Water surface atmospheric environment monitoring and tracing system based on multi-source data fusion

Abstract

The invention relates to the technical field of atmospheric environment monitoring, in particular to a water surface atmospheric environment monitoring and tracing system based on multi-source data fusion. According to the system, a monitoring layer adopts a dynamic sampling frequency adaptive algorithm to realize multi-source data synchronous acquisition, and the data precision of a high-pollution area is improved; the data transmission and fusion layer solves the problems of time-space dislocation and standardization of multi-source data through a time attention mechanism and an inverse distance weighting algorithm; the analysis layer improves a forward matrix decomposition model, realizes automatic pollution source identification by introducing a Bayesian optimization factor number and combining with random forest classification, and quantifies secondary pollution contribution by matching with a gradient lifting regression tree model; and the application layer realizes pollution source quantitative traceability and high-value area positioning through a geographically weighted regression model and hotspot analysis, and outputs visual products and prevention and control suggestions. The system effectively solves the problems of low water surface atmosphere monitoring data quality and insufficient traceability precision, and provides technical support for cross-regional joint defense and joint control of the Yangtze river basin.

Description

Technical Field

[0001] This invention relates to the field of atmospheric environment monitoring technology, specifically to a water surface atmospheric environment monitoring and tracing system based on multi-source data fusion. Background Technology

[0002] As the core economic area and an important ecological barrier of my country, the Yangtze River Basin has a dense concentration of industrial clusters such as petrochemicals, docks, and shipbuilding along its banks. It also carries more than 70% of the total cargo volume of the Yangtze River. The superposition of ship navigation and industrial emissions has led to unique problems in the atmospheric environment on the water surface, such as "coupling of pollution sources along the river and water surface transmission, complex wind fields on the river surface, and interference of ship attitude and turbulence with monitoring". The pressure of controlling volatile organic compounds (VOCs) and secondary organic aerosols (SOA) pollution is prominent.

[0003] Existing water surface detection methods employ multi-source data, generative adversarial networks, and spatiotemporal random forest algorithms to output high-precision, seamless spatiotemporal particulate matter concentration data hourly; or they train AI models based on historical component data and industry labels to detect air pollution sources and construct source tracing paths based on distribution information, matching pollution sources in related industries within the path range. These location-based early warning methods primarily target urban or regional atmospheric environments, failing to fully consider water surface turbulence, complex river wind fields, and the coupling relationship between pollution sources along the riverbank and water surface transmission. Furthermore, the algorithms lack intelligence and coordination, mainly relying on similarity comparisons and simple statistical methods for source tracing. Summary of the Invention

[0004] To address the problems existing in current technologies, this invention provides a water surface atmospheric environment monitoring and source tracing system based on multi-source data fusion. The system comprises: a monitoring layer employing a dynamic sampling frequency adaptive algorithm to achieve synchronous acquisition of multi-source data, improving data accuracy in highly polluted areas; a data transmission and fusion layer using a spatiotemporal attention mechanism and an inverse distance weighting algorithm to solve the spatiotemporal misalignment and standardization problems of multi-source data; an analysis layer improving the forward matrix factorization model by introducing Bayesian optimization factors and combining it with random forest classification to achieve automatic pollution source identification, coupled with a gradient boosting regression tree model to quantify the contribution of secondary pollution; and an application layer using a geographic weighted regression model and hotspot analysis to achieve quantitative source tracing of pollution sources and location of high-value areas, outputting visualized products and prevention and control recommendations. The system effectively solves the problems of low data quality and insufficient source tracing accuracy in water surface atmospheric monitoring, providing technical support for cross-regional joint prevention and control in the Yangtze River Basin.

[0005] This invention adopts the following technical solution: a water surface atmospheric environment monitoring and tracing system based on multi-source data fusion, comprising: The monitoring layer, based on the monitoring equipment on board the ship, uses a dynamic sampling frequency adaptive algorithm to collect multi-source data in real time during the ship's navigation; the multi-source data includes at least: mobile monitoring data, national control point data, meteorological data, and industrial cluster data; The data transmission and fusion layer is used to perform data cleaning on the multi-source data, and to perform spatiotemporal alignment and standardization on the multi-source data using a time attention mechanism and an inverse distance weighted interpolation algorithm to obtain a standardized three-dimensional mesh dataset. The intelligent analysis layer processes the three-dimensional grid dataset using a forward matrix factorization model based on Bayesian optimization factors to obtain a factor-species concentration matrix; a classification model is constructed based on the random forest algorithm, and the classification model is used to identify the factor-species concentration matrix to obtain pollution source type identification results; based on the factor-species concentration matrix and the pollution source type identification results, a gradient boosting regression tree is used to obtain secondary organic aerosol generation potential prediction values. At the application layer, based on the three-dimensional grid dataset, the pollution source type identification results, and the predicted value of secondary organic aerosol generation potential, a geographically weighted regression model combined with a hotspot analysis algorithm is used to obtain visualized source tracing results.

[0006] Furthermore, a dynamic sampling frequency adaptive algorithm is used to collect multi-source data in real time during ship navigation, specifically: Real-time acquisition of ship location information; The location information of the ship is compared with the geographical location of the preset industrial cluster to determine whether the ship has entered the industrial cluster. When the comparison results show that the ship enters the industrial cluster, the frequency of the equipment collecting multi-source data will be increased from the first frequency to the second frequency. When the comparison results show that the ship has left the industrial cluster, the frequency of the equipment collecting multi-source data will be restored to the first frequency.

