Machine learning based ecosystem for water soluble heavy metal detection and analysis system

Abstract

This invention discloses a machine learning-based detection and analysis system for water-soluble heavy metals in ecosystems, used to detect water-soluble harmful heavy metal elements in atmospheric particulate matter. The system includes a sampling preprocessing module, an electrochemical detection module, a machine learning analysis module, and a data verification and output module connected in sequence. The sampling preprocessing module uses a quartz fiber filter membrane to collect PM2.5. 2.5 and PM 10 Particulate matter is pretreated by dissolution and filtration; an electrochemical detection module enriches and detects heavy metals; a machine learning analysis module uses an improved CNN network to automatically identify voltammetric spectra and establish a nonlinear mapping model of the relationship between metal ion concentration and peak current; a data verification and output module uses the standard addition method for quantitative analysis and identifies pollution sources by comparing with a multi-element fingerprint feature database. This invention can simultaneously detect eight heavy metal elements, including Cd, V, Cr, Ni, Se, As, Mn, and Pb, with a detection range of 0.1–1000 μg / L, realizing intelligent and automated heavy metal detection.

Description

Technical Field

[0001] This invention relates to the field of heavy metal pollutant analysis technology, and in particular to a machine learning-based detection and analysis system for water-soluble heavy metals in ecosystems. Background Technology

[0002] Traditional heavy metal detection methods mainly rely on laboratory analysis and periodic sampling at fixed monitoring stations, which suffers from low efficiency, high cost, and limited spatial coverage. Existing methods are limited by complex sample pretreatment and long detection times, making rapid and accurate on-site detection difficult, and the timeliness and accuracy of data acquisition need improvement. Current heavy metal detection methods often focus on single-element analysis, lacking comprehensive quantitative analysis systems that simultaneously detect multiple elements, particularly automated detection methods that combine machine learning technology and intelligent recognition algorithms. This results in highly subjective detection results, making it difficult to provide a scientific basis for environmental pollution control.

[0003] Identifying heavy metal pollution sources is directly related to the targeting and effectiveness of pollution control. However, current methods mainly rely on expert experience and simple concentration comparisons, lacking scientific analysis of multi-element fingerprint characteristics and pollution source composition profiles. In particular, the comprehensive impact assessment of key pollution sources such as motor vehicles, industrial emissions, and coal combustion is not accurate enough, leading to significant uncertainties in pollution control. Furthermore, the lack of scientific methods for predicting pollution levels and a quantitative analysis system for pollution source contribution rates hinders the rational allocation of control resources and the scientific formulation of control strategies.

[0004] Therefore, there is an urgent need for an intelligent heavy metal detection method based on machine learning to achieve scientific decision-making for simultaneous detection of multiple elements and automatic identification of pollution sources. Summary of the Invention

[0005] The purpose of this invention is to achieve scientific decision-making for simultaneous detection of multiple elements and automatic identification of pollution sources.

[0006] To achieve the above objectives, the present invention provides a machine learning-based detection and analysis system for water-soluble heavy metals in an ecosystem, used to detect water-soluble harmful heavy metal elements in atmospheric particulate matter. The system includes, in sequence, a sampling preprocessing module, an electrochemical detection module, a machine learning analysis module, and a data verification and output module.

[0007] The sampling preprocessing module is used to collect environmental samples and perform dissolution and filtration preprocessing.

[0008] The electrochemical detection module uses stripping voltammetry for heavy metal enrichment and detection, and includes an electrolytic cell, working electrode, reference electrode, auxiliary electrode, and electrochemical workstation.

[0009] The machine learning analysis module uses a deep convolutional neural network to automatically identify and extract features from the current-voltage spectrum. It preprocesses the original spectrum using wavelet denoising and baseline correction algorithms, establishes a nonlinear mapping model of the relationship between metal ion concentration and peak current, and outputs predicted values ​​of heavy metal element concentration.

[0010] The data processing output module includes a quantitative analysis unit and a pollution source identification unit. It uses the standard addition method to perform quantitative concentration calculation and identifies pollution sources by comparing multi-element fingerprint features with a pollution source database.

[0011] Furthermore, the water-soluble harmful heavy metal elements detected by the system include: Cd, V, Cr, Ni, Se, As, Mn, and Pb, with a detection concentration range of 0.1-1000 μg / L.

[0012] The electrochemical detection was performed using differential pulse stripping voltammetry (DPASV). The detection parameters were set as follows: enrichment potential voltage from -1.2V to -0.8V, enrichment time from 60 to 300 s, scan rate from 20 mV to s, pulse amplitude from 50 mV, and pulse width from 50 ms.

[0013] Furthermore, the sampling preprocessing module specifically includes: an atmospheric particulate matter sampling unit, a dissolution treatment unit, and a filtration and purification unit.

[0014] The atmospheric particulate matter sampling unit uses a quartz fiber filter membrane to collect PM2.5. 2.5 and PM 10 Particulate matter, sampling flow rate 16.7 L / min, sampling time 24 hours.

[0015] The dissolution unit dissolves the collected particulate matter samples in an acetic acid-sodium acetate buffer solution with a pH of 4.5 ± 0.1. The buffer solution consists of 0.1% acetic acid (CH3COOH) + 0.1% sodium acetate (CH3COONa) + ultrapure water. The dissolution temperature is 25 ± 2℃, followed by ultrasonic treatment for 30 minutes and oscillation extraction for 2 hours.

[0016] The filtration and purification unit uses a 0.45μm polyethersulfone membrane filter to remove insoluble particles.

[0017] Furthermore, the working electrode is a suspended mercury electrode or a bismuth film electrode with an electrode area of ​​2 mm². 2 The reference electrode is an Ag / AgCl electrode, and the auxiliary electrode is a platinum wire electrode.

[0018] The electrolytic cell has a volume of 10 mL and is equipped with a magnetic stirrer with a stirring speed of 400 rpm and a nitrogen deoxygenation time of 300 s.

[0019] The electrochemical workstation includes a potential control unit and a scanning detection unit. The potential control unit uses a potentiostat to control the enrichment potential, which is set according to the target metal. The scanning detection unit uses a differential pulse mode to scan from the enrichment potential to the positive potential direction up to +0.5V, records the current-potential curve, and has a sampling interval of 2mV and a data acquisition frequency of 1000Hz.

[0020] Furthermore, the feature extraction network adopts an improved CNN network structure, including an input layer, multiple convolutional layers, pooling layers and fully connected layers connected in sequence. The network outputs the concentration prediction values ​​of eight metal ions. The training uses the Adam optimizer with a learning rate of 0.001 and the loss function is the mean squared error.

[0021] The machine learning analysis module first preprocesses the original voltammetric spectrum using the wavelet transform function: Where f(t) is the original volt-ampere signal, ψ * Here, is the Daubechies wavelet basis function, where a is the scaling parameter used to control the frequency characteristics of the wavelet, and b is the translation parameter used to control the temporal positioning of the wavelet.

[0022] The objective function for baseline correction is: Where y i For the observed signal, z i For the corrected baseline, w i λ is the weighting coefficient used to control the importance of each point, λ is the smoothing parameter used to control the smoothness of the baseline, and Δ 2 z i It is a second-order difference operator used to constrain the continuity of the baseline.

[0023] Furthermore, the data verification output module includes a quantitative analysis unit and a pollution source identification unit.

[0024] The quantitative analysis unit employs the standard addition method, sequentially adding target metal standard solutions of known concentrations to the sample solution. A linear relationship is established between the added amount C and the peak current response I: I = a·C + b. The regression coefficients a and b are obtained using the least squares method, thus the original sample concentration is C. x = -b / a, and calculate the combined standard uncertainty.

[0025] The pollution source identification unit establishes a multi-element fingerprint feature database, including feature fingerprints of six types of pollution sources: motor vehicle emissions, coal combustion, waste incineration, metal smelting, electronics manufacturing, and dust. A positive definite matrix factorization model is used. Perform source parsing, where X ij Let G be the concentration value of the j-th element in the i-th sample. ik For the source contribution matrix, Fkj For the source component spectral matrix, E ij The residual matrix represents the error term of the model fitting.

