Abstract: Due to the problems existing in the process of hydraulic pump fault diagnosis, the difficulty and the complexity to extract weak feature of the failure hydraulic pump and to automate, a sparse coding method was proposed for the hydraulic pump fault diagnosis. Firstly, the vibration signals were demodulated and transformed to frequency domain, then the K-SVD algorithm was used to obtain the dictionary from the learning of training samples, at last, the orthogonal matching pursuit algorithm was used to decompose and reconstruct the test signals. The classification of the failure of the hydraulic pump was achieved according to the reconstruction rate of the testing signal in different types of dictionary. Compared with BP neural network and support vector machine (SVM), the proposed sparse coding method shows a faster recognition speed, higher accuracy and better stability, and it can realize the fault diagnosis of hydraulic pump effectively.