[0007] Furthermore, the data cleaning process for the multi-source data includes: For the aforementioned navigation monitoring data, data outside the preset monitoring period are removed; data collected when the ship's lateral roll angle is greater than 15 degrees or the longitudinal roll angle is greater than 10 degrees are removed; data with a concentration value of zero are replaced with half of the maximum range value of the monitoring equipment; data exceeding the range of the monitoring equipment are identified using the 3σ criterion. For the national monitoring point data, data collected during equipment failures are removed, and data overlapping with the preset monitoring period is retained; for national monitoring point data with missing data, the average of the data from its adjacent national monitoring points is used to complete the data. For the meteorological data, data that overlaps with the preset monitoring period are filtered out, and data from periods with less than 90% data integrity are removed; For the industrial cluster data, retain the industrial cluster data associated with the mobile monitoring area, and remove pollution sources whose geographical location is outside the mobile monitoring area.

[0008] Furthermore, a temporal attention mechanism and an inverse distance weighted interpolation algorithm are used to perform spatiotemporal alignment and standardization processing on the multi-source data, specifically as follows: Based on the timestamp of the mobile monitoring data, the hourly timestamps of the national control point data and meteorological data are mapped to the midpoint of that hour; Windowing is constructed for the mapped national monitoring point data and meteorological data, centered on the set minute-level timestamps; The temporal attention score for each data point within the window is calculated based on the time decay coefficient and the region weight. The minute-level interpolated data of national control point data and meteorological data are obtained by weighting the time attention scores of each data point in the window; A three-dimensional spatial grid is established based on the ship's navigation route, and the initial spatial weight of each grid is calculated using an inverse distance weighted interpolation algorithm. Set pollution source matching weights within each grid, and calculate the spatial attention score for each grid based on the initial spatial weights and the pollution source matching weights; Based on the spatial attention score, pollution source attributes are matched to each grid, and the data format and units in each grid are standardized to obtain a standardized 3D grid dataset.

[0009] Furthermore, the three-dimensional grid dataset is processed using a forward matrix factorization model based on Bayesian optimization factors to obtain a factor-species concentration matrix, including: The input matrix for the positive matrix decomposition model is established based on the volatile organic compounds with corresponding pollution source attributes contained in the three-dimensional mesh dataset. Define the range of factors for the forward matrix factorization model, and run the Bayesian optimization algorithm to iteratively search within this range to obtain the optimal number of factors for the forward matrix factorization model. The forward matrix factorization model performs factorization calculations on the input matrix based on the optimal number of factors to obtain a factor-species concentration matrix.

[0010] Furthermore, a classification model is constructed based on the random forest algorithm, and the factor-species concentration matrix is ​​identified using the classification model to obtain pollution source type identification results, including: Based on the pre-defined correspondence between pollution source attributes and volatile organic compound species, a training dataset containing factor component data and corresponding pollution source type labels is established. The factor component data in the training dataset are standardized. The random forest algorithm is used to train the classification model based on the standardized training dataset. The factor-species concentration matrix is ​​input into the trained classification model, and the pollution source type identification result is output.

[0011] Furthermore, based on the factor-species concentration matrix and the pollution source type identification results, a gradient boosting regression tree is used to obtain the predicted value of secondary organic aerosol generation potential, including: The model input features are constructed based on the factor-species concentration matrix and the pollution source type identification results. The corresponding secondary organic aerosol generation potential is obtained as label data based on the model input features. A prediction model is constructed using a gradient boosting regression tree, and the prediction model is trained using labeled data; The input features of the model are input into the trained prediction model to obtain the predicted value of the secondary organic aerosol generation potential.

[0012] Furthermore, based on the three-dimensional grid dataset, the pollution source type identification results, and the predicted value of secondary organic aerosol generation potential, a geographically weighted regression model combined with a hotspot analysis algorithm is used to obtain visualized source tracing results, specifically: A geographic weighted regression model is constructed based on the aforementioned 3D grid dataset, and the spatial weight matrix of each monitoring point in the geographic weighted regression model is calculated using the Gaussian kernel function. The geographic weighted regression model is fitted using the local weighted least squares method based on the spatial weight matrix of each monitoring point to obtain the local regression coefficients for each monitoring point. Calculate the contribution weight of each pollution source type to each monitoring point based on the local regression coefficient of each monitoring point, and simulate the pollution transmission trajectory based on the contribution weight. Based on the pollution source type identification results and the predicted value of secondary organic aerosol generation potential, the hotspot analysis algorithm is used to locate the pollution source aggregation area in the three-dimensional grid dataset. Visual source tracing results are obtained based on the pollution transmission trajectory and the pollution value clustering areas in the three-dimensional grid dataset.

