An on-line detection system and method for atmospheric aerosol acidity

Abstract

The application discloses an online detection method and system for atmospheric aerosol acidity, and belongs to the field of atmospheric pollutant detection, and comprises the following steps: acquiring aerosol data and meteorological data of an atmospheric observation station, and pre-processing the acquired data; adopting a Metastable-Forward mode thermodynamic model ISORROPIA II to calculate the pre-processed data, so as to obtain a theoretical value pH of aerosol acidity; taking preset measurement parameters as characteristic variables and taking the theoretical value pH of aerosol acidity as a target variable, a plurality of machine learning models are constructed, cross validation is carried out, and the machine learning model with the optimal performance is selected; adopting mass spectrum analysis to acquire real-time aerosol component data and current meteorological data, and utilizing the optimal machine learning model, real-time atmospheric aerosol acidity is calculated. In view of the poor real-time performance of the atmospheric aerosol acidity detection by physical sampling, the application combines a thermodynamic model and machine learning, and does not need to replace detection consumables, so that the real-time performance is improved.

Description

Technical Field

[0001] This application relates to the field of atmospheric pollutant detection, and more specifically, to an online detection system and method for atmospheric aerosol acidity. Background Technology

[0002] Atmospheric aerosols are small solid or liquid particles suspended in the atmosphere. Their acidity (pH value), as one of the most fundamental properties of aerosols, varies significantly under different spatiotemporal environments. Aerosol acidity has many important impacts on the environment, climate, and human health: it can significantly promote the formation of secondary organic aerosols, alter the light absorption characteristics of brown carbon aerosols and thus affect radiation balance, and change the hygroscopic growth characteristics and surface tension of aerosols by influencing gas-liquid partition equilibrium and phase separation processes. Accurately measuring and understanding the variation patterns of atmospheric aerosol acidity has significant scientific and practical value for environmental quality assessment, climate change prediction, and health risk assessment.

[0003] However, direct detection of aerosol pH in ambient air faces numerous technical challenges. Traditional research relies on collecting aerosols onto a filter membrane, extracting them with deionized water in the laboratory, and then analyzing the colorimetric phenomena using acid-base indicators. This offline method generally has low temporal resolution, typically obtaining only one data point every 24 hours in routine observations, focusing primarily on detecting average atmospheric acidity, which does not meet real-time monitoring requirements. In some intensive observations, multiple filter membranes are used to control sampling time and switch between them to improve temporal resolution; however, there is usually a long interval between obtaining the sample and final detection, generally requiring a low-temperature environment to ensure the relative stability of volatile substances in the sample.

[0004] Currently, there are very few online in-situ detection technologies for aerosol acidity, and most rely on optical recognition technologies such as indicator color development and Raman spectroscopy (e.g., CN 116202919 A ​​and CN 119246489 A). These technologies suffer from the following main problems: Low temporal resolution: aerosols still need to be collected through sampling membranes, and the acidity characterization is actually an average value over a certain period of time; High consumable consumption: to achieve higher temporal resolution, sampling membranes need to be replaced frequently, generating a large amount of waste consumables; High measurement uncertainty: measurements based on acid-base indicators are affected by factors such as the quality of indicator preparation, the accuracy of optical signal detection, and ambient temperature, resulting in significant uncertainty; Limited detection conditions: a high relative humidity environment needs to be maintained at the sampling interface to ensure that aerosols enter the liquid phase to meet the indicator color development conditions; Limited detection range: the color change range of a single acid-base indicator is limited (e.g., methyl yellow only covers pH = 2.9 to 4.0), while atmospheric aerosol acidity can be widely distributed between pH = 1 and 8; Insufficient resolution: pH test strips with a large measurement range typically have a resolution of only 1 pH unit, which cannot meet the requirements for accurate measurement; Lack of mechanistic analysis: existing technologies only remain at the level of acidity numerical measurement and fail to provide an explanation of the formation and change mechanism of acidity.

[0005] Taking the "Aerosol Acidity Detection Device and Method Based on Imaging System" disclosed in CN 116202919 A ​​as an example, although this technology achieves a certain degree of online detection, it still adopts the traditional acid-base indicator principle. It requires the collection of samples through components such as a sampler, gas flow sensor, micro air pump, and particulate matter detection sensor, and then uses indicator color development combined with LED light source illumination for detection. This method is limited by various factors such as sampling membrane replacement, environmental humidity control, and indicator color change range, making it difficult to achieve high-precision, wide-range, real-time, and continuous aerosol acidity monitoring. Summary of the Invention

[0006] To address the poor real-time performance of atmospheric aerosol acid detection through physical sampling, this application provides an online detection system and method for atmospheric aerosol acidity. By combining thermodynamic models and machine learning, the system improves real-time performance without requiring replacement of detection consumables.

[0007] One aspect of this application provides a method for detecting atmospheric aerosol acidity, comprising: acquiring aerosol data and meteorological data from atmospheric observation stations, and preprocessing the acquired data; the aerosol data including the mass concentrations of ammonium salts, sulfates, nitrates, and chlorides; the meteorological data including atmospheric temperature and relative humidity; calculating the theoretical aerosol acidity pH using the Metastable-Forward thermodynamic model ISORROPIAII on the preprocessed data; constructing multiple machine learning models using preset measurement parameters as feature variables and the theoretical aerosol acidity pH as the target variable, and selecting the machine learning model with the best performance through cross-validation; acquiring real-time aerosol component data using mass spectrometry analysis and acquiring current meteorological data, and calculating the real-time atmospheric aerosol acidity using the optimal machine learning model.

[0008] The Metastable-Forward mode is an operating mode of the ISORROPIA II thermodynamic model. "Forward" indicates the direction of calculation starting from the total aerosol concentration (gas phase + particulate phase), while "Metastable" signifies that the system is in a metastable state, assuming all inorganic salts exist only in the liquid phase and neglecting solid phase formation. This mode is particularly suitable for aerosol studies in high-humidity environments, as aerosols are more likely to remain in a liquid state under these conditions.