[0026] Furthermore, the characteristic ratio ranges for various pollution sources are as follows: Motor vehicle emission sources: Pb / Cd > 5 and Zn / Cu = 2-8; Coal combustion sources: As / Se > 10 and V / Ni = 0.5-2; Waste incineration sources: Pb / As < 2 and Cd / Zn > 0.01; Metal smelting sources: Cu / Pb > 3 and Ni / Cr = 0.2-2; Electronic manufacturing sources: Se / As > 0.5 and Cd / Pb > 0.1; Dust sources: uniform distribution of element concentrations. The ratios for coal combustion, metal smelting, and electronic manufacturing sources are influenced by regional geology and technological characteristics, and parameters can be adjusted after actual verification.

[0027] Furthermore, the distance value D between the sample fingerprint and the standard pollution source fingerprint is calculated using Euclidean distance. k , Where: r sample,i Let r be the ratio of the i-th element in the sample. source,k,i Let n be the ratio of the i-th element in the k-th pollution source, and n be the logarithm of the element ratio; select distance value D. k The smallest pollution source is considered the primary pollution source type.

[0028] Furthermore, the standard addition method quantitative calculation process includes the following steps:

[0029] (1) Measure the peak current I0 of the original sample;

[0030] (2) Add volumes of V to the same sample solution sequentially. add Concentration of C std The standard solution was used to measure the peak current I after its addition. add ;

[0031] (3) Establish a linear relationship between peak current and added concentration: I add =I0+k×C add ×V add / (V0+V add ), where k is the response coefficient, V0 is the original sample volume, and C add This represents the concentration increment after adding the standard solution;

[0032] (4) The original sample concentration was calculated using the linear extrapolation method: C sample =|I0 / k|×(V0+V add ) / V0;

[0033] (5) Perform at least three standard additions and verify the reliability of the linear relationship through regression analysis, requiring a correlation coefficient R0.2 ≥0.995.

[0034] Furthermore, the system also includes a calibration module, which establishes the concentration-current response relationship using a multi-point standard curve method. The standard solution concentration gradients are 0.1 μg / L, 0.5 μg / L, 1.0 μg / L, 5.0 μg / L, 10 μg / L, 50 μg / L, 100 μg / L, 500 μg / L, and 1000 μg / L, covering a linear range of more than three orders of magnitude.

[0035] Compared with the prior art, the beneficial effects of the present invention are:

[0036] This invention achieves rapid simultaneous detection of eight water-soluble heavy metal elements by combining electrochemical detection with deep learning, covering a concentration range of 0.1-1000 μg / L. It has advantages such as high detection accuracy, high degree of automation, and simultaneous identification of pollution source types, providing an efficient and accurate technical means for environmental monitoring and pollution source control; and has broad application prospects. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the composition of the machine learning-based ecosystem detection and analysis system of the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only a part of the embodiments of this invention, not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0039] Example

[0040] like Figure 1 The diagram shows the composition of the machine learning-based detection and analysis system for water-soluble heavy metals in an ecosystem according to the present invention. The system is used to detect water-soluble harmful heavy metal elements in atmospheric particulate matter. The system includes, in sequence, a sampling preprocessing module, an electrochemical detection module, a machine learning analysis module, and a data verification and output module.

[0041] The sampling preprocessing module is used to collect environmental samples and perform dissolution and filtration preprocessing.

[0042] The atmospheric particulate matter sampling unit employs a staged sampling technique and a precision flow control system to ensure the accuracy and representativeness of the sampling. The main body of the sampler is made of stainless steel and is internally equipped with PM2.5 sampling equipment. 2.5 and PM10 The two-stage cutting head, with its geometry designed based on Stokes number theory, ensures precise separation of particles of different sizes. The quartz fiber filter membrane is a 47mm diameter Whatman QMA membrane with a pore size of 0.7μm, exhibiting excellent particulate matter retention efficiency and chemical stability. Before use, the filter membrane is baked in a muffle furnace at 450℃ for 4 hours to remove organic contaminants, and then equilibrated in a desiccator for 24 hours to constant weight.

[0043] The sampling flow control system uses a mass flow controller (MFC) for precise control, with the flow rate set at 16.7 L / min, equivalent to sampling 1 m³ per hour. 3 Air. Flow control accuracy reaches ±1%, and the built-in pressure and temperature compensation function automatically corrects for the impact of environmental condition changes on flow. An oil-free diaphragm pump is used for sampling to ensure that no additional contamination is introduced during the sampling process. The sampling tubing is made of polytetrafluoroethylene (PTFE), with an inner diameter of 6mm and a wall thickness of 2mm, exhibiting excellent chemical inertness and low adsorption characteristics.

[0044] The sampling period was set to 24-hour continuous sampling, starting at 00:00 and ending at 00:00 the following day, to ensure complete diurnal variation data were collected. Environmental parameters such as flow rate, temperature, humidity, and atmospheric pressure were recorded every hour during the sampling process for subsequent data quality control. The ambient temperature was controlled within the range of 15-35℃, and the relative humidity was less than 85%. In case of rainfall, the sampler automatically activated its rain protection device to prevent rainwater from entering the sampling system.

[0045] After sampling, the filter membrane must be replaced in a clean bench. Operators should wear disposable polyethylene gloves and use stainless steel tweezers to handle the filter membrane, avoiding direct contact. The sampled filter membrane should be immediately placed in a pre-cleaned polyethylene filter membrane box, labeled with the sampling date, time, location, and flow rate, and then stored in a 4°C refrigerator for no more than 30 days. Blank filter membranes are processed under the same conditions to remove background interference.

[0046] The sampling preprocessing module specifically includes: an atmospheric particulate matter sampling unit, a dissolution treatment unit, and a filtration and purification unit;

[0047] The atmospheric particulate matter sampling unit uses a quartz fiber filter membrane to collect PM2.5. 2.5 and PM 10 Particulate matter, sampling flow rate 16.7 L / min, sampling time 24 hours.

[0048] The dissolution unit dissolves the collected particulate matter samples in an acetic acid-sodium acetate buffer solution with a pH of 4.5 ± 0.1. The buffer solution consists of 0.1% acetic acid (CH3COOH) + 0.1% sodium acetate (CH3COONa) + ultrapure water. The dissolution temperature is 25 ± 2℃, followed by ultrasonic treatment for 30 minutes and oscillation extraction for 2 hours.

[0049] The dissolution treatment unit employs a simulated lung fluid extraction method to accurately reflect the bioavailability of water-soluble heavy metals in particulate matter. The preparation of the acetate-sodium acetate buffer solution strictly follows standard operating procedures: first, prepare 1L of 0.1% acetic acid solution by dissolving 1mL of glacial acetic acid (purity ≥99.7%) in 999mL of ultrapure water; then prepare 1L of 0.1% sodium acetate solution by dissolving 1g of anhydrous sodium acetate (analytical grade) in 1L of ultrapure water. Mix the two solutions in equal volumes and adjust the pH to 4.5±0.1 using a pH meter (accuracy ±0.01). If necessary, fine-tune with 0.1mol / L acetic acid or 0.1mol / L sodium hydroxide.

[0050] After the buffer solution is prepared, it is filtered through a 0.22μm polyethersulfone membrane filter to remove any possible particulate impurities. It is then dispensed into pre-cleaned polypropylene containers and stored at 4°C for later use. Each batch of buffer solution has a shelf life of one month; the pH and ionic strength must be retested before use. Ultrapure water requires a resistivity ≥18.2 MΩ·cm, total organic carbon (TOC) content <4 ppb, and heavy metal content <0.1 μg / L.

[0051] Sample dissolution was performed in a cleanroom at a controlled temperature of 25±2℃ and relative humidity of 45-65%. The sampled filter membrane was removed from its storage container and cut into 2mm×2mm pieces using ceramic scissors, avoiding contamination with metal tools. All the shredded filter membrane was transferred to 50mL polypropylene centrifuge tubes, and 30mL of buffer solution was added to completely cover the membrane fragments. The centrifuge tubes were then sealed and placed in an ultrasonic cleaner at 300W power, 40kHz frequency, and for 30 minutes. During ultrasonication, the centrifuge tubes were manually shaken every 10 minutes to ensure even dissolution.

[0052] After ultrasonic treatment, the centrifuge tubes were transferred to a constant-temperature shaker for further extraction at 150 rpm for 2 hours, maintaining a temperature of 25 ± 2℃. The centrifuge tubes were kept sealed during the shaking process to prevent solution evaporation and external contamination. After shaking, the samples were allowed to stand at room temperature for 30 minutes to allow insoluble particles to settle completely. Direct sunlight was avoided throughout the dissolution process to prevent photochemical reactions from affecting the heavy metal speciation.

