The research focuses on engine knock detection and intensity evaluation using vibration signal. Based on the generalized orthogonal matching pursuit， a K-mean singular value decomposition （K-SVD） method， classified as sparse representation theory， is introduced into the engine knock feature extraction. The cylinder block vibration signal is sparse decomposed to the sparse dictionary covering the knock characteristics and the sparse coefficient for each single signal. On the basis of the above， a knock intensity evaluation index is proposed by using the natural logarithm of the fourth-order cumulant of the reconstructed signal. This method shows good noise reduction ability and feature extraction ability for cylinder vibration signal mixed with strong background noise. The results show that the heavy knock， light knock and normal combustion can be effectively distinguished by using this method. It proves the application value of this method in the field of engine knock identification.