Groundwater pollution tracing method and system based on deep learning and multi-source data fusion

Abstract

The application provides a groundwater pollution tracing method and system based on deep learning and multi-source data fusion, the method uses multi-source data fusion to accurately depict the characteristics of the pollution source, constructs a multi-source data set of the pollution source, and constructs a pollution source identification model based on a CNN convolutional neural network and an LSTM long short-term memory network, trains by using the multi-source data set, fully excavates the characteristics and potential relationships of the multi-source data, so that the tracing model can accurately identify the main pollution source characteristics and the contribution, overcomes the limitations of single-dimensional data tracing and the difficulty of multi-source data fusion, and improves the accuracy and efficiency of groundwater pollution tracing.

Description

Technical Field

[0001] This invention relates to the field of groundwater pollution prevention and control technology, and in particular to groundwater pollution source tracing technology. Specifically, it relates to a groundwater pollution source tracing method and system based on deep learning and multi-source data fusion. Background Technology

[0002] With rapid industrialization and urbanization, groundwater pollution has become an increasingly serious problem. Source tracing of groundwater pollution is a crucial step in pollution control. Traditional methods for tracing groundwater pollution sources rely primarily on limited monitoring data and simple statistical analysis, which have many limitations. For example, source tracing methods based on a single type of monitoring data (such as chemical indicators) often lead to results that deviate from reality due to data incompleteness and uncertainty. Furthermore, statistical analysis methods struggle to address the complex relationships between pollution sources and receptors when dealing with complex groundwater pollution issues.

[0003] Against the backdrop of rapid development in artificial intelligence technology, existing technologies attempt to apply it to groundwater pollution tracing. For example, Chinese patent application number 2024116403398 discloses a watershed pollution tracing method based on spatiotemporal sequence environmental DNA data. However, this method for surface water pollution tracing only characterizes the biological (eDNA) dimension of the pollution source. For groundwater systems, since biological information may change during groundwater migration, using only this dimension of data for groundwater pollution tracing still leads to misjudgments in the tracing results, and its accuracy and applicability are insufficient. Summary of the Invention

[0004] In view of the shortcomings and defects of existing technologies, the purpose of this invention is to provide a groundwater pollution source tracing method based on deep learning and multi-source data fusion. This method uses multi-source data fusion to accurately characterize the features of pollution sources, constructs a multi-source dataset of pollution sources, builds a pollution source identification model and trains it using the multi-source dataset, and fully explores the features and potential relationships of multi-source data. This enables the source tracing model to accurately identify the features of major pollution sources and their contributions, overcomes the limitations of single-dimensional data source tracing and the difficulties of multi-source data fusion, and improves the accuracy and efficiency of groundwater pollution source tracing.

[0005] According to a first aspect of the present invention, a groundwater pollution source tracing method based on deep learning and multi-source data fusion is proposed, comprising the following steps:

[0006] Step 1: Based on historical groundwater monitoring data at the target location, determine the pollutant indicators that need to be traced.

[0007] Step 2: Investigate and mark potential pollution sources in the vicinity of the target location, and determine the background value area;

[0008] Step 3: Perform continuous periodic sampling on potential pollution sources and background value areas to obtain potential pollution source samples;

[0009] Step 4: Test the potential pollution source samples taken continuously in cycles, including three-dimensional fluorescence spectroscopy, eDNA and target pollutant isotope test analysis, and obtain the three-dimensional fluorescence water quality fingerprint data, eDNA data and target pollutant isotope data of the potential pollution source samples respectively.

[0010] Step 5: Based on the three-dimensional fluorescence water quality fingerprint data, eDNA data and isotopic data of the target pollutant from the potential pollution source sample, perform multi-source data fusion to obtain the fusion features of linear combination and construct the feature set of potential pollution sources at the target location.

[0011] Step 6: Construct a source tracing model based on multi-layer convolutional neural network (CNN) and long short-term memory network (LSTM), and train the model based on the feature set of potential pollution sources at the target location to obtain the source tracing model of potential pollution sources at the target location.

[0012] Step 7: For newly sampled groundwater samples at the target location, after feature extraction and analysis through three-dimensional fluorescence spectroscopy, eDNA, and target pollutant isotope testing, the samples are input into the potential pollution source tracing model of the target location, and a multi-dimensional vector is output. Each vector element represents the probability that the data belongs to each potential pollution source.

[0013] Step 8: Based on the characteristics of multi-source data of the target location and the prediction results of the potential pollution source tracing model of the target location, establish a probability model based on Bayes' theorem to obtain the conditional probability distribution under the prior probability of each pollution source.

[0014] Step 9: Combine Monte Carlo simulation and the conditional probability distribution to establish a potential pollution source contribution ratio model and obtain the contribution ratio of each potential pollution source.

[0015] As an optional embodiment, step 2, which involves investigating and marking potential pollution sources within the vicinity of the target location and determining the background value area, includes:

[0016] Step 2-1: Using the target location as the center, investigate the potential pollution sources within the preset radius and mark the specific locations and scale information of domestic pollution sources, agricultural production sources, and industrial production sources.

[0017] Step 2-2: Select an area that is more than a preset range away from the target point, where the surrounding pollution source indicators are lower than the set standards and the geological conditions are uniform as the background value area.

[0018] As an optional embodiment, in step 5, for the test data obtained for each type of potential pollution source, namely three-dimensional fluorescence water fingerprint data, eDNA data, and isotopic data of the target pollutant, a unified data matrix is ​​constructed to characterize its features. The rows of the data matrix represent different sample information of the pollution source, each sample corresponds to a unique identifier, and the columns correspond to the feature parameters of the three-dimensional fluorescence water fingerprint data, eDNA data, and isotopic data of the target pollutant, respectively. For numerical features, they are directly used as matrix elements, and for categorical features, one-hot encoding is used to transform the information as feature parameters.

[0019] As an optional embodiment, in step 5, for the data matrix constructed for each type of potential pollution source, principal component analysis is first used to reduce the dimensionality of the data matrix. By calculating the covariance matrix, eigenvalues ​​and eigenvectors of the data, principal components with a cumulative contribution rate of more than a preset threshold are selected and the number of principal components to be retained is determined.

[0020] Then, the Pearson correlation coefficients between different types of preserved feature parameters are calculated, and different weights are assigned to the feature parameters according to the magnitude of the absolute value of the Pearson correlation coefficients.

[0021] Finally, the feature matrix is ​​weighted according to the assigned weights, and the weighted features are linearly combined to obtain the fused feature matrix.

[0022] As an optional embodiment, for the data matrix constructed for each type of potential pollution source, principal component analysis is first used to reduce the dimensionality of the data matrix. By calculating the covariance matrix, eigenvalues, and eigenvectors of the data, principal components with a cumulative contribution rate exceeding a preset threshold are selected and the number of principal components to be retained is determined, including:

[0023] For any type of potential pollution source, the data matrix is ​​defined as data matrix X, with shape n*p, where n and p represent the number of samples and the number of features, respectively.

[0024] For a data matrix X, calculate its covariance matrix and solve for the eigenvalues ​​and corresponding eigenvectors;

[0025] The calculated eigenvalues ​​and corresponding eigenvectors are arranged in descending order;

[0026] For the sorted feature values, calculate the cumulative contribution rate of the first k principal components, and find the minimum value of k such that the cumulative contribution rate of the first k principal components reaches or exceeds a preset threshold; and

[0027] Select the first k sorted feature vectors to form a p*k matrix V, and project the original data matrix X onto the k principal components to obtain the dimensionality-reduced data matrix Y: Y = X·V.

