A fault diagnosis scheme based on ensemble empirical mode decompositi on (EEMD) entropy of singular values and support vector machine is put forward. Firstly, original acceleration vibration signals are decomposed into a finite number of stationary intrinsic mode functions (IMFs), and the initial feature vector matrixes are automatically formed by the IMFs．Ｓecondly, to apply the singular value decomposition to the initial feature vector matrixes, the singular values, as the fault characteristic vectors, are obtained. By normalizing the vectors and getting the entropies of singular values, fault patterns and conditions of gear cases can be identified. Singular values extracted from a number of IMFs containsing the most dominant fault information can serve as input vectors of su p port vector machine. Practical examples show that the diagnosis approach can identify gear fault patterns when the numbers of samples are small.