A new method was proposed which is multiscale localmaximum sample entropy (MLSE) to measure the complexity of time series. The noise and interference of vibration signals were suppressed by it, at the same time, the precision of sample entropy on each time scale was improved as compared with multiscale entropy. By treated the MLSE of hydraulic pump under different status as feature vectors, the extension theory was applied to the fault model identification, and it was compared with the other two methods. The results show that the fault identification rate of this method is the highest, and it takes the shortest time, so the superiority of this method is verified.