The method was studied of acquisition of internal information realized by encoder and grating ruler which built in servo feed system of whole closedloop computer numbercial control (CNC) system of machine tool. On basis of analyzing characteristics of nonlinearity and nonstability of signal of the ball screw, a method was proposed of the energy eigenvalue extracted based on wavelet packets decomposing of fault state signal of the ball screw, and the energy eigenvalue with peak, frequency and square diference of the characteristic parameter of timefrequency domain were used to form the original characteristic set for fault diagnosis of the ball screw. Due to the redundancy of the original characteristic set, kernel principal component analysis (KPCA) was used to get rid of the characteristic parameter which was not obviously reflected the contribution rate of characteristics set in fault diagnosis. Fault model of the ball screw was established based on KPCALVQ neural nework (NN). The test results comparison between KPCALVQ and KPCABP of the two NN is analyzed, which illustrates that the studied method is effective and feasible in the application of fault diagnosis of ball screw.