The Research About Mixed-Gases Concentration Detection Based On Convolution Neural Network

In the real gas environment,gas sensors cannot detect specific gases,which resulting in the concentration accuracy of inversion often being interfered by other gas components.Therefore,the improvement of the accuracy and precision of mixed gas detection has become an important issue.The spectroscopic analysis method is based on the absorption properties of the gas,thus there is no phenomenon such as sensor poisoning and the interference of gas to other gas detectors has less influence on concentration prediction,it has been to detect SO₂,NH₃ and NO₂ mixed gases.The detection wavelength is 200-350nm ultraviolet band,and all three gases have characteristic absorption in this band.Then the absorption spectra of different concentrations of gases were measured.By comparing with the reference absorption cross sections in the molecular database,the standard absorption cross sections of gases under the spectrometer were obtained.The mixed spectral data required by the algorithm are obtained by numerical fitting the standard absorption cross sections of different gases,in which the fitting concentration of gases is in the linear range of beer's law.When the absorption spectrum at the characteristic wavelength is used for concentration fitting,the inversion error will increase due to noise or interference.Aiming at this problem,a method utilizing Convolution Neural Network(CNN)is putting forward to the quantitative analysis of mixed gases.Based on the quasi periodicity of gas spectrum in UV band,this method realizes the information integration of peaks and feature extraction by convolution and pooling operation.It also uses the full connection layer to fit the function between the characteristics and the gas concentration.In order to improve the stability of CNN and accelerate the convergence of algorithm,this paper suggests that regard the weak absorption gas as noise,and designed a cascade network structure for strong absorption and weak absorption gas.In the prediction results,the relative errors of 0.074%,1.127%and 0.318%can be achieved by CNN algorithm in the concentration inversion of SO₂,NH₃ and NO₂.For single component gas,the best relative error of SO₂ can reach 0.118%.When there is strong background gas SO₂,and the absorption of NO₂is only one third of SO₂,the calculation is carried out that the relative errors of SO₂ and NO₂ are 0.194%and 1.088%respectively.When there is a low concentration of preferential gas in the mixture,the influence of the hierarchical network structure on the inversion results of NO₂ is also discussed.The maximum error of NO₂is less than 2.5ppm,and the error of mixed gas is within the acceptable range.It shows that the hierarchical network has influence on the inversion of weak absorption of priority gas,but the model still has high inversion accuracy and strong stability.In addition,the inversion effects of CNN,multiple regression,principal component regression and partial least squares in noisy data sets are compared.The inversion error of CNN is far less than the other three algorithms,which means CNN still has strong robustness in the presence of noise or gas interference than other algorithms.