[0013] The beneficial effects of this invention are as follows: This invention achieves accurate fusion and pollution source tracing of multi-source data on the atmospheric environment of the Yangtze River surface. Specifically, it simplifies the overall process of data processing and source tracing in complex water surface environments through a four-layer architecture of "monitoring, fusion, analysis, and application," and reduces the difficulty of solving the interference of ship turbulence and the spatiotemporal misalignment of multi-source data. Furthermore, through dynamic optimization using technologies such as dynamic sampling frequency adaptive algorithms, Bayesian and random forest algorithms to optimize the forward matrix factorization model, and gradient boosting regression trees, it improves the accuracy of data collection in highly polluted areas, the accuracy of pollution source identification, and the scientific nature of quantifying the contribution of secondary organic aerosols. Ultimately, it significantly improves the monitoring accuracy of volatile organic compound species concentrations in the atmospheric environment of the water surface and the efficiency of quantitative source tracing of pollution sources. It can effectively meet the actual needs of cross-regional joint prevention and control in the Yangtze River Basin for accurate environmental data and prevention and control recommendations, and provides efficient and reliable technical support for the governance and ecological protection of atmospheric pollution on the Yangtze River surface. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a schematic diagram of the water surface atmospheric environment monitoring and tracing system based on multi-source data fusion according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the results of mobile monitoring of volatile organic compounds in the Yangtze River section of Jiangsu Province in autumn, according to an embodiment of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] A schematic diagram of the water surface atmospheric environment monitoring and tracing system based on multi-source data fusion according to an embodiment of the present invention is shown below. Figure 1 As shown, it includes: The monitoring layer, based on the monitoring equipment on board the ship, uses a dynamic sampling frequency adaptive algorithm to collect multi-source data in real time during the ship's navigation. In this embodiment of the invention, the multi-source data collected by the monitoring layer includes at least mobile monitoring data, national control point data, meteorological data, and industrial cluster data. Specifically, the mobile monitoring data includes ship position and navigation data, as well as collected data on water surface particulate matter and volatile organic compounds (VOCs). The ship position and navigation data can be obtained through GPS or Beidou dual-mode positioning and ship attitude sensors to obtain ship latitude and longitude, sailing speed, and roll and pitch angle data with an accuracy of 1 meter. Water surface particulate matter can be obtained through point-type particulate matter analysis units or particulate matter remote sensing lidar, and the collected data includes PM2.5 / PM10 concentrations, 0.1-10μm particulate matter size distribution, etc. The volatile organic compound data can be obtained through point-type gas spectral detection units or boundary layer detection lidar, and the specific data can include the content of total volatile organic compounds in the air (TVOCs concentration) and the component proportion of characteristic species such as trichloroethane, ethylbenzene, and pentene.

[0018] Data from national monitoring stations and industrial clusters can be obtained through the national monitoring station interface and the geographic database of industrial clusters. This typically includes hourly PM2.5 / PM10 data from national monitoring stations along the coast, as well as location and industry type data for 20 industrial clusters. This data supplements fixed-point data from mobile monitoring and supports pollution source correlation.

[0019] Meteorological data can be acquired through wind profile lidar or boundary layer detection lidar, obtaining data such as wind speed, wind direction (vertical profile + near-surface), boundary layer height, temperature and humidity, and cloud thickness. During real-time acquisition, ship attitude data can also be combined to correct wind speed deviations.

[0020] In one specific embodiment of the present invention, the monitoring equipment carried by the ship determines the ship's position information in real time through the monitoring layer during navigation. At the same time, the monitoring equipment continuously collects data such as VOCs concentration in the atmosphere. Therefore, the present invention designs a dynamic sampling frequency adaptive algorithm to realize intelligent collection of multi-source data. It obtains the ship's position information in real time; compares the ship's position information with the geographical location of a preset industrial cluster to determine whether the ship has entered the industrial cluster; when the comparison result shows that the ship has entered the industrial cluster, due to frequent industrial activities, the atmospheric pollution emissions in the industrial cluster are complex and the concentration is high, making it a key monitoring area. At this time, it is necessary to increase the frequency of the equipment collecting multi-source data from the first frequency to the second frequency. In this embodiment of the present invention, it can be increased from the conventional once every 5 minutes to once every 1 minute; when the comparison result shows that the ship has left the industrial cluster, the frequency of the equipment collecting multi-source data is restored to once every 5 minutes to balance the amount of data collected and the consumption of system resources. This dynamic sampling frequency adaptive algorithm design can collect data more intensively, capture instantaneous changes in pollution source concentration, and obtain richer and more accurate data, providing strong support for subsequent pollution analysis and source tracing.

[0021] The data transmission and fusion layer is used to clean the multi-source data and to perform spatiotemporal alignment and standardization on the multi-source data using a time attention mechanism and an inverse distance weighted interpolation algorithm to obtain a standardized three-dimensional mesh dataset. In this embodiment of the invention, after collecting multi-source data, the first step is to clean the obtained mobile monitoring data, national control point data, meteorological data, and industrial cluster data. The specific cleaning process is as follows: For onboard monitoring data, based on the pre-set monitoring period of 7:00-17:00, all sampling data outside this period are first removed. Simultaneously, abnormal data with a roll angle >15° or pitch angle >10° are removed, as the sampling flow rate of the flight mass spectrometer and particulate matter size spectrometer is affected by turbulence during these periods, leading to concentration measurement deviations. Such data is marked as "invalid" and deleted. When the concentration of volatile organic compounds (VOCs) is 0, this does not mean there are no actual emissions, but rather that it is below the maximum range of the monitoring equipment. Therefore, this data value is automatically replaced with half of the maximum range of the monitoring equipment, and an uncertainty is set for this data, which can be set to 5 / 6 of the maximum range. When the concentration of VOCs exceeds the instrument's range, a 3... The criteria identify and calculate the mean and standard deviation of the concentration of volatile organic compounds (VOCs) during a certain period of navigation. If a sampling point for a certain VOC is missing, linear interpolation is used to fill in the missing data. Typically, the data is filled in by weighting the concentrations of the two valid points before and after the missing point according to the time interval. For example, if the missing point is at 10:00, the concentrations at 9:59 and 10:01 are used for linear interpolation. If more than three consecutive points are missing, the data is filled in by combining the ERA5 wind speed data of the same period with the concentration of the previous sampling point.

[0022] For national control point data, since the national control points are hourly data, this embodiment of the invention, after removing the data marked by equipment failure, only retains the hourly data that overlaps with the monitoring period (7:00-17:00) of the mobile monitoring data. At the same time, if the national control point data for a certain hour is missing, it can be supplemented by the average of adjacent national control points.

[0023] For meteorological data, in this embodiment of the invention, meteorological data that matches the mobile monitoring area and the monitoring period of the mobile monitoring data are also selected, grid data that exceeds the area are removed, and time periods with less than 90% data integrity are deleted.

