To address the issue that traditional convolutional neural networks struggle with achieving high recognition accuracy for electronic nose data of agricultural products such as Cabernet Sauvignon due to complex features like high aspect ratio and dual-stage asymmetry, an ACNet(Asymmetric Convolutional Neural Network) is proposed for identifying VOCs(Volatile Organic Compounds) from Cabernet Sauvignon grapes. By analyzing the adsorption-desorption kinetics of the electronic nose, the convolutional kernels in ACNet are structurallyoptimized. Experiments show that ACNet effectively captures the odor characteristics released by Cabernet Sauvignon grapes during quality changes, adaptively adjusts its attention distribution, and demonstrates differentiated focusing capabilities. The model achieves an accuracy of 0. 901 4, macro-precision of 0. 858 6, macro-recall of 0. 856 1, and macro-F1 of 0. 857 1. Using conservative and lenient transition-state strategies these metrics are further improved to 0. 954 3,0. 956 1,0. 954 1,0. 954 9 and 0. 932 7,0. 932 2,0. 930 7, 0. 931 3, respectively. This study advances the field by providing a new solution for non-destructive grape testing and a theoretical basis for designing asymmetric kernels for high-aspect-ratio electronic nose data.

Cabernet Sauvignon, electronic nose, volatile organic compounds, asymmetric convolutional neural network