Abstract: In order to determine the independent variables of multiple linear regression and the input layer neurons of BP network, factor analysis is used to select out the 12 physical and chemical indicators with much impact on quality of wine as their independent variables and input layer neurons, respectively. Two models are established by using multiple linear regression and improved BP neural network, respectively, which show the relationships between the physical-chemical indicators and the quality of wine. The comparison of generalization performance for the both models, draws that average relative error of multiple linear regression model for the prediction of new samples is 1.93%, while the average relative error of the BP neural network model is 0.37%. The simulations show that the generalization capability and stability of the BP neural network are better than those of the multiple regression model.

Key words: factor analysis, multiple linear regression, Back Propagation（BP） neural network, physical and chemical indicator, generalization performance

摘要： 利用因子分析法筛选出对葡萄酒质量影响较大的12种理化指标，将其作为多元线性回归的自变量和BP网络输入层神经元，分别用多元线性回归和改进的BP神经网络两种方法建立葡萄酒和酿酒葡萄的主要理化指标与葡萄酒质量的关系模型。比较了两种模型的泛化能力，得出多元线性回归模型对新样本预测的平均相对误差是1.93%，而BP神经网络模型的平均相对误差是0.37%。仿真实验表明，BP神经网络的泛化能力和稳定性明显优于多元回归模型。

关键词: 因子分析法, 多元线性回归, 反向传播（BP）神经网络, 理化指标, 泛化能力