Pattern Recognition-Artificial Neural Network Method Study And Application To Material Field

A new method for the optimization of materials design—PatternRecognition-Artificial Neural Network(PR-ANN) was studied in this dissertation, and applied it to the solution of problems with long cycle of experiment, complex affecting factors and uncertainty reaction mechanism. Qualitative analysis and sample selection of data were treated by Pattern recognition method, while quantitative analysis and exact predict was managed with special designed Artificial Neural Network model, and display directly the distribution of objective function values on a two-dimensional plane. The fundamental principle of the method is that mapping multi-dimension samples to a two-dimensional plane firstly, and next extending to a multi-dimension space by nonlinear functions base on the rule that the topology structure can be maintained throughout the mapping procedure. It can prevent relations among samples from being covered or shielded in transfer process. Thus the hidden rule in data can be showed and the optimum point and region can be found intuitively from this mapping plane. That provided a effective method for optimum design of experiment and composition.Main work in this dissertation is described as follows:An artificial neural network nonlinear mapping model for Pattern Recognition is proposed. By combining Line-up Competition Algorithm and Speediest Descent method, the local minimum problem of errors has been solved in training artificial neural networks, which is favourable to balance local and global search. The main role for the Line-Up Competition Algorithm LCA) is in global research, while Speediest Descent Algorithm (SDA) is in local search. The computations show that the hybrid algorithm is superior to single one. The hybrid algorithm can get higher speed and minimum error in training artificial neural networks. The mapping model can reflect the real rules.Computer-aided design for material and experiment optimization with PR-ANN method was studied in this dissertation. A PR-ANN system was developed. The functions of sample selection, variables selection, Pattern Recognition and optimum point prediction were realized by the systemSingle objective and multiple objective problems were studied with PR-ANN system to examine its performance. By men's intuitive judgement from the contours on the mapping plane, complex multiple objective problems can transform to simple ones. The PR-ANN system was applied to the design of material and optimum experiments, and the satisfactory results were obtained. The computations show that PR-ANN system is superior to other optimization methods.