[0028] As an optional embodiment, for each type of potential pollution source, the data matrix constructed, for its dimensionality-reduced data matrix Y, calculates each pair of features Y in the matrix. i and Y j The Pearson correlation coefficient r between them ij , i,j=1,2,3,…,k;

[0029] Pearson correlation coefficient r among all feature pairs ij Form a k*k correlation matrix R, where the elements of the matrix R are... ij =r ij ;

[0030] Different weighting intervals are set based on the absolute value of the correlation coefficient;

[0031] Traverse each element R in the correlation matrix R ij Based on the weight range, a weight value w is randomly selected for each element. ij This ultimately forms a k*k weight matrix W;

[0032] The data matrix Y is weighted by the k*k weight matrix W, and the weighted features are linearly combined to obtain the fused feature matrix.

[0033] As an optional embodiment, step 6, which involves constructing a source tracing model based on a multi-layer convolutional neural network (CNN) and a long short-term memory network (LSTM), includes:

[0034] A source tracing model combining a multi-layer convolutional neural network (CNN) and a long short-term memory network (LSTM) is constructed, and the specific design is as follows:

[0035] The multi-layer convolutional neural network (CNN) consists of an input layer, convolutional layers, and pooling layers:

[0036] The convolutional layers consist of a first convolutional layer, a second convolutional layer, and a third convolutional layer: The first convolutional layer uses a 3×3 kernel with a stride of 1 and 32 channels. The kernel slides across the input data, extracting preliminary spatial features through convolution operations. Each kernel learns a specific set of feature patterns. The second convolutional layer uses a 5×5 kernel with a stride of 1 and 64 channels. Based on the features extracted by the first convolutional layer, it further captures local features. The third convolutional layer uses a 7×7 kernel with a stride of 1 and 128 channels. This layer aims to obtain deep-level global feature information. After the convolutional layers, two pooling layers are connected: The first pooling layer uses max pooling with a 2×2 window and a stride of 2. By taking the maximum value within the window, the dimensionality of the feature map is reduced while retaining the most important feature information. The second pooling layer uses max pooling with a 3×3 window and a stride of 2, further compressing the data dimensionality.

[0037] The Long Short-Term Memory (LSTM) network consists of two memory units, each containing an input gate, a forget gate, and an output gate:

[0038] The first memory unit has 256 neurons and is used to process time-series data. It controls the input of new information through the input gate, the forget gate determines whether to retain or discard the memory information of the previous moment, and the output gate determines the output of the current moment in order to capture the short-term dynamic changes of the data in the time dimension. The second memory unit has 512 neurons and is used to capture long-term dynamic change patterns.

[0039] As an optional embodiment, in step 8, based on the characteristics of multi-source data of the target location and the prediction results of the potential pollution source tracing model of the target location, a probability model based on Bayes' theorem is established to obtain the conditional probability distribution under the prior probability of each pollution source, including:

[0040] Given that there are n potential pollution sources at a target location, determine the prior probabilities of different types of pollution sources;

[0041] Combining the prediction results of the potential pollution source tracing model at the target location, i.e., outputting a multi-dimensional vector, where each vector element represents the probability that the data belongs to each potential pollution source, the vector elements are normalized, and the softmax function is used to convert the vector into a probability distribution, obtaining the conditional probability distribution under the prior probability of each pollution source:

[0042]

[0043] Among them, z i It is the i-th element of the model's output vector.

[0044] As an optional embodiment, in step 9, the step of establishing a potential pollution source contribution ratio model by combining Monte Carlo simulation and the conditional probability distribution to obtain the contribution ratio of each potential pollution source includes:

[0045] Random samples are generated through Monte Carlo simulation: the number of simulations is set to N. For each simulation, a sample is randomly selected from the conditional probability distribution, and the probability of the sample belonging to each pollution source is calculated. The above simulation process is repeated N times, and the number of times each pollution source is selected is counted, thereby calculating the probability distribution of each pollution source.

[0046] The contribution ratio of each pollution source is calculated based on its probability distribution. For the i-th pollution source, the contribution ratio is calculated using the following formula: Where P i avg It is the average probability of the i-th pollution source, which is obtained by averaging the probability distribution of the pollution source obtained from Monte Carlo simulation; It is the sum of the average probabilities of all pollution sources.

[0047] According to a second aspect of the present invention, a computer system is provided, comprising:

[0048] One or more processors;

[0049] The memory stores operable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, including the steps of the aforementioned method.

[0050] It should be understood that all combinations of the foregoing concepts and the additional concepts described in more detail below may be considered part of the inventive subject matter of this disclosure, provided that such concepts do not contradict each other. Furthermore, all combinations of the claimed subject matter are considered part of the inventive subject matter of this disclosure.

[0051] The foregoing and other aspects, embodiments, and features of the teachings of the present invention will be more fully understood from the following description in conjunction with the accompanying drawings. Other additional aspects of the invention, such as features and / or beneficial effects of exemplary embodiments, will become apparent from the following description or may be learned through practice of specific embodiments according to the teachings of the present invention. Attached Figure Description

[0052] The accompanying drawings are not intended to be drawn to scale. In the drawings, each identical or nearly identical component shown in the various figures may be denoted by the same reference numeral. For clarity, not every component is labeled in each figure. Embodiments of various aspects of the invention will now be described by way of example and with reference to the accompanying drawings.

[0053] Figure 1 This is a flowchart illustrating a groundwater pollution tracing method based on deep learning and multi-source data fusion according to an embodiment of the present invention.

[0054] Figure 2 This is a schematic diagram of the eDNA monitoring results of potential contaminant sources according to an embodiment of the present invention.

[0055] Figure 3 This is a schematic diagram of the three-dimensional fluorescence spectral monitoring results of potential pollution sources according to an embodiment of the present invention.

[0056] Figure 4 This is a schematic diagram of the monitoring results of nitrate isotopes, a potential source of pollution according to an embodiment of the present invention. Detailed Implementation

[0057] To better understand the technical content of the present invention, specific embodiments are described below in conjunction with the accompanying drawings.

[0058] Various aspects of the invention are described in this disclosure with reference to the accompanying drawings, which illustrate numerous illustrative embodiments. The embodiments of this disclosure are not necessarily intended to encompass all aspects of the invention. It should be understood that the various concepts and embodiments described above, as well as those described in more detail below, can be implemented in any of many ways, because the concepts and embodiments disclosed herein are not limited to any particular implementation. Furthermore, some aspects of the invention disclosed may be used alone or in any suitable combination with other aspects of the invention disclosed.

[0059] {Example 1}

[0060] This embodiment aims to propose a groundwater pollution tracing method based on deep learning and multi-source data fusion. It utilizes multi-source data fusion to accurately characterize pollution source features, constructs a multi-source pollution source dataset, and builds a pollution source identification model based on CNN convolutional neural network and LSTM long short-term memory network. The model is trained using the multi-source dataset to fully explore the features and potential relationships of the multi-source data, enabling the tracing model to accurately identify the characteristics and contributions of major pollution sources. This overcomes the limitations of single-dimensional data tracing and the challenges of multi-source data fusion, improving the accuracy and efficiency of groundwater pollution tracing.

