Faults of reciprocating compressor are observed as impacts in the vibration signal， but their information such as number and location within a period is difficult to identify because of the severe background noise. Therefore， this paper presents an innovative approach to extract impact features from the vibration signal based on combining ensemble empirical mode decomposition （EEMD） method with an adaptive variable scale morphological filter （AVSMF）. The method can extract impact components without previous fault information. Since it is difficult to accurately locate the impact components in a signal， an adaptive impact positioning algorithm is proposed. After smoothing the Hilbert envelope of the signal filtered with Teager energy operator， the location of the impacts can be determined. Repeated simulation experiments are carried out 50 times on impact signals with different frequencies and amplitudes. The results show that the algorithm has high accuracy， precise and stability. The effectiveness and accuracy of the proposed algorithm for early fault diagnosis is also verified by impact detection of the low pressure cylinder mid-body acceleration signal of a flash gas double-acting reciprocating compressor.