Deep Learning Modelling and Model Transfer for Near-Infrared Spectroscopy Quantitative Analysis

Near-infrared spectroscoqy analysis technologyrelies on Chemometric methods that characterize the relationships between the spectral matrix and the chemical or physical properties. However, the samples' spectra are composed of signals and various noises. It is difficult for traditional Chemometric methods to extract the effective features of the spectra and establish a calibration model with strong generative performance for a complex assay. Furthermore, the same quantitative analysis results cannot be achieved when the calibration model established on one instrument is applied to another because of the differences between the instruments. Hence, this paper presents a quantitative analysis modeling and model transfer frameworkbased on convolution neural networks and transfer learning to improve model prediction performance on one instrument and across the instrument. An advanced model named MSRCNN is presented based on a convolutional neural network, which integrates multi-scale feature fusion and residual structure and shows outstanding model generalization performance on the master instrument. Then, four transfer learning methods based on fine-tuning are proposed to transfer the MSRCNN established on the master instrument to the slave instrument. The experimental results on open accessed datasets of drug and wheat show that the RMSE and R2 of MSRCNN on the master instrument are 2. 587, 0. 981, and 0. 309, 0. 977, respectively, which outperforms PLS, SVM, and CNN. Byusing 30 slave instrument samples, the transfer of the convolutional layer and fully connected layer in the MSRCNN model is the most effective among the four fine-tune methods, with RMSE and R-2 2. 289, 0. 982, and 0. 379, 0. 965, respectively. The performance can be further improved by increasing the sample of slave instruments that participated in model transferring.