[0009] ISORROPIA II is a thermodynamic equilibrium model specifically designed to simulate the gas-liquid-solid three-phase distribution of inorganic components in atmospheric aerosols. ISORROPIA means "equilibrium" in Greek. This model can calculate the distribution of components such as ammonium salts, sulfates, nitrates, and chlorides among the gas, liquid, and solid phases under specific temperature and relative humidity conditions. ISORROPIA II is an improved version of this model, offering enhanced computational power and a wider range of applications.

[0010] Aerosol acidity refers to the acidity or alkalinity of the liquid phase of atmospheric particulate matter, usually expressed as pH. Similar to aqueous solutions, aerosol acidity is the negative logarithm of the hydrogen ion concentration (or activity) in the liquid phase of the aerosol. Aerosol acidity is a key parameter influencing atmospheric chemical processes, secondary aerosol formation, and the health and climate effects of aerosols. Since it cannot be directly measured, it is typically calculated using thermodynamic models (such as ISORROPIAII) combined with aerosol component concentrations and environmental parameters.

[0011] Furthermore, machine learning models include: linear regression, fine-grained trees, ensemble trees, neural networks, and Gaussian process regression models.

[0012] Furthermore, ensemble trees include bagged trees and boosting trees; Gaussian process regression includes various basis functions and kernel functions.

[0013] Furthermore, the preset measurement parameters include: local time, atmospheric temperature, relative humidity, ammonium salt mass concentration, sulfate mass concentration, nitrate mass concentration, and chloride mass concentration.

[0014] Furthermore, the acquired data is preprocessed, including: sorting the acquired aerosol data and meteorological data in chronological order according to the timestamps of each data point, and aligning the time nodes to obtain a consistent data time series; identifying time periods with no data records based on the data time series, and marking the data points in the data segments as non-NaN; and using the moving window method to detect outliers based on the data time series marked with non-NaN, calculating the local median and the local equivalent absolute median difference (MAD) within N1 moving windows, and marking data points whose difference from the local median is greater than or equal to N2 times the MAD as non-NaN, thus obtaining the preprocessed data.

[0015] In this application, N1 ranges from 3 to 5, and N2 ranges from 2 to 4. "Local" refers to the MAD calculation being performed only within a moving window of size N1, not the entire dataset; "conversion" means multiplying the original MAD value by a constant factor (typically 1.4826) to make it comparable to the standard deviation under the assumption of a normal distribution. The technical advantage of MAD over standard deviation lies in its inherent resistance to outliers, being less affected by extreme values, making it particularly suitable for scenarios with intermittent interference, such as environmental monitoring. When the difference between a data point and the local median exceeds N² times the MAD, it is marked as an outlier (NaN), achieving efficient and robust outlier detection.

[0016] Specifically, this application explicitly identifies missing data instead of imputing or ignoring it by default, thus maintaining the authenticity of the data. This prevents subsequent analysis from incorrectly interpreting missing data as zero or other default values. The use of NaN markers makes missing data traceable, which helps in assessing the completeness of the model's input data.

[0017] Furthermore, the pretreated data was analyzed using the Metastable-Forward thermodynamic model ISORROPIA II to obtain the theoretical aerosol acidity value (pH). This included using the pretreated data as input and calculating the hydrogen ion mass concentration in the aerosol using the Metastable-Forward thermodynamic model ISORROPIA II. And the mass concentration of liquid water (LWC); utilizing the mass concentration of hydrogen ions Given the liquid water mass concentration (LWC), the theoretical aerosol acidity value (pH) is calculated using the following formula:

[0018] Furthermore, using preset measurement parameters as feature variables and the theoretical aerosol acidity value (pH) as the target variable, multiple machine learning models are constructed. Through cross-validation, the machine learning model with the best performance is selected, including: establishing a time series dataset with a one-to-one correspondence between feature variables and target variables; constructing multiple machine learning models based on the time series dataset; performing N-fold cross-validation on each constructed machine learning model; validating the performance of each machine learning model; and selecting the model with the best performance as the optimal machine learning model.

[0019] Furthermore, mass spectrometry was used to obtain real-time aerosol component data, along with current meteorological data. An optimal machine learning model was then used to calculate the real-time atmospheric aerosol acidity (pH). This process included: acquiring real-time aerosol and meteorological data using a mass spectrometer, and preprocessing the acquired data; using the preprocessed data as input, the optimal machine learning model was employed to calculate the real-time atmospheric aerosol acidity (pH) using the following formula. i pH i =f(x) i1 ,x i2 ,.....,x i7 ); where pH i Let f be the real-time atmospheric aerosol acidity at the i-th data point, f be the optimal machine learning model, and x be the value of the aerosol. ij Let be the value of the j-th feature variable for the i-th data point; the feature variables are the mass concentrations of ammonium salt, sulfate, nitrate, and chloride, as well as the local time, atmospheric temperature, and relative humidity.

[0020] In particular, the ISORROPIA II thermodynamic model, based on the thermodynamic equilibrium principle of aerosol systems, can simulate complex thermodynamic processes in gas-liquid-solid multiphase systems. Its calculation results have a sound theoretical foundation and scientific interpretability, and their reliability has been verified in numerous studies. It is one of the most widely used methods for calculating aerosol acidity globally. However, such complex thermodynamic models have certain limitations. For example, they require a large number of input parameters, necessitating the simultaneous acquisition of data from 11 aerosol, gas, and meteorological elements, which is often impossible in many observations. Furthermore, because the sampling periods for aerosols and gases are not consistent, while a few instruments can integrate simultaneous detection of aerosols and gases, the time resolution is often limited to one hour and requires frequent manual maintenance, resulting in insufficient resolution and convenience.

