Boosting Gas Classification with Attention-based Mechanism

Jiwei Qin, Xiang Li, Meiqi Pan, Gaofei Liu, Zhiyuan Ma

Abstract: With the developments of sensor technologies, Electronic Nose (E-Nose) has attracted increasing attentions. In the scenario of gas recognition using E-Nose, both traditional machine learning and deep learning-based approaches have been used. Most traditional methods rely on manually craft features, while deep-learning approaches uses complex structures that are costly in both time and money. In view of the problems, we propose a novel approach to recognize gas types using a generalized model based on CNN and attention mechanism that can extract concentration related features automatically. It significantly improves recognition accuracy and simplifies data processing procedures for E-Nose. Experimental evaluations are conducted on UCI Gas Sensor array drift dataset, and the results show that our proposed model obtains 99.5% accuracy on average. Visualization of extracted features also confirms that our model extracts distinct features among diverse gas classes.