[0024] For industrial cluster data, this embodiment of the invention retains only the data of 20 industrial clusters associated with the mobile survey area, removes pollution sources whose geographical location exceeds the range, and verifies the completeness of pollution source attribute data, such as industry type and pollution discharge declaration information. Pollution sources that are missing major emission sources are marked as to be supplemented, so that they do not participate in subsequent spatial association operations.

[0025] This invention further addresses the spatiotemporal differences of multi-source data by employing a temporal attention mechanism and an inverse distance weighted interpolation algorithm to perform spatiotemporal alignment and standardization processing on the multi-source data, as detailed below: In the time dimension, this embodiment of the invention uses the timestamp of the ship's underway monitoring data as the benchmark, i.e., the ship's own GPS timestamp, to map the hourly times of the national control point data and meteorological data to the midpoint of that hour, for example, mapping 14:00 to 14:30, while unifying the time format to UTC+8. Since both the national control point data and meteorological data are hourly data, they need to be further interpolated to minute-level data. Therefore, this embodiment of the invention introduces a time attention mechanism. A window is constructed for the mapped national control point data and meteorological data, centered on a set minute-level timestamp. Each window contains one hour of data, for example, the hourly data from 08:00 to 09:00, corresponding to 08:30 within the window. The time attention score for each data point within the window is calculated based on the time decay coefficient and regional weight, expressed as: In the formula, Indicates the first in the window The time attention score for each data point to be interpolated. For the first in the window The time for the data to be interpolated For the set minute-level timestamp, The time decay coefficient is set to 15 minutes in this embodiment of the invention to ensure that the closer the distance, the higher the weight. The regional weight refers to the weight given when the target time t falls within the sampling period of an industrial cluster. For non-industrial areas, the value is 1.0, highlighting the time weight of key areas.

[0026] After calculating the time attention score for each data point, weighted interpolation is performed to generate minute-level data. If there are multiple interpolated data points in the window, the target minute-level data can be calculated by weighting the time attention scores.

[0027] Regarding the spatial dimension, this embodiment of the invention introduces a spatial attention scoring mechanism based on the inverse distance weighted interpolation algorithm, and highlights the grid association weight of industrial clusters according to the pollution source matching weight, solving the problem of concentration and pollution source mismatch caused by location misalignment. First, based on the latitude and longitude coordinates of the mobile sampling route, a 100m×100m×10m three-dimensional spatial grid is constructed to ensure that each grid can cover the spatial range of the mobile sampling point, national control point, and industrial cluster. For each spatial grid, the distance to the surrounding national control point is calculated, and the influence radius is set to 5km, with only national control points within the radius participating in the interpolation. Next, the initial spatial weight of each grid is calculated using the following formula: ,in, The straight-line distance (km) from the grid center to the national control point is given. The closer the distance, the greater the weight. Finally, the data concentration in each grid is interpolated and expressed as: Data concentration of a grid = (Concentration of national control point A × 1 + Concentration of national control point B × 0.25) / (1 + 0.25).

[0028] Furthermore, a spatial attention scoring mechanism is introduced based on the initial spatial weights, and a pollution source matching weight is set within each grid. The spatial attention score for each grid is calculated based on the initial spatial weights and the pollution source matching weights, as follows: in, This refers to the pollution source matching weight, which indicates that if the volatile organic compound species in the grid match the species emitted by the pollution source, such as olefins emitted from a dock, and the olefin content in the grid is high, then... Finally, the pollution source attributes are associated with the grid. The associated information of the coastal pollution sources for each grid is: (Pollution source 1 attribute × +Pollution Source 2 Attributes × ) / This lays a spatial foundation for matching concentrations with sources during subsequent application-level source tracing. It links mobile monitoring data points, fixed national control points, and industrial cluster locations to a unified spatial grid, resolving the mismatch between data concentrations and pollution sources caused by location discrepancies. Finally, it standardizes the format and units of all data, unifying the units for volatile organic compound species and particulate matter concentrations. / The time format is UTC+8, and meteorological parameters are retained to one decimal place. After data cleaning and spatiotemporal alignment, a concentration file and an uncertainty file that meet the input requirements of the forward matrix factorization model are generated. At the same time, a standardized grid dataset of 100m×100m×10m three-dimensional grid is output. Each grid contains: time stamp, mobile monitoring data, national control point interpolation data, meteorological data, pollution source association data, and data quality labels. This dataset fully meets the input requirements of the subsequent intelligent analysis layer and solves the problems of low quality and spatiotemporal disorder of multi-source water surface data.

[0029] The intelligent analysis layer uses a forward matrix factorization model based on Bayesian optimization factors to process the three-dimensional grid dataset to obtain a factor-species concentration matrix; a classification model is constructed based on the random forest algorithm, and the classification model is used to identify the factor-species concentration matrix to obtain the pollution source type identification results; based on the factor-species concentration matrix and the pollution source type identification results, a gradient boosting regression tree is used to obtain the predicted value of the secondary organic aerosol generation potential. In this embodiment of the invention, the forward matrix factorization model can identify the different characteristics of each factor and thus distinguish them among pollution sources. Therefore, it is first necessary to screen out volatile organic compounds with clear pollution source orientation based on a standardized grid dataset as model inputs. For example, tracers from petrochemical energy combustion sources include: 1,3-butadiene (exclusive emissions from petrochemical combustion) and styrene (byproducts of refining processes); tracers from port and dock sources include: pentene (vaporization from ship fuel oil) and toluene (vaporization from container coatings); tracers from shipbuilding industry sources include: trichloroethane (main component of metal cleaning agents) and dichloromethane (rust removal solvent); and tracers from solvent use sources include: ethylbenzene (printing ink) and ethyl acetate (coating solvent). At the same time, the input data must meet the formal requirements of the US forward matrix factorization model, that is, it must include data concentration files and uncertainty files, and the data time series must be on the minute level.

