The invention discloses a bearing fault diagnosis method based on feature enhancement, which can effectively and rapidly extract bearing fault vibration signal shock characteristics while the signal data amount is reduced. Firstly, the bearing fault vibration signals are decomposed by variational mode decomposition (VMD), a kurtosis value and a component with the maximum cross-correlation functionwith the original signal are selected as the optimal components which have better block sparse characteristics. On the basis of the traditional online dictionary learning constraint model, l2, 1 normconstraints of a sparse coefficient are added. Under a new constraint model, sparse representation and dictionary learning are carried out alternatively, inter-block sparse characteristics of the newconstraint can be matched with block sparse characteristics of vibration signals in a sparse representation process, the redundant component in the signals is further removed, l2, 1 norm constraintsare added during the dictionary learning process at the same time, and an experimental result shows that dictionary atoms acquired from the dictionary learning process with new constraints added are more robust against noise interference. The dictionary obtained based on learning and the sparse coefficient are subjected to signal reconstruction, the signal shock characteristics with the redundantcomponents enhanced such as noise in the signals can be removed, and the fault information of the signals is further extracted to complete fault diagnosis.