[0021] Therefore, this application simplifies the input parameters by using the theoretically calculated value of aerosol acidity as the training target of the machine learning model, thus preserving the accuracy of the theoretical model and overcoming its practicality shortcomings: the theoretical model provides a reliable "standard value" as a training benchmark; the machine learning model provides efficient computing power and real-time response; and it reduces the dependence on comprehensive observation of atmospheric components.

[0022] Furthermore, after calculating the real-time atmospheric aerosol acidity, the following steps are also included: (1) analyzing the spatiotemporal distribution of aerosol acidity based on the preset measurement parameters and the calculated real-time atmospheric aerosol acidity; (2) analyzing the transport path of atmospheric aerosols using the Lagrange source model, and calculating the contribution (CWT) of different air mass source regions to the aerosol acidity at the observation station using the concentration weighted trajectory method. i,j : Among them, CWT i,j C represents the contribution of grid i to the data at detection location j. j,l Let τ be the magnitude of data j in trajectory l when it reaches the detection location. i,j,l Let N be the dwell time of data j in trajectory l at grid i, and N be the total number of trajectories; (3) Calculate the Shapley value using the game theory-based SHAP method; Quantitatively evaluate the contribution of preset measurement parameters to atmospheric aerosol acidity φ through the Shapley value. i,j :

[0023]

[0024]

[0025] Among them, y i Let φ be the acidity value of the i-th aerosol obtained from the model inversion. i,j Here are the Shapley values, φ0 is a constant, and φ i,j For the j-th feature variable pair y i The Shapley value contribution is given by y(S∪{j}), where S is a subset of the input parameter set in the i-th calculation that does not contain variable j. i -y(S) i The impact of including or not including feature variable j in the model calculation on the model results;

[0026] (4) Based on the analysis results of steps (1) to (3), analyze the spatiotemporal variation of atmospheric aerosol acidity.

[0027] Another aspect of this application provides a system for detecting atmospheric aerosol acidity, comprising: a data acquisition module for acquiring aerosol data and meteorological data from atmospheric observation stations and preprocessing the acquired data; a theoretical value module for calculating the theoretical aerosol acidity pH using the Metastable-Forward thermodynamic model ISORROPIA II; a machine learning module for constructing multiple machine learning models using preset measurement parameters as feature variables and the theoretical aerosol acidity pH as the target variable, and selecting the best-performing machine learning model through cross-validation; and a detection module for acquiring real-time aerosol data and meteorological data using a mass spectrometer and detecting the real-time atmospheric aerosol acidity pH using the optimal machine learning model. i The simulation module calculates the pH value of atmospheric aerosols based on real-time atmospheric aerosol levels. i By analyzing the spatiotemporal distribution, Lagrange source analysis, and SHAP analysis based on game theory, the spatiotemporal variation law of atmospheric aerosol acidity was obtained.

[0028] Compared to existing technologies, the advantages of this application are:

[0029] (1) Compared with traditional acidity detection based on acid-base indicators, this application does not require dynamically maintaining a high relative humidity (e.g., 90%) at the sampling interface so that atmospheric aerosols can wet the sampling membrane or test paper after entering the liquid phase to meet the detection conditions of acid-base indicators. It also eliminates the need to frequently replace disposable sampling membranes or test papers and other consumables, further reducing the potential uncertainty of aerosol acidity detection during the adjustment of atmospheric relative humidity, simplifying the internal environmental settings of the instrument, reducing unnecessary material consumption, and helping to improve the sustainability of observation.

[0030] (2) Compared with traditional online detection technology for atmospheric aerosol acidity, this application does not only measure aerosol acidity, but also integrates multiple measurement systems to provide more comprehensive and synchronous meteorological elements and aerosol component data, which helps to analyze the mechanism and process of atmospheric aerosol acidity evolution in a more systematic, comprehensive and detailed way.

[0031] (3) Compared with traditional online detection technology for atmospheric aerosol acidity, this application has a larger detection range and higher detection sensitivity for atmospheric aerosol acidity, which can completely cover the common variation range of atmospheric aerosol acidity (e.g., pH=1 to pH=8), and the pH detection sensitivity can reach 0.000000000000001.

[0032] (4) Compared with the traditional thermodynamic model, this application further simplifies and adjusts the input parameters required for operation, eliminates the observation requirements of various metal ions and gas data, and can be matched with the data products of more traditional atmospheric aerosol observation instruments. It solves the problem that in many scenarios, there is a lack of simultaneous observation data from multiple instruments, which makes it impossible to calculate atmospheric aerosol acidity based on the thermodynamic model. It significantly improves the inversion potential of atmospheric aerosol acidity and has better adaptability, compatibility and mobility.

[0033] (5) Compared with traditional online detection technology and thermodynamic model of atmospheric aerosol acidity, this application has a faster acidity calculation speed. It can calculate more than 50,000 data points per second of atmospheric aerosol acidity on a laptop with an average absolute percentage error of less than 4% and a model size of about 83,000 bytes. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the overall structure of the online detection system for atmospheric aerosol acidity of this application;

[0035] Figure 2 This is an exemplary flowchart of a method for detecting the acidity of atmospheric aerosols according to this application;

[0036] Figure 3 The closure of the atmospheric aerosol acidity inversion model results constructed in this embodiment with the ISORROPIA model results;

[0037] Figure 4 The closure of the atmospheric aerosol acidity inversion model results constructed in this embodiment with the ISORROPIA model results;

[0038] Figure 5 This represents the contribution of different factors to the average Shapley value of atmospheric aerosol acidity retrieved by the model in this embodiment.