[0030] Traditional forward matrix factorization models decompose the concentration matrix of volatile organic compounds into a source contribution matrix and a source component spectrum matrix through factorization. This process requires manually setting the number of factors and judging the goodness of fit through Q-values ​​(Qtrue / Qrobust), where Qtrue represents the true Q-value and Qrobust represents the stable Q-value. Manually setting the number of factors has the problems of large subjective bias and low trial-and-error efficiency. Therefore, this invention introduces a Bayesian optimization algorithm to achieve automated search for the number of factors.

[0031] In this embodiment of the invention, a parameter space is first defined, and the search range for the number of factors is set to 2-6. This range is obtained by considering the complexity of the types of volatile organic compounds on the river surface and excluding unreasonable cases of single factors or more than 6 factors. The optimization objective is to minimize the absolute deviation of the Q value (Qtrue / Qrobust) from 1. The formula is as follows: At the same time, add constraint residuals The standard deviation is less than 15% to avoid ignoring data rationality in pursuit of goodness of fit; Gaussian process regression (GPR) is used as a surrogate model, which has strong fitting ability for discrete data and can output prediction confidence intervals, adapting to the uncertainty characteristics of water surface data. The kernel function is a squared exponential kernel, and the initial parameter is set to the signal variance. To avoid excessive initial bias in the surrogate model, the embodiment of this invention also uses Latin hypercube sampling to randomly select three initial factor numbers from the parameter space, substitutes them into the forward matrix factorization model to calculate the corresponding Q value and residual, and inputs the three initial sampling points into the Gaussian process regression model to fit the initial probability distribution of the factor number and loss value, thereby constructing the initial surrogate model.

[0032] Next, the Bayesian optimization algorithm is iteratively optimized. Each iteration selects points using the data acquisition function, performs forward matrix factorization calculations, and then updates the surrogate model, gradually narrowing the search range for the optimal number of factors. The specific process is as follows: Data collection function selection: Data collection points can be selected using the expected improvement function, which calculates the number of data points at a certain factor. The expected probability that sampling at a given location will reduce the loss function is given by the formula: ,in The current minimum loss function value (initial) The point with the highest expected value is selected as the next sampling point.

[0033] In the initial iteration, the expected value of p=3 is 0.032, the predicted loss function value is 0.12, and there is a 32% probability that the loss function value will be lower than the current minimum value of 0.06. The expected value of p=5 is 0.018, and the predicted loss function value is 0.15. There is only an 18% probability that the loss function value will decrease, so p=3 is chosen as the sampling point. Forward matrix factorization calculation: Substituting p=3 into the forward matrix factorization model, we get a Q value of 1.18 (loss function value of 0.12) and a residual of 16.2%. Since the constraint is not met, this point is marked as an invalid sampling point and the loss function value is assigned to infinity to avoid misleading the model. Proxy model update: Based on the results of forward matrix factorization, new sampling points and historical valid points are merged, and the parameters of the Gaussian process regression model are adjusted using maximum likelihood estimation to improve the model's fitting accuracy to the water surface data. In subsequent iterations, sampling p=5 (Q value is 1.15, loss function value is 0.15, residual is 14.5%) is used to verify the stability of p=4. When the termination condition of loss function value fluctuation being less than 0.02 for three consecutive rounds is finally met, the optimal number of factors p is output. After p is determined, the goodness of fit is judged by the Q value, and the factor-species concentration matrix is ​​obtained.

[0034] After identifying the different characteristics of each factor, different pollution sources need to be classified. However, traditional source type identification requires manual inference of the source type based on the proportion of characteristic species in the factor composition spectrum, which is inefficient and relies on experience. Therefore, in this embodiment of the invention, a random forest classification model is trained to achieve automated identification.

[0035] First, based on the characteristics of industrial pollution during the navigation section, a characteristic species definition system for four types of pollution sources was pre-constructed to provide a basis for the model's judgment. The specific correspondence is shown in Table 1. Table 1. Correspondence between pollution source types and volatile organic compound species In this embodiment of the invention, based on the pre-defined correspondence between pollution source attributes and volatile organic compound species in the superscript, each factor parsed from the factor-species concentration matrix is ​​manually labeled, thereby forming a training dataset containing factor component data and corresponding pollution source type labels. This training dataset contains monitoring data from different flight segments and time periods to ensure the model's generalization ability. At the same time, since the concentration units of different volatile organic compound species are consistent, but the numerical ranges vary greatly, the factor component data needs to be Z-score standardized to eliminate the influence of dimensions and avoid excessive interference from high-concentration species on model training.

[0036] Further training is performed using the random forest algorithm on the standardized training dataset to obtain a trained classification model. The core of this model is to reduce the overfitting risk of a single decision tree by ensembling multiple decision trees, thereby improving recognition stability. In this embodiment of the invention, the key parameters of the random forest algorithm are set for the water surface air pollution source identification scenario, including: Number of decision trees: The optimal value was determined through 5-fold cross-validation, and 200 decision trees were ultimately selected to ensure model accuracy while avoiding waste of computational resources; Maximum number of features: Each tree was randomly selected. Split based on features to ensure diversity in each tree and reduce correlation between trees; Maximum depth: set to 8 to prevent the model from overfitting the training data; Minimum number of sample splits: set to 5 to ensure sufficient sample size when splitting the decision tree and avoid splitting out nodes with no statistical significance.

