Deep Learning-based Detection Technology For SO₂ And NO Mixed Gas Concentration

With the continuous development of modern industry,atmospheric pollution is becoming more and more serious,among which SO2 and NO mixture is one of the main pollutants causing atmospheric pollution.In order to combat atmospheric pollution and improve the detection accuracy of SO2 and NO mixed gas concentration,it is of great practical significance and application value to use deep learning technology for research.Deep learning is a neural network-based machine learning technique,which has the advantages of automation,efficiency and reliability.In SO2 and NO mixed gas concentration detection,deep learning technique can improve the detection accuracy,reduce the error and improve the reliability.In this study,we will use deep learning technology,combined with sensor technology and absorption law of large light,to establish a differential optical joint density separation model and differential optical density identification model for SO2 and NO gas mixture.The model will be studied in terms of data pre-processing,feature extraction,model construction and training,and model evaluation to simplify the algorithmic process of DOAS technology and to improve the detection accuracy and precision.The main research contents include:(1)Principle of differential absorption spectroscopy technique and experimental system construction.According to the requirements in absorption spectroscopy for gas concentration in actual production,the spectrometer,gas chamber,UV light source and gas distribution instrument are selected,and the gas concentration detection platform is built,and the absorption spectra are collected and experimental data are obtained according to the principle of Lambert-Bier law.(2)Study of slow-varying separation algorithm based on morphological operations.In order to achieve a stable slow-variation separation effect under different parameters,the upper and lower contours of the absorption spectrum are firstly extracted using the onedimensional operation of morphology,and the slow-variation separation is realized by the average operation and smoothing filter.The algorithm can eliminate the reliance on a priori experience,reduce the time cost of algorithm search and manual parameter tuning,and achieve adaptive separation termination through iteration with better robustness compared with the traditional polynomial fitting.(3)Blind source separation algorithm for SO2 and NO mixed differential optical density based on deep learning.In order to solve the problem that the traditional DOAS algorithm needs to obtain the standard differential absorption cross section of known gas concentration to calculate the gas concentration,a deep learning algorithm is chosen to simplify the separation process of differential optical density and achieve a single differential optical density of SO2 and NO in the gas mixture simultaneously.This method is applied to the Wave-U-Net network in the first dimension of the base U-Net network,and the up-sampling of the network is modified to introduce the Bi-LSTM network,and the experiment proves that the accuracy of the network separation is improved compared with the Wave-U-Net,and the accuracy of the recognition is improved for the subsequent concentration inversion.(4)Differential optical density concentration recognition based on BP neural network.The method designs a BP network model with a hidden layer of 10,an activation function of Sig Moid,and an activation function of Re LU in the output layer.Experiments show that the mean square error of the model for concentration recognition is only 0.7905,and for the single gas differential optical density separated based on the deep learning algorithm in the previous chapter also decreases the relative error by 5.9% on average compared to the least squares method.