Application Of Three-Dimensional Fluorescence Spectrum Identification Method Based On Convolution Neural Network In Water Pollution Traceability

In recent years,with the rapid growth of China’s economy and the rapid development of industry,the number of factories producing and using chemicals and dangerous goods has increased rapidly,and the potential hazard sources of accidents have also increased,resulting in the frequent occurrence of water pollution events and the destruction of river basin water environment,which also makes the situation of water shortage in China more severe.In order to quickly and accurately trace the source of pollutants in water,block pollution in time,rectify and supervise the main body responsible for pollution,so as to prevent the recurrence of water pollution events and better protect water resources and watershed ecological environment.In this study,Tanglangchuan watershed in Anning City was taken as a demonstration watershed to carry out water pollution traceability research.Collect water samples from typical enterprises in various industries and some rivers in Anning City for three-dimensional fluorescence measurement,and obtain the original fluorescence spectrum;The original fluorescence spectra were pretreated by interpolation and target method;Based on computer vision neural network convolution neural network(Conv Net),the water pollution traceability model is constructed and simulated;A real water pollution event in a tributary of Tanglangchuan is used to test the practical application value of the water pollution traceability model.The main conclusions are as follows:(1)In order to avoid the interference of Rayleigh scattering and Raman scattering in the original three-dimensional fluorescence spectrum to the identification spectrum of Conv Net,the interpolation method and target method are respectively used to pretreat the original three-dimensional fluorescence spectrum of sewage raw water and diluted water 10 times,and 50 Conv Net are set to test the improvement effect of the two methods on the accuracy of Conv Net identification of three-dimensional fluorescence spectrum.The results show that the interpolation method can not well remove the Rayleigh scattering in the three-dimensional fluorescence spectrum,but also enhance other interference peaks that are not obvious in the spectrum,but it can improve the recognition accuracy of Conv Net to a certain extent;The target method can not only effectively eliminate Rayleigh scattering and Raman scattering in the three-dimensional fluorescence spectrum,but also enhance the fluorescence characteristics weakened by water dilution,which has a more obvious effect on improving the recognition accuracy of Conv Net.(2)In order to evaluate the performance of three-dimensional fluorescence spectrum recognition algorithm,the sample set is composed of 306 fluorescence spectra of 45 water samples collected in March 2021.95% of the samples in the sample set are set as training samples and the remaining 5% of the samples are model test samples.Different three-dimensional fluorescence spectrum recognition models are established and trained by changing the super parameters such as image resolution,convolution kernel size and convolution kernel number,and 1425 groups of training results are obtained,Three network models with traceability accuracy of 87.5% are selected,and the performance of these three models is evaluated with loss and accuracy as evaluation indexes.The evaluation results show that the three models can converge well after 200 iterations,and the accuracy of their training sets can reach 100%.(3)In order to verify the effectiveness of the three-dimensional fluorescence spectrum recognition algorithm,the three-dimensional fluorescence spectra of water samples from 8 enterprises collected in April 2021 are used as simulation test samples,and the 306 fluorescence spectra obtained in March are used to form the water sample fluorescence spectrum database.The test samples are input into model a,model B and model C for simulation test.The experimental results show that the three models have good recognition effect for water samples containing high concentrations of organic pollutants,and the corresponding convolution network recognition score is high.It is difficult to recognize water samples with unclear fluorescence characteristics in three-dimensional fluorescence spectrum,so it is difficult to recognize them correctly.In general,the traceability model of model a has the best performance,with the traceability accuracy of 87.5%,and the traceability accuracy of model B and model C has also reached 75%,indicating that the water pollution traceability model constructed in this study can accurately identify the three-dimensional fluorescence spectrum after pretreatment,and is an effective water pollution traceability method.(4)A water pollution incident occurred in a tributary of Tanglangchuan in October2021 was used to trace the source.The three-dimensional fluorescence spectrum was obtained by collecting the polluted river water,and the target method was used to input it into the water pollution model for tracing the source.The source was successfully traced to the illegal sewage enterprises.It shows that the water pollution tracing method based on convolution neural network to identify the three-dimensional fluorescence spectrum proposed in this study has high practical application value.