[0037] After setting the parameters, Bootstrap resampling is used to randomly extract several sub-samples from the training set, allowing repeated sampling. Each sub-sample corresponds to the training data of a decision tree. For each decision tree, nodes are recursively split based on randomly selected features and the Gini coefficient is used as the splitting criterion. No pruning is performed on each decision tree until all samples within a node are of the same pollution source type, or the stopping condition is met. The above steps are repeated to construct 200 independent decision trees, forming a random forest classification model.

[0038] By inputting the factor-species concentration matrix into the trained classification model, new factor component data can be automatically identified. At the same time, the prediction results of 200 decision trees are integrated using the majority voting method. That is, the number of votes obtained for each type of pollution source is counted, and the type with the most votes is the pollution source type corresponding to that factor. The percentage of votes for the type with the most votes is calculated as the confidence level of the identification result. If the confidence level is less than 50%, the model will trigger an early warning, indicating that further verification is needed in combination with the geographical information of pollution sources along the coast to ensure the reliability of the identification. The pollution source type identification result is then output.

[0039] This invention further utilizes the factor-species concentration matrix and pollution source type identification results to obtain secondary organic aerosol generation potential (SOAP) prediction values ​​using a gradient boosting regression tree, including the following steps: First, a training dataset needs to be constructed to train the gradient boosting regression tree. This dataset needs to contain both input features and label values, and include samples from different flight segments, time periods, and pollution source types, with a minimum of 500 sets to ensure the model's generalization ability. The sources for constructing this training dataset can be found in Table 2. Table 2. Sources of training datasets for gradient boosting regression trees The gradient boosting regression tree model uses 5-fold cross-validation to select the optimal parameters, balancing accuracy and computational efficiency. Its core parameters include the number of decision trees, learning rate, depth, and minimum number of sample splits. Iteratively, the regression trees are built starting with the mean SOAP value in the training set as the initial value. The residuals of the current model are calculated, and a regression tree is built based on the residuals. The split nodes of the tree are determined by minimizing the mean squared error. Weights corresponding to the learning rate are assigned to the regression tree, and its prediction results are superimposed with the prediction results of all previous trees to update the model's total prediction value. The process of iteratively building regression trees is repeated until 300 trees are built, or the MSE of the model on the validation set no longer decreases, and finally a trained prediction model is formed. Then, the input features are input into the trained prediction model to obtain the secondary organic aerosol generation potential prediction value, which is also associated with the corresponding pollution source type.

[0040] At the application layer, based on the three-dimensional grid dataset, the pollution source type identification results, and the predicted value of secondary organic aerosol generation potential, a geographically weighted regression model combined with a hotspot analysis algorithm is used to obtain visualized source tracing results.

[0041] In this embodiment of the invention, a geographic weighted regression model is first constructed based on a three-dimensional grid dataset, and a Gaussian kernel function is used to calculate the spatial weight matrix of each monitoring point in the geographic weighted regression model. Based on the spatial weight matrix of each monitoring point, the geographic weighted regression model is fitted using the local weighted least squares method to obtain the local regression coefficients for each monitoring point, expressed as: in, Let represent the concentration of volatile organic compounds at the i-th monitoring point, which is the dependent variable of the model; represents the local regression coefficient of the i-th monitoring point, reflecting the strength of the influence of the independent variable on the dependent variable at that point; The residuals represent the values ​​that must follow a normal distribution and have a standard deviation of less than 15%. The independent variables of the model are the data contained in the 3D grid dataset, specifically the distance to the pollution source. Meteorological parameters, including wind speed. The wind direction is Boundary layer height The model iteration process is as follows: Initial bandwidth Then, calculate the spatial weight matrix for each monitoring point. For each monitoring point, the weighted least squares method is used based on the weights. Fit its local regression coefficients Then calculate the model goodness of fit. ,like Adjust the bandwidth h, repeat the iteration until... Finally, through local regression coefficients Calculate the contribution percentage of each pollution source to the monitoring point, output the pollution source contribution weight, and combine it with the wind direction coefficient to simulate the transmission path of the pollution source from the source to the monitoring point, and output the pollution transmission trajectory.

[0042] To address the types of pollution sources and the risk of secondary pollution, this embodiment of the invention further employs a hotspot analysis algorithm to cluster and locate high-value pollution clusters. First, based on the navigation route, the river surface is divided into a 100m × 100m planar grid. The concentration value of each grid is the mean of all navigation sampling points within that grid. Then, a distance threshold weight is applied to set a threshold. This threshold represents the effective radius of influence of pollution sources spreading across the river surface. If the distance between grid i and grid j is less than or equal to 3 km, they are considered adjacent grids and have a spatial correlation. Their weights are determined accordingly. ,otherwise After standardizing the concentration values ​​of all grids, the local hotspot index is calculated for each grid i. The formula is: in, This represents the standardized concentration value of grid j; This represents the average concentration value across all grid cells. The variance of the concentration values ​​across all grid cells is represented; n represents the total number of grid cells involved in the calculation. The local hotspot index is calculated. After the value, if This grid is a hotspot, indicating that the surrounding grids have high concentrations, forming a high-value cluster; if If the grid is a cold spot, it means that the concentration of the surrounding grids is low, forming a low-value cluster. Finally, the spatial distribution of pollution hotspots is obtained and the industrial type along the coast of the hotspot area is automatically matched.