[0061] Combination Figure 1 As shown, the groundwater pollution tracing method based on deep learning and multi-source data fusion according to an embodiment of the present invention includes the following steps:

[0062] Step 1: Based on historical groundwater monitoring data at the target location, determine the pollutant indicators that need to be traced.

[0063] Step 2: Investigate and mark potential pollution sources in the vicinity of the target location, and determine the background value area;

[0064] Step 3: Perform continuous periodic sampling on potential pollution sources and background value areas to obtain potential pollution source samples;

[0065] Step 4: Test the potential pollution source samples taken continuously in cycles, including three-dimensional fluorescence spectroscopy, eDNA and target pollutant isotope test analysis, and obtain the three-dimensional fluorescence water quality fingerprint data, eDNA data and target pollutant isotope data of the potential pollution source samples respectively.

[0066] Step 5: Based on the three-dimensional fluorescence water quality fingerprint data, eDNA data and isotopic data of the target pollutant from the potential pollution source sample, perform multi-source data fusion to obtain the fusion features of linear combination and construct the feature set of potential pollution sources at the target location.

[0067] Step 6: Construct a source tracing model based on multi-layer convolutional neural network (CNN) and long short-term memory network (LSTM), and train the model based on the feature set of potential pollution sources at the target location to obtain the source tracing model of potential pollution sources at the target location.

[0068] Step 7: For newly sampled groundwater samples at the target location, after feature extraction and analysis using three-dimensional fluorescence spectroscopy, eDNA, and target pollutant isotope testing, the samples are input into the potential pollution source tracing model at the target location, and a multi-dimensional vector is output. Each vector element represents the probability that the data belongs to each potential pollution source.

[0069] Step 8: Based on the characteristics of multi-source data of the target location and the prediction results of the potential pollution source tracing model of the target location, establish a probability model based on Bayes' theorem to obtain the conditional probability distribution under the prior probability of each pollution source.

[0070] Step 9: Combine Monte Carlo simulation and conditional probability distribution to establish a model of the contribution ratio of potential pollution sources and obtain the contribution ratio of each potential pollution source.

[0071] As an optional embodiment, in step 2, potential pollution sources within the surrounding area of ​​the target location are marked, and background value areas are determined, including:

[0072] Step 2-1: Using the target location as the center, investigate the potential pollution sources within the preset radius and mark the specific locations and scale information of domestic pollution sources, agricultural production sources, and industrial production sources.

[0073] Step 2-2: Select an area that is more than a preset range away from the target point, where the surrounding pollution source indicators are lower than the set standard (there are no obvious pollution sources in the surrounding area), and where the geological conditions are uniform as the background value area.

[0074] As an optional embodiment, in step 5, for the test data obtained for each type of potential pollution source, namely three-dimensional fluorescence water fingerprint data, eDNA data, and isotopic data of the target pollutant, a unified data matrix is ​​constructed to characterize its features. The rows of the data matrix represent different sample information of the pollution source, each sample corresponds to a unique identifier, and the columns correspond to the feature parameters of the three-dimensional fluorescence water fingerprint data, eDNA data, and isotopic data of the target pollutant, respectively. For numerical features, they are directly used as matrix elements, and for categorical features, one-hot encoding is used to transform the information as feature parameters.

[0075] As an optional embodiment, in step 5, for the data matrix constructed for each type of potential pollution source, principal component analysis is first used to reduce the dimensionality of the data matrix. By calculating the covariance matrix, eigenvalues ​​and eigenvectors of the data, principal components with a cumulative contribution rate of more than a preset threshold are selected and the number of principal components to be retained is determined.

[0076] Then, the Pearson correlation coefficients between different types of preserved feature parameters are calculated, and different weights are assigned to the feature parameters according to the magnitude of the absolute value of the Pearson correlation coefficients.

[0077] Finally, the feature matrix is ​​weighted according to the assigned weights, and the weighted features are linearly combined to obtain the fused feature matrix.

[0078] As an optional embodiment, for the data matrix constructed for each type of potential pollution source, principal component analysis is first used to reduce the dimensionality of the data matrix. By calculating the covariance matrix, eigenvalues, and eigenvectors of the data, principal components with a cumulative contribution rate exceeding a preset threshold are selected and the number of principal components to be retained is determined, including:

[0079] For any type of potential pollution source, the data matrix is ​​defined as data matrix X, with shape n*p, where n and p represent the number of samples and the number of features, respectively.

[0080] For a data matrix X, calculate its covariance matrix and solve for the eigenvalues ​​and corresponding eigenvectors to reflect the correlation between the various features. The covariance can be calculated as follows:

[0081]

[0082] In the formula, This represents the mean vector of each column of matrix X; the eigenvectors can be solved using existing algorithms.

[0083] The calculated eigenvalues ​​and corresponding eigenvectors are arranged in descending order. The purpose of this sorting is to prioritize the direction with the largest variance (i.e., the largest eigenvalue) when selecting principal components later.

[0084] For the sorted feature values, calculate the cumulative contribution rate CR(k) of the first k principal components, and find the minimum value of k such that the cumulative contribution rate of the first k principal components reaches or exceeds a preset threshold, where:

[0085]

[0086] In the formula λ i This represents the i-th eigenvalue;

[0087] Select the first k sorted feature vectors to form a p*k matrix V, and project the original data matrix X onto the k principal components to obtain the dimensionality-reduced data matrix Y: Y = X·V.

[0088] As an optional embodiment, for each type of potential pollution source, the data matrix constructed, for its dimensionality-reduced data matrix Y, calculates each pair of features Y in the matrix. i and Y j The Pearson correlation coefficient r between them ij , i,j=1,2,3,…,k;

[0089] Pearson correlation coefficient r among all feature pairs ij Form a k*k correlation matrix R, where the elements of the matrix R are... ij =r ij ;

[0090] Different weighting intervals are set based on the absolute value of the correlation coefficient;

[0091] Traverse each element R in the correlation matrix R ij Based on the weight range, a weight value w is randomly selected for each element. ij This ultimately forms a k*k weight matrix W;

[0092] The data matrix Y is weighted by the k*k weight matrix W, and the weighted features are linearly combined to obtain the fused feature matrix.

[0093] As an optional embodiment, step 6 involves constructing a source tracing model based on a multi-layer convolutional neural network (CNN) and a long short-term memory network (LSTM), including:

[0094] A source tracing model combining multilayer convolutional neural networks (CNN) and long short-term memory networks (LSTM) is constructed.

[0095] A multilayer convolutional neural network (CNN) consists of an input layer, convolutional layers, and pooling layers.

[0096] The input layer takes into account three-dimensional fluorescence water fingerprint data in the form of feature matrices, eDNA data, and isotopic data of the target pollutants.

[0097] The convolutional layers consist of a first convolutional layer, a second convolutional layer, and a third convolutional layer: The first convolutional layer uses a 3×3 kernel with a stride of 1 and 32 channels. The kernel slides across the input data, extracting preliminary spatial features through convolution operations. Each kernel learns a specific set of feature patterns. The second convolutional layer uses a 5×5 kernel with a stride of 1 and 64 channels. Based on the features extracted by the first convolutional layer, it further captures local features. The third convolutional layer uses a 7×7 kernel with a stride of 1 and 128 channels. This layer aims to obtain deep-level global feature information. After the convolutional layers, two pooling layers are connected: The first pooling layer uses max pooling with a 2×2 window and a stride of 2. By taking the maximum value within the window, the dimensionality of the feature map is reduced while retaining the most important feature information. The second pooling layer uses max pooling with a 3×3 window and a stride of 2, further compressing the data dimensionality.

