The invention discloses an improved variation mode decomposition diagnosis method for engine main shaft bearing fault diagnosis. Firstly, a fault original signal is input into inherent time scale decomposition, the signal is decomposed into a plurality of inherent rotation components and a residual term, the residual term is filtered out, and key components of the signal are completely retained while the original signal is denoised; secondly, each inherent rotation component is subjected to variation mode decomposition, the optimal component in each group of IMFs is selected according to the kurtosis principle, and the signal is reconstructed; finally, the reconstructed signal is subjected to hilbert envelope transformation to diagnose the fault types of bearings. According to the method,on the one hand, the signal is denoised by means of inherent time scale decomposition, and the signal-to-noise ratio is increased; on the other hand, each inherent rotation component is self-adaptively decomposed to be close to the respective center frequency by means of variable mode decomposition, and the optimal components are selected to reconstruct the signal. The method has good denoising capability, completely retains fault information, and has strong fault diagnosis advantages.