Research On Key Technologies Of Online Water Quality Multi-parameter Detection Based On Spectral Fusion

Compared with the traditional chemical analysis method,the multi-parameter water quality detection method based on spectral analysis has significant advantages such as fast speed,no pollution(no chemical reagents),simple operation,and low cost.As a green detection technology,it is suitable for automatic online water quality detection equipment and represents the future development direction of water quality monitoring.However,at this stage,the detection accuracy of this method fluctuates greatly,and generalization ability is weak,which seriously restricts the technical development and market application expansion of the spectral analysis method.This thesis takes the comprehensive pollution index of water quality organic matter as research object,in order to realize the “spectral fusion online water quality multi-parameter detection”,focus on four key technologies: “spectral characteristics analysis and singularity detection”,“spectral cleaning method”,“multi-parameter fusion modeling by spectral method” and “domain adaptation to spectral distribution differences”,analyze the complementary principle of multi-source spectral information,explore information fusion methods and deep domain adaptation frameworks based on deep learning,realize rapid detection of Chemical Oxygen Demand(abbreviation as COD,unit: mg/L),Dissolved Organic Phosphorous(abbreviation as DOP,unit: mg/L),Dissolved Organic Nitorgen(abbreviation as DON,unit: mg/L).The main research content of this thesis includes the following aspects:(1)The on-line multi-parameter water quality detection system and its basic hardware based on spectral fusion have been studied.In order to meet the needs of online water quality multi-parameter detection,the current water quality multi-parameter detection scene's demand for spectral fusion technology is analyzed.On this basis a research system oriented to detection process is proposed,and optical equipment required for online detection is selected,then the structure of key parts are designed.Finally,combined with detection process,the key technology needed to improve detection accuracy is determined.(2)The characteristic analysis and singularity of water sample spectrum have been studied.By using Chebyshev's theorem of large numbers and Taylor series,the sensitive bands of real water sample spectrum to COD,DOP and DON of water quality are explored,and the singularity of normal and abnormal collection of water sample spectrum is studied with the help of wavelet transform and Lie index,on this basis,the discriminant logic of whether water sample spectrum is collected normally is proposed.(3)The cleaning method of water sample spectrum has been studied.A denoising autoencoder for fluorescence spectrum and Ultraviolet-Visible(abbreviation as UV-Vis)spectrum is designed,meanwhile an improved soft threshold function is proposed for spectral cleaning based on wavelet transform.Then the applicability and effectiveness of two methods are verified by experiments.After comparative analysis,the most suitable spectral cleaning method for online engineering requirements has been determined.(4)The fusion modeling method based on spectroscopy to detect water quality multiparameters has been studied.The fusion modeling paradigm is investigated,on this basis a water quality multi-parameter detection model based on data-level fusion,a water quality multi-parameter detection model based on feature-level fusion,and a water quality multiparameter detection model based on multi-level fusion are successively proposed.Then the feasibility and effectiveness of three models are verified by experiment.After comparative and analysis,the most suitable spectral fusion detection model for online engineering requirements has been determined.(5)The depth domain adaptation method for water spectral distribution difference has been studied.The depth domain adaptation framework is analyzed,on this basis the water quality multi-parameter domain adversarial training network for feature level fusion and water quality multi-parameter domain separation training network for feature level fusion are successively proposed.Then the feasibility and effectiveness of two models are verified by experiments.After comparison and analysis,the most suitable depth domain adaptation method for online engineering requirements has been determined.(6)The deployment of fusion detection model and performance testing.On the basis of the above key technologies,a fusion detection model deployment scheme is proposed,and the performance of detection model is evaluated by testing set.