In modern industry,the fault of machinery may bring giant maintenance costs and industrial loss.As the nonstationarity of signal,diversity of mechan ical fault and comprehensiveness of causes of mechanical failure,the mapping relationship between signal characteristics and failure cannot be determined easily.Data mining technology is able to extract valuable information from large data sets which seem to be irrelevant and dispersed then find out rules.This paper proposes an improved ripper algorithm applied to the field of mechanical fault diagnosis.The algorithm performance of improved ripper is compared with that of ripper,SVM and BP neural network.Table 1 shows the summary of performance in 10-fold cross-validation using the same training dataset.Through the analysis,we can conclude that the SVM algorithm has the lowest accuracy,the accuracy of BP neural net work can meet our requirement,but the time complexity is too high,especial ly when dealing with massive dataset.The rapidity and validity of ripper made it a good choice of fault diagnosis,and the improved ripper further enhanced its accuracy without raising time complexity when producing simple and clear knowledge rules,hence has advantages in mechanical fault diagnosis.