Empirical Modal Decomposition (EMD) and the methods based on EMD have been widely used in the field of fault diagnosis. The selection of Intrinsic Mode Function (IMF) after decomposition is important for accurate extraction of fault features. To solve such problem more effectively, the gearbox local fault optimal feature extraction algorithm based on Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) combined with Signal Quality Index (SQI) algorithm and Singular Value Decomposition (SVD) is proposed in this study. The method is evaluated by experiments on local crack of gear with different fault levels. Firstly, the original data are obtained by experiment. Then, they are decomposed by CEEMDAN. The effective IMF is decomposed by SVD to obtain the optimal feature vector, which is the input of BP neural network for training and test. Finally, the test results are compared with several common methods. Experimental results show that the proposed CEEMDANSQISVD algorithm has high recognition accuracy and is better than several conventional methods for local fault of gearbox.