[0053] The filtration and purification unit uses a 0.45μm polyethersulfone membrane filter to remove insoluble particles.

[0054] The filtration and purification process employs a staged filtration strategy. First, coarse filtration removes large particulate impurities, followed by fine filtration to obtain a clear test solution. The coarse filtration uses a 47mm diameter glass fiber membrane (1.0μm pore size), which is pre-washed with ultrapure water and dried. The dissolved sample solution is then slowly passed through the coarse membrane using gravity filtration, with the filtration rate controlled at 5-10 mL / min to avoid membrane damage. The filtration apparatus, including the funnel, membrane holder, and collection bottle, is made entirely of plastic to prevent metal ion contamination.

[0055] Fine filtration uses a 0.45 μm polyethersulfone (PES) membrane filter, which exhibits low protein adsorption and good chemical compatibility, with very low non-specific adsorption of heavy metal ions. Before use, the filter is pre-rinsed with 50 mL of ultrapure water to remove any residual surfactants and preservatives that may have remained from the manufacturing process. The pre-rinse solution is discarded, and then the coarsely filtered sample solution is slowly injected into the filter using a syringe at a filtration rate controlled at 2-3 mL / min.

[0056] The filtration process employed positive pressure filtration, using a 50mL disposable syringe as the propulsion device. Pressure was controlled between 0.1-0.2MPa to avoid excessive pressure that could damage the filter membrane or force heavy metal ions into its pores. The filtrate was collected in pre-cleaned 50mL polypropylene centrifuge tubes, avoiding prolonged contact with air during collection to prevent oxidation or volatilization. The filtration volume for each sample was accurately recorded for volume correction in subsequent concentration calculations.

[0057] After filtration, the pH of the filtrate should be tested immediately to ensure it remains within the range of 4.5 ± 0.1. If the pH deviates, the buffer solution must be prepared again and the dissolution and extraction process repeated. The filtrate should be stored at 4°C for a maximum of 7 days. During storage, the container should be tightly sealed to prevent evaporation, concentration, and external contamination.

[0058] The electrochemical detection module uses stripping voltammetry for heavy metal enrichment and detection, and includes an electrolytic cell, working electrode, reference electrode, auxiliary electrode, and electrochemical workstation.

[0059] The electrolytic cell employs a three-electrode system design. The cell body is made of high borosilicate glass, with a volume precisely controlled at 10.00±0.05mL. The inner wall is smooth and scratch-free, ensuring uniform current distribution. The bottom of the electrolytic cell is conical with a 60° cone angle, facilitating sediment collection and cleaning. Five standard ground glass joints are located at the top of the cell, used for installing the working electrode, reference electrode, auxiliary electrode, vent pipe, and temperature probe, respectively. Each joint is sealed with a PTFE sealing ring to ensure system airtightness.

[0060] The magnetic stirring system is built into the bottom of the electrolytic cell. The stir bar is a PTFE-coated magnetic rotor, 15mm in length and 5mm in diameter, with a smooth, defect-free surface. The stirrer is driven by a brushless DC motor with continuously adjustable speed ranging from 50-800 rpm and a speed stability of ±2 rpm. The stirring rate is set to 400 rpm, which ensures thorough mixing of the solution without generating excessive convection that could affect the diffusion layer on the electrode surface. During stirring, contact between the stir bar and the electrodes is avoided; the electrode insertion depth is controlled to 3-5mm from the bottom of the cell.

[0061] The ventilation system removes dissolved oxygen from the solution using high-purity nitrogen (≥99.99%). Moisture and organic matter are further removed through a two-stage purifier consisting of molecular sieves and activated carbon. The gas flow rate is controlled by a precision gas flow meter, set at 100 mL / min, with a ventilation time of 300 seconds. The ventilation tube is made of polytetrafluoroethylene (PTFE) with an inner diameter of 2 mm. The tube opening is designed with a porous distributor to generate fine, uniform bubbles, improving deoxygenation efficiency. Monitoring begins immediately after ventilation to prevent the re-dissolution of oxygen from the air.

[0062] The electrolytic cell temperature control system employs a circulating water bath, with the temperature set at 25.0 ± 0.1℃. A temperature probe is inserted into the electrolyte for real-time monitoring. Temperature control accuracy is achieved through a PID controller with a response time of <30 seconds. The outer wall of the electrolytic cell is covered with insulation material to reduce heat loss. Temperature variations during monitoring do not exceed ±0.2℃.

[0063] The water-soluble harmful heavy metal elements detected by the system include: Cd, V, Cr, Ni, Se, As, Mn, and Pb, with a detection concentration range of 0.1-1000 μg / L;

[0064] The electrochemical detection was performed using differential pulse stripping voltammetry (DPASV). The detection parameters were set as follows: enrichment potential voltage from -1.2V to -0.8V, enrichment time from 60 to 300 s, scan rate from 20 mV to s, pulse amplitude from 50 mV, and pulse width from 50 ms.

[0065] The electrochemical workstation includes a potential control unit and a scanning detection unit. The potential control unit uses a potentiostat to control the enrichment potential, which is set according to the target metal. The scanning detection unit uses a differential pulse mode to scan from the enrichment potential to the positive potential direction up to +0.5V, records the current-potential curve, and has a sampling interval of 2mV and a data acquisition frequency of 1000Hz.

[0066] The working electrode is a suspended mercury electrode or a bismuth film electrode, with an electrode area of ​​2 mm². 2The reference electrode is an Ag / AgCl electrode, and the auxiliary electrode is a platinum wire electrode. The electrolytic cell has a volume of 10 mL and is equipped with a magnetic stirrer at a stirring speed of 400 rpm. Nitrogen gas is used for deoxygenation for 300 s. The suspended mercury electrode is a suspended mercury drop electrode (HMDE) with the mercury drop volume controlled at 2.0 ± 0.1 mm. 3 The corresponding electrode area is approximately 2.0 mm². 2 The mercury droplets are controlled using a precision micro-syringe, and a new droplet is replaced before each test to ensure the electrode surface is clean. The mercury material used is triple-distilled high-purity mercury (purity ≥99.9999%), stored in a sealed container under inert gas protection. After the mercury droplets form, surface cleaning is performed in a blank electrolyte solution by applying a +0.3V potential for 30 seconds to remove surface oxides and organic contaminants.

[0067] Bismuth film electrodes, as an environmentally friendly alternative to mercury electrodes, utilize glassy carbon electrodes (3 mm in diameter) as the substrate electrode, with surfaces polished to a mirror finish using 0.3 μm and 0.05 μm alumina. The bismuth film is prepared via in-situ electrodeposition: in a substrate containing 50 μg / L Bi... 3+ In a 0.1 mol / L acetic acid solution, a constant potential of -1.2 V was applied for deposition for 120 seconds, with the solution stirred during the deposition process to ensure uniform deposition. The deposited bismuth film was approximately 50-100 nm thick and exhibited good conductivity and catalytic activity.

[0068] The reference electrode is either a saturated calomel electrode (SCE) or a silver-silver chloride electrode (Ag / AgCl). The electrode filling solution is replaced periodically to maintain potential stability. The Ag / AgCl electrode is filled with 3 mol / L KCl solution, and its electrode potential relative to the standard hydrogen electrode is +0.197V (25℃). Before use, the electrode needs to be activated by soaking in deionized water for 30 minutes. During use, the potential is calibrated periodically, with a deviation not exceeding ±2mV.

[0069] The auxiliary electrode is a spiral platinum wire electrode with a diameter of 0.5 mm, a spiral length of 30 mm, and a surface area of ​​approximately 10 mm². 2 Before use, clean the platinum electrode by soaking it in dilute nitric acid, and then rinse it thoroughly with ultrapure water. The platinum electrode surface is activated by cyclic voltammetry: in 0.5 mol / L sulfuric acid, at a potential range of -0.2 V to +1.2 V, at a scan rate of 100 mV / s, for 20 cycles until the cyclic voltammogram is stable.

