Reciprocating compressors are characterized by complex structures and numerous excitation sources, leading to a severe coupling of fault characteristics in their vibration signals. Therefore, selecting the most discriminative fault features from high-dimensional feature sets is considered highly important. In traditional genetic algorithms, fixed-length encoding is used, which tends to trap the algorithm in local optima. This often results in issues such as over-selection or loss of features within the selected feature subsets. To address these limitations, a feature selection method based on BVL-SCGA (bounded variable length sign coding genetic algorithm) was proposed. In this approach, sign coding was applied to high-dimensional datasets. Bounded intervals were set to constrain the search scope, and a repair operator was incorporated to prevent feature redundancy, combined with feature evaluation criteria, the method accomplishes diagnostic tasks. Experimental results on public datasets indicate that the BVL-SCGA algorithm effectively overcomes the problems of genetic algorithms becoming trapped in local optima and feature subset redundancy during feature selection in high-dimensional spaces. Furthermore, the application of this feature selection method to vibration signals from reciprocating compressor bearings enables accurate diagnosis of large bearing clearance faults. The superiority of the proposed approach is thus validated.