[0039] Figure 6 This represents the Shapley values ​​of atmospheric aerosol acidity retrieved by the model for different factors in this embodiment. Detailed Implementation

[0040] The present application will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0041] Example 1

[0042] like Figure 1As shown, the online detection system for atmospheric aerosol acidity provided in this application specifically includes a data processing system, a sampling pump, a vacuum pump, an atmospheric temperature sensor, a relative humidity sensor, a drying tube, a flow meter, a flow restrictor, an aerodynamic lens, a heating device and an ionization device, and a mass spectrometer for analyzing aerosol components based on the aerosol mass-to-charge ratio (m / z), all interconnected by pipelines. The ambient temperature and relative humidity of the atmospheric aerosol environment are acquired in situ by atmospheric temperature and relative humidity sensors. Subsequently, the polydisperse aerosols in the atmosphere are dried in a drying tube at a rated sample flow rate by a sampling pump. The sampling flow rate is monitored by a flow meter. A flow-limiting orifice restricts the aerosol particle diameter to below 2.5 micrometers. Then, under the action of an aerodynamic lens, the aerosols are progressively accelerated, causing them to converge into extremely fine particle beams that move at high speed along the centerline to a vacuum system formed by three turbine vacuum pumps. This effectively separates the gas from the aerosol and eliminates interference from gaseous components in the detection of aerosol components. The aerosols then fly towards a heating device continuously heated to 600°C. Non-refractory aerosols are vaporized or vaporized at the heating device, and the gaseous aerosol chemical components are further ionized by electron bombardment in an ionization device. The ionized aerosol fragments then enter a mass spectrometer. Based on the correspondence between the aerosol mass-to-charge ratio (m / z) and the aerosol components, ammonium salts are analyzed. nitrates sulfates Analysis of the mass concentrations of aerosol chemical components such as chloride (Chl). The measured atmospheric temperature, relative humidity, ammonium salt mass concentration, nitrate mass concentration, sulfate mass concentration, and chloride mass concentration data were all transmitted to the data processing system via Ethernet. The data processing system, based on seven data points—atmospheric temperature, relative humidity, ammonium salt mass concentration, nitrate mass concentration, sulfate mass concentration, chloride mass concentration, and the local time at which the data was saved—was fed into a pre-trained aerosol acidity machine learning inversion model to achieve rapid online measurement and mechanism interpretation of aerosol acidity.

[0043] like Figure 2As shown, the online detection method for atmospheric aerosol acidity provided in this application acquires aerosol data and meteorological data from atmospheric observation stations, and preprocesses the acquired data. The aerosol data includes the mass concentrations of ammonium salts, sulfates, nitrates, and chlorides; the meteorological data includes atmospheric temperature and relative humidity. The preprocessed data is calculated using the Metastable-Forward thermodynamic model ISORROPIA II to obtain the theoretical value of aerosol acidity, pH. Multiple machine learning models are constructed using preset measurement parameters as feature variables and the theoretical value of aerosol acidity, pH, as the target variable. Through cross-validation, the machine learning model with the best performance is selected. Real-time aerosol component data is acquired using mass spectrometry analysis, and current meteorological data is also acquired. The real-time atmospheric aerosol acidity is calculated using the optimal machine learning model.

[0044] Detailed historical data acquisition and quality control: Acquiring historical data on atmospheric aerosols, gases, and meteorological elements from atmospheric observation stations and conducting data quality control; atmospheric aerosol data includes sodium (Na+), magnesium (Mg2+), potassium (K+), calcium (Ca2+), and ammonium salts. sulfates nitrates Mass concentration data for eight different aerosol components, including chloride (Chl), gaseous data for ammonia (NH3), and meteorological data for atmospheric temperature and relative humidity. All data include timestamps.

[0045] Data quality control includes time series matching, identification and removal of missing values ​​and outliers. Specifically, it includes sorting various types of data in chronological order, aligning time nodes, assigning the missing data periods in the time series data to Not a Number (NaN) at each time resolution to represent undefined or unrepresentable values ​​in the computer, and defining outliers as points within 5 moving windows that differ from the local median by more than 3 times the local transformed absolute deviation (MAD) and assigning them the value NaN.

[0046] The mathematical formula for assigning NaN to outliers is as follows:

[0047]

[0048]

[0049] In the formula, n represents the total number of data points, t represents the t-th time point, and y t,j Let y represent the j-th type of data at time t. m,j This represents the median (median y) of data of class j within the moving window. m,j (located at time m), where median represents the median;

[0050] An atmospheric aerosol acidity inversion model was established. The ISORROPIA II thermodynamic model of the Metastable-Forward mode was used to analyze the mass concentrations of sodium (Na+), magnesium (Mg2+), potassium (K+), calcium (Ca2+), and ammonium salts after quality control. mass concentration, sulfate mass concentration, nitrate Eleven data points, including mass concentration of chloride (Chl), mass concentration of ammonia (NH3), atmospheric temperature, and relative humidity, were processed and calculated to obtain atmospheric aerosol acidity data (pH value). The mathematical formula for calculating atmospheric aerosol acidity in the ISORROPIA II thermodynamic model is as follows: In the formula, pH is the pH value of the aerosol. The mass concentration of hydrogen ions in atmospheric aerosols (unit: μg / m3) is given by the model, and LWC is the mass concentration of liquid water in atmospheric aerosols (unit: μg / m3).

[0051] Furthermore, to simplify the calculation method of aerosol acidity and improve the speed and feasibility of operation, the aerosol pH value obtained by the ISORROPIA II thermodynamic model under the Metastable-Forward mode is simplified by inversion and the mechanism is analyzed. This is combined with the diurnal variation characteristics and important physicochemical mechanisms of aerosols in the actual atmosphere, and fully considers the main components of current domestic and international atmospheric aerosol observation data. Based on data timestamps, local time, atmospheric temperature, relative humidity, and ammonium salts are identified. mass concentration, sulfate mass concentration, nitrate A model for atmospheric aerosol acidity retrieval was constructed using seven data types, including mass concentration and chloride (Chl) mass concentration, which are widely and synchronously measured both domestically and internationally. The specific process of model construction includes:

[0052] After data quality control, local time, atmospheric temperature, relative humidity, and ammonium salts were collected. mass concentration, sulfate mass concentration, nitrate Mass concentration and chloride (Chl) mass concentration are used as characteristic variables, and aerosol acidity (pH value) calculated by the corresponding thermodynamic model is used as the target variable; the time series of characteristic variables and target variables correspond one-to-one.

