| Dissolved Organic Carbon(DOC)refers to the carbon content of organic matter in natural waters,which responds to the content of organic matter in water and can reflect the degree of organic pollution.In order to meet the needs of seawater environment safety ensuring and urban smart water system development,this dissertation explores the advantages of spectral inspection method,such as sensitivity,real-time and in-situ monitoring.Studies in the dissertation mainly focus on the photochemical mechanisms,sample processing,feature extraction and quantitative analysis.Approximate estimation of dissolved organic carbon content in different situations is investigated and new quantitative methods are proposed,which support the technical demand for the detection of water environments.The major work and methodological innovations are as follows:(1)The basic principles of DOC spectral detection techniques are studied since conventional detection methods are not easy to meet the real-time demand of DOC on-line monitoring.The generation mechanism,spectral characteristics and environmental factors of different spectral detection methods are focused on.The spectral data processing and analysis technology in DOC detection are disscussed.The characteristics and applicability of different spectral detection methods are analyzed.Aiming at the limitations of single spectra detection method in different DOC concentration range,the feasibility of multi-source spectra fusion is studied,and the basic strategy of multi-source spectra fusion is proposed,which lays a theoretical foundation for the follow-up work.(2)The detection method of DOC based on UV-Vis spectra is studied.The local feature information extracted is difficult to fully reflect the absorption characteristics of DOC in the UV-Vis spectrum.A spectral feature extraction technology based on self-organizing map to reduce the dimension and extract the spectral features is proposed.In order to reduce the nonlinear effect caused by scattering interference and the absorption of non effective samples in DOC measurements,an UV-Vis detection model of DOC based on regularized greedy forest is established,which is combined with structural regularization and other optimization strategies to enhance the generalization ability of the model.The applicability of the detection method is analyzed,and the detection ability of UV-Vis detection method for different concentrations of DOC is discussed.(3)The fluorescence property of DOC is also studied to provide another alternative spectral detection techniques with high sensitivity.In order to improve the interpolation fitting effect and reduce the noise interference,a spectral smoothing method based on bilinear interpolation and second-order derivative is proposed.In view of the characteristics of high-dimensional data and the overlapping of component features,an alternating least square decomposition method is proposed to reduce the spectral dimension and extract the feature from the perspective of tensor decomposition.The effectiveness of spectral component extraction is improved by the boundary constraint optimization method.Then,nonlinear analysis model of DOC is constructed by using the component information obtained from spectral decomposition and regularized greedy forest.The detection ability of fluorescence for different concentrations of DOC is discussed by comparing the results of different detection methods.(4)In order to utilize the complementarity between UV-Vis and fluorescence spectra,a multi-source spectral data fusion method is studied.In view of the imbalance problem of different spectra data,based on data-driven and feature embedding,a hybrid spectral matrix construction method is proposed on the basis of data latent variable interaction to realize data fusion.Then,a feature extraction method using convolution neural network is proposed.And Long Short-Term Memory(LSTM)analysis method is introduced to establish the detection model of DOC.Experimental analysis shows that the method has a larger detection range than single spectral detection method,and can be used as a supplement to single spectral detection method.Finally,in order to adapt to the changes of sample concentration in the actual online monitoring process and improve the detection performance,a spectral detection method selection mechanism is proposed based on the non-probabilistic convex model.Based on the spectral feature space of different detection methods,a bounded discrimination model is established to realize the automatic discrimination of detection methods under different sample concentrations.The experimental results show that the proposed adaptive discriminant model is helpful to improve the performance of online detection.In summary,this dissertation has highlighted both single and multiple spectroscopy taken for detecting DOC in natural water.This research provides a general technique for data preprocessing,feature extraction and quantification.In order to overcome the physical constraints and limitations of single spectroscopy,a mixture of individual spectral components is used as a robust technique for DOC determination.This study could be an effective technique for the application of spectral detection technology in real-time online monitoring scene of DOC. |
