Research On Multi-components Concentration Retrieving Methods Based On Artificial Neural Network

RS-FTIR(Remote sensing-Fourier Transform infrared spectroscopy)can realize on-line dynamic monitoring of atmospheric environment in non-contact condition.The traditional monitoring methods have low speed and need on-site sampling.In contrast,RS-FTIR overcome these shortcomings and it is very appropriate for the concentration detection of air pollutants.ANN(Artificial Neural Network)has a preferable performance in nonlinear mapping ability and fitting precision compared with traditional methods.The absorption spectra of multi-component air pollutants are often seriously overlapped with each other.To solve this problem,we use the spectra measured by a FTIR spectrometer and choose BP-ANN(Back-Propagation Artificial Neural Network)algorithm to retrieve the concentrations of multi-components in this paper.Considering practical needs of the project,we select the spectral range according to the test gases and the characteristic peaks in the spectra,and we determine the measuring range in view of the type and concentration range of air pollutants exhausted by factories.On this basis,we build the model of multi-components and conduct a research on the concentration retrieving of air pollutants.The work and research results of this paper are listed as follows:Firstly,we combine the spectra of air pollutants obtained from standard database with test field conditions on the basis of Lambert-Beer law and build the model of mixed absorbance spectra to carry out simulation experiments.Secondly,we preprocess the spectra acquired from FTIR spectrometer and choose appropriate data processing algorithm according to the overlap character of characteristic peaks.Furthermore,we adjust the parameters in the algorithm to achieve optimal concentration retrieving model.Finally,we conduct noise prediction and error analysis on the concentration retrieving model and determine the optimal algorithm.Result shows that PLS can solve multicollinearity problems and with less samples,but it is greatly influenced by the system noise and the precision of PLS is lower comparing with ANN and CNN.BP-ANN performs best in the three methods mentioned above.This method can avoid noise distribution in the process of characteristic peak selection and applies to non-linear problems,thus making the final concentration retrieving results reach relatively high precision.CNN is mainly used in image classification and identification and performs well in the classification of one-dimensional spectra information,and it don’t need to conduct feature extraction to spectral and have good fault tolerance.