GIS mechanical defects are one of the key factors leading to equipment failures, and timely detection and accurate localization of defects are of great significance for ensuring equipment safety. In this paper, we propose a diagnostic method for GIS mechanical defects based on modal decomposition and improved kernel extreme learning machine. First of all, the GIS vibration signals were decomposed into multiple modal components by the improved complete ensemble empirical mode decomposition with adaptive noise and the introduction of cross-correlation coefficients to screen out the effective components; then the energy moment eigenvalues were extracted for the effective components; finally, the GIS mechanical defect diagnostic model was constructed by using the improved kernel extreme learning machine with genetic algorithm to classify and identify the four operating states. The results show that the method in this paper can effectively eliminate redundant signals, and the genetic algorithm optimized kernel extreme learning machine (GA-KELM) model can significantly improve the diagnostic accuracy of GIS mechanical defects, and the overall diagnostic accuracy reaches more than 97%.