Research On Identification Of Urban River Water Pollution Sources With Conventional Water Quality Indicators

River water pollution incidents have become more frequent with the acceleration of urbanization.To effectively prevent river water pollution,both end treatment and source control are absolutely necessary.At present,many countries have been gradually establishing water quality monitoring hardware facilities which are working more real-time and intelligent.However,the conventional water quality indicators mainly indirectly reflect the water quality from the perspective of pH,turbidity and organic content,and is difficult to directly distinguish the abnormal water quality and identify the cause of pollution.In view of this situation,this thesis analyzed the realationship between the conventional water quality indicators and the classes of pollution sources,and after this,the studies on key technologies of pollution source identification methods based on conventional water quality index,the optimization method for identification of pollution sources with low concentration and the design and development of the pollution sources identification prototype system were carried out.The main work and innovative points are listed as follows:(1)The response characteristics of conventional water quality indexes to different pollution sources were analyzed.Taking common river pollution sources as study objects,this thesis explored the response characteristics of conventional water quality indexes(e.g.,electrical conductivity,pH and turbidity)to different river water pollution sources.The response's specificity and regularity of conventional water quality indexes caused by different pollution sources was analyzed,which laid the foundation for the following work of river water pollution sources identification.(2)Pollution source identification method based on conventional water quality indexes was studied.Aiming at the problems of water quality monitoring data being interfered by many external factors and the multi-indicator joint response regularity is not remarkable,a feature extraction method based on information entropy and the Gradient Boosting Decision Tree algorithm was proposed.To improve the robustness to noise,the coarse-grained discretization features for each water quality index based on information entropy were constructed.And the nonlinear correlation between water quality indexes and pollution source classes was excavated by the Gradient Boosting Decision Tree algorithm,which was utilized to acquire combined features.Taking the Logistic Regression algorithm which has strong interpretability and the Support Vector Machine algorithm which has good classification performance for small amount of data as classifiers,the comparison experiments with different feature extraction methods were conducted.The results reveal the proposed method has a better identification performance for industrial wastewater,domestic seawage and salt tide water,and can effectively identify pollution sources at medium to high concentration pollution.(3)The optimization method for identification of pollution sources under low concentrations was studied.Aiming at the problems of the water features extracted under low concentration of polluted water are less distiguishable,the cosine distance features which are not sensitive to the change of pollution degree were added.And the DeepFM classification method based on factorization machine and neural network was introduced to improve the generalization ability of the identification model.The water samples at different pollution levels were obtained by mixing normal river water and high-concentration polluted water samples,and they were employed as test sets to compare the discriminative performance of different models.The experimental results showed that the optimized pollution soureces identication model can effectively improve the performance under low concentration pollution,which is more applicable in actual river pollution scenarios.Based on the technical research,the prototype system for river pollution sources identification was developed,the overall architecture and some interfaces of the system were introduced.And the advantages and applicable scenarios of different identification algorithms were also discussed in the practical application cases.The results reveal the DeepFM model can identify water pollution sources with a small size of training set more effectively,and the logistic regression model is still more effective when there is a long interval between training cycles.