Research of Properties of Deformed Steel Bar Based on Multiple Stepwise Regression and BP Neural Network

Excluding secondary factors and reserving decisive factors of steel properties, relationship between the properties of deformed steel bar and chemical elements is analyzed by applying stepwise multiple linear regressions. The scatter diagram is drafted with SPSS; possible fitted curves of various categories are established (e.g., exponential model, logarithmic model, logistic model, linear model). Then the quantitative curve relationship is determined. Research shows that the influence added by chemical elements is neither a simple linear relationship nor a nonlinear one, but belongs to a complex multivariable coupled system. B.P. (abbr. for Backpropagation) neural network, a feedforward network formed of non - linear transformation units, is able to realize adjustment of multi-layer feedforward neural weight and handle highly-nonlinear problems intelligently. Therefore, the paper utilized MATLAB to design and modify B.P. neural network model. Reflection between steel properties and chemical elements are trained by the model repeatedly. According to the requirements of problem, the mathematics model of is obtained on condition that the error of training is extremely small. The modified model is inspected with systematic-sampling extracted data from attachment 1. The high degree of fit of the modified model is proved with the comparison between the predictor and practical value.