[0070] The electrochemical workstation includes a potential control unit, a current measurement unit, a data acquisition unit, and a software control system. The potential control unit employs a high-precision digital-to-analog converter (DAC), achieving a potential resolution of 0.1mV, potential stability of ±0.5mV, and potential scan linearity >99.9%. The current measurement unit is equipped with a multi-range current amplifier, providing a current measurement range of 10... -12 A to 10 -3 A. Resolution reaches 0.1 pA, noise level <0.5 pA. The differential pulse stripping voltammetry (DPASV) detection procedure consists of three stages: pretreatment, enrichment, and detection. The pretreatment stage maintains a +0.2 V potential for 30 seconds to remove impurities and oxides from the electrode surface. In the enrichment stage, a negative potential is applied to reduce and enrich the target metal ions on the electrode surface. The enrichment potential is determined based on the reduction potential of different metals: Cd -0.8 V, Pb -0.7 V, Cu -0.3 V, and Zn -1.2 V. The enrichment time is adjusted according to the sample concentration: 300 seconds for concentrations below 1 μg / L, 180 seconds for concentrations of 1-10 μg / L, and 60 seconds for concentrations of 10-100 μg / L.

[0071] The detection phase employs a differential pulse mode, scanning from the enrichment potential towards the positive potential up to +0.5V. Pulse parameters are set as follows: pulse amplitude 50mV, pulse width 50ms, sampling time 17ms, step potential 2mV, and scan rate 20mV / s. Current sampling is performed before the pulse ends, using dual sampling technology to measure the current difference before and after the pulse, effectively eliminating interference from background current and charging current; the data acquisition frequency is set to 1000Hz. During detection, the system automatically records the current-potential curve and displays it on the computer screen in real time. The software features peak identification and peak area calculation functions, automatically identifying characteristic peaks of heavy metals and calculating peak current and peak potential.

[0072] The atmospheric particulate matter sampling unit employs a staged sampling technique and a precision flow control system to ensure the accuracy and representativeness of the sampling. The main body of the sampler is made of stainless steel and is internally equipped with PM2.5 sampling equipment. 2.5 and PM 10 The two-stage cutting head, with its geometry designed based on Stokes number theory, ensures precise separation of particles of different sizes. The quartz fiber filter membrane is a 47mm diameter Whatman QMA membrane with a pore size of 0.7μm, exhibiting excellent particulate matter retention efficiency and chemical stability. Before use, the filter membrane is baked in a muffle furnace at 450℃ for 4 hours to remove organic contaminants, and then equilibrated in a desiccator for 24 hours to constant weight.

[0073] The sampling flow control system uses a mass flow controller (MFC) for precise control, with the flow rate set at 16.7 L / min, equivalent to sampling 1 m³ per hour. 3 Air. Flow control accuracy reaches ±1%, and the built-in pressure and temperature compensation function automatically corrects for the impact of environmental condition changes on flow. An oil-free diaphragm pump is used for sampling to ensure that no additional contamination is introduced during the sampling process. The sampling tubing is made of polytetrafluoroethylene (PTFE), with an inner diameter of 6mm and a wall thickness of 2mm, exhibiting excellent chemical inertness and low adsorption characteristics.

[0074] The sampling period was set to 24-hour continuous sampling, starting at 00:00 and ending at 00:00 the following day, to ensure complete diurnal variation data were collected. Environmental parameters such as flow rate, temperature, humidity, and atmospheric pressure were recorded every hour during the sampling process for subsequent data quality control. The ambient temperature was controlled within the range of 15-35℃, and the relative humidity was less than 85%. In case of rainfall, the sampler automatically activated its rain protection device to prevent rainwater from entering the sampling system.

[0075] After sampling, the filter membrane must be replaced in a clean bench. Operators should wear disposable polyethylene gloves and use stainless steel tweezers to handle the filter membrane, avoiding direct contact. The sampled filter membrane should be immediately placed in a pre-cleaned polyethylene filter membrane box, labeled with the sampling date, time, location, and flow rate, and then stored in a 4°C refrigerator for no more than 30 days. Blank filter membranes are processed under the same conditions to remove background interference.

[0076] The dissolution treatment unit employs a simulated lung fluid extraction method to accurately reflect the bioavailability of water-soluble heavy metals in particulate matter. The preparation of the acetate-sodium acetate buffer solution strictly follows standard operating procedures: first, prepare 1L of 0.1% acetic acid solution by dissolving 1mL of glacial acetic acid (purity ≥99.7%) in 999mL of ultrapure water; then prepare 1L of 0.1% sodium acetate solution by dissolving 1g of anhydrous sodium acetate (analytical grade) in 1L of ultrapure water. Mix the two solutions in equal volumes and adjust the pH to 4.5±0.1 using a pH meter (accuracy ±0.01). If necessary, fine-tune with 0.1mol / L acetic acid or 0.1mol / L sodium hydroxide.

[0077] After the buffer solution is prepared, it is filtered through a 0.22μm polyethersulfone membrane filter to remove any possible particulate impurities. It is then dispensed into pre-cleaned polypropylene containers and stored at 4°C for later use. Each batch of buffer solution has a shelf life of one month; the pH and ionic strength must be retested before use. Ultrapure water requires a resistivity ≥18.2 MΩ·cm, total organic carbon (TOC) content <4 ppb, and heavy metal content <0.1 μg / L.

[0078] Sample dissolution was performed in a cleanroom at a controlled temperature of 25±2℃ and relative humidity of 45-65%. The sampled filter membrane was removed from its storage container and cut into 2mm×2mm pieces using ceramic scissors, avoiding contamination with metal tools. All the shredded filter membrane was transferred to 50mL polypropylene centrifuge tubes, and 30mL of buffer solution was added to completely cover the membrane fragments. The centrifuge tubes were then sealed and placed in an ultrasonic cleaner at 300W power, 40kHz frequency, and for 30 minutes. During ultrasonication, the centrifuge tubes were manually shaken every 10 minutes to ensure even dissolution.

[0079] After ultrasonic treatment, the centrifuge tubes were transferred to a constant-temperature shaker for further extraction at 150 rpm for 2 hours, maintaining a temperature of 25 ± 2℃. The centrifuge tubes were kept sealed during the shaking process to prevent solution evaporation and external contamination. After shaking, the samples were allowed to stand at room temperature for 30 minutes to allow insoluble particles to settle completely. Direct sunlight was avoided throughout the dissolution process to prevent photochemical reactions from affecting the heavy metal speciation.

[0080] The filtration and purification process employs a staged filtration strategy. First, coarse filtration removes large particulate impurities, followed by fine filtration to obtain a clear test solution. The coarse filtration uses a 47mm diameter glass fiber membrane (1.0μm pore size), which is pre-washed with ultrapure water and dried. The dissolved sample solution is then slowly passed through the coarse membrane using gravity filtration, with the filtration rate controlled at 5-10 mL / min to avoid membrane damage. The filtration apparatus, including the funnel, membrane holder, and collection bottle, is made entirely of plastic to prevent metal ion contamination.

[0081] Fine filtration uses a 0.45 μm polyethersulfone (PES) membrane filter, which exhibits low protein adsorption and good chemical compatibility, with very low non-specific adsorption of heavy metal ions. Before use, the filter is pre-rinsed with 50 mL of ultrapure water to remove any residual surfactants and preservatives that may have remained from the manufacturing process. The pre-rinse solution is discarded, and then the coarsely filtered sample solution is slowly injected into the filter using a syringe at a filtration rate controlled at 2-3 mL / min.

[0082] The filtration process employed positive pressure filtration, using a 50mL disposable syringe as the propulsion device. Pressure was controlled between 0.1-0.2MPa to avoid excessive pressure that could damage the filter membrane or force heavy metal ions into its pores. The filtrate was collected in pre-cleaned 50mL polypropylene centrifuge tubes, avoiding prolonged contact with air during collection to prevent oxidation or volatilization. The filtration volume for each sample was accurately recorded for volume correction in subsequent concentration calculations.

[0083] After filtration, the pH of the filtrate is immediately measured to ensure it remains within the range of 4.5 ± 0.1. If the pH deviates, the buffer solution must be prepared again and the dissolution and extraction process repeated. The filtrate is stored at 4°C for a maximum of 7 days. During storage, the container must be tightly sealed to prevent evaporation, concentration, and external contamination. A blank control is simultaneously processed for each batch of samples; this is a buffer solution without a filter membrane that undergoes the same treatment process to remove background and contaminants.