[0098] Long Short-Term Memory (LSTM) networks consist of two memory units, each containing an input gate, a forget gate, and an output gate.

[0099] The first memory unit has 256 neurons. This unit can process time-series data, controlling the input of new information through an input gate, deciding whether to retain or discard memory information from the previous moment through a forget gate, and determining the output for the current moment through an output gate. It can capture short-term dynamic changes in data over time, such as the changing trends of pollution source characteristics over different time periods.

[0100] The second memory unit has 512 neurons. Compared to the first memory unit, it has stronger memory and feature extraction capabilities, and can process more complex time-series information, capturing long-term dynamic change patterns, such as seasonal or annual pollution source changes.

[0101] As an optional embodiment, in step 8, based on the characteristics of multi-source data of the target location and the prediction results of the potential pollution source tracing model of the target location, a probability model based on Bayes' theorem is established to obtain the conditional probability distribution under the prior probability of each pollution source, including:

[0102] Given that there are n potential pollution sources at a target location, determine the prior probabilities of different types of pollution sources;

[0103] Combining the prediction results of the potential pollution source tracing model at the target location, i.e., outputting a multi-dimensional vector, where each vector element represents the probability that the data belongs to each potential pollution source, the vector elements are normalized, and the softmax function is used to convert the vector into a probability distribution, obtaining the conditional probability distribution under the prior probability of each pollution source:

[0104]

[0105] Among them, z i It is the i-th element of the model's output vector.

[0106] As an optional embodiment, in step 9, a potential pollution source contribution ratio model is established by combining Monte Carlo simulation and conditional probability distribution to obtain the contribution ratio of each potential pollution source, including:

[0107] Random samples are generated through Monte Carlo simulation: the number of simulations is set to N. For each simulation, a sample is randomly selected from the conditional probability distribution, and the probability of the sample belonging to each pollution source is calculated. The above simulation process is repeated N times, and the number of times each pollution source is selected is counted, thereby calculating the probability distribution of each pollution source.

[0108] The contribution ratio of each pollution source is calculated based on its probability distribution. For the i-th pollution source, the contribution ratio is calculated using the following formula: Where P i avg It is the average probability of the i-th pollution source, which is obtained by averaging the probability distribution of the pollution source obtained from Monte Carlo simulation; It is the sum of the average probabilities of all pollution sources.

[0109] Therefore, the groundwater pollution source tracing method of this invention, by integrating multi-source data and no longer relying on a single type of data, greatly enriches the information sources for source tracing, reduces the possibility of misjudgment and omission, and thus significantly improves the accuracy of source tracing. Simultaneously, the use of artificial intelligence deep learning technology to achieve automatic source tracing greatly improves work efficiency and saves manpower and time costs. Furthermore, this invention can further accurately calculate the contribution ratio of different pollution sources, providing a scientific basis for formulating targeted governance strategies.

[0110] {Example 2}

[0111] In this embodiment, we will further elaborate on the specific implementation process of the aforementioned Embodiment 1.

[0112] The groundwater pollution tracing method based on deep learning (CNN and LSTM networks) and multi-source data fusion, as described in Example 1 above, includes the following steps:

[0113] 1. Determine the source pollutant indicators. Based on the water quality monitoring results of the groundwater monitoring points at the target locations (i.e., the monitoring points), obtain detailed water quality data reports.

[0114] Furthermore, based on national standards and industry-related groundwater quality standards, as well as the environmental background values ​​of the target location area, pollutants exceeding the standard range are screened out, and pollutant indicators that need to be traced are identified, such as sulfate, ammonia nitrogen, nitrate, and organic matter.

[0115] 2. Investigate potential pollution sources within the surrounding area of ​​the target location. Using the target location as the center, conduct a comprehensive and detailed investigation of potential pollution sources within a 1km radius, marking the specific locations and scale of: domestic pollution sources (such as septic tanks and garbage disposal stations in residential areas), agricultural production sources (such as large-scale livestock farms and farmland fertilization areas), and industrial production sources (such as factory discharge outlets and raw material storage areas).

[0116] Meanwhile, areas that are relatively far from the target location (e.g., about 1-5km away depending on the actual situation), have no obvious pollution sources in the surrounding area, and have relatively uniform geological conditions are selected as background value areas.

[0117] 3. Continuous monitoring and sampling of potential pollution sources. For the potential pollution sources and background value areas identified in step 2, continuous sampling and monitoring will be conducted for 6 days, with monitoring once per day. Representative samples will be taken from each potential pollution source to reflect the changing characteristics of different pollution sources over time. In this example, the representative samples include discharged domestic sewage, industrial wastewater, groundwater from concentrated agricultural production areas, and groundwater from the background value area.

[0118] 4. Testing of potential pollution source samples. For samples collected from each potential pollution source, eDNA, three-dimensional fluorescence spectroscopy, and target pollutant isotope testing and analysis were performed to obtain eDNA data, three-dimensional fluorescence water fingerprint data, and target pollutant isotope data of the potential pollution source samples.

[0119] (1) eDNA testing. Collected DNA samples were rapidly transferred to cryopreservation equipment to ensure sample stability during transportation and storage. In the molecular biology laboratory, high-quality DNA extraction kits were used for DNA extraction, strictly following the kit instructions to ensure the purity and integrity of the extracted DNA. Agarose gel electrophoresis and a nucleic acid concentration analyzer were used to perform quality testing on the extracted DNA; only samples meeting the quality standards were allowed to proceed to subsequent experimental steps.

[0120] Based on the characteristics of the target microorganism, specific primers were designed for PCR amplification, and PCR reaction conditions, including reaction temperature, time, and number of cycles, were optimized to ensure efficient amplification of the target gene fragment. The amplified products were then subjected to high-throughput sequencing using a reliable sequencing platform, such as Illumina. After sequencing, bioinformatics software, such as Mothur or QIIME, was used to process and analyze the sequencing data. The data was compared with known microbial databases (such as NCBI and UNITE) to accurately identify the microbial species, and the sequence information and relative abundance of each microorganism were recorded in detail.

[0121] (2) Three-dimensional fluorescence spectroscopy. The collected water samples were pretreated by filtering through a 0.45 μm filter membrane to remove suspended particulate matter. The filtered water sample was then injected into the sample cell of the fluorescence spectrometer. Appropriate instrument parameters were set, such as the excitation wavelength range (typically 200-800 nm), emission wavelength range (adjusted according to the target fluorescent substance), and scan speed. The Savitzky-Golay filtering algorithm was used to smooth the original three-dimensional fluorescence spectral data. Rayleigh and Raman scattering were corrected by repeatedly measuring blank water samples and subtracting the subtracted values. The Parallel Factor Analysis (PARAFAC) algorithm was used to decompose the corrected spectrum and determine the independent fluorescent components.

[0122] When calculating characteristic parameters such as the maximum value, integral value, and peak position of fluorescence intensity, professional data analysis software, such as Origin, can be used to ensure the accuracy of the calculation. These characteristic parameters can be standardized to a value between [0,1]. One optional standardization method is to subtract the minimum value of each parameter from its value and then divide by the difference between the maximum and minimum values.

