Research And Application Of Water Quality COD Prediction Based On UV-visible Spectroscop

Water is the source of life and is necessary for the survival of all lives.Chemical oxygen demand(COD)is an important indicator to evaluate the quality of water,and the rapid detection of COD is of great significance to water management.COD detection method based on UV-Vis spectroscopy can detect water COD quickly,accurately and without pollution.It has become one of the important technologies for COD detection.This thesis focused on the construction of COD predictive model which was based on UVvisible spectroscopy.The spectral data were processed by first-order derivative and SG filtering.The characteristic wavelengths were selected by competitive adaptive reweighting algorithms.This thesis researched to establish COD predictive model which was based on artificial neural network and deep learning algorithm.The main elements of the study are as follows:(1)First-order derivative and SG filtering algorithms were used for processing spectral data.The processed data formed a spectral data set.A competitive adaptive reweighting algorithm was used for selecting characteristic wavelength,and 13 characteristic wavelengths were selected.According to the characteristic wavelength,the multi-wavelength COD predictive model based on PLS and BP algorithm was established,and the CNN algorithm was introduced to establish the multi-wavelength COD predictive model.(2)For BP algorithm,this thesis proposed Sparrow Search Algorithm(SSA)to optimize BP neural network parameters,the SSA-BP algorithm was applied to the field of water quality modeling and a multi-wavelength COD predictive model based on SSA-BP algorithm was established.The SSA-BP model solved the problem of poor accuracy and over-fitting of the BP model in the low concentration interval,and greatly improved the overall predictive accuracy of the model.(3)For the CNN algorithm,this thesis proposed the dual optimization of data information and network structure.For data information,the characteristic wavelength was replaced by the characteristic band.That could increase the number of valid data for a single sample.For the structure of network,firstly,the residual network was introduced into the field of COD modeling,and the residual blocks were constructed on the basis of the original CNN network.Next,the maximum pooling layer in the original CNN network was removed,resulting in a residual optimization network.The multi-band COD predictive model based on residual optimization network also effectively solved the problem of poor accuracy of prediction and over-fitting in the low concentration interval that occurs in CNN,and the predictive accuracy is similar to that of SSA-BP model.(4)In this thesis,the effect of turbidity on the detection of COD by spectrometry was investigated both theoretically and experimentally.Experiments had shown that turbidity interferes seriously with the detection of COD by spectrometry.In order to suppress turbidity interference,the PLS-based multi-wavelength turbidity compensation model and the visible full-spectrum turbidity compensation model based on the residual optimization network were established respectively.The predictive results shown that the visible full-spectrum turbidity compensation model based on the residual optimization network was more suitable for the spectral method to detect water COD.(5)A web-based water quality monitoring and early warning system was established.The prediction model established in this paper was applied to water quality early warning,which made the model research more practical.