Abstract: The characteristics of worsted roving procedure and BP neural network have been analyzed.Based on interdependence of all parameters,the principal component analysis(PCA) has been proposed to preprocess the roving procedure′s data collected from a worsted mill.The five general indexes: fiber characteristic,top weight unevenness,top drafting state,top impurities content and top drawing state have been gained,thus eliminating the original variables′interdependence.After PCA the data were inputted to BP network for modeling.The results indicate that the number of node decreased;the network structure simplified;the performance and learning rate of network enhanced.The relative mean error percent between the forecast values ofCVand weight of 20 groups of test samples of roving and the measured values are 2.24% and 1.95% compared to 4.86% and 3.35% before PCA respectively;the precision is improved.The correlation coefficient between the predict value and measured value is also distinctly enhanced.