[0123] (3) Target pollutant isotope testing. The collected water samples underwent chemical pretreatment, and the target pollutants were separated and enriched using methods such as ion exchange and extraction. A high-precision isotope mass spectrometer was then used for measurement. The mean and standard deviation of the measured isotope ratio data were calculated using statistical analysis software (such as SPSS). Data exceeding the mean plus or minus three times the standard deviation were considered outliers and removed.

[0124] The isotope ratio data were standardized using the Z-score standardization method, and the calculation formula is as follows:

[0125] Z = (x - μ) / σ;

[0126] Where x is the original data value, μ is the mean, and σ is the standard deviation, so that the standardized data has a mean of 0 and a standard deviation of 1.

[0127] 5. Multi-source data fusion, including data matrix construction, dimensionality reduction and weight calculation, and the final data representation after linear combination.

[0128] (1) Data Matrix Construction. For each potential pollution source, a unified data matrix is ​​constructed to characterize its features. Rows represent different samples from the pollution source, with each sample corresponding to a unique identifier containing detailed information such as sampling time and location. Columns correspond to the feature parameters of three-dimensional fluorescence water fingerprint data, eDNA data, and isotope data, respectively.

[0129] For numerical features, such as the maximum fluorescence intensity, relative abundance of microorganisms, and standardized isotope ratios, they are directly incorporated into the matrix. For categorical features (such as the type of microorganism), one-hot encoding is used for conversion. That is, for each type of microorganism, a new column is created in the matrix. If the sampling point contains that type of microorganism, the value of the corresponding column is 1; otherwise, it is 0.

[0130] (2) Dimensionality Reduction and Weight Calculation. Principal Component Analysis (PCA) is used to reduce the dimensionality of the fused data matrix. By calculating the covariance matrix, eigenvalues, and eigenvectors of the data, principal components that can explain most of the data variation are selected, and the number of principal components is determined. In this embodiment, principal components with a cumulative contribution rate of over 85% are retained to achieve data dimensionality reduction.

[0131] Then, calculate the Pearson correlation coefficient between different types of preserved data features. This can be done using Python's Pandas library or the correlation function in R. Based on the Pearson correlation coefficient, a correlation matrix can be formed.

[0132] Furthermore, different weights are assigned based on the absolute value of the Pearson correlation coefficient.

[0133] In this embodiment, the weight interval is divided as follows:

[0134] Features with an absolute correlation coefficient greater than 0.7 are given higher weights (e.g., 0.8-1.0), features with an absolute correlation coefficient between 0.3 and 0.7 are given medium weights (e.g., 0.4-0.7), and features with an absolute correlation coefficient less than 0.3 are given lower weights (e.g., 0.1-0.3).

[0135] Accordingly, each element in the correlation matrix is ​​traversed, and a suitable weight value is randomly selected for each element based on the pre-defined weight range, thus forming a weight matrix.

[0136] The weighted matrix is ​​obtained by weighting the matrix formed by the weighted matrix and the matrix formed by the retained features.

[0137] (3) Final data representation. The weighted features are linearly combined to obtain the final fused data representation.

[0138] The specific calculation formula is as follows:

[0139] Where FinalData represents the final data, w i The weight of the i-th feature. i Let be the value of the i-th feature, and n be the total number of features.

[0140] 6. Source tracing model construction.

[0141] (1) CNN+LSTM Model Structure. A source tracing model based on deep neural networks is constructed, using an architecture that combines multi-layer convolutional neural networks (CNN) and long short-term memory networks (LSTM).

[0142] The CNN consists of three parts: input layer, convolutional layer, and pooling layer.

[0143] 1) Input layer

[0144] Three-dimensional fluorescence spectral data (e.g., in the form of a feature matrix).

[0145] Feature matrix encoded by eDNA sequence (such as One-Hot encoding).

[0146] Isotope data (e.g., δ) 13 C、δ 15 (Standardized value of N).

[0147] 2) Convolutional layer design

[0148] The first convolutional layer uses a 3×3 kernel with a stride of 1 and 32 channels. The kernel slides across the input data, extracting preliminary spatial features through convolution operations. Each kernel learns a specific set of feature patterns. For example, it might learn local features related to the distribution of certain fluorescent components or short fragments of the eDNA sequence.

[0149] The second convolutional layer uses a 5×5 kernel with a stride of 1 and 64 channels. Building upon the features extracted by the first convolutional layer, this layer further captures more complex and globally relevant local features. For example, it may identify combinations of different microbial species with specific fluorescence characteristics.

[0150] The third convolutional layer uses a 7×7 convolutional kernel with a stride of 1 and 128 channels. This layer aims to obtain deep-level global feature information and can uncover potential relationships between isotopic features and other data features, which are of great significance for the identification of pollution sources.

[0151] 3) Pooling layer design

[0152] Pooling Layers: Following the convolutional layers, two pooling layers are connected. The first pooling layer uses max pooling with a 2×2 window and a stride of 2. By taking the maximum value within the window, the dimensionality of the feature map is reduced, decreasing computational cost while retaining the most important feature information. The second pooling layer also uses max pooling with a 3×3 window and a stride of 2, further compressing the data dimensionality and allowing the model to focus more on key features.

[0153] The LSTM part consists of two memory units, each containing an input gate, a forget gate, and an output gate.

[0154] The design is as follows:

[0155] The first memory unit has 256 neurons. This unit can process time-series data, controlling the input of new information through an input gate, deciding whether to retain or discard memory information from the previous moment through a forget gate, and determining the output for the current moment through an output gate. It can capture short-term dynamic changes in data over time, such as the changing trends of pollution source characteristics over different time periods.

[0156] The second memory unit has 512 neurons. Compared to the first memory unit, it has stronger memory and feature extraction capabilities, and can process more complex time-series information, capturing long-term dynamic change patterns, such as seasonal or annual pollution source changes.

[0157] (2) Model Training

[0158] During the model training phase, the fused multi-source data is first randomly divided into training, validation, and test sets at a ratio of 80%, 10%, and 10%, respectively. In this embodiment, to ensure the randomness and balance of the data partitioning, a random number generator is used to shuffle the data.

[0159] Training data is input into the model in batches, with each batch size set to 64.

[0160] The model parameters were optimized using the stochastic gradient descent (SGD) algorithm, with an initial learning rate of 0.005, a momentum parameter of 0.9, and a weight decay coefficient of 0.0005 to prevent overfitting.

[0161] In each training iteration, a batch of data is input into the model, and the model performs forward propagation based on the current parameters to calculate the predicted value.

[0162] The mean squared error (MSE) is used as the loss function to calculate the error between the predicted and actual values. The specific calculation formula is as follows:

[0163] Where n is the number of samples, y i p is the true value i These are predicted values.

[0164] Then, the gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm. Based on the gradient information, the weights and bias parameters of the model are updated using the SGD algorithm.

[0165] After each training epoch (training the entire training dataset through the model once), the model's performance is evaluated on the validation set. The loss value on the validation set and other evaluation metrics, such as accuracy and recall, are calculated. Training is stopped if the loss on the validation set does not decrease over 10 consecutive epochs to avoid overfitting.

[0166] (3) Calculate the contribution ratio of different pollution sources

[0167] Probabilistic model establishment: Based on the characteristics of multi-source data and the prediction results obtained from model training, a probabilistic model based on Bayes' theorem is established.