[0053] Various atmospheric aerosol acidity inversion models were constructed, including linear regression, fine-grained trees, ensemble trees, neural networks, and Gaussian process regression. Multi-fold cross-validation (5-fold, 10-fold, etc.) was performed based on the data size. The ensemble tree model included bag trees and LSBoost trees, and Bayesian optimization was used for expected improvement per second. The Gaussian process regression model included constant, zero, and linear basis functions, and kernel functions such as Nonisotropic exponent, Nonisotropic quadratic exponent, Nonisotropic quadratic rational, Nonisotropic Matern3 / 2, Nonisotropic 5 / 2, Isotropic exponent, Isotropic quadratic exponent, Isotropic quadratic rational, Isotropic Matern3 / 2, and Isotropic Matern 5 / 2.

[0054] In multi-fold cross-validation, if we assume that N-fold cross-validation is performed, the dataset Y is randomly divided into N mutually exclusive subsets of the same size. In each model construction, N-1 subsets are used as the training set and 1 subset is used as the validation set. After N rounds of training and validation, the model performance is evaluated.

[0055] Preferably, multi-round N-fold cross-validation is used to further improve the reliability of model evaluation;

[0056] The model performance was evaluated using metrics such as the coefficient of determination (R²), root mean square error (RMSE), mean squared error (MSE), mean absolute error (MAE), mean absolute percentage error (MAPE), prediction speed, training time, and model size. The R², RMSE, MSE, MAE, and MAPE evaluations included both training and validation set results. Among the evaluation metrics, a higher R² value (closer to 1), lower RMSE, lower MSE, lower MAE, lower MAPE, faster prediction speed, shorter training time, and smaller model size resulted in a better evaluation. Furthermore, the model with the best overall evaluation was selected and trained using all historical datasets to further improve model performance. Preferably, a Gaussian process regression model with constant basis functions and isotropic exponential kernel functions was used.

[0057] The atmospheric aerosol acidity online detection system based on this application acquires eight types of data via Ethernet, including local time, atmospheric temperature, relative humidity, ammonium salt mass concentration, nitrate salt mass concentration, sulfate salt mass concentration, chloride salt mass concentration, and flow rate, and performs data quality control. The data quality control includes time series matching, identification and removal of missing values ​​and outliers. Specifically, it includes sorting various types of data in chronological order, aligning time nodes, assigning NaN to missing data periods in the time series data at each time resolution, defining outliers as points within 5 moving windows that differ from the local median by more than 3 times the local converted absolute median (MAD) and assigning them NaN, and assigning NaN to data during periods of abnormal flow rate.

[0058] Seven data points—local time, atmospheric temperature, relative humidity, ammonium salt mass concentration, nitrate mass concentration, sulfate mass concentration, and chloride mass concentration—were acquired online and quality-controlled. An atmospheric aerosol acidity inversion model was used to calculate the atmospheric aerosol acidity (pH value), mathematically expressed as follows: y i =f(x) i1 ,x i2 ,x i3 ,x i4 ,x i5 ,x i6 ,x i7 In the formula, y i Let x be the atmospheric aerosol acidity (pH value) at the i-th data point, f be the constructed atmospheric aerosol acidity inversion model, and x be the pH value. i,j Let j be the value of the feature variable at the i-th data point;

[0059] Atmospheric aerosol acidity analysis is performed based on the changes in the numerical values ​​of seven types of data, including local time, atmospheric temperature, relative humidity, ammonium salt mass concentration, nitrate mass concentration, sulfate mass concentration, and chloride mass concentration, as well as the changes in the model inversion results.

[0060] Preferably, the atmospheric transport path can be analyzed at the case, cluster, and global levels based on the Lagrange source tracing model, and the atmospheric source can be analyzed using the Concentration Weight Trajectory (CWT) method. The mathematical formula is as follows: In the formula, CWT i,j C represents the contribution of grid i to the data at detection location j. j,l Let τ be the magnitude of data j in trajectory l when it reaches the detection location. i,j,l Let N be the dwell time of data j in trajectory l at grid i, and N be the total number of trajectories.

[0061] Preferably, based on seven types of data—local time, atmospheric temperature, relative humidity, ammonium salt mass concentration, nitrate mass concentration, sulfate mass concentration, and chloride mass concentration—the SHAP (SHapley Additive exPlanation) method, based on game theory, can be used to quantitatively analyze the contribution of Shapley values ​​to atmospheric aerosol acidity, both globally and locally. When all features are present, the mathematical model is as follows:

[0062]

[0063] In the formula, y i Let φ be the i-th value obtained from model inversion, φ be the Shapley value, φ0 be a constant, and φ i,j For the j-th feature variable pair y i Shapley value contribution, In the i-th calculation, set S refers to the set of input parameters N in the i-th calculation. i A subset of N and containing no N i The variable j in (y(S∪{j}) i -y(S) i ) represents the impact of including the j feature variable in the i-th calculation of the model on the model inversion result compared to not including it.

[0064] Example 2

[0065] Historical data acquisition and quality control: Historical data on atmospheric aerosols, gases, and meteorological elements from November 20, 2019 to January 13, 2020 were acquired from the Nanjing atmospheric observation station, and data quality control was conducted. Atmospheric aerosol data included sodium (Na+), magnesium (Mg2+), potassium (K+), calcium (Ca2+), and ammonium salts. sulfates nitrates Mass concentration data for eight different aerosol components, including chloride (Chl), gaseous data for ammonia (NH3), and meteorological data for atmospheric temperature and relative humidity. All data include timestamps.

[0066] Data quality control includes time series matching, identification and removal of missing values ​​and outliers. Specifically, it includes sorting various types of data in chronological order, aligning time nodes, assigning the missing data periods in the time series data to Not a Number (NaN) at each time resolution to represent undefined or unrepresentable values ​​in the computer, and defining outliers as points within 5 moving windows that differ from the local median by more than 3 times the local transformed absolute deviation (MAD) and assigning them the value NaN.