[0084] The electrolytic cell employs a three-electrode system design. The cell body is made of high borosilicate glass, with a volume precisely controlled at 10.00±0.05mL. The inner wall is smooth and scratch-free, ensuring uniform current distribution. The bottom of the electrolytic cell is conical with a 60° cone angle, facilitating sediment collection and cleaning. Five standard ground glass joints are located at the top of the cell, used for installing the working electrode, reference electrode, auxiliary electrode, vent pipe, and temperature probe, respectively. Each joint is sealed with a PTFE sealing ring to ensure system airtightness.

[0085] The magnetic stirring system is built into the bottom of the electrolytic cell. The stir bar is a PTFE-coated magnetic rotor, 15mm in length and 5mm in diameter, with a smooth, defect-free surface. The stirrer is driven by a brushless DC motor with continuously adjustable speed ranging from 50-800 rpm and a speed stability of ±2 rpm. The stirring rate is set to 400 rpm, which ensures thorough mixing of the solution without generating excessive convection that could affect the diffusion layer on the electrode surface. During stirring, contact between the stir bar and the electrodes is avoided; the electrode insertion depth is controlled to 3-5mm from the bottom of the cell.

[0086] The ventilation system removes dissolved oxygen from the solution using high-purity nitrogen (≥99.99%). Moisture and organic matter are further removed through a two-stage purifier consisting of molecular sieves and activated carbon. The gas flow rate is controlled by a precision gas flow meter, set at 100 mL / min, with a ventilation time of 300 seconds. The ventilation tube is made of polytetrafluoroethylene (PTFE) with an inner diameter of 2 mm. The tube opening is designed with a porous distributor to generate fine, uniform bubbles, improving deoxygenation efficiency. Monitoring begins immediately after ventilation to prevent the re-dissolution of oxygen from the air.

[0087] The electrolytic cell temperature control system employs a circulating water bath, with the temperature set at 25.0±0.1℃. A temperature probe is inserted into the electrolyte for real-time monitoring. Temperature control accuracy is achieved through a PID controller with a response time of <30 seconds. The outer wall of the electrolytic cell is covered with insulation material to reduce heat loss. Temperature variations during testing do not exceed ±0.2℃, ensuring the stability and reproducibility of the electrode reactions.

[0088] The suspended mercury electrode adopts the form of a suspended mercury drop electrode (HMDE), with the mercury drop volume controlled at 2.0±0.1 mm. 3 The corresponding electrode area is approximately 2.0 mm². 2 The mercury droplets are controlled using a precision micro-syringe, and a new droplet is replaced before each test to ensure the electrode surface is clean. The mercury material used is triple-distilled high-purity mercury (purity ≥99.9999%), stored in a sealed container under inert gas protection. After the mercury droplets form, surface cleaning is performed in a blank electrolyte solution by applying a +0.3V potential for 30 seconds to remove surface oxides and organic contaminants.

[0089] Bismuth film electrodes, as an environmentally friendly alternative to mercury electrodes, utilize glassy carbon electrodes (3 mm in diameter) as the substrate electrode, with surfaces polished to a mirror finish using 0.3 μm and 0.05 μm alumina. The bismuth film is prepared via in-situ electrodeposition: in a substrate containing 50 μg / L Bi... 3+ In a 0.1 mol / L acetic acid solution, a constant potential of -1.2 V was applied for deposition for 120 seconds, with the solution stirred during the deposition process to ensure uniform deposition. The deposited bismuth film was approximately 50-100 nm thick and exhibited good conductivity and catalytic activity. The bismuth film needed to be redeposited before each detection to ensure stable electrode performance.

[0090] The reference electrode is either a saturated calomel electrode (SCE) or a silver-silver chloride electrode (Ag / AgCl). The electrode filling solution is changed periodically to maintain potential stability. The Ag / AgCl electrode is filled with 3 mol / L KCl solution, and its electrode potential relative to the standard hydrogen electrode is +0.197 V (25℃). Before use, the electrode needs to be activated by soaking in deionized water for 30 minutes. During use, the potential is calibrated periodically, with a deviation not exceeding ±2 mV. The electrode is stored in the appropriate storage solution, avoiding drying and contamination.

[0091] The auxiliary electrode is a spiral platinum wire electrode with a diameter of 0.5 mm, a spiral length of 30 mm, and a surface area of ​​approximately 10 mm². 2 Before use, the platinum electrode should be cleaned by soaking in dilute nitric acid, followed by thorough rinsing with ultrapure water. The platinum electrode surface was activated using cyclic voltammetry: in 0.5 mol / L sulfuric acid, at a potential range of -0.2 V to +1.2 V, a scan rate of 100 mV / s, for 20 cycles until the cyclic voltammogram stabilized. The activated platinum electrode exhibited good electrocatalytic activity and stability.

[0092] The electrochemical workstation adopts a modular design, including a potential control unit, a current measurement unit, a data acquisition unit, and a software control system. The potential control unit uses a high-precision digital-to-analog converter (DAC) with a potential resolution of 0.1mV, potential stability of ±0.5mV, and potential scan linearity >99.9%. The current measurement unit is equipped with a multi-range current amplifier, with a current measurement range of 10⁻¹² A to 10⁻³ A, a resolution of 0.1pA, and a noise level <0.5pA.

[0093] The differential pulse stripping voltammetry (DPASV) detection procedure consists of three stages: pretreatment, enrichment, and detection. The pretreatment stage involves holding at +0.2V for 30 seconds to remove impurities and oxides from the electrode surface. In the enrichment stage, a negative potential is applied to reduce and enrich the target metal ions on the electrode surface. The enrichment potential is determined based on the reduction potential of different metals: -0.8V for Cd, -0.7V for Pb, -0.3V for Cu, and -1.2V for Zn. The enrichment time is adjusted according to the sample concentration: 300 seconds for concentrations below 1 μg / L, 180 seconds for concentrations of 1-10 μg / L, and 60 seconds for concentrations of 10-100 μg / L.

[0094] The detection phase employs a differential pulse mode, scanning from the enriched potential towards the positive potential up to +0.5V. Pulse parameters are set as follows: pulse amplitude 50mV, pulse width 50ms, sampling time 17ms, step potential 2mV, and scan rate 20mV / s. Current sampling is performed before the pulse ends, using a dual-sampling technique to measure the current difference before and after the pulse, effectively eliminating interference from background current and charging current. The data acquisition frequency is set to 1000Hz to ensure sufficient acquisition of the current response signal.

[0095] During the testing process, the system automatically records the current-potential curve and displays it on the computer screen in real time. The software has peak identification and peak area calculation functions, automatically identifying the characteristic peaks of heavy metals and calculating peak current and peak potential. The system also has data storage, export, and post-processing functions; the test data is saved in a standard format for easy subsequent analysis and processing. After each test, the electrodes are automatically cleaned in preparation for the next test.

[0096] Simultaneous multi-element detection employs a wide-window scanning technique using differential pulse voltammetry, with a potential scan range from -1.3V to +0.6V, covering the oxidation peak potentials of all target metals. The characteristic peak potentials for the eight metals are: Zn (-1.05V), Cd (-0.65V), Pb (-0.45V), Cu (-0.15V), As (+0.05V), Se (+0.18V), V (+0.35V), and Cr (+0.52V). These peak potentials vary within ±20mV, depending on the pH, ionic strength, and coexisting ions of the solution.

[0097] To avoid interference between metals, a segmented enrichment technique is employed: first, easily reduced metals such as Zn, Cd, and Pb are enriched at -1.2V for 180 seconds; then, Cu is enriched at -0.3V for 60 seconds; finally, easily oxidized metals such as As and Se are enriched at +0.1V for 60 seconds. Immediately after each enrichment stage, the corresponding potential range is measured to obtain clearly separated metal peaks. When the distance between adjacent metal peaks is less than 100mV, a Gaussian-Lorentz mixture function is used to fit the peak shape, and overlapping peaks are separated through mathematical deconvolution. The software incorporates a standard peak shape database containing peak shape parameters for each metal under different conditions. For severely overlapping peaks, a second derivative technique is used to enhance peak separation.

[0098] The machine learning analysis module uses a deep convolutional neural network to automatically identify and extract features from the current-voltage spectrum. It preprocesses the original spectrum using wavelet denoising and baseline correction algorithms, establishes a nonlinear mapping model of the relationship between metal ion concentration and peak current, and outputs predicted values ​​of heavy metal element concentration.

