In order to improve the speed and accuracy of diesel engine fault diagnosis, a method based on improved multiscale kernel independent component analysis (MSKICA) and kernel extreme learning machine optimized by quantum particle swarm optimization (QPSOKELM) was proposed. The cylinder head vibration signal was first decomposed into several timefrequency bands by intrinsic timescale decomposition and the effective components were selected according to the fault sensitivity in order to enhance the vibration characteristics. Then the frequency aliasing between different effective components was eliminated by using kernel independent component analysis in order to find the fault sensitive frequency bands. And the AR model parameters, multiscale fuzzy entropy and standardized energy moment of each band were extracted to build the structural feature vector. The kernel extreme learning machine optimized by quantum particle swarm optimization was finally constructed to diagnose diesel engine fault. The tests results indicate that the proposed method effectively enhances the features sensitive to engine fault in cylinder head vibration signal and the fault classification accuracy is higher than 98.45%, which improves the speed and accuracy of diesel engine fault diagnosis.