Dual‐Band Near‐Infrared Spectroscopy Combined With Machine Learning for Rapid Identification of Processing and Mold Condition in Paeoniae Radix Alba

With the rapid development of the Traditional Chinese Medicine (TCM) industry, ensuring the safety and standardization of TCM has become increasingly important. This study develops an integrated "processing-mold prevention-detection" method for rapid quality assessment of TCM using miniature near-infrared (NIR) spectrometers and machine learning. Focusing on Paeoniae Radix Alba (PRA), the research aims to discriminate processing parameters (methods and degrees) and detect mold contamination. A comprehensive sample set, including various processing methods (crude, stir-fry, bran-fry, and wine-moist), processing degrees (early, middle, and late stages), and mold conditions (normal and moldy), was collected from Jiangxi University of Chinese Medicine. Spectral data were acquired using two portable miniature NIR spectrometers with spectral ranges of 900–1700 nm and 1350–2550 nm, respectively, providing a data basis for multi-band analysis. For mold detection, the Savitzky-Golay first derivative with random forest model (SGFD_RF) achieved 100% accuracy on the test set. In processing method identification, the standard normal variate first derivative with random forest model (SNVFD_RF) in the 1350–2550 nm band performed best (92.4% accuracy), with only 16 wavelengths required after competitive adaptive reweighted sampling feature selection. In contrast, the 900–1700 nm band achieved a maximum accuracy of 78.5%. For processing degree discrimination, the SGFD_RF model in the 900–1700 nm band achieved 96.7% accuracy, significantly outperforming the 1350–2550 nm band (86.7%). The proposed method is portable, efficient, and accurate, offering a practical solution for TCM quality control across the supply chain, from processing and storage to transportation and end-use consumption.