Component Prediction of Antai Pills Based on One-Dimensional Convolutional Neural Network and Near-Infrared Spectroscopy

Abstract

Convolutional neural networks (CNNs) are widely used for image recognition and text analysis and have been suggested for application on one-dimensional data as a way to reduce the need for preprocessing steps. In this study, the performance of one-dimensional convolutional neural network (1DCNN) machine learning algorithm was investigated for regression analysis of Antai pills spectral data. This algorithm was compared with other chemometric methods, including support vector machine regression (SVR) and partial least-square regression (PLSR) methods. The results showed that the 1DCNN model outperformed the PLSR and SVR models with similar data preprocessing for the three analytes (wogonoside, scutellarin, and ferulic acid) in Antai pills. Taking wogonoside as an example, the indices such as the correction coefficient of determination (Rv2), the root mean-squared error of cross validation (RMSECV) for calibration set, the prediction coefficient of determination (Rp2), and the root mean-squared error of prediction (RMSEP) obtained by PLSR modeling were 0.9340, 0.5568, 0.9491, and 0.5088; the indices obtained by SVR modeling were 0.9520, 0.4816, 0.9667, and 0.4117; and the indices obtained by 1DCNN modeling were 0.9683, 0.3397, 0.9845, and 0.2807, respectively. The evaluation metrics of 1DCNN are better than those of PLSR and SVR, and the prediction effect is the best, proving that 1DCNN has a good generalization ability. Especially with outliers of spectra, PLSR’s Rp2 decreased by 0.0181, SVR’s Rv2 decreased by 0.01, and 1DCNN’s Rv2 increased by 0.0009 and Rp2 decreased by 0.0057. The evaluation indices of 1DCNN have no significant change in comparison with no outliers and can still show good performance, which reflects the inclusiveness of the 1DCNN model for outliers. Simultaneously, the feasibility and robustness of the 1DCNN model in the application of near-infrared spectroscopy was verified, which has a certain application value.