[0067] The mathematical formula for assigning NaN to outliers is as follows:

[0068]

[0069] In the formula, n represents the total number of data points, t represents the t-th time point, and y t,j Let y represent the j-th type of data at time t. m,j This represents the median (median y) of data of class j within the moving window. m,j (located at time m), where median represents the median;

[0070] An atmospheric aerosol acidity inversion model was established. The ISORROPIA II thermodynamic model of the Metastable-Forward mode was used to analyze the mass concentrations of sodium (Na+), magnesium (Mg2+), potassium (K+), calcium (Ca2+), and ammonium salts after quality control. mass concentration, sulfate mass concentration, nitrate Eleven data points, including mass concentration of chloride (Chl), mass concentration of ammonia (NH3), atmospheric temperature, and relative humidity, were processed and calculated to obtain atmospheric aerosol acidity data (pH value). The mathematical formula for calculating atmospheric aerosol acidity in the ISORROPIA II thermodynamic model is as follows:

[0071]

[0072] In the formula, pH is the pH value of the aerosol. The mass concentration of hydrogen ions in atmospheric aerosols (unit: μg / m3) is given by the model, and LWC is the mass concentration of liquid water in atmospheric aerosols (unit: μg / m3).

[0073] Furthermore, to simplify the calculation method of aerosol acidity and improve the speed and feasibility of operation, the aerosol pH value obtained by the ISORROPIA II thermodynamic model under the Metastable-Forward mode is simplified by inversion and the mechanism is analyzed. This is combined with the diurnal variation characteristics and important physicochemical mechanisms of aerosols in the actual atmosphere, and fully considers the main components of current domestic and international atmospheric aerosol observation data. Based on data timestamps, local time, atmospheric temperature, relative humidity, and ammonium salts are identified. mass concentration, sulfate mass concentration, nitrate A model for atmospheric aerosol acidity retrieval was constructed using seven data types, including mass concentration and chloride (Chl) mass concentration, which are widely and synchronously measured both domestically and internationally. The specific process of model construction includes:

[0074] After data quality control, local time, atmospheric temperature, relative humidity, and ammonium salts were collected. mass concentration, sulfate mass concentration, nitrate Mass concentration and chloride (Chl) mass concentration are used as characteristic variables, and aerosol acidity (pH value) calculated by the corresponding thermodynamic model is used as the target variable; the time series of characteristic variables and target variables correspond one-to-one.

[0075] We constructed various atmospheric aerosol acidity inversion models, including linear regression, fine-grained tree, ensemble tree, neural network, and Gaussian process regression, and conducted 5-fold cross-validation.

[0076] The ensemble tree includes bag trees and LSBoost trees, using Bayesian optimization for expected improvement per second; Gaussian process regression includes constant, zero, and linear basis functions, and kernel functions such as Nonisotropic exponent, Nonisotropic quadratic exponent, Nonisotropic quadratic rational, Nonisotropic Matern 3 / 2, Nonisotropic 5 / 2, Isotropic exponent, Isotropic quadratic exponent, Isotropic quadratic rational, Isotropic Matern 3 / 2, and Isotropic Matern 5 / 2.

[0077] Five-fold cross-validation involves randomly dividing the dataset Y into five mutually exclusive subsets of equal size. In each model build, four subsets are used as the training set and one subset is used as the validation set. The model performance is evaluated after five rounds of training and validation.

[0078] Preferably, multiple rounds of 5-fold cross-validation are used to further improve the reliability of model evaluation;

[0079] The model performance was evaluated using metrics such as the coefficient of determination (R²), root mean square error (RMSE), mean squared error (MSE), mean absolute error (MAE), mean absolute percentage error (MAPE), prediction speed, training time, and model size. The R², RMSE, MSE, MAE, and MAPE evaluations included both training and validation set results. Among the evaluation metrics, a higher R² value (closer to 1), lower RMSE, lower MSE, lower MAE, lower MAPE, faster prediction speed, shorter training time, and smaller model size resulted in a better evaluation. Furthermore, the model with the best overall evaluation was selected and trained using all historical datasets to further improve its performance.

[0080] Preferably, a Gaussian process regression model with constant basis functions and isotropic exponential kernel functions is used (in this embodiment, R2 = 0.9882, prediction speed is 52747 predictions / second, model size is 82882 bytes, and training time is 21.386 seconds).

[0081] Online detection of atmospheric aerosol acidity: An online detection system for atmospheric aerosol acidity is constructed, which specifically includes a data processing system, a sampling pump, a vacuum pump, an atmospheric temperature sensor, a relative humidity sensor, a drying tube, a flow meter, a flow restrictor, an aerodynamic lens, a heating device and an ionization device, as well as a mass spectrometer for analyzing aerosol components based on the aerosol mass-to-charge ratio (m / z), all interconnected through pipelines.

[0082] The ambient temperature and relative humidity of the atmospheric aerosol environment are acquired in situ by atmospheric temperature and relative humidity sensors. Subsequently, the polydisperse aerosols in the atmosphere are dried in a drying tube at a rated sample flow rate by a sampling pump. The sampling flow rate is monitored by a flow meter. A flow-limiting orifice restricts the aerosol particle diameter to below 2.5 micrometers. Then, under the action of an aerodynamic lens, the aerosols are progressively accelerated, causing them to converge into extremely fine particle beams that move at high speed along the centerline to a vacuum system formed by three turbine vacuum pumps. This effectively separates the gas from the aerosol and eliminates interference from gaseous components in the detection of aerosol components. The aerosols then fly towards a heating device continuously heated to 600°C. Non-refractory aerosols are vaporized or vaporized at the heating device, and the gaseous aerosol chemical components are further ionized by electron bombardment in an ionization device. The ionized aerosol fragments then enter a mass spectrometer. Based on the correspondence between the aerosol mass-to-charge ratio (m / z) and the aerosol components, ammonium salts are analyzed. nitrates sulfates Analysis of the mass concentration of aerosol chemical components such as chloride (Chl).

