A new method (BSSVM) using binary tree and sphere-structured support vector machines (SSVM) is presented for fault diagnosis. It constructs the hyper-planes step by step according to binary tree, partitions a class in every step and eliminates blind areas which are not insuperable for other multi-class methods of SVM. 4 common faults are created by a test-bed of rotor, their vibration signals are collected and transformed to frequency domain by FFT, then the spectrum energy in 8 bands divided by their total energy are taken as the energy distributions. With PCA, the 8-dimensional energy distributions are converted to 2-dimensional fault samples which can hold more than 80% useful information of the primary data. With the fault samples, the new method is tested and compared with several other multi-class methods of SVM, and experimental results show that the new method has higher speed and better accuracy than other similar ones.