Many subsystems in CNC machine tool have only a few failures, so they are small samples. In this case, classical reliability modeling method often has large errors. For the failure data of small sample, maximum likelihood estimation method is used to estimate the Weibull distribution model parameters of time between failures of the subsystems. Then, the Weibull distribution model parameters are modified by parameter bias correction method. Finally, the effect of bias correction parametersis tested by the D test and the error area ratio test method. The test results show that the modified models are better.