[0043] In one specific embodiment of the present invention, the scenario is a mobile monitoring of volatile organic compounds in the Yangtze River section of Jiangsu Province (Nanjing to Nantong section, 250km in length) in October 2023. This mobile monitoring route passes through 20 industrial parks, including: the Qixia Chemical Industrial Park (K0-K20 km) and the Jiangdu Shipbuilding Base (K80-K100 km); the Yangzhou urban riverside area (K40-K60 km); and the Yizheng Yangtze River Wetland (K20-K40 km). Schedule: December 26-30, 2023, daily 08:00-18:00, with additional morning peak emission monitoring from 05:00-07:00 in industrial parks; Baseline data: Simultaneously accessing data from 5 national monitoring stations along the coast (hourly concentration data), 3 online monitoring stations in industrial parks (15-minute data), and ship AIS navigation data. The detailed schedule is shown in Table 3. Table 3. Underway Timetable The monitoring equipment carried by the monitoring vessel includes: a PTR-TOF MS 6000 proton transfer reaction time-of-flight mass spectrometer; a meteorological five-parameter sensor, a GPS positioning module, and a 4G or 5G transmission terminal; through dynamic sampling frequency control, the sampling frequency in normal areas is once every 2 minutes; in industrial concentrated areas, after the system identifies the area through GPS, it automatically triggers high-frequency sampling once every 30 seconds.

[0044] After collecting multi-source data, data preprocessing was performed, including: anomaly removal: automatically filtering data from periods of equipment failure (such as sensor drift data at K50 km from 10:30 to 11:00 on December 28); time alignment: using the monitoring vessel's GPS timestamp as a benchmark, the hourly data from national control points and the 15-minute data from the industrial park were mapped to a unified time axis. Then, the weighted interpolation method proposed in this invention was used for time dimension interpolation; the spatial dimension fusion adopted an improved inverse distance weighted difference algorithm, assigning a spatial weight of 1.3 times to the industrial cluster grid.

[0045] By inputting the fused data of 120 volatile organic compounds, meteorological data, and pollution source type data into the model constructed in this invention, the classification results output by the random forest model are as follows: the main pollution source in the Qixia section is chemical coating emissions (contributing 42%), and the main pollution source in the Jiangdu section is shipbuilding and repair exhaust gas (contributing 38%).

[0046] Further assessment based on the predicted potential of secondary organic aerosol formation revealed that the Qixia section had the highest contribution to diethylbenzene (35%), while the Jiangdu section had the highest contribution to toluene (28%), thus providing a basis for key controlled components.

[0047] Finally, through hotspot analysis, K15 km (near the sewage outlet of the chemical industrial park) and K90 km (near the ship repair yard) were identified as high-value hotspot areas. When using geographically weighted regression quantification, it was found that when the wind speed was <2 m / s, the influence coefficient of the chemical industrial park emissions on the concentration in the downwind 5 km area reached 0.72. Therefore, a dynamic map was used to display the thermal distribution of volatile organic compound concentrations and the movement trajectory of high values ​​in real time. At the same time, the source tracing revealed 3 key emitting enterprises, thus providing control suggestions such as optimizing production restriction periods and improving waste gas collection efficiency.

[0048] In one specific embodiment of the present invention, monitoring results show that the concentration and composition characteristics of volatile organic compounds in the Yangtze River section exhibit a trend of low concentration upstream and high concentration downstream. Figure 2 As shown, the concentration of volatile organic compounds (VOCs) in the Yangtze River section is higher in the east and lower in the west. The concentration trends of VOCs in the outbound and return journeys are basically consistent, with an average of 74.26 in the eastern section (from Jiangyin to Taicang). g / The average value for the western section (Nanjing to Jiangyin) is 26.92. g / On October 9, there was fog on the Yangtze River section, with many low-altitude clouds and high humidity, which was not conducive to the diffusion of pollution sources, resulting in a relatively high overall concentration. On October 11, the weather was sunny, and the overall concentration was relatively low, but the spatial trend was basically the same.

[0049] Figure 2 Further analysis revealed a strong correlation between high concentrations of volatile organic compounds (VOCs) and the distribution of volatile organic compounds at the dock and power plant. Data were selected chronologically based on VOC concentrations exceeding 500 during the transit period. g / The table shows the six time periods and the top five dominant species by mass concentration, as shown in Table 4. Table 4. Dominant Species in Areas with High Volatile Organic Compound Concentrations The dominant species in the first three high-value points on the outbound journey were all halogenated hydrocarbons, with trichloroethane and tetrachloroethane generally accounting for more than 30% in total. The content of volatile organic compounds in the Changshu section did not differ much, with diethylbenzene having the highest content. On the return journey, two high-value points were measured in the Taizhou-Jingjiang section. The first high-value point had pentene accounting for 40%, and the second high-value point had ethylbenzene accounting for 60%.

[0050] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A water surface atmospheric environment monitoring and tracing system based on multi-source data fusion, characterized in that, include: The monitoring layer, based on the monitoring equipment on board the ship, uses a dynamic sampling frequency adaptive algorithm to collect multi-source data in real time during the ship's navigation. The multi-source data includes at least: mobile monitoring data, national monitoring point data, meteorological data, and industrial cluster data; The data transmission and fusion layer is used to perform data cleaning on the multi-source data, and to perform spatiotemporal alignment and standardization on the multi-source data using a time attention mechanism and an inverse distance weighted interpolation algorithm to obtain a standardized three-dimensional mesh dataset. The intelligent analysis layer processes the three-dimensional grid dataset using a forward matrix factorization model based on Bayesian optimization factors to obtain a factor-species concentration matrix; a classification model is constructed based on the random forest algorithm, and the classification model is used to identify the factor-species concentration matrix to obtain pollution source type identification results; based on the factor-species concentration matrix and the pollution source type identification results, a gradient boosting regression tree is used to obtain secondary organic aerosol generation potential prediction values. At the application layer, based on the three-dimensional grid dataset, the pollution source type identification results, and the predicted value of secondary organic aerosol generation potential, a geographically weighted regression model combined with a hotspot analysis algorithm is used to obtain visualized source tracing results.

