The feedstock in an adsorption tower in a p-xylene unit is mainly composed of ethylbenzene(EB), p-xylene(PX), m-xylene(MX), o-xylene(OX) and other non-aromatic hydrocarbon components (denoted as NA). The content of NA is about 3~10% of feedstock, which will directly affect the adsorption tower operation and the purity of PX product, so it needs to be detected online quickly. However, NA is a mixture of different alkanes, whose composition is unknown, so it is difficult to analyze NA in the feedstock. Therefore, this paper proposed a novel method based on two spectral decomposition algorithms for two Raman spectral regions. According to the difference of Raman spectra between aromatic components and NA, two characteristic regions are chosen. NA spectral components can be neglected in the region 650 ~ 900 cm-1, so a direct decomposition algorithm is applied to get their corresponding peak areas based on spectra of pure components EB, PX, MX and OX. In the region 1350 ~ 1550 cm-1, another spectral decomposition algorithm based on Lorentzian function is adopted to obtain the spectral peak of NA which is fitted by several Lorentz peaks, then the corresponding peak area is calculated. Finally, the linear models between peak areas and concentrations of each component are built with all training samples. The experimental results show that the standard prediction error for NA is 0.242 %, and the correlation coefficient reaches 0.993. This method can provide more precision than PLS and direct spectral decomposition algorithms, so the new method can be widely applied, especially for the mixture composed of several known components and some unknown substances.