Abstract
The adulteration of Zanthoxylum bungeanum powder presents a complex challenge, as current near-infrared spectroscopy (NIRS) models are typically designed for specific adulterants and require extensive preprocessing, limiting their practical utility. To overcome these limitations, this study proposes IncepSpect-CBAM, an end-to-end one-dimensional convolutional neural network that integrates multi-scale Inception modules, a Convolutional Block Attention Module (CBAM), and residual connections. The model directly learns features from raw spectra while maintaining robustness across multiple adulteration scenarios, focusing specifically on quantifying Zanthoxylum bungeanum powder content. When evaluated on a dataset containing four common adulterants (corn flour, wheat bran powder, rice bran powder, and Zanthoxylum bungeanum stem powder), the model achieved a Root Mean Square Error of Prediction (RMSEP) of 0.058 and a coefficient of determination for prediction (RP2) of 0.980, demonstrating superior performance over traditional methods including Partial Least Squares Regression (PLSR) and Support Vector Regression (SVR), as well as deep learning benchmarks such as 1D-CNN and DeepSpectra. The results establish that the proposed model enables high-precision quantitative analysis of Zanthoxylum bungeanum powder content across diverse adulteration types, providing a robust technical framework for rapid, non-destructive quality assessment of powdered food products using near-infrared spectroscopy.