In view of the interference of cross-term in diesel engine vibration signal time-frequency distribution and the difficulty in feature extraction, a new diesel engine intelligent fault diagnosis method based on rapid vibration signal sparse decomposition and two-dimensional time-frequency f0eature encoding recognition are proposed. First, a modified matching pursuit method called adaptive matching pursuit (AMP) whose dictionary can be updated according to the residual signal adaptively in the decomposition is put forward to obtain the vibration time-frequency images with good time-frequency aggregation. Then, the signals are decomposed into a series of atoms based on AMP algorithm and the time-frequency distribution of the original signals are obtained based on the Wigne-Ville distribution of these atoms. The feature coding of the time-frequency image amplitude matrix is captured using a two-directional two-dimensional NMF (TD2DNMF) algorithm, which can acquire the low dimensional characteristics contained within the higher dimensional time-frequency image. Finally, the nearest neighbor classifier is used to realize the automatic classification of time-frequency image recognition. The proposed method is applied to extract the fault features from four different state diesel engine valve faults, and the results verify that this method works well in eliminating the interference of cross terms in diesel engine vibration signal time-frequency distribution, and it improves the traditional image characteristic parameter extraction method of pattern recognition, the quick and effective fault diagnosis of diesel engine valve is realized by this proposed method.