[0168] As an optional embodiment, it is assumed that there are n potential pollution sources, and the prior probability P(C) of each pollution source is... i Preliminary assumptions are made based on historical data or expert experience. For example, for a region known to have multiple industrial pollution sources and a small number of agricultural pollution sources, prior probabilities are set for different types of pollution sources based on past pollution records and understanding of industrial activities in the region. For new observed data Q, the conditional probability P(Q|C) under each pollution source assumption is calculated using the model's predicted output. i ).

[0169] Specifically, suppose the model outputs an n-dimensional vector for a given observation, where each element represents the probability that the data belongs to each pollution source. This vector is normalized, and the softmax function is used to convert it into a probability distribution, yielding the conditional probability distribution P(Q|C) under each pollution source hypothesis. i ).

[0170] As an example, the formula for calculating the aforementioned softmax function is:

[0171]

[0172] Among them, z i It is the i-th element of the model's output vector.

[0173] Monte Carlo simulation: Generates a large number of random samples through Monte Carlo simulation.

[0174] As an optional embodiment, the number of simulations is set to N, for example, N = 10000. For each simulation, the conditional probability distribution P(Q|C) is used. i A sample is randomly selected from the sample, and the probability that the sample belongs to each pollution source is calculated. The simulation and calculation process is repeated 10,000 times to obtain the probability distribution of each pollution source.

[0175] For example, after 10,000 simulations, the number of times each pollution source was selected is counted, and then the probability distribution of each pollution source is calculated.

[0176] Contribution ratio calculation: The contribution ratio of each pollution source is calculated based on the probability distribution of each pollution source.

[0177] Specifically, for the i-th pollution source, its contribution ratio is calculated using the following formula:

[0178]

[0179] Where P i avg It is the average probability of the i-th pollution source, which is obtained by averaging the probabilities of the pollution source obtained from Monte Carlo simulation; It is the sum of the average probabilities of all pollution sources.

[0180] Based on the groundwater pollution tracing method designed above, in application, groundwater samples are collected at the target location according to the same specifications and methods as those used for potential pollution sources. eDNA, three-dimensional fluorescence spectroscopy, and target pollutant isotope testing and analysis are performed to obtain eDNA data, three-dimensional fluorescence water quality fingerprint data, and target pollutant isotope data of the groundwater samples at the target location. The newly sampled data is preprocessed and feature extracted using a multi-source data fusion method to obtain data consistent with the model input format. The processed data is input into the tracing model trained in step 6. The model makes predictions based on the learned features and patterns, outputting a multi-dimensional vector, where each vector element represents the probability that the data belongs to each potential pollution source. Furthermore, a probability model based on Bayes' theorem calculates and outputs the probability distribution of each potential pollution source.

[0181] Therefore, the main pollution sources can be determined based on the probability distribution. For example, a probability threshold (such as 0.1) can be set, and pollution sources with a probability greater than this threshold are considered to be the main pollution sources.

[0182] Meanwhile, based on the calculation method of the potential pollution source contribution ratio model, the contribution ratio of the main pollution sources can be further calculated and output, thus realizing the source tracing of the main pollution sources.

[0183] {Example 3}

[0184] In this embodiment, we will use a specific example to trace the source of pollution in the groundwater surrounding a chemical industrial park according to the method described in the foregoing embodiments of the present invention.

[0185] 1. Determine the pollutant indicators for source tracing.

[0186] A comprehensive water quality test was conducted on the groundwater at the target location. Based on the national "Groundwater Quality Standard" (GB / T14848-2017) and the environmental background values ​​of the area, the test results showed that the nitrate content was 50 mg / L (the standard limit is 20 mg / L), the ammonia nitrogen content was 3 mg / L (the standard limit is 0.5 mg / L), and the trichloroethylene content in organic matter was 50 μg / L (the background value in this area is usually less than 10 μg / L). These three pollutants were identified as indicators that require source tracing.

[0187] 2. Investigate and identify potential sources of pollution.

[0188] A survey was conducted within a 1km radius of the target location. Using GIS technology, a pollution source distribution map was created, revealing one chemical plant (industrial production source) primarily producing organic chemical products with multiple wastewater discharge outlets; two residential areas (domestic pollution sources) equipped with septic tanks and small-scale sewage treatment facilities; and a farmland (agricultural production source) where nitrogen-containing fertilizers and organic pesticides had recently been used.

[0189] Meanwhile, a region 1.5km away from the target location, with no obvious pollution sources and uniform geological conditions, was selected as the background value area. The geological type of this area was recorded as silty clay, and the groundwater level was 2.5m deep.

[0190] 3. Continuous monitoring and sampling of potential pollution sources

[0191] Develop detailed sampling plans for identified potential pollution sources and background value areas.

[0192] Chemical companies: Set up sampling points at their main wastewater discharge outlets, use corrosion-resistant water sampling devices, and collect wastewater samples every morning at 9:00 AM.

[0193] Residential areas: Sampling was conducted at the outlet of domestic sewage treatment facilities in residential areas using a sewage sampler, and domestic sewage samples were collected at fixed times every day.

[0194] Farmland: Water samples are collected from groundwater monitoring wells in crop planting areas once a day.

[0195] Background value area: Set up a groundwater sampling well in the background value area and collect groundwater samples once a day.

[0196] During the six-day monitoring period, samples were taken on time each day, and the sampling time, location, and other information were recorded in detail.

[0197] 4. Testing of samples from potential pollution sources

[0198] (1) eDNA testing

[0199] DNA Extraction and Quality Control: Collected samples were rapidly placed in a -20°C cryopreservation device and transported to the laboratory. High-quality DNA extraction kits were used for DNA extraction. After extraction, DNA integrity was assessed by agarose gel electrophoresis, and concentration was measured using a nucleic acid concentration analyzer to ensure DNA purity (A260 / A280) was between 1.8 and 2.0, and concentration was not less than 50 ng / μL.

[0200] PCR amplification and sequencing: Based on the known microorganisms that may exist in the chemical industrial park, specific primers were designed to amplify the target gene fragment by PCR. The optimized PCR reaction conditions were: 95℃ pre-denaturation for 5 min; 95℃ denaturation for 30 s, 55℃ annealing for 30 s, 72℃ extension for 30 s, for a total of 35 cycles; and a final extension at 72℃ for 10 min. After the amplified products passed the tests, high-throughput sequencing was performed using the Illumina sequencing platform.

[0201] Data Analysis: Sequencing data were processed using Mothur software and compared with the NCBI microbial database to determine the types and relative abundance of microorganisms. Pseudomonas species were detected at a relative abundance of 30% and Bacillus species at 20% in chemical plant wastewater; Escherichia coli was detected at a relative abundance of 40% in residential sewage; and nitrogen-fixing bacteria were detected at a relative abundance of 35% in agricultural groundwater. The background area had relatively few microbial species in its groundwater, dominated by common soil microorganisms, such as Streptomyces species at a relative abundance of 25%.

[0202] like Figure 2 As shown, the results of eDNA monitoring of potential contaminants are illustrated.

[0203] (2) Three-dimensional fluorescence spectroscopy test

[0204] Water sample pretreatment: The collected water sample was filtered through a 0.45μm filter membrane and then injected into the sample cell of the fluorescence spectrometer.

