o improve the accuracy of engine fault diagnosis， a fault diagnosis method based on synchrosqueezing generalized S-transform （SSGST） and center-symmetric local binary pattern （CSLBP） is proposed. Firstly， aiming at the problems of energy leakage， spectrum smearing， frequency band aliasing and low time-frequency resolution in time-frequency analysis of signals， SSGST is proposed based on the synchrosqueezing algorithm and generalized S-transform. Two-dimensional time-frequency images with high time-frequency aggregation are obtained by the analysis of cylinder head vibration signals based on SSGST. Then， texture spectrum features of time-frequency images are extracted by CSLBP and input into the kernal extream learning machine （KELM） with cross validation optimization to diagnose engine faults. The results of the simulated signal analysis and engine fault diagnosis experiments show that SSGST has good energy aggregation performance， high time-frequency resolution， narrow frequency band distribution and no aliasing， and can effectively separate each component from the non-linear mixed signal. CSLBP-based texture spectrum features has low dimension， good intra-class aggregation and inter-class discreteness. The KELM with cross validation optimization is used to classify the fault features， and diagnose the engine faults， which obtains high diagnosis speed and accuracy.