Aiming at the non-linear and non-stationary characteristics of the reciprocating compressor valve vibration signal,a method for reciprocating compressor valve fault feature extraction was proposed based on grey wolf optimization algorithm(GWO) optimized smoothness priors approach(SPA) and combined with multi-scale sample entropy (MSE).Taking the mean value of multi-scale sample entropy and the square of the skewness as the fitness function,the gray wolf algorithm was used to optimize the parameter λ of SPA,and the optimized parameter λ was brought into the SPA to adaptively decompose the vibration acceleration signal at the valve of the reciprocating compressor to obtain the trenditem and detrend item of the signal.And then,the mean value of the entropy and the square of the skewness of the multi-scale sample of the detrend item data were obtained,and they were input into the SVM as the feature vector of the reciprocating compressor valve signal.After training and testing,through analyzing experimental results,this method can be used to effectively extract the fault characteristics of the reciprocating compressor valve.