In consideration of the problem that the general clustering algorithm is ineffec tive in detection of reciprocating compressor early fault data with complex shap e clusters, a novel double evolution genetic clustering algorithm is put forward in this work. The new approach employs geodesic distance to measure the similar ity of data samples, and encodes each chromosome as a sequence of real integer n umbers representing the cluster representatives. Based on the selforganizing, selflearning and selfadapting of evolution, a power law selecting operator, double evolution crossover operator and selfadapting generation strategy are d esigned to execute the clustering detection of fault data. The results of the ex periments on early fault detection of reciprocating compressor’s valve leakage show that the new algorithm is efficient and effective. Its performance of recog nition is better than that of the Kmeans algorithm and generic genetic cluster ing algorithm, and can be adopted for detecting machine fault with complex data distribution.