Electronic nose integrating an adaptive collaborative classification network for rice adulteration identification

Abstract

Rice adulteration, particularly the mixing of aged rice with fresh rice, poses significant challenges to market quality control and consumer health. This study addresses the challenge of rice adulteration by proposing an Adaptive Collaborative Classification Network (ACC-Net) combined with electronic nose (e-nose). The e-nose captures volatile organic compounds from rice samples, but its detection data exhibit characteristics of cross-sensitivity, time dependence, and complex signal patterns. An effective gas information classification method, spanning from the data end to decision-making end. Firstly, gas information is detected from two varieties (Daohuaxiang and Changlixiang) with varying adulteration ratios. Secondly, an Adaptive Collaborative Calculation Module (ACCM) is proposed, which integrates lightweight convolution, convolution attention and cross self-attention mechanism to extract and fuse bidirectional features from sensor and time directions. Thirdly, combining the lightweight design concept of ACCM and grouped convolution, ACC-Net is designed to identify the rice adulteration. Finally, Gradient-weighted class activation mapping (Grad-CAM) is introduced to visualize the important features calculated and classified by ACC-Net, and to observe the contribution degrees of different sensors and different time points to the classification performance. After demonstrating the rationality of ACC-Net's design through ablation experiments, we compare it with other deep learning methods and state-of-the-art gas information classification approaches. The results show that ACC-Net achieves optimal performance, proving its superior advantages. In the Daohuaxiang rice dataset, ACC-Net achieves an accuracy of 98.67 %, a precision of 98.56 %, and a recall of 98.97 %. In the Changlixiang rice dataset, the accuracy is 98.33 %, precision is 98.26 %, and recall is 98.06 %.