[0083] The measured local time, atmospheric temperature, relative humidity, ammonium salt mass concentration, nitrate mass concentration, sulfate mass concentration, chloride mass concentration, and flow rate data were all transmitted to the data processing system via Ethernet and subjected to data quality control. Data quality control included time series matching, identification and removal of missing and outlier values. Specifically, this included sorting all types of data in chronological order, aligning time nodes, assigning NaN values ​​to missing data periods at each time resolution, defining outliers as points within 5 moving windows that differ from the local median by more than 3 times the local transformed absolute median (MAD) and assigning them NaN values, and assigning NaN values ​​to data from periods of abnormal flow rates. After quality control, the atmospheric aerosol acidity (pH value) of the seven data points (local time, atmospheric temperature, relative humidity, ammonium salt mass concentration, nitrate mass concentration, sulfate mass concentration, and chloride mass concentration) was calculated using an atmospheric aerosol acidity inversion model. The mathematical form is as follows:

[0084] y i =f(x) i1 ,x i2 ,x i3 ,x i4 ,x i5 ,x i6 ,x i7 )

[0085] In the formula, y i Let x be the atmospheric aerosol acidity (pH value) at the i-th data point, f be the constructed atmospheric aerosol acidity inversion model, and x be the pH value. ij Let j be the value of the feature variable at the i-th data point;

[0086] like Figure 3 , Figure 4 As shown, local time, atmospheric temperature, relative humidity, and ammonium salts are used to determine the data. mass concentration, sulfate mass concentration, nitrate Atmospheric aerosol acidity results derived by using sodium (Na+), magnesium (Mg2+), potassium (K+), calcium (Ca2+), and ammonium salt mass concentrations as input variables were compared with those derived by using sodium (Na+), magnesium (Mg2+), potassium (K+), calcium (Ca2+), and ammonium salt mass concentrations as input variables. mass concentration, sulfate mass concentration, nitrate The thermodynamic model results with mass concentration, chloride (Chl) mass concentration, ammonia (NH3) mass concentration, atmospheric temperature, and relative humidity as input variables are highly consistent.

[0087] Step 4: Atmospheric aerosol acidity analysis

[0088] The acidity of atmospheric aerosols is analyzed based on the numerical changes of seven types of data, including local time, atmospheric temperature, relative humidity, ammonium salt mass concentration, nitrate mass concentration, sulfate mass concentration, and chloride mass concentration, as well as the changes in the model inversion results.

[0089] Preferably, the atmospheric transport path can be analyzed at the case, cluster, and global levels based on the Lagrange source tracing model, and the atmospheric source can be analyzed using the Concentration Weight Trajectory (CWT) method. The mathematical formula is as follows:

[0090]

[0091] In the formula, CWT i,j C represents the contribution of grid i to the data at detection location j. j,l Let τ be the magnitude of data j in trajectory l when it reaches the detection location. i,j,l Let N be the dwell time of data j in trajectory l at grid i, and N be the total number of trajectories.

[0092] Preferably, the parameters can be based on local time (Hour), atmospheric temperature (Temperature), relative humidity (RH), and ammonium salt. mass concentration, sulfate mass concentration, nitrate Seven types of data, including mass concentration and chloride (Chl) mass concentration, were analyzed using the SHAP (SHapley Additive exPlanation) method based on game theory. The contribution of Shapley values ​​to atmospheric aerosol acidity was quantitatively analyzed globally and locally. The mathematical model, assuming all characteristics are present, is as follows:

[0093]

[0094] In the formula, y i Let φ be the i-th value obtained from model inversion, φ be the Shapley value, φ0 be a constant, and φ i,j For the j-th feature variable pair y i Shapley value contribution, In the i-th calculation, set S refers to the set of input parameters N in the i-th calculation. i A subset of N and containing no N i The variable j in (y(S∪{j}) i -y(S) i The effect of including the j feature variable in the i-th calculation of the model compared to not including it on the model inversion result;

[0095] like Figure 5 , Figure 6 As shown, in this embodiment, the local time (Hour), atmospheric temperature (Temperature), relative humidity (RH), and ammonium salt are... mass concentration, sulfate mass concentration, nitrate The contributions of seven characteristic variables, including seven types of data such as mass concentration and chloride (Chl) mass concentration, to the Shapley value of atmospheric aerosol acidity differed significantly. The contributions, ranked from largest to smallest, were: sulfate... mass concentration, nitrate Mass concentration, atmospheric temperature, relative humidity (RH), local time (Hour), nitrate (NO3) - Mass concentration of chloride (Chl); mass concentration of sulfate The higher the mass concentration, the greater the negative contribution of the Shapley value; ammonium salts The greater the mass concentration, the greater the positive contribution of the Shapley value. This is consistent with the corresponding cation-anion balance of H+ ions and indicates that changes in the mass concentration of sulfate and ammonium salts in the atmosphere have a significant impact on changes in aerosol acidity. The more sulfate, the more acidic the aerosol tends to be, and the more ammonium salt, the more neutral or alkaline it tends to be.

[0096] The foregoing illustrative description of the present application and its embodiments is not restrictive and can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application. The accompanying drawings are only one embodiment of the present application, and the actual structure is not limited thereto. Therefore, if a person skilled in the art, inspired by this description, designs a similar structure and embodiment without departing from the spirit of the present application, such design should fall within the scope of protection of this application. Furthermore, the word "comprising" does not exclude other elements or steps, and the word "a" preceding an element does not exclude the inclusion of "a plurality" of that element. Multiple elements stated in the product claims may also be implemented by a single element through software or hardware. The terms "first," "second," etc., are used to indicate names and do not indicate any specific order.

