TransCNN: a hybrid deep learning model for detecting honey adulteration by LED-induced fluorescence spectroscopy

Abstract

In recent years, honey products have faced increasing issues of adulteration, posing significant challenges to their authenticity and quality. LED-induced fluorescence (LED-IF) has the characteristics of being non-destructive, rapid and efficient, offering significant advantages in detecting honey adulteration. A hybrid architecture integrating a transformer module and CNN (TransCNN) is proposed in this paper for processing spectroscopic fluorescence data. For honey adulteration detection tasks, a lightweight transformer module is introduced before the fully connected layer of a convolutional neural network, leveraging multi-head self-attention to enhance global feature modeling capabilities in spectroscopic fluorescence data. The TransCNN model demonstrated the best performance, with an average accuracy of 98.75% and a root mean square error (RMSE) of 3.91% to 4.33%, outperforming the traditional CNN (93.25% and 5.54% to 6.24%) and SVM (88.5% and 8.07% to 9.88%). This study demonstrates that the proposed TransCNN framework provides an effective analytical strategy for modeling long-range spectral dependencies in fluorescence-based detection.