Abstract: Rub-impact is the ordinary failure of rotating machinery, and its position detection is a problem pending deeply study. Sphere support vector machines, a kind of multi-class support vector machines, is improved after the influence of the size of sphere on the classification is taken into account and a effective and simple rule of classification is derived from rational analysis and simulation experiment. The rub-impacts in different parts of shaft are regarded as different faults, and their position detection becomes a large-scale multi-class fault diagnosis and can be solved by the improved sphere support vector machines. The experiment of rub-impact position detection of simulation shaft shows that the improved sphere support vector machines distinctly improves the fault recognition accuracy, the diagnosis speed and the data processing capability, and is more suitable for practical application of large-scale multi-class fault diagnosis.