A Research On The Inversion Of Mixed Gas Concentrations Based On The Multi-feature Information Fusion Neural Network

The wide range of gas detection applications covers many aspects of life.The prevention and control of air pollution has been paid more and more attention,and the monitoring of pollution gas is related to the quality of daily life.Based on the ultraviolet absorption spectroscopy and convolutional neural network（CNN）,the thesis takes sulfur dioxide（SO₂）,nitrogen dioxide（NO₂）and ammonia（NH₃）as examples to carry out relevant research on the inversion of mixed gas concentrations.In the detection of gas mixtures,cross-interference between different components are inevitable,especially the strong absorption gas would have a serious overlap with the weak one,and the precision will be deteriorated.In the thesis,a two-module hierarchical CNN model based on multi-task learning is proposed,Module 1 and Module 2 perform concentration inversion of strong and weak absorption gas respectively.With the hierarchical structure,the model can better learn the shared features applicable to both tasks and improve the prediction accuracy of the weak absorption gas.At the same time,setting different loss weights can make the modules which are expected to be optimized receive intensive training and improve the model performance.The extracted features are crucial when performing gas concentration inversion based on the hierarchical CNN.CNN extracts shallow features with high resolution,accurate location information and rich detail information,but with much noise.Although the deep features filter out noise and are more abstract,they are not able to perceive the details of the original input.Deep and shallow features reflect sample information from different perspectives,an effective fusion of the two features can lead to the more robust feature representation and improve the generalization ability of the model.In the thesis,several information fusion models based on the hierarchical CNN are proposed.The model works even if there are measurement errors between the simulated spectra and the actual samples,the robustness is enhanced.The models proposed in the thesis were applied to the detection of the mixtures of SO₂,NO₂ and NH₃.SO₂ and NO₂ were the detected gases while NH₃ was the interfering gas.After comparison,the feature channel concatenate hierarchical CNN model has a more superior performance among the several proposed models.Compared with the model without the hierarchical structure and information fusion,the mean absolute errors of SO₂ and NO₂decreased by 80.3%and 87.0%in the simulated spectral samples.For the same experimental samples,the prediction accuracy of the feature channel concatenate hierarchical CNN model is much higher,indicating that the information fusion as well as the hierarchical structure also improves prediction accuracy in practical applications.In addition,the applicability of the model in the presence of more interfering gases is explored in the thesis.The results show that the proposed model also have significant advantages in the case of extended interfering gas species.Both simulation and experimental results confirm that the hierarchical structure and information fusion can significantly improve the performance of the CNN model,which provides a new idea for the quantitative analysis of gas mixtures.