Research on data-driven rapid nondestructive quality evaluation method, Calculus Bovis as an example

To address the low accuracy of traditional methods for animal-derived products due to spatial heterogeneity and limited portable tools resolution. This study built a data-driven rapid non-destructive framework integrating portable near infrared, multi-location spectral fusion, and machine learning. It applied low-level (raw spectral stitching) and mid-level (feature stitching) fusion for Calculus bovis heterogeneity, combined with variable selection. Qualitatively, the accuracy of the optimized linear model achieved 96.70%. Quantitatively, the mid-level model demonstrated superior performance compared to the other models, achieving the R² value of 0.9450 and the RPD value of 2.89. This framework meets the demands of analytical chemistry for on-site efficacy and qualitative/quantitative precision and provides a transferable paradigm for complex natural products.