The research on support vector machine in fault diagnose has already obtained a lot of breakthroughs, such as the mode identify problems in small sample, nonlinearity, high dimension and so on. However, there are some limitations in the traditional support vector machine. In this paper, in allusion to the current rotating machinery fault diagnosis problem, the basic principles of support vector machine are studied. According to the complex characteristics of rotating machinery vibration fault, a fault extraction method is proposed based on the K-L transform. Multi-classification algorithm of support vector machine is improved, and the algorithm is used to analyze the rotating machinery vibration. By using its capabilities of model identification and system modeling, the initial symptom, occurrence, development of the typical faults are dynamically analyzed. These provide new ideas and methods for fault diagnosis of rotating machinery.