2. The water surface atmospheric environment monitoring and tracing system based on multi-source data fusion according to claim 1, characterized in that: During ship navigation, a dynamic sampling frequency adaptive algorithm is used to collect multi-source data in real time, specifically: Real-time acquisition of ship location information; The ship's location information is compared with the geographical location of the preset industrial cluster to determine whether the ship has entered the industrial cluster. When the comparison results show that the ship enters the industrial cluster, the frequency of the equipment collecting multi-source data will be increased from the first frequency to the second frequency. When the comparison results show that the ship has left the industrial cluster, the frequency of the equipment collecting multi-source data will be restored to the first frequency.

3. The water surface atmospheric environment monitoring and tracing system based on multi-source data fusion according to claim 1, characterized in that: The data cleaning process for the multi-source data includes: For the aforementioned navigation monitoring data, data outside the preset monitoring period are removed; data collected when the ship's lateral roll angle is greater than 15 degrees or the longitudinal roll angle is greater than 10 degrees are removed; data with a concentration value of zero are replaced with half of the maximum range value of the monitoring equipment; data exceeding the range of the monitoring equipment are identified using the 3σ criterion. For the national monitoring point data, data collected during equipment failures are removed, and data overlapping with the preset monitoring period is retained; for national monitoring point data with missing data, the average of the data from its adjacent national monitoring points is used to complete the data. For the meteorological data, data that overlaps with the preset monitoring period are filtered out, and data from periods with less than 90% data integrity are removed; For the industrial cluster data, retain the industrial cluster data associated with the mobile monitoring area, and remove pollution sources whose geographical location is outside the mobile monitoring area.

4. The water surface atmospheric environment monitoring and tracing system based on multi-source data fusion according to claim 1, characterized in that: The multi-source data is spatiotemporally aligned and standardized using a time attention mechanism and an inverse distance weighted interpolation algorithm, specifically as follows: Based on the timestamp of the mobile monitoring data, the hourly timestamps of the national control point data and meteorological data are mapped to the midpoint of that hour; Windowing is constructed for the mapped national monitoring point data and meteorological data, centered on the set minute-level timestamps; The temporal attention score for each data point within the window is calculated based on the time decay coefficient and the region weight. The minute-level interpolated data of national control point data and meteorological data are obtained by weighting the time attention scores of each data point in the window; A three-dimensional spatial grid is established based on the ship's navigation route, and the initial spatial weight of each grid is calculated using an inverse distance weighted interpolation algorithm. Set pollution source matching weights within each grid, and calculate the spatial attention score for each grid based on the initial spatial weights and the pollution source matching weights; Based on the spatial attention score, pollution source attributes are matched to each grid, and the data format and units in each grid are standardized to obtain a standardized 3D grid dataset.

5. The water surface atmospheric environment monitoring and tracing system based on multi-source data fusion according to claim 1, characterized in that: The three-dimensional grid dataset was processed using a forward matrix factorization model based on Bayesian optimization factors to obtain a factor-species concentration matrix, including: The input matrix for the positive matrix decomposition model is established based on the volatile organic compounds with corresponding pollution source attributes contained in the three-dimensional mesh dataset. Define the range of factors for the forward matrix factorization model, and run the Bayesian optimization algorithm to iteratively search within this range to obtain the optimal number of factors for the forward matrix factorization model. The forward matrix factorization model performs factorization calculations on the input matrix based on the optimal number of factors to obtain a factor-species concentration matrix.

6. The water surface atmospheric environment monitoring and tracing system based on multi-source data fusion according to claim 1, characterized in that: A classification model is constructed based on the random forest algorithm. This model is then used to identify the factor-species concentration matrix, yielding pollution source type identification results, including: Based on the pre-defined correspondence between pollution source attributes and volatile organic compound species, a training dataset containing factor component data and corresponding pollution source type labels is established. The factor component data in the training dataset are standardized. The random forest algorithm is used to train the classification model based on the standardized training dataset. The factor-species concentration matrix is ​​input into the trained classification model, and the pollution source type identification result is output.

7. The water surface atmospheric environment monitoring and tracing system based on multi-source data fusion according to claim 1, characterized in that: Based on the factor-species concentration matrix and the pollution source type identification results, a gradient boosting regression tree is used to obtain the predicted value of secondary organic aerosol generation potential, including: The model input features are constructed based on the factor-species concentration matrix and the pollution source type identification results. The corresponding secondary organic aerosol generation potential is obtained as label data based on the model input features. A prediction model is constructed using a gradient boosting regression tree, and the prediction model is trained using labeled data; The input features of the model are input into the trained prediction model to obtain the predicted value of the secondary organic aerosol generation potential.

8. The water surface atmospheric environment monitoring and tracing system based on multi-source data fusion according to claim 1, characterized in that: Based on the three-dimensional grid dataset, the pollution source type identification results, and the predicted value of secondary organic aerosol generation potential, a geographically weighted regression model combined with a hotspot analysis algorithm is used to obtain visualized source tracing results, specifically: A geographic weighted regression model is constructed based on the aforementioned 3D grid dataset, and the spatial weight matrix of each monitoring point in the geographic weighted regression model is calculated using the Gaussian kernel function. The geographic weighted regression model is fitted using the local weighted least squares method based on the spatial weight matrix of each monitoring point to obtain the local regression coefficients for each monitoring point. Calculate the contribution weight of each pollution source type to each monitoring point based on the local regression coefficient of each monitoring point, and simulate the pollution transmission trajectory based on the contribution weight. Based on the pollution source type identification results and the predicted value of secondary organic aerosol generation potential, the hotspot analysis algorithm is used to locate the pollution source aggregation area in the three-dimensional grid dataset. Visual source tracing results are obtained based on the pollution transmission trajectory and the pollution value clustering areas in the three-dimensional grid dataset.

Images