Abstract: Based on standardization of the roving working procedure data gathering from worsted mill，in view of the BP neural network input variables effecting the result，the correlation analysis and Multivariate Stepwise Regression Analysis（MSRA） have been proposed respectively to select important parameters that influence the roving CV（R₁） and weight（R₂）.According to the important degree，the chosen parameters were inputted to BP network from large to small in turn；the four BP network models were established with the sub-network way for multi-input single output.The relative Mean Error Percent（MEP） between the forecast value of the 10 groups of testing samples and the observed value are all below 4%.Using the 20 groups of data that do not participate for modeling to forecast the roving quality，the results indicate that：the absolute average precision respectively are 2.63% and 2.98% for R₁ and the R₂ by the correlation analysis；also the correlation coefficients between the forecast and observed value respectively are 0.884 and 0.958；these targets are all better than using MSRA to select parameters for modeling.

Key words: worsted, roving working procedure, BP neural network, correlation analysis, multivariate stepwise regression analysis

摘要： 在标准化企业粗纱工序生产数据的基础上，针对神经网络输入端参数组会影响最终预报结果的特点，提出分别利用相关性分析法和多元逐步回归分析法筛选对粗纱CV值（R₁）和单重（R₂）影响较大的参数。将筛选出的参数按重要程度由大到小依次输入BP网络，采用多输入单输出子网组方式建立了4个网络模型。训练好的模型经10组检验样本检验，其预报结果和实测结果的平均相对误差（MEP）都低于4%。用20组未参与建模的验证数据进行预报表明：相关性分析法筛选参数建立的模型对R₁和R₂的绝对值平均预报精度分别为2.63%和2.98%，且预报值与实测值间的相关系数分别为0.884和0.958，这些指标都优于采用多元逐步回归分析法筛选参数建立的模型。

关键词: 精毛纺, 粗纱工序, BP网络, 相关性分析, 多元逐步回归