the panorama of targets,heterogeneous information fusion is in great need of relevant context information.Taking into accout a good many factors residing in the context might increase the complexity of the reasoning process,but it is the indispensable stage to reach the intelligent goal of target recognition,and it is analogous to the cognition of humans.Humans make use of associative memory to perceive objective things,and the relevant context is adopted to help people form concepts about external objects.Objiects of interest are connected to the existing concept structure by analyzing their similarities and differences.Form this point of view,studying the techniques of target recognition should be devoted to building concept structures incorporating prior information about a certain domain to reason on the basis of concept relations and solve problems which are not encountered before.For target recognition,establishing comprehensive concept structure needs the aid of artificial intelligence.In general,two key points should be emphasized in the research.On one hand,the reasoning process of target recognition is brought into concept level,so the conversion method of reasoning from data level to concept level needs to be studied.On the other hand,the organization of concepts in the context needs to be studied to achieve a comprehensive understanding of target recognition result.The dissertation is organized as follows:Firstly,the research begins with the conceptual representation of information.The conceptual model of data and the ontology methodology of concept structure are studied respectively.Training data in the repository is modeled by type-1 fuzzy set,while experiences from experts are modeled by interval type-2 fuzzy sets.The properties and operations of fuzzy logic relevant to reasoning are introduced.For the concept level,two representation models of domain context are proposed.One adopts formal concept analysis,and aims at constructing the relationship between intensions and extensions of a concept.Some definitions and examples are given to show the process through which concept lattice can be generated from formal context.The other adopts lexical concept analysis to establish lexical chains in the concept intension.Some definitions and examples are illustrated to help make the concept representation more flexible.Secondly,a target recognition method based on fuzzy reasoning is proposed to form conceptual rules by learning from training data.It uses probabilistic type-1 fuzzy logic system to generate conceptual rules with good interpretability,and it paves the way for the uniform reasoning format between training data in the repository and experiences from experts.The method of probabilistic type-1 fuzzy rule generation and reasoning is given on the basis of fuzzy uncertainty model,and it proves that the proposed type-1 fuzzy logic system is capable of approximating any probability distribution function on a compact set to arbitrary accuracy.Furthermore,the optimization of probabilistic type-1 fuzzy logic system based on genetic algorithm is analyzed,and its indicators,models and steps are illustrated.According to the simulation experiment of radar emitter identification scenario,compared with the two high precision method based on neural networks,the proposed method not only gets better identification results,but also possesses better interpretation.Thirdly,a concept reasoning approach based on interval type-2 fuzzy logic system is presented to form conceptual rules by learning from the experiential interval data of experts.It starts with some general principles for collecting experiential interval data from experts.The construction of interval type-2 fuzzy sets from a set of experiential interval data is surveyed in detail.The method of interval type-2 fuzzy rule generation and reasoning is also given on the basis of fuzzy uncertainty model,and the matching between the antecedents of interval type-2 rules and different kinds of input data is discussed.After that,a fused fuzzy reasoning system is designed by making a combination of probabilistic type-1 fuzzy logic system and interval type-2 fuzzy logic system,both of which are useful for target identification.According to the simulation experiment of radar emitter identification scenario,the fused fuzzy reasoning system gets much better results than neural networks.Lastly,two reasoning methods based on ontology models are proposed by using background information of specific domain.On one hand,aiming at establishing relationships between intensions and extensions of a concept,formal concept analysis is used to generate concept lattice from formal context provided by experts.The shipboard radar function identification scenario is utilized as an example to show the fusion process under this ontology model.Heterogeneous information from airborne radar warning receiver and airborne image sensor is fused by concept matching,and the identification is achieved through concept reasoning within the generated concept structures.On the other hand,aiming at establishing connections among the elements in concept intensions,lexical concept analysis is used to select terminologies and define semantic relations for the domain ontology.Seven basic semantic relations of nouns and adjectives are defined to help build lexical chains in concept intensions.The recognition of warship action intentions is utilized as an example to show the multi-source information fusion process under this ontology model.Message format conversion of hard and soft sensors is surveyed,and the matching between sensor messages and concept intensions are analyzed.The recognition is finally achieved by calculating the degree of concept registration.With the help of ontology based concept reasoning methods,the effectiveness of heterogeneous information fusion for target recognition is verified through the two examples of maritime surveillance.