Study On Some Problems For Genetic Algorithms And Wavelet Neural Networks

In this dissertation, some problems for genetics algorithms and wavelet neuralnetworks are deeply studied, and a series of methods to solve the problems are presented. Firstly, the functions and perfected method of traditional genetic operators arestudied. The diversity measure and the mature degree of population are defined. Theactions of three traditional genetic operators are demonstrated through quantitativeanalysis theoretically. Then one novel genetic operator named the diffusion operator ispresented, and the diffusion probability is given on the basis of simulated annealingalgorithm. The results of simulation and theoretical analysis all show its goodperformance. Secondly, the prevention of premature convergence in GA is studied. Thereasons for premature convergence and the general prevention strategies are introduced.Multipopulation parallel genetic algorithm based on even partition is presented byanalyzing some distributive characters of samples in schemas. It is proved that theprobability of premature convergence for the algorithm declines by exponent withoutaffecting convergence rate. The results of experiment show that it appears goodcharacteristic in optimal problems of multimodal function. Thirdly, the method to enhance the convergence precision of GA is studied. Azooming genetic algorithm is presented based on dynamic coding. Under the prerequisiteof no increasing the length of code and no expanding the size of population, it stores theiterative information and brings into the new constantly. The results of experiment showthat the method is very effective for dealing with high precision optimization problem. Fourthly, the structure optimal method of wavelet neural network (WNN) andits stability are studied. The essence causes of curse of dimensionality and lack ofrobustness for WNN are demonstrated through theoretical analysis. An extended waveletneural network (EWNN) is presented on the basis of the theory of Principal ComponentAnalysis (PCA), and its structure is optimized by Oja algorithm. The simulation resultsshow that EWNN has small scale of nods and high quality of anti-interference comparedwith general wavelet network. Fifthly, the optimal structure of WNN based on GA is studied. For dealing withof EWNN, an evolutionary wavelet network (EWN), whose structure is optimized by ahybrid genetic algorithm, is presented. The global searching characteristic of geneticalgorithm is used to search an area of optimal solution in order to lock it in a smalldomain, and the strong searching ability of Oja algorithm in local field is used to fast findthe optimum in the local area. Thus, the better combination of global search and fastsearch is obtained. The EWN which structure is trained by a hybrid genetic algorithm issuperior to the EWNN. Finally, the fault diagnosis method based on evolution FCM algorithm isstudied. In order to enhance the effect of fault diagnosis, an improved discrete Fouriertransform approach is used to extract fault features, and then a fault diagnosis method,combining the diffusion genetic algorithm introduced in chapter 2 with Fuzzy C-Mean(FCM), is proposed. Through selecting optimally a group of virtual standard samples offline, fault can be recognized rapidly and accurately on line. So the phenomena of localconvergence of FCM can be overcome effectively. The proposed scheme is demonstratedusing the model of a fighter, and the results show that the method is superior to FCM andBP neural network.