Infrared spectroscopy analysis of SF6 and its derivative is an important method for operating state assessment and fault diagnosis of the gas insulated switchgear (GIS). Traditional methods are complicated and inefficient, and the results can vary with different subjects. In the present work, the feature extraction methods in machine learning are recommended to solve such diagnosis problem, and a multiscale weighted principal component analysis method is proposed. The proposed method combines the advantage of standard principal component analysis and multiscale decomposition to maximize the feature information in different scales, and modifies the importance of the eigenvectors in classification. The classification performance of the proposed method was demonstrated to be 3 to 4 times better than that of the standard PCA for the infrared spectra of SF6 and its derivative provided by Guangxi Research Institute of Electric Power.