[0099] The feature extraction network employs an improved CNN network structure, comprising a sequentially connected input layer, multiple convolutional layers, pooling layers, and fully connected layers. The network outputs predicted concentrations of eight metal ions. Training utilizes the Adam optimizer with a learning rate of 0.001 and a mean squared error loss function. The network input layer receives one-dimensional spectral data of length 2048, corresponding to a potential range of -1.3V to +0.6V, with a sampling interval of approximately 0.9mV. The first convolutional layer uses 32 kernels with a kernel size of 15, a stride of 1, and the same padding method. The ReLU activation function is used. Larger kernels can capture the overall shape characteristics of electrochemical peaks.

[0100] Batch normalization is applied after convolutional operations to accelerate network convergence and improve training stability. Max pooling is used in the pooling layers with a window size of 2 and a stride of 2, halving the feature map size to 1024. The second convolutional layer uses 64 kernels with a kernel size of 11 to capture more detailed peak features and inter-peak relationships. The third convolutional layer uses 128 kernels with a kernel size of 7. Each convolutional layer is followed by ReLU activation and batch normalization. Pooling layers are applied after the second and third convolutional layers, resulting in a final feature map size of 256×128. Residual connections (ResNet structure) are introduced into the network to mitigate the degradation problem of deep networks.

[0101] The residual block structure is: x→Conv→BN→ReLU→Conv→BN→Add(x)→ReLU, where Add(x) represents a skip connection to the input. The network contains 3 residual blocks, and each residual block contains 2 convolutional layers.

[0102] The fully connected layer is designed as a three-layer structure: the first layer has 512 neurons, the second layer has 256 neurons, the third layer has 128 neurons, and finally the output layer has 8 neurons, corresponding to the concentration prediction of 8 metal ions. Each fully connected layer is followed by a Dropout layer with a dropout rate set to 0.3 to prevent overfitting; the output layer does not use an activation function and directly outputs the concentration prediction value.

[0103] The training dataset was constructed using a combination of experimental and synthetic data. The experimental data consisted of detection results from 1000 real-world environmental samples, while the synthetic data was generated by superimposing metal peaks of different concentrations onto standard spectra. The dataset was divided into training, validation, and test sets in a 7:2:1 ratio to ensure a uniform distribution of metal concentrations across all sets.

[0104] Data augmentation techniques include: (1) Time-shifting augmentation: random shifting of the spectrum along the potential axis by ±5mV to simulate potential calibration error; (2) Amplitude scaling: random scaling of the peak current by 0.8-1.2 times to simulate electrode state changes; (3) Noise injection: adding Gaussian white noise with a signal-to-noise ratio range of 20-50dB; (4) Baseline drift: adding polynomial baseline drift with polynomial order 1-3. Each original sample generates 10 variants through data augmentation, increasing the number of training samples to 10,000.

[0105] The machine learning analysis module first preprocesses the original voltammetric spectrum using the wavelet transform function: Where f(t) is the original volt-ampere signal, ψ * Here, is the Daubechies wavelet basis function, a is the scaling parameter used to control the frequency characteristics of the wavelet, and b is the translation parameter used to control the temporal positioning of the wavelet.

[0106] The objective function for baseline correction is: Where y i For the observed signal, z i For the corrected baseline, w i λ is the weighting coefficient used to control the importance of each point, λ is the smoothing parameter used to control the smoothness of the baseline, and Δ 2 z i It is a second-order difference operator used to constrain the continuity of the baseline.

[0107] The data processing output module includes a quantitative analysis unit and a pollution source identification unit. It uses the standard addition method to perform quantitative concentration calculation and identifies pollution sources by comparing multi-element fingerprint features with a pollution source database.

[0108] The data verification output module includes a quantitative analysis unit and a pollution source identification unit;

[0109] The quantitative analysis unit employs the standard addition method, sequentially adding target metal standard solutions of known concentrations to the sample solution. A linear relationship is established between the added amount C and the peak current response I: I = a·C + b. The regression coefficients a and b are obtained using the least squares method, thus the original sample concentration is C. x = -b / a, and calculate the combined standard uncertainty;

[0110] The pollution source identification unit establishes a multi-element fingerprint feature database, including feature fingerprints of six types of pollution sources: motor vehicle emissions, coal combustion, waste incineration, metal smelting, electronics manufacturing, and dust. A positive definite matrix factorization model is used. Perform source parsing, where X ij Let G be the concentration value of the j-th element in the i-th sample. ik For the source contribution matrix, F kj For the source component spectral matrix, E ij The residual matrix represents the error term of the model fitting.

[0111] The characteristic ratio ranges for various water-soluble metal pollution sources are as follows: Motor vehicle emission sources: Pb / Cd > 5 and Zn / Cu = 2-8; Coal combustion sources: As / Se > 10 and V / Ni = 0.5-2 (this ratio needs to be recalibrated regularly depending on the coal type in different study areas); Waste incineration sources: Pb / As < 2 and Cd / Zn > 0.01; Metal smelting sources: Cu / Pb > 3 and Ni / Cr = 0.2-2 (this ratio needs to be judged comprehensively in conjunction with smelting processes and raw material composition); Electronic manufacturing sources: Se / As > 0.5 and Cd / Pb > 0.1 (this ratio needs to be judged comprehensively considering process characteristics and detection methods); Dust sources: uniform distribution of element concentrations. In practical applications, multi-indicator collaborative analysis is used, and thresholds are calibrated using local background data to improve source identification accuracy.

[0112] The distance D between the sample fingerprint and the standard pollution source fingerprint was calculated using Euclidean distance. k , Where: r sample,i Let r be the ratio of the i-th element in the sample. source,k,i Let n be the ratio of the i-th element in the k-th pollution source, and n be the logarithm of the element ratio; select distance value D. k The smallest pollution source is considered the primary pollution source type.

[0113] The standard addition method quantitative calculation process includes the following steps:

[0114] (1) Measure the peak current I0 of the original sample;

[0115] (2) Add volumes of V to the same sample solution sequentially. add Concentration of C std The standard solution was used to measure the peak current I after its addition.add ;

[0116] (3) Establish a linear relationship between peak current and added concentration: I add =I0+k×C add ×V add / (V0+V add ), where k is the response coefficient, V0 is the original sample volume, and C add This represents the concentration increment after adding the standard solution;

[0117] (4) The original sample concentration was calculated using the linear extrapolation method: C sample =|I0 / k|×(V0+V add ) / V0;

[0118] (5) Perform at least three standard additions and verify the reliability of the linear relationship through regression analysis, requiring a correlation coefficient R0. 2 ≥0.995.

[0119] The system also includes a calibration module, which establishes the concentration-current response relationship using a multi-point standard curve method. The standard solution concentration gradients are 0.1 μg / L, 0.5 μg / L, 1.0 μg / L, 5.0 μg / L, 10 μg / L, 50 μg / L, 100 μg / L, 500 μg / L, and 1000 μg / L, covering a linear range of more than three orders of magnitude. The standard solutions are prepared using a stepwise dilution method, starting with a 1000 μg / mL standard stock solution, and then sequentially preparing intermediate stock solutions of 100 μg / mL, 10 μg / mL, and 1 μg / mL. Each dilution is made up to volume using volumetric flasks with an accuracy grade A and a volume tolerance of ±0.03 mL (50 mL). The dilution water is ultrapure water with a resistivity ≥18.2 MΩ·cm. Different concentrations of mixed standard solutions are sequentially added to the sample to establish a linear relationship between the amount added and the peak current. Each metal should be added at least 5 times its concentration, with the amount added covering 0.5 to 5 times the sample concentration; the correlation coefficient of linear regression should be greater than 0.995, otherwise the standard solution should be prepared again or the instrument status should be checked.

[0120] Measurement uncertainty was assessed using the GUM (Guide to Uncertainty) method to identify and quantify all sources of uncertainty. The main sources of uncertainty included: (1) uncertainty of standard solution concentration u(C_std); (2) uncertainty of volume measurement u(V); (3) uncertainty of instrument reproducibility u(rep); (4) uncertainty of regression analysis u(reg); and (5) uncertainty of sample homogeneity u(hom).

