The highly accurate on-line detection of solute concentration by using near-infrared spectroscopy analysis technology is of great significance for production optimization. Establishing a detection model needs to extract relevant information from the near-infrared spectrum, more representative samples will extract more effective information, and the model will also be more accurate. However, the purity of products has been continuously improved, and sample discrimination is reduced. The coefficient of variation of the sample is small, which leads to the diversity of samples being insufficient. Moreover,there are measurement errors when measuring the sample concentration in the laboratory, which will lead to the lack of correlation between the solute concentration and the spectrum. It is hard to establish a reliable and highly accurate near-infrared detection model. A new transfer learning modeling method based on PLS subspace alignment is proposed and applied in the near-infrared on-line detection of 2,6-dimethylphenol high purity in the product tower of the distillation purification process. In preparing the chemical monomer 2,6-dimethylphenol, there are side reactions and unreacted impurities, the materials after reaction must flow through different rectifying towers in sequence, and finally, the product with purity higher than 99% is obtained in the product tower. The product quality inspection of the product tower is particularly important. Due to the lack of diversity of the detection data of product tower detection point, the generalization ability of the detection model is weak. We create subspaces for the data sets of different detection points in the separation and purification of 2,6-dimethylphenol using the partial least squares method. Then, a mapping that aligns the other tower subspace into the product tower is learned by minimizing a Bregman matrix divergence function, reducing the feature distribution discrepancy between other towers and product towers. It avoids information loss in product tower data when projecting to a common subspace and makes full use of the feature information in other towers. The partial least squares regression modeling is completed on the transferred subspace and the final model coefficients are determined by the winner-takes-all-based weighting method. After the above method, the product tower detection model's performance has improved. Finally, the effectiveness of this method is verified with the near-infrared detection data set of 2,6-dimethylphenol. We discuss the influence of transferring different amounts of data from other detection points on the performance of the product tower detection model. The performance improvement of the product tower detection model is improved by 52.19% in the best case, and the root means square error of prediction (RMSEP) dipped from 0.059 4 to 0.028 4. Compared with the traditional modeling methods such as support vector machine regression (SVR) and Back-Propagation (BP) neural network, it has better performance.