A new method to build melt index soft sensor is proposed based on improved orthogonal least squares (IOLS) for nonlinear polypropylene process. OLS model has good generalization and sparseness by combining parameter local regularization and leave-one-out mean square error in cost function. Orthogonal signal correction(OSC) is applied to preprocess OLS model in order to reduce the noise information which is uncorrelated with output variables. Considering multi-grade operation in polypropylene plant, model parameter adaptive updating strategy is presented for updating the OLS model parameters online. The application results on real industrial process data show that IOLS can predict polypropylene melt index more accurately than partial least squares (PLS) and OLS.