A Multitask Learning Framework for Multi-Property Detection of Wine

The electronic nose (E-nose) is a bionic olfactory system and a powerful tool in many fields. Sample classification and parameter prediction are the core functions of the E-nose. We present two algorithms for simultaneous recognition of four properties (wine region, grape variety, vintage, and fermentation processes) based on a back-propagation neural network (BPNN) and convolutional neural network (CNN), respectively, where the tasks (i.e., identification of the four properties) share underlying features. These algorithms exploited synergy among tasks to enhance their individual performance. Experimental results show that the model based on BPNN achieved the best performance with accuracies of 94.5%, 83.7%, 75.1%, and 76.9% in identifying wine region, grape, vintage, and fermentation processes, respectively. Furthermore, the results reveal that the models can capture global and local information and perform better than single-task models.