[0121] The concentration uncertainty of the standard solution comes from the expanded uncertainty given in the standard substance certificate and is treated as a normal distribution. For a 1000 μg / mL cadmium standard solution, the certificate gives an expanded uncertainty of 5 μg / mL (k = 2), then the standard uncertainty u(C_std) = 5 / 2 = 2.5 μg / mL, and the relative standard uncertainty u_r(C_std) = 2.5 / 1000 = 0.0025. The uncertainty of the stepwise dilution process is calculated according to the uncertainty propagation law.

[0122] The uncertainty of volume measurement includes the effects of pipette, volumetric flask and temperature; the uncertainty of pipette is given according to the calibration certificate and is treated as a rectangular distribution; the permissible error of the Class A 50 mL volumetric flask is ±0.05 mL, treated as a rectangular distribution, with a standard uncertainty u(V) = 0.029 mL.

[0123] Taking an environmental monitoring project in an industrial park as an example, the park contains electronics manufacturing companies, metal processing plants, and a coal-fired power plant, surrounded by residential areas and major traffic arteries. Environmental protection departments need to conduct long-term monitoring of water-soluble heavy metals in atmospheric particulate matter in the park and surrounding areas to identify major pollution sources and assess health risks. Traditional monitoring methods suffer from problems such as long detection cycles (7-10 days), high costs (1500 yuan per sample), and the inability to simultaneously detect multiple elements, making it difficult to meet the needs of real-time monitoring and rapid emergency response.

[0124] Three monitoring points were set up within the industrial park for system deployment: Point A was located 200 meters from the electronics factory within the industrial zone; Point B was located 500 meters from the metal processing plant at the park boundary; and Point C was located 100 meters from the main road in a residential area. Each monitoring point was equipped with a complete detection system, including a sampling pretreatment module, an electrochemical detection module, a machine learning analysis module, and a data verification and output module. System calibration was performed using national standard reference materials, containing a mixed standard solution of eight target metals: cadmium (10 μg / L), vanadium (15 μg / L), chromium (25 μg / L), nickel (20 μg / L), selenium (8 μg / L), arsenic (12 μg / L), manganese (50 μg / L), and lead (30 μg / L). Calibration results showed that the limits of detection were all below 0.05 μg / L, the linear correlation coefficient was greater than 0.998, the recovery rate was in the range of 95-105%, and the relative standard deviation was less than 3%. The machine learning model has a root mean square error of 2.8% and a mean absolute percentage error of 4.5% on the validation set, which is significantly better than the 8-12% error range of traditional electrochemical methods.

[0125] Continuous monitoring was conducted for 30 days, with 24-hour sample collection per day, resulting in 90 valid sample data points. The sampling process was strictly performed according to technical procedures. 2.5The sampling flow rate was 16.7 L / min. The quartz fiber membrane was pretreated at 450°C, and the sample was extracted using an acetate-sodium acetate buffer solution at pH 4.5 ± 0.1. Ultrasonic extraction was performed for 30 minutes, followed by purification via 0.45 μm membrane filtration. Electrochemical detection employed differential pulse stripping voltammetry with a bismuth membrane electrode as the working electrode. The enrichment time was 180 seconds, and the scan rate was 20 mV / s. The detection time, from sample pretreatment to result output, was only 2 hours, significantly shorter than the 7-10 days of traditional ICP-MS methods. A machine learning analysis module automatically identified the voltammetric spectra, enabling simultaneous quantification of eight metals without manual intervention.

[0126] The test results in the industrial area at point A showed that the concentrations of cadmium, vanadium, chromium, nickel, selenium, arsenic, manganese, and lead were 2.3 ± 0.4 μg / L, 5.8 ± 1.2 μg / L, 8.9 ± 1.8 μg / L, 6.7 ± 1.1 μg / L, 12.5 ± 2.1 μg / L, 3.4 ± 0.7 μg / L, 15.2 ± 2.8 μg / L, and 4.6 ± 0.9 μg / L. The detection results at the boundary of the park at point B were relatively low, with cadmium concentrations of 1.8 ± 0.3 μg / L, vanadium concentrations of 4.2 ± 0.9 μg / L, chromium concentrations of 6.1 ± 1.2 μg / L, nickel concentrations of 4.9 ± 0.8 μg / L, selenium concentrations of 8.7 ± 1.6 μg / L, arsenic concentrations of 2.8 ± 0.5 μg / L, manganese concentrations of 11.4 ± 2.1 μg / L, and lead concentrations of 3.2 ± 0.6 μg / L. The lowest test results were found in the residential area at point C, with cadmium concentrations of 0.9 ± 0.2 μg / L, vanadium concentrations of 2.1 ± 0.4 μg / L, chromium concentrations of 3.5 ± 0.7 μg / L, nickel concentrations of 2.8 ± 0.5 μg / L, selenium concentrations of 4.3 ± 0.8 μg / L, arsenic concentrations of 1.6 ± 0.3 μg / L, manganese concentrations of 7.8 ± 1.4 μg / L, and lead concentrations of 5.8 ± 1.1 μg / L.

[0127] Through multi-element fingerprint feature analysis and PMF source apportionment, the main pollution sources at each location were successfully identified. At point A, the pollution source contribution rates were distributed as follows: electronics manufacturing 45%, metal smelting 28%, vehicle emissions 15%, coal combustion 8%, and dust 4%. Feature ratio verification results showed a selenium-arsenic ratio of 3.68 (greater than 0.5) and a cadmium-lead ratio of 0.5 (greater than 0.1), consistent with electronics manufacturing source characteristics. Distance calculation results showed a fingerprint distance of 0.23 from the electronics manufacturing source, with a confidence level of 95%. At point B, the pollution source contribution rates were: metal smelting 52%, coal combustion 23%, vehicle emissions 13%, electronics manufacturing 8%, and dust 4%. Feature ratio verification results showed a copper-lead ratio of 2.87 (close to 3) and a nickel-chromium ratio of 0.8 (within the 0.2-2 range), consistent with metal smelting source characteristics. The PMF apportionment Q value was 156, meeting the requirement that Q-theory is less than Q-robust. At point C, the pollution source contribution rates were: vehicle emissions 58%, dust 22%, coal combustion 12%, metal smelting 6%, and electronics manufacturing 2%. The characteristic ratio verification results show that the lead-cadmium ratio is 6.44, which is greater than 5, and the zinc-copper ratio is 3.2, which is within the range of 2-8, consistent with the characteristics of vehicle emission sources. Time-varying analysis shows that the contribution rate of vehicle emissions increases to 75% during morning and evening peak hours.

[0128] To verify the accuracy of the system, some samples were collected at the same time and sent to an authoritative third-party laboratory for ICP-MS comparative analysis, as shown in Table 1.

[0129] Table 1. Experimental Comparison and Analysis Results

[0130]

[0131] The comparison results show that the relative deviations between this system and ICP-MS are both less than 2%, while the relative deviations of traditional electrochemical methods are in the range of 9.8-16.2%, which fully demonstrates the high accuracy of the system.

[0132] From an economic perspective, the cost of testing a single sample using this system is approximately 150 yuan, while the traditional ICP-MS method requires 1500 yuan, representing a 90% cost reduction. Regarding the testing cycle, this system requires only 2 hours, while the traditional method takes 7-10 days, improving efficiency by more than 40 times. Automated testing reduces manual operation by 80%, significantly reducing human error and labor costs. In terms of equipment investment, the initial investment is 800,000 yuan, while traditional ICP-MS equipment requires 3 million yuan, resulting in a 73% cost saving.

[0133] Analysis of 30 days of continuous monitoring data revealed significant environmental issues. Selenium and arsenic concentrations at monitoring point A exceeded WHO recommendations, necessitating strengthened emission controls for electronics manufacturing enterprises. The significant difference in heavy metal concentrations between weekdays and weekends at monitoring point B confirmed the substantial contribution of industrial emissions. Lead concentrations at monitoring point C increased 2.5 times during peak traffic hours, providing a scientific basis for traffic control measures. The pollution source identification accuracy rate at the three monitoring points reached 94%, providing reliable technical support for precise pollution control.

[0134] System stability verification results showed that the system maintained a stable operating rate of 99.2% during 30 days of continuous operation, with only a 0.8% interruption due to power outages. The recovery rate of quality control samples consistently remained within the range of 95-105%, and no contamination was detected in blank sample testing, demonstrating good system cleanliness. The relative standard deviation of the precision of standard sample testing was less than 3%, fully meeting the stringent requirements of environmental monitoring.

