Near infrared spectroscopy (NIR) using multi-optical lengths is a very effective technique for quantitative analysis of mixture composition detection. Traditional chemometric methods build multi-optical lengths NIR quantitative model by representing multi-optical lengths NIR spectral data serially, which will also introduce noise and affect the predictive ability. In this paper, we demonstrate for the first time a parallel express and feature extraction method based on quaternion signal processing techniques, which can greatly improve the performance of the multi-optical lengths NIR quantitative model. We measured the NIR spectra of the mixture solutions of Intra-lipid and India ink under 9 optical lengths that were 1, 1.15, 1.3, 1.45, 1.6, 1.75, 1.9, 2.05, and 2.2 mm, then we randomly selected three optical lengths NIR spectra to constitute an integral representation model by a pure quaternion matrix. Baseline correction pretreatment and data enhancement were performed and followed by Quaternion principal component analysis (QPCA) to obtain quaternion feature vectors. Support vector regression (SVR) was then employed to establish the quantitative model of Intra lipid in the mixture solution with volume percent from 6% to 10%. Compared with PCA-SVR for single optical length NIR data method, QPCA-SVR method for random three optical lengths NIR data significantly improved the accuracy of quantification. The determination coefficients of prediction (Rp2) are in the range of 0.9713 to 0.9882, which is remarkably higher than that of the PCA-SVR method for single optical length NIR data. The root mean square error of prediction (RMSEP) are in the range of 0.1365 to 0.2215, which is appreciably lower than that of the PCA-SVR method. The research results show that the QPCA method can effectively improve the accuracy of multi-optical lengths NIR quantitative analysis, and has huge potential in multi-optical lengths quantitative applications.