According to feature extraction of high order cumulant,a method is proposed to solve the quantitatively analyzing of five types of gas,where the primary and secondary absorbed peaks are seriously overlapped. The absorbed spectrogram is mapped to four-order cumulant space and detached from each other. A technique is presented based on support vector machine of regularization theory to construct multi-dimension model,and accuracy and iterated rate are improved in small scale samples. The result hints the referenced error is no more than 4%,and satisfied with outfield demands.