A Novel Few-Shot Gas Classification Technique Based on Relation Network

Abstract

Accurate classification of gas can help people to identify abnormal gases and thus protect their health. The performance of electronic nose (E-nose) for gas classification depends heavily on the strength of its algorithm. In recent years, machine learning methods that require a large number of samples for training have developed rapidly, but gas collection is a tedious task, and it is not easy to obtain large gas samples in some cases, so techniques that use few-shot gases for research are essential. As one of the few-shot learning methods, relation networks can be trained to find the matching degree of any two gas samples, thus classifying unknown gases with only a few reference samples for each gas class. In this paper, we present a relation network for small sample gas classification based on residual blocks and a bi-directional long short-term memory (BiLSTM) block and added squeeze-and-excitation blocks to improve the attention of the model to key information. After this, the applicability of each structure is determined by ablation experiments, and controlled experiments are also conducted. The experimental results show that the proposed relation network has good classification ability in the few-shot gas classification task.

Keywords: Few-shot gas classification, Electronic nose, Relation network, Squeeze-and-excitation blocks