Fusion the electronic tongue and electronic nose with a graph neural network-Mamba hybrid models for the rapid traceability of wine origin

Abstract

The origin of wine has a decisive impact on its quality and market pricing. Existing techniques for tracing the origin of wine involve complex instruments and redundant analytical procedures, which limit their rapid and on-site application. This study proposes a rapid wine provenance detection method based on the fusion information of electronic tongue (ET) and electronic nose (EN) combined with a graphical convolutional neural network (GCN)-Mamba hybrid model. First, the ET and EN are employed to collect the taste and olfactory fingerprint information of wine samples from different regions, respectively. The collected ET and EN signals are then converted into two-dimensional time-frequency spectrograms by the Stockwell transform (ST) to reveal the potential intrinsic dynamic features of the signals. Subsequently, a GCN-Mamba hybrid model is proposed to achieve comprehensive extraction of both local and global features from the spectrograms of different red wine samples. A feature interaction module and a fusion module are further proposed to reduce the heterogeneities between ET and EN, thereby achieving accurate recognition of fusion features. The experiments indicate that the proposed method demonstrates better classification performance compared to using a single sensor device for distinguishing the origin of red wine. The average accuracy, precision, recall, and F1-score of the test set across five experiments reached 99.20%, 99.22%, 99.20%, and 99.20%, with standard deviations of 0.25, 0.24, 0.26, and 0.25, respectively. This study provides a low-cost, fast, and direct method for tracing the origin of wine, offering broad application prospects for rapid or on-site measurements.