Claims

1. An online detection method for atmospheric aerosol acidity, characterized in that, include: The process involves acquiring aerosol and meteorological data from atmospheric observation stations and preprocessing the acquired data. The aerosol data includes the mass concentrations of ammonium salts, sulfates, nitrates, and chlorides. The meteorological data includes atmospheric temperature and relative humidity. The preprocessing process includes sorting the acquired aerosol and meteorological data in chronological order according to the timestamps of each data point and aligning the time nodes to obtain a time-consistent data time series. The theoretical aerosol acidity pH was obtained by using the Metastable-Forward thermodynamic model ISORROPIA II to calculate the pre-processed data. Using preset measurement parameters as feature variables and the theoretical aerosol acidity value (pH) as the target variable, multiple machine learning models were constructed. Through cross-validation, the machine learning model with the best performance was selected. The preset measurement parameters included: local time, atmospheric temperature, relative humidity, ammonium salt mass concentration, sulfate mass concentration, nitrate mass concentration, and chloride mass concentration. Mass spectrometry analysis was used to obtain real-time aerosol composition data and current meteorological data. The optimal machine learning model was then used to calculate the real-time atmospheric aerosol acidity. Using preset measurement parameters as feature variables and the theoretical aerosol acidity value (pH) as the target variable, multiple machine learning models were constructed. Through cross-validation, the machine learning model with the best performance was selected, including: Establish a time series dataset with a one-to-one correspondence between feature variables and target variables; Build various machine learning models based on time series datasets; Perform N-fold cross-validation on each constructed machine learning model; Verify the performance of each machine learning model; Choose the model with the best performance as the optimal machine learning model; Mass spectrometry analysis was used to obtain real-time aerosol composition data, and current meteorological data was acquired. An optimal machine learning model was then used to calculate real-time atmospheric aerosol acidity, including: Real-time aerosol and meteorological data were acquired using a mass spectrometer, and the acquired data were preprocessed. Using the preprocessed data as input, and employing the optimal machine learning model, the real-time atmospheric aerosol acidity is calculated using the following formula. : in, Let f be the real-time atmospheric aerosol acidity at the i-th data point, and f be the optimal machine learning model. Let be the value of the j-th feature variable for the i-th data point.

2. The online detection method for atmospheric aerosol acidity according to claim 1, characterized in that: Machine learning models include: linear regression, fine-grained trees, ensemble trees, neural networks, and Gaussian process regression models.

3. The online detection method for atmospheric aerosol acidity according to claim 2, characterized in that: Ensemble trees include bagged trees and boosting trees; Gaussian process regression includes various basis functions and kernel functions.

4. The online detection method for atmospheric aerosol acidity according to claim 1, characterized in that: The acquired data is preprocessed, including: Based on the data time series, identify the time periods with no data records, and mark the data point positions in the data segments as non-Number NaN; Based on the time series of data after marking non-NaN values, outlier detection is performed using the moving window method. Within N1 moving windows, the local median and the local equivalent absolute median difference (MAD) are calculated. Data points whose difference from the local median is greater than or equal to N2 times the MAD are marked as non-NaN values, thus obtaining the preprocessed data.

5. The online detection method for atmospheric aerosol acidity according to claim 4, characterized in that: The ISORROPIA II thermodynamic model in the Metastable-Forward mode was used to calculate the theoretical aerosol acidity pH value, including: Using the preprocessed data as input, the mass concentration of hydrogen ions in the aerosol was calculated using the ISORROPIA II thermodynamic model in the Metastable-Forward mode. and liquid water mass concentration (LWC); Using hydrogen ion mass concentration Given the liquid water mass concentration (LWC), the theoretical aerosol acidity value (pH) is calculated using the following formula: 。 6. The online detection method for atmospheric aerosol acidity according to any one of claims 1 to 5, characterized in that: After calculating the real-time atmospheric aerosol acidity, the following steps are also included: (1) Analyze the spatiotemporal distribution of aerosol acidity based on the preset measurement parameters and the calculated real-time atmospheric aerosol acidity. (2) Using the Lagrange source model, the transport path of atmospheric aerosols was analyzed, and the contribution of different air mass source regions to the aerosol acidity at the observation station was calculated using the concentration-weighted trajectory method. : in, The contribution of grid i to the data at detection location j. For trajectory The size of the j data when it arrives at the detection location. For trajectory The dwell time of data j in grid i, where N is the total number of trajectories; (3) Calculate the Shapley value using the game theory-based SHAP method; The contribution of pre-defined measurement parameters to atmospheric aerosol acidity was quantitatively assessed using Shapley values. : in, Let i be the aerosol acidity value obtained from the model inversion. Shapley value, It is a constant. For the j-th feature variable pair The Shapley value contribution, where S refers to the subset of the input parameter set in the i-th calculation that does not contain variable j. The impact of including or not including feature variable j in the model calculation on the model results; (4) Based on the analysis results of steps (1) to (3), analyze the spatiotemporal variation of atmospheric aerosol acidity.

7. An online detection system for atmospheric aerosol acidity, used to perform the method according to any one of claims 1 to 6, characterized in that, include: The data acquisition module acquires aerosol and meteorological data from atmospheric observation stations and preprocesses the acquired data. The theoretical value module uses the Metastable-Forward thermodynamic model ISORROPIA II to calculate the theoretical value of aerosol acidity pH from the preprocessed data. The machine learning module uses preset measurement parameters as feature variables and the theoretical value of aerosol acidity (pH) as the target variable to construct multiple machine learning models, and selects the machine learning model with the best performance through cross-validation. The detection module acquires real-time aerosol and meteorological data through a mass spectrometer and uses an optimal machine learning model to detect real-time atmospheric aerosol acidity values. ; The simulation module, based on real-time atmospheric aerosol acidity values... By analyzing the spatiotemporal distribution, Lagrange source analysis, and SHAP analysis based on game theory, the spatiotemporal variation law of atmospheric aerosol acidity was obtained.

Images