In order to improve the speed and accuracy of diesel engine fault diagnosis by using the cylinder head vibration signal, a diesel engine fault diagnosis method of extracting the fault sensitive frequency bands is presented based on multi-scale kernel independent component analysis. Firstly, the singular value energy standard spectrum is proposed to enhance the weak shock characteristics of the vibration signal. Then the signal is decomposed into several different frequency bands by intrinsic time-scale decomposition, and the effective frequency components are selected according to the correlation criterion. Finally, the frequency aliasing between different effective components is eliminated by using kernel independent component analysis in order to obtain independent components, which contain the fault sensitive information, and AR model parameters, fuzzy entropy and standardized energy moment of each independent component are extracted to be feature vectors. They are input into the kernel extreme learning machine (KELM) in order to diagnose different running faults of the diesel engine. The test result indicates that the proposed method can effectively extract the fault sensitive frequency bands in cylinder head vibration signal, enhance the fault features and improve the fault diagnosis accuracy to 99.65%.