A new chemometric method for determining the contents of fat and protein in milk powder was established by using near infrared(NIR) spectroscopy combined with chemometric methods. The calibration and prediction sets were partitioned by Kernard-Stone algorithm. Wavelet transform(WT) was used for de-noising and compressing signals. The radical basis function neural networks(RBFNN) model for the contents of fat and protein was built by combining with the reconstruction spectral signal. Three parameters, i.e., the spread value of RBF network, the wavelet functions, and decomposition levels were discussed in detail. The results show that the precision of the prediction model is the best when wavelet function, compression level and spread value are db2, 4, and 3.5 for fat. In the same way, the precision is the best when wavelet functin, compression level and spread value are db8, 4, and 6 for protein. Correlation coefficients(Rp) of prediction set for the correction model of fat and protein are 0990 and 0994, and root mean square error prediction(RMSEP) is 0.007 or 0.004, respectively. The results also show that the model is easy and robust, and the prediction accuracy is improved by using RBFNN combined with WT for building NIR models. This method is suitable for determining the content of fat and protein in milk powder rapidly and nondestructively.