To solve the problems that early weak fault signals are susceptible to noise interference and difficult to extract and identify, a piston pump weak fault diagnosis method based on variational mode decomposition (VMD), multiscale dispersion entropy (MDE) and extreme learning machine (ELM) is proposed. First, the vibration signals of various states to perform VMD are collected to obtain several modal components. According to the feature frequency energy contribution rate in the Hilbert envelope spectrum of each modal component, the variational modal decomposition feature energy reconstruction method (VMDF) with normalized feature energy ratio (FER) as the reconstruction criterion is proposed to reconstruct the signal of each modal component. Then, the MDE of the reconstructed signals are calculated. After analyzing the dispersion entropy at each scale, the effective scale dispersion entropy is selected as the feature vector. Finally, the feature vector is input to the ELM for pattern recognition. The verification results of the examples of the sliding shoe surface wear fault to varying degrees show that the proposed method can not only improve the efficiency of pattern recognition, but also better reflect the change law of fault degree. So, the proposed method has better applicability.