Rapid detection of traditional Chinese Baijiu with edible alcohol by an electronic nose with MS1D CNN-TE hybrid model

An electronic nose system based on MS1D CNN-TE hybrid model was proposed to distinguish traditional Chinese Baijiu with edible alcohol. A self-developed gas testing system was prepared to collect gas data from different Baijiu samples, which were mixed with different proportions of the edible alcohol. To capture the flavor features of the sample dataset, the hybrid model based on multi-scale one-dimensional convolutional neural network (MS1D CNN) and transformer encoder (TE) was constructed. The hybrid model of MS1D CNN-TE achieved a classification accuracy of 99.26 %, a precision of 99.33 %, a recall of 99.29 % and an F1-score of 99.28 % with a faster and more stable convergence, while the accuracies of the single-model architectures of MS1D CNN and TE were only 94.81 % and 95.56 %. Besides, the performances of MS1D CNN-TE were significantly better than those of traditional machine learning models and other hybrid models based on recurrent neural networks. An accuracy of 95.56 % was achieved for the test data of the first 10 s after gas introduction with an inference time of 26.65 ms, indicating the potential application in rapid detection. The study was expected to be further applied to develop online and rapid detection system for food quality and safety monitoring. • A hybrid model of MS1D CNN-TE was proposed for electronic nose. • A collaborative modeling of local and global features in temporal signals of electronic nose was achieved. • The traditional Chinese Baijiu with edible alcohol was accurately distinguished with the hybrid model. • The rapid detection and identification were achieved with the data of the first 10 s after gas introduction.