[0205] Spectral Measurement and Processing: The excitation wavelength range was set to 200-800 nm, and the emission wavelength range was adjusted to 250-600 nm according to the target fluorescent substance. The scan rate was 12000 nm / min. The Savitzky-Golay filtering algorithm was used to smooth the raw spectral data. Rayleigh scattering and Raman scattering were corrected by subtracting from multiple measurements of blank water samples. The PARAFAC algorithm was used to decompose the spectrum and determine the independent fluorescent components. Three main fluorescent components were decomposed from wastewater from chemical enterprises. The maximum fluorescence intensities were calculated to be 1000, 1500, and 1200, respectively, and the integral values ​​were 8000, 12000, and 10000, respectively. The peak positions were 300 nm, 350 nm, and 400 nm, respectively. The maximum fluorescence intensity of the main fluorescent component in domestic sewage from residential areas was 800, the integral value was 6000, and the peak position was 320 nm. The maximum fluorescence intensity of the fluorescent component in groundwater from farmland was 600, the integral value was 5000, and the peak position was 330 nm. The fluorescence intensity in groundwater in the background value area was relatively weak, with a maximum value of 200, an integral value of 1500, and a peak position of 310 nm. After standardizing these characteristic parameters, the standardized values ​​of the three fluorescent components in chemical enterprise wastewater are [0.8, 0.9, 0.85], those in residential sewage are [0.6, 0.7, 0.65], those in farmland groundwater are [0.5, 0.6, 0.55], and those in the background area groundwater are [0.2, 0.3, 0.25].

[0206] like Figure 3 As shown, the results of three-dimensional fluorescence spectral monitoring of potential pollution sources are illustrated.

[0207] (3) Target pollutant isotope testing

[0208] Water sample pretreatment and measurement: The collected water samples were separated and enriched for nitrates, ammonia nitrogen, and trichloroethylene using ion exchange and extraction methods. Measurements were then performed using a high-precision isotope mass spectrometer, which was calibrated before each measurement.

[0209] Data processing: The nitrogen isotope ratio (δ) in nitrate was measured. 15 N-NO3 - The data, with a mean of +5‰ and a standard deviation of 0.5‰, were standardized using the Z-score method after removing data exceeding three times the mean standard deviation. The δ¹⁸O value of nitrates in chemical wastewater was also analyzed. 15 N-NO3 - The original value was +6‰, and the standardized value was +2; the δ of nitrate in domestic sewage in residential areas 15 N-NO3 - The original value was +4‰, and the standardized value was -2; the δ of nitrate in farmland groundwater 15 N-NO3- The original value was +4.5‰, and the standardized value was -1; the background value was the δ of nitrate in the groundwater of the region. 15 N-NO3 - The original value was +3‰, which was standardized to -4. Similar processing was performed on the isotopic data for ammonia nitrogen and trichloroethylene.

[0210] like Figure 4 As shown, this example illustrates the results of isotope monitoring.

[0211] 5. Multi-source data fusion

[0212] Data matrix construction: A unified data matrix is ​​established, with rows representing different sampling points and columns corresponding to different data feature parameters. For example, for wastewater sampling points in chemical enterprises, numerical features such as the relative abundance of microorganisms in eDNA data, the feature parameters after standardization of three-dimensional fluorescence spectroscopy data, and the ratios after standardization of isotope data are directly incorporated into the matrix; for the categorical feature of microbial species, one-thermal coding conversion is used, such as Pseudomonas spp. being coded as [1,0,0,…] and Bacillus spp. being coded as [0,1,0,…].

[0213] Dimensionality Reduction and Weight Calculation: PCA was used to reduce the dimensionality of the fused data matrix. The cumulative contribution rate of the first three principal components reached 90%, and these three principal components were selected for retention. Pearson correlation coefficients were calculated between different types of data features. It was found that the absolute value of the correlation coefficient between a certain fluorescent component in chemical wastewater and a specific microbial species was 0.8, and it was assigned a high weight of 0.9. The absolute value of the correlation coefficient between the nitrate isotope ratio and the fluorescence intensity integral value in residential sewage was 0.6, and it was assigned a medium weight of 0.6.

[0214] The final data is represented as follows: based on the weighted features, a linear combination is performed. For example, for wastewater sampling points of chemical enterprises, the final data is represented as: FinalData=0.9×Feature1+0.6×Feature2+….

[0215] 6. Source tracing model construction

[0216] (1) Model Structure

[0217] A deep learning-based source tracing model is constructed, employing an architecture that combines CNN and LSTM.

[0218] CNN section:

[0219] Input the processed three-dimensional fluorescence spectrum, eDNA, and isotope characteristic data.

[0220] The first convolutional layer uses a 3×3 convolutional kernel with a stride of 1 and 32 channels to perform convolution operations on the input fused data and extract preliminary spatial features.

[0221] The second convolutional layer uses a 5×5 convolutional kernel with a stride of 1 and 64 channels to further capture complex local features.

[0222] The third convolutional layer uses a 7×7 convolutional kernel with a stride of 1 and 128 channels to obtain deep-level feature information.

[0223] Pooling layers: The first pooling layer uses 2×2 max pooling with a stride of 2; the second pooling layer uses 3×3 max pooling with a stride of 2 to reduce the dimension of the feature map.

[0224] The LSTM part: The first memory unit has 256 neurons and the second memory unit has 512 neurons, which process time series data.

[0225] (2) Model Training

[0226] The data after fusion of multiple sources is divided into training set, validation set and test set according to the proportions of 80%, 10% and 10% respectively.

[0227] Training data was input into the model in batches of 64 samples. The SGD algorithm was used to optimize the parameters, with an initial learning rate of 0.005, a momentum parameter of 0.9, and a weight decay coefficient of 0.0005.

[0228] The MSE (Mean Sequence Equation) is used as the loss function. In each training iteration, the error between the predicted and actual values ​​is calculated, and the model parameters are updated via backpropagation. Performance is evaluated on the validation set after each epoch, and changes in metrics during training are recorded. Training is stopped when the validation set loss fails to decrease for 10 consecutive epochs after 50 epochs of training.

[0229] 7. Pollution source identification and contribution ratio calculation

[0230] Groundwater samples were collected at the target locations according to the same specifications. eDNA, three-dimensional fluorescence spectroscopy, and target pollutant isotope testing and analysis were performed. The relevant data were obtained, preprocessed, and feature extracted before being input into the trained source tracing model.

[0231] Probabilistic model establishment: Assuming there are three potential pollution sources: chemical enterprises, residential areas, and farmland, based on historical data and expert experience, the prior probability of chemical enterprises is set to 0.6, the prior probability of residential areas is set to 0.2, and the prior probability of farmland is set to 0.2.

[0232] For newly observed input data, the model output is a three-dimensional vector [0.7, 0.2, 0.1], which, after being normalized by the softmax function, yields a conditional probability distribution of [0.72, 0.22, 0.06].

[0233] Monte Carlo simulation: The simulation is set to run 10,000 times. Each time, a sample is randomly drawn from the conditional probability distribution, and the probability of belonging to each pollution source is calculated. The simulation is repeated to obtain the probability distribution of each pollution source.

[0234] Contribution ratio calculation: The contribution ratio of each pollution source is calculated based on the probability distribution. The model output results show that the probability of chemical enterprises as the main pollution source is 0.75, and the contribution ratio is P1 = 70%; the probability of residential areas as pollution sources is 0.15, and the contribution ratio is P2 = 15%; the probability of farmland as pollution sources is 0.1, and the contribution ratio is P3 = 15%.

