Eleven milk powers were diluted into 77 samples.Choosing fat content as the index,the near-infrared spectrum quantitative analysis combined with the partial least squares method were used.Abnormal samples(14,52 and 76) shall be recognized and removed through twice abnormal spectrum removals.Seventy four milk power samples were studied by six spectrum pretreatments such as smoothing,derivative and standard variable transformation.The best pretreatment method was the standard normal variable transformation combined with Norris first order derivative,which root-mean-square error of cross validation was 0.354 7 and correlation coefficient square of cross validation reached 0.990 8.The established model performance improved by optimal pretreatment spectrum with three band selections.Random Frog(RF) was the optimal band selection method after comparing with the full-spectrum model,which correlation coefficient square of training set and test set were 0.997 2 and 0.997 0 respectively.The root-meansquare of training set and test set were 0.186 2 and 0.198 2 respectively.The result is that Monte-Carlo sampling(MCS),spectrum pretreatment and band optimization technology can improve the generalization and prediction in milk power fat dectection near-infrared quantitative model.