Estimating concentrations of atmospheric pollutants in mixed gases based on deep convolutional network with time series decomposition

With the increasing gas detection applications, single-gas detection can no longer fulfill the requirements of most scenarios. As a result, analyzing gas mixtures, whether used to identify the types of component gases or estimate their concentrations, has become an important research focus. This paper concentrates on four major atmospheric pollutants: CO, NO₂, SO₂, and HCHO, along with their gas mixtures. First, we developed an automated gas mixture acquisition system that generates gas mixtures according to preset ratios, this system enables unattended data collection, providing a foundational dataset for studying gas detection in mixed gases. To estimate the concentrations of mixed gases and identify their constituent types, we propose a deep learning model based on convolutional neural networks (CNNs) with high-low frequency decomposition of time series. The original signal is decomposed into rough and detailed components, which are then employed to estimate the mean and bias of gas concentrations, respectively. Additionally, it completes the identification of gas types. Given that multiple sets of sensors can reduce measurement uncertainty, we also explored the application of sensor arrays for analyzing gas mixtures. Comparative experiments demonstrate that our proposed method achieves highly satisfactory results in terms of estimation accuracy.