Because virtual metrology (VM) can achieve real-time and on-line total inspection, it is a promising way for measuring machining precision of machine tools. However, the machining processes possess the characteristics of severe vibrations. Thus, how to effectively handle signals with low signal/noise ratios and extract key features from them is a challenging issue for successfully applying VM to the machine tools. In this paper, a novel VM scheme for predicting machining precision of machine tools is proposed based on several previously developed methods for data quality evaluation, model reliance evaluation, and machining precision prediction. Besides, for data preprocess, we propose a Wavelet-based de-noising method to improve the S/N ratio of sensor data. In addition, we base on the stepwise technique to develop an automatic feature selection method that can extract key features related to machining operations in time, frequency, and time-frequency domains, and can reduce the dimension of essential features. Testing results of a 3-axis CNC machine center machining standard workpieces show that the VMS can achieve the performance that the maximum average error of machining-precision conjecture is less than 2 um and the conjecture of 20 machining-precision items can be completed within 3.8 sec.