Construction Of A New Method For Simultaneous Determination Of Multiple Indicators Of Water Environment

Water is the material basis for human and biological survival.In the past few years,water quality has been threatened by various pollutants.Since 1998,China has set up automatic surface water quality monitoring stations in the key watersheds of seven major water systems.According to the local management needs,a number of local-level surface water quality monitoring points have been established.Water monitoring indicators including temperature,colour,turbidity,pH,conductivity,suspended matter and dissolved oxygen can be measured directly using simple and convenient instruments.The monitoring of total nitrogen,total phosphorus and total organic carbon requires a lot of manpower and resources for sampling,analysis and testing.Conventional testing methods are costly and complex.Therefore,building models and predicting water quality are very important to control water pollution.We propose an efficient method for detecting the concentration of different indicators in complex water environments,initially building a spectroscopic device and combining high-throughput techniques and machine learning to achieve this purpose.The experimental algorithm is constructed by testing actual water samples in the Chang'an district as an example.By inputting the characteristic spectral images with different index concentration information into the established neural network model,the type and concentration of mixed water samples corresponding to the map can be obtained.This method can be used for basic research work on regional environmental smart detection networks to build a method for rapid detection of contaminants in the study area.It also provides methodological guidance for the development of similar work in other regionsThe main research process and results are as follows:(1)The water samples are collected and preserved,and the concentrations of the different indicators of the water samples are measured using large instruments.The content of total nitrogen,nitrate nitrogen,total organic carbon,total phosphorus,Ca2+,Na+,Cl-,F-,SO42-in water samples was determined by different large instruments.(2)The 20 kinds of chromogenic agent were screened according to colour reactions and the properties of the colorants.The color difference before and after the color reaction was calculated and the suitable single color reagent was selected through the hole plate titration color development experiment.The effect of chromogenic reagents under different water background solutions and different index solutions was studied.According to the experimental results,nine chromogenic reagents including neutral red,chromium azure S,bromothymol blue,bromophenol blue,malachite green,bromocresol green,calcium reagent sodium carboxylate,cresol red and bromocresol purple were selected as compound chromogenic reagents.(3)Set up and commission spectroscopy equipment for proper operation and conduct spectroscopy experiments.The characteristic spectral image was obtained by spectroscopic instrument.During the operation of the device,the raw light is modulated into a number of scattered beams that can simultaneously pass through a pool of samples containing complex water samples.At the end of the optical path,the residual beam is captured by the high-definition camera to produce an image of the spectroscopic experiment.The whole equipment is composed of reaction system,sampling system and data collection system.The reaction system consists of four parts:light source,frosted glass sheet,filter and sample reaction cell.The sample injection system consists of a peristaltic pump and a sample pipe.The data collection system consists of a camera and a computer.The digitized images containing information about the solution are saved by the computer.(4)Feature images is standardized and the mapping information is digitized.The spectral images are greyed out and the grey values of the images are extracted to create a contour map.The difference between the spectral images before and after the response is directly observed through the contour map.The obtained spectral images are rotated,positioned and cropped by batch processing operations to obtain spectral images with 275×275 pixels for uniform numbering.The calculated values of the concentration of different indexes of the water sample are summarized into the total dimension labels,and the label values correspond to pictures one by one.(5)The machine learning model is established.The images and the corresponding label values are input into a network for machine learning,with 60%of the data used as the training and 20%as validation set set 20%as the test set.A model is built to predict the prediction of the concentration of different indicators in the water body based on the learning results.At the same time,the impact of different learning networks,feature mapping resolutions and the amount of special diagnosis mapping data on machine learning is also analyzed.The results show that different networks have a great influence on the learning results of each indicator.The results show that different networks have a great influence on the learning results of each index.The GoogLeNet Inception v1 and ResNet-50 have better learning effects and similar results.The learning results shows R2 for total phosphorus was 0.97,and the R2 for the remaining indicators ranged from 0.98 to 0.99.The R2 of the network SqueezeNet V1 for each metric was between 0.91 and 0.94.The spectral image resolution had less influence on the fitting results.The GoogLeNet Inception v1 has the best fitting effect on 14400 spectral images with 275 X 275 pixels,and the R2 of all indicators in the prediction set is greater than 0.99,indicating that the prediction results are more accurate.