Mixed gas concentration inversion based on the hierarchical feature fusion convolutional neural network

The analysis of gas mixtures is of great relevance for environmental protection, human life and health [1], [2], [3], [4], [5], [6]. The common methods include chemical measurement and spectral analysis [7]. The chemical measurement is simple and easy to operate, but it is susceptible to the influence of the detection environment [8]. Traditional chemical measurement includes electrochemical analysis, electronic nose and gas chromatography-mass spectrometry (GC-MS). In the electrochemical analysis and electronic nose systems, the sensors require a more stringent working condition, and the prediction accuracy would decrease with the loss of electrolyte [9], [10], [11], [12]. Meanwhile, GC-MS is complicated and not suitable for on-site analysis [13], [14], [15]. Compared with chemical measurement, spectral analysis has a rapid response and high efficiency, which is suitable for continuous analysis. It has been widely used in the field of gas detection [16]. Spectral analysis is based on the characteristic spectra of gases and can be divided into emission spectroscopy, Raman scattering spectroscopy and absorption spectroscopy [17], [18], [19]. In the absorption spectroscopy, different gases absorb light at specific wavelengths and the absorption intensity follows Beer–Lambert law. According to different measurement wavebands, absorption spectroscopy can be divided into infrared absorption spectroscopy and ultraviolet absorption spectroscopy [20], [21].

Based on the spectral analysis, more and more studies have applied artificial neural networks to gas detection [22], [23], [24], [25], [26]. The main idea of these networks is to select the spectral data at several peaks or to manually extract features before performing concentration inversion. If the spectral lines are shifted or there is cross-interference between gases, it will lead to a decrease in prediction accuracy. The most suitable method is to perform automatic extraction of features from the spectra. In Convolutional Neural Network (CNN), the convolutional layer extracts feature automatically [27], [28], [29], [30], [31], [32], [33], [34]. Therefore, the spectral data or images of the full waveband can be fed into the CNN, which greatly reduces the workload. Compared with human selection of spectral features, automatic feature extraction can avoid the loss of information and reduce the impact of individual outliers. Thus, CNN has unique advantages in the field of spectral analysis. In the concentration inversion of gases based on CNN, the extracted spectral features are the crucial factor for prediction accuracy. Improving the ability of CNN to extract spectral features and reducing the cross-interference of components in the gas mixture becomes an urgent problem to be solved.

In this paper, a hierarchical feature fusion convolutional neural network (CNN) model is proposed for concentration inversion of gas mixtures based on the ultraviolet absorption spectroscopy. The mixtures of SO2, NO2 and NH3 were analyzed in our experiment since these three gases have strong absorption in the ultraviolet band [35]. Compared with the model without hierarchical structure and feature fusion, both the numerical and the experimental results demonstrate that the proposed model could effectively reduce the cross-interference and improve the accuracy of gas concentration inversion.

留言 (0)

沒有登入
gif