[0135] This technology has achieved several significant innovative breakthroughs. In simultaneous multi-element detection, traditional methods require separate determination of each element, while this system can obtain the concentrations of eight metals in a single detection, increasing efficiency by eight times. Intelligent identification technology automatically identifies voltammetric peaks through a machine learning model, achieving an accuracy rate of 98% and completely eliminating subjective errors from manual identification. Real-time pollution source analysis combined with the PMF algorithm enables automatic identification of pollution sources, providing timely decision-making support for environmental management departments. In terms of detection performance, the detection limits for all eight metals are below 0.05 micrograms per liter, and the relative standard deviation of precision is less than 3%, meeting the stringent requirements of environmental monitoring. The system is highly adaptable, capable of accurately detecting various types of environmental samples, from highly polluted samples in industrial areas to low-concentration samples in clean areas. This case fully demonstrates the technical advantages and economic benefits of a machine learning-based water-soluble heavy metal detection system in practical applications, providing important reference value for the development of environmental monitoring technology.

[0136] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A machine learning-based detection and analysis system for water-soluble heavy metals in an ecosystem, used to detect water-soluble harmful heavy metal elements in atmospheric particulate matter, characterized in that... The system comprises, in sequence: a sampling preprocessing module, an electrochemical detection module, a machine learning analysis module, and a data verification output module; The sampling preprocessing module is used to collect environmental samples and perform dissolution and filtration preprocessing. The electrochemical detection module uses stripping voltammetry for heavy metal enrichment and detection, and includes an electrolytic cell, working electrode, reference electrode, auxiliary electrode, and electrochemical workstation. The machine learning analysis module uses a deep convolutional neural network to automatically identify and extract features from the voltammetric spectrum. It preprocesses the original spectrum using wavelet denoising and baseline correction algorithms, establishes a nonlinear mapping model between metal ion concentration and peak current, and outputs predicted values ​​of heavy metal element concentration. The data verification output module includes a quantitative analysis unit and a pollution source identification unit. It uses the standard addition method to perform quantitative concentration calculation and identifies pollution sources by comparing multi-element fingerprint features with various pollution source databases. The quantitative analysis unit employs the standard addition method, sequentially adding target metal standard solutions of known concentrations to the sample solution to establish the addition volume. With peak current response linear relationship The regression coefficients were obtained using the least squares method. and The original sample concentration is And calculate the combined standard uncertainty; The pollution source identification unit establishes a multi-element fingerprint feature database, including feature fingerprints of six types of pollution sources: motor vehicle emissions, coal combustion, waste incineration, metal smelting, electronics manufacturing, and dust. A positive definite matrix factorization model is used. Perform source resolution, where For the first The first sample Concentration values ​​of the elements, For the source contribution matrix, The source component spectral matrix, The residual matrix represents the error terms of the model fit. The distance between the sample fingerprint and the standard pollution source fingerprint was calculated using Euclidean distance. , in: For sample number The ratio of each element For the first Type of pollution source The ratio of each element Logarithm of element ratio; select distance value The smallest pollution source is considered the primary pollution source type.

2. The machine learning-based detection and analysis system for water-soluble heavy metals in an ecosystem according to claim 1, characterized in that, The system detects water-soluble harmful heavy metal elements including Cd, V, Cr, Ni, Se, As, Mn, and Pb, with a detection concentration range of 0.1-1000 μg / L. The electrochemical detection was performed using differential pulse stripping voltammetry (DPASV). The detection parameters were set as follows: enrichment potential voltage from -1.2V to -0.8V, enrichment time from 60 to 300 s, scan rate from 20 mV to s, pulse amplitude from 50 mV, and pulse width from 50 ms.

3. The machine learning-based detection and analysis system for water-soluble heavy metals in an ecosystem according to claim 1, characterized in that, The sampling preprocessing module specifically includes: an atmospheric particulate matter sampling unit, a dissolution treatment unit, and a filtration and purification unit; The atmospheric particulate matter sampling unit uses a quartz fiber filter membrane to collect PM2.

5. 2.5 and PM 10 Particulate matter, sampling flow rate 16.7 L / min, sampling time 24 hours; The dissolution unit dissolves the collected particulate matter sample in an acetate-sodium acetate buffer solution with a pH of 4.5 ± 0.

1. The buffer solution consists of 0.1% acetate (CH3COOH) + 0.1% sodium acetate (CH3COONa) + ultrapure water. The dissolution temperature is 25 ± 2℃, followed by ultrasonic treatment for 30 minutes and oscillation extraction for 2 hours. The filtration and purification unit uses a 0.45μm polyethersulfone membrane filter to remove insoluble particles.

4. The machine learning-based detection and analysis system for water-soluble heavy metals in an ecosystem according to claim 2, characterized in that, The working electrode is a suspended mercury electrode or a bismuth film electrode with an electrode area of ​​2 mm². The reference electrode is an Ag / AgCl electrode and the auxiliary electrode is a platinum wire electrode. The electrolytic cell has a volume of 10 mL and is equipped with a magnetic stirrer with a stirring speed of 400 rpm and a nitrogen deoxygenation time of 300 s. The electrochemical workstation includes a potential control unit and a scanning detection unit. The potential control unit uses a potentiostat to control the enrichment potential, which is set according to the target metal. The scanning detection unit uses a differential pulse mode to scan from the enrichment potential to the positive potential direction up to +0.5V, records the current-potential curve, and has a sampling interval of 2mV and a data acquisition frequency of 1000Hz.

5. The machine learning-based detection and analysis system for water-soluble heavy metals in an ecosystem according to claim 4, characterized in that, The feature extraction employs an improved CNN network structure, comprising an input layer, multiple convolutional layers, pooling layers, and fully connected layers connected in sequence. The network outputs predicted concentration values ​​for eight metal ions. Training uses the Adam optimizer with a learning rate of 0.001 and a loss function of mean squared error. The machine learning analysis module first preprocesses the original voltammetric spectrum using the wavelet transform function: ,in, The original volt-ampere signal, For Daubechies wavelet basis functions, The scaling parameter is used to control the frequency characteristics of the wavelet. The translation parameters are used to control the time positioning of the wavelet; The objective function for baseline correction is: ,in For observing signals, For the corrected baseline, The weighting coefficients are used to control the importance of each point. The smoothing parameter is used to control the degree of smoothness of the baseline. It is a second-order difference operator used to constrain the continuity of the baseline.

6. The machine learning-based detection and analysis system for water-soluble heavy metals in an ecosystem according to claim 5, characterized in that, The characteristic ratio ranges for various pollution sources are as follows: for motor vehicle emission sources, Pb / Cd > 5 and Zn / Cu = 2-8; for coal combustion sources, As / Se > 10 and V / Ni = 0.5-2; for waste incineration sources, Pb / As < 2 and Cd / Zn > 0.01; for metal smelting sources, Cu / Pb > 3 and Ni / Cr = 0.2-2; for electronic manufacturing sources, Se / As > 0.5 and Cd / Pb > 0.1; and for dust sources, the concentrations of each element are evenly distributed.

7. The machine learning-based detection and analysis system for water-soluble heavy metals in an ecosystem according to claim 6, characterized in that, The standard addition method quantitative calculation process includes the following steps: (1) Determine the peak current of the original sample ; (2) Add volumes of [volume] to the same sample solution sequentially. Concentration is The standard solution was used to measure the peak current after its addition. ; (3) Establish the linear relationship between peak current and added concentration: ,in For the response coefficient, The original sample volume, This represents the concentration increment after adding the standard solution; (4) Calculate the original sample concentration using the linear extrapolation method: ; (5) Perform at least 3 standard additions and verify the reliability of the linear relationship through regression analysis, requiring a correlation coefficient R² ≥ 0.

995.

8. The machine learning-based detection and analysis system for water-soluble heavy metals in an ecosystem according to claim 7, characterized in that, The system also includes a calibration module, which uses a multi-point standard curve method to establish the concentration-current response relationship. The standard solution concentration gradients are 0.1 μg / L, 0.5 μg / L, 1.0 μg / L, 5.0 μg / L, 10 μg / L, 50 μg / L, 100 μg / L, 500 μg / L, and 1000 μg / L, covering a linear range of more than three orders of magnitude.

Images