Rapid on-site detection of Tartary buckwheat powder adulteration using three-way XRF data combined with transformer-based feature extraction.

Tartary buckwheat powder (TBP) is highly valued for its nutritional benefits, yet its premium price often leads to adulteration with cheaper cereal powders, necessitating a rapid and non-destructive detection method. This study proposes, an intelligent authentication strategy combining three-way portable X-ray fluorescence (p-XRF) data with Transformer-based feature extraction. Three-way XRF data were acquired by varying measurement voltages and currents for pure TBP and samples adulterated with sorghum, oat, and wheat powders. The Transformer was initially employed to extract deep features from the high-dimensional three-way data, which were subsequently used as inputs for Random Forest (RF), Partial Least Squares Discriminant Analysis (PLS-DA), and Convolutional Neural Network (CNN) models to perform binary (authenticity) and multi-class (adulterant type) classifications. The results show that directly input of three-way XRF data leads to lengthy training times and misclassifications. In contrast, using Transformer-extracted features achieved 100% accuracy in binary classification and improved the accuracy of PLS-DA and CNN to 100% in identifying specific adulterants, while significantly reducing computational time. This approach offers a convenient, accurate, and efficient solution for on-site monitoring of TBP quality, providing a novel reference for food authenticity control.