Abstract: In neural network modeling， it may build a wrong model using original data without any treatment or only eliminating the abnormal value， for it could contain the law not to follow the physical metallurgy principle. To make the regularity significant， the original data need to be processed before modeling. In this work， based on the data of the C-Mn steel derived from a large number of data collected from different steel grades， a set of method for data preprocessing was proposed and a model for predicting yield strength of the C-Mn steel was established using LM-BP neural network， which could make the prediction accuracy meet the requirement (94.21%). The effects of the elements content and processing parameters on the yield strength were analyzed by the mean impact value (MIV). The results showed that the yield strength increased with the increase of carbon content and increased with the decrease of final rolling thickness and coiling temperature.

Key words: big data, modeling, data preprocessing, mean impact value (MIV), C-Mn steel