Lightweight and High-Precision Gas Identification Neural Network for Embedded Electronic Nose Devices

To solve the problem that the deep learning models are difficult to realize real-time detection in embedded electronic nose devices, this work designed a lightweight high-precision convolutional neural network (CNN) named SeparateNet and a data transformation method called dislocation-stack (DS) adapting to the convolutional operations. According to the experimental results of 14 electronic nose datasets, the DS method reduced FLOPs by an average of up to 54.91% for CNNs compared to the common method, with an average accuracy and F1 score of over 98%, and the SeparateNet achieved an average reduction in FLOPs exceeding 84% compared to CNN with the same depth and width, while improving the average accuracy and F1 score to more than 99% when combined with the DS method. The research results prove the effectiveness of our proposed methods in precision and lightweight and provide a new pattern recognition solution for embedded electronic nose devices.