摘要：神经网络在系统辨识、模式识别、智能控制等领域有着广泛的吸引人的前景，而且神经网络也是解决自动控制中控制器适应能力这个难题的关键钥匙之一。同时，实验设计（DOE）已广泛运用到航天业和一般生产制造业的产品质量改善、工艺流程优化，甚至已运用到医学界。根据实际需求，判别与选择不同的实验设计种类，设计实验步骤，发现如何控制各种影响因素，以最少的投入，换取最大的收益，从而使产品质量得以提升，工艺流程最优化。22464
本次设计研究了实验设计和神经网络两者之间的联系，本文先对实验设计方法的选取进行了深入的研究，结合现在工业、制造业、医学界等广泛运用的实验设计方法：拉丁超立方实验设计、正交实验设计和均匀实验设计方法，来对数据样本进行筛选。首先找出影响样本的因素，以最少的试验次数，获得最优的样本数据，为了对比不同实验设计所选取样本的建模误差，本文所选取的数据样本大小一样，对所选取的数据样本进行BP神经网络的建模。对于神经网络训练函数的建立，本文研究不同实验设计方法所选取的数据对相同神经网络建模函数的误差对比，所以神经网络训练函数选数进取相同的函行训练，并且使用相同的训练目标误差、显示结果的周期、最大迭代次数、学习率、训练次数，观察不同数据样本的神经网络训练性能曲线，比较误差率的大小，最后得出结论，对于数据的筛选与神经网络建模有一个最优化的运用。
毕业论文关键词： 实验设计；神经网络；建模预测；误差
The research of Neural Network theory based on The experiment design
Abstract：Neural network has a broad and attractive prospects in system identification, pattern recognition, intelligent control and other fields, and the neural network is regarded as one of the keys to solve the adaptability of controllers in the automatic control to the problem. At the same time, design of experiment (DOE) has been widely used to improve, from the aerospace industry to the general manufacturing industry product quality and process optimization and has been applied to the medical profession. According to the actual demand, discrimination and selection of different types of experimental design, experimental design steps you, find out how to control the effects of various factors, with minimal investment, for maximum benefits, so that product quality can be improved, process optimization.
The design of a connection between the two, this article first selected for the experimental design method was studied, the experimental design method in combination with the now industrial, manufacturing, medicine and other widely used: Latin hypercube experiment design, orthogonal design and uniform design of experiment method, the data sample selection. First, find out the influence of the factors on the test samples, the least number of sample data, obtained the optimal experimental design, in order to compare different selected modeling error of the sample, data sample size the same, modeling the selected data samples for the BP neural network. For the establishment of neural network training function, error comparison on the same neural network modeling function of the different experimental design method of the selected data, so the neural network training function to choose the same function for training, and use the same training error, display the results of the cycle, the maximum number of iterations, the learning rate, number of training, the neural network training performance curve of different data samples, comparing the error rate, finally draws the conclusion, by using an optimization for the screening of neural network modeling and data.
KeyWords : the experiment design; neural network theory; modeling prediction；error
目录
1. 绪论    1 基于实验设计方法的神经网络建模研究+源程序:http://www.751com.cn/zidonghua/lunwen_15114.html