Determination of Tea Quality: A Separate Domain Graph Convolution Network Combined With an Electronic Nose

The quality of tea is directly influenced by its harvesting periods. It is a common fraudulent practice for lower-quality tea to be passed off as higher quality. In this study, an electronic nose (e-nose) system was used to capture information about gases emitted by tea leaves harvested over six different periods. To process this information, we propose a separate domain node-level graph convolutional network (SDN-GCNet) with two main components: a separate domain graph convolutional network (SD-GCN) and a node-level attention mechanism (NLAM). Initially, the model extracts feature values and creates topological graphs by quantifying their similarity. Subsequently, the SD-GCN processes the feature graphs, and the NLAM characterizes the complex interrelationships within the gas data. Furthermore, during the graph convolution process, multiple parameter matrices are utilized to capture various types of feature information. The performance of the proposed SDN-GCNet model was compared to those of other advanced gas classification methods. The SDN-GCNet model achieved superior classification performance, with an accuracy of 94.58%, F1-score of 94.76%, and a Kappa coefficient of 93.50%. The proposed algorithm not only enhances the detection performance of the e-nose, but also offers an improved method for determining the quality of tea leaves harvested in different periods.