Research On Intelligent Learning Approachs And Application For Fuzzy Model

In order to enhance learning and applied capability of fuzzy model when it treats with complicated objects, it often needs to combine the fuzzy model with other intelligent technology, well then the complemental hybrid intelligent fuzzy model is developed. It is difficult to learn fuzzy model via single intelligent technology because of the problem is multi-constraint and multi-target optimization. It is necessary to make use of the interaction and cooperation of multiform intelligence techniques for learning fuzzy model. In the dissertation, the different intelligent intermix theory and technology are utilized to complemental learning fuzzy model from the different point of view, and the frameworks based on multi-intelligence intermix for different applied and formal fuzzy models are constructed. The frameworks request scarcely any previous information and can learn different characteristic fuzzy models (generalized fuzzy model, linguistic model, hierarchical fuzzy model).The showing forms and respective feature and learning contents and modeling process of the fuzzy models are discussed and summarized. The contemporary research status and the existing problems of the technology of constructing fuzzy models are summarized, and the basic ideas and the contemporary research status of the relevant computational intelligence are surveyed.Based on various fuzzy models application mainly are described uniformly as generalized fuzzy model and then a co-evolution generalized fuzzy model (GFM-COE) is put forward. In co-evolution framework, the generalized fuzzy model is decomposed as twain species, the first species describes the structure of fuzzy model and fuzzy rules and adopts flexible matrix encoding, the second species describes the parameters of membership functions of respective partitions and adopts tree-like structure encoding. In accordance with each species character, the various evolution strategies are adopted, and the fuzzy model is formed by various structure species coevolved. GFM-COE can identify multiplicate fuzzy models and behaves well in compact andaccuracy.Another characteristic of the GFM-COE requests a little of previous information about the object. The validity of the model has been demonstrated by examples of function approximation and prediction of chaotic time series and typical classification.Interpretable fuzzy model (linguistic model) behaves well in cognition and description, but the accuracy of linguistic model is not a sufficient degree to a complicated object. The thesis develops the research on trade-off between interpretation and accuracy, and proposes two strategies for enhancing accuracy for linguistic model. The one is that making linguistic terms and its membership function parameters adapt to the essential characteristics of each object variable, and the other strategy is that the collaborative rule is inserted into each subspace. The particle swarm optimization(PSO) is used to optimize each orthodoxy membership function of linguistic terms from each variable, and a incorporative strategy for linguistic terms is adopted, and candidate rule bases and rough linguistic model are formed. SA is utilized to choose excellent candidate rule from each subspace and reconstruct accurate linguistic models. In the process of applying PSO, the modified particle swarm optimization(MPSO) is proposed,MPSO adopts a strategy of dynamical self-adjusting inertia parameter based on illuminating information of the model fitness and iterative order characteristic, and behaves well in locally searching and ail-roundly exploring abilities.Hierarchical fuzzy model can express high-dimension complicated objects availably, but it is very difficult to obtain the model of the objects. A new GA-DBP hybrid intelligent algorithm is presented for learning hierarchical fuzzy model (include ascending and accumulating model) with high-dimension complicated objects. In the best framework of the models, hierarchical fuzzy model can be satisfied the features of relationship of variables and inherent structure of the object by GA. The author points out that the hierarchical fuzzy model is a feedforward network connected with a series of sub-models. The tactics connected each sub-model to hierarchical fuzzy model are presented. The error rate between respective sub-model propagation relation theorem is presented and proved in dynamic structure, so dynamic BP(DBP) can adjust theoperation parameters of hierarchical fuzzy model with varying structure inGA.Based on the data feature of dissolved gas analysis(DGA), the transformer condition features are organized and their previous handling way is decided. The application problem of linguistic model and hierarchical fuzzy model for condition recognition(classification) are studied, so two condition recognition models are established based on DGA, every model behaves well in accuracy and generalization, and these condition recognition models make up for the defect of IEC way. A method based on the sub-spaces support each condition degree to obtain linguistic classification rules is presented.