[0235] {Example 4}

[0236] Based on the groundwater pollution tracing methods based on deep learning and multi-source data fusion described in the above examples, this disclosure also proposes a computer system, including:

[0237] One or more processors;

[0238] The memory stores operable instructions that, when executed by one or more processors, cause one or more processors to perform operations, including the steps of the aforementioned groundwater pollution tracing method based on deep learning and multi-source data fusion.

[0239] While the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Those skilled in the art can make various modifications and refinements without departing from the spirit and scope of the invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A groundwater pollution source tracing method based on deep learning and multi-source data fusion, characterized in that, Includes the following steps: Step 1: Based on historical groundwater monitoring data at the target location, determine the pollutant indicators that need to be traced. Step 2: Investigate and mark potential pollution sources in the vicinity of the target location, and determine the background value area; Step 3: Perform continuous periodic sampling on potential pollution sources and background value areas to obtain potential pollution source samples; Step 4: Test the potential pollution source samples taken continuously in cycles, including three-dimensional fluorescence spectroscopy, eDNA and target pollutant isotope test analysis, and obtain the three-dimensional fluorescence water quality fingerprint data, eDNA data and target pollutant isotope data of the potential pollution source samples respectively. Step 5: Based on the three-dimensional fluorescence water quality fingerprint data, eDNA data and isotopic data of the target pollutant from the potential pollution source sample, perform multi-source data fusion to obtain the fusion features of linear combination and construct the feature set of potential pollution sources at the target location. Step 6: Construct a source tracing model based on a multi-layer convolutional neural network and a long short-term memory network, and train the model based on the feature set of the potential pollution sources at the target location to obtain the source tracing model of the potential pollution sources at the target location. Step 7: For newly sampled groundwater samples at the target location, after feature extraction and analysis through three-dimensional fluorescence spectroscopy, eDNA and target pollutant isotope testing, the samples are input into the potential pollution source tracing model of the target location, and a multi-dimensional vector is output. Each vector element represents the probability that the data belongs to each potential pollution source. Step 8: Based on the characteristics of multi-source data of the target location and the prediction results of the potential pollution source tracing model of the target location, establish a probability model based on Bayes' theorem to obtain the conditional probability distribution under the prior probability of each pollution source. Step 9: Combine Monte Carlo simulation and the conditional probability distribution to establish a potential pollution source contribution ratio model and obtain the contribution ratio of each potential pollution source.

2. The groundwater pollution tracing method based on deep learning and multi-source data fusion according to claim 1, characterized in that, In step 2, potential pollution sources within the surrounding area of ​​the target location are marked, and background value areas are determined, including: Step 2-1: Using the target location as the center, investigate the potential pollution sources within the preset radius and mark the specific locations and scale information of domestic pollution sources, agricultural production sources, and industrial production sources. Step 2-2: Select an area that is more than a preset range away from the target point, where the surrounding pollution source indicators are lower than the set standards and the geological conditions are uniform as the background value area.

3. The groundwater pollution source tracing method based on deep learning and multi-source data fusion according to claim 1, characterized in that, In step 5, for the test data obtained for each type of potential pollution source, namely three-dimensional fluorescence water fingerprint data, eDNA data, and isotopic data of the target pollutant, a unified data matrix is ​​constructed to characterize its features. The rows of the data matrix represent different sample information of the pollution source, each sample corresponds to a unique identifier, and the columns correspond to the feature parameters of the three-dimensional fluorescence water fingerprint data, eDNA data, and isotopic data of the target pollutant, respectively. For numerical features, they are directly used as matrix elements, and for categorical features, one-heat encoding conversion information is used as feature parameters.

4. The groundwater pollution source tracing method based on deep learning and multi-source data fusion according to claim 3, characterized in that, In step 5, for the data matrix constructed for each type of potential pollution source, principal component analysis is first used to reduce the dimensionality of the data matrix. By calculating the covariance matrix, eigenvalues ​​and eigenvectors of the data, principal components with a cumulative contribution rate of more than a preset threshold are selected and the number of principal components to be retained is determined. Then, the Pearson correlation coefficients between different types of preserved feature parameters are calculated, and different weights are assigned to the feature parameters according to the magnitude of the absolute value of the Pearson correlation coefficients. Finally, the feature matrix is ​​weighted according to the assigned weights, and the weighted features are linearly combined to obtain the fused feature matrix.

5. The groundwater pollution tracing method based on deep learning and multi-source data fusion according to claim 4, characterized in that, For each type of potential pollution source, the data matrix is ​​first reduced in dimensionality using principal component analysis. This is done by calculating the covariance matrix, eigenvalues, and eigenvectors of the data. Principal components with a cumulative contribution rate exceeding a preset threshold are selected and the number of retained principal components is determined, including: For any type of potential pollution source, the constructed data matrix is ​​defined as a data matrix. X , shape is n × p , n and p These represent the number of samples and the number of features, respectively. For data matrix X Calculate its covariance matrix and solve for the eigenvalues ​​and corresponding eigenvectors; The calculated eigenvalues ​​and corresponding eigenvectors are arranged in descending order; For sorted feature values, before calculation k The cumulative contribution rate of each principal component is calculated, and the minimum contribution rate is found. k Values ​​that satisfy: the previous k The cumulative contribution rate of each principal component reaches or exceeds a preset threshold; and Before selection k The sorted feature vectors form a p × k matrix V and the original data matrix X Projected to k Based on the principal components, a dimensionality-reduced data matrix is ​​obtained. Y : Y = X·V .

6. The groundwater pollution source tracing method based on deep learning and multi-source data fusion according to claim 5, characterized in that, For each type of potential pollution source, a data matrix is ​​constructed, and its dimensionality is reduced accordingly. Y Calculate each pair of features in the matrix Y i and Y j Pearson correlation coefficient between r ij i,j=1,2,3,… k ; Pearson correlation coefficient between all feature pairs r ij Form a k × k correlation matrix R Elements in the matrix R ij = r ij ; Different weighting intervals are set based on the absolute value of the correlation coefficient; Traversing the correlation matrix R Each element in R ij Based on the weight range, a weight value is randomly selected for each element. w ij Ultimately forming a k × k weight matrix W ; according to k × k weight matrix W For data matrix Y The features are weighted and then linearly combined to obtain the fused feature matrix.

7. The groundwater pollution source tracing method based on deep learning and multi-source data fusion according to claim 3, characterized in that, In step 6, a source tracing model based on a multi-layer convolutional neural network and a long short-term memory network is constructed, including: A source tracing model combining a multi-layer convolutional neural network and a long short-term memory network is constructed, and the specific design is as follows: The multi-layer convolutional neural network consists of an input layer, convolutional layers, and pooling layers: The input layer takes into account three-dimensional fluorescence water quality fingerprint data in the form of feature matrices, eDNA data, and isotopic data of the target pollutants. The convolutional layers consist of a first convolutional layer, a second convolutional layer, and a third convolutional layer: The first convolutional layer uses a 3×3 kernel with a stride of 1 and 32 channels. The kernel slides across the input data, extracting preliminary spatial features through convolution operations. Each kernel learns a specific set of feature patterns. The second convolutional layer uses a 5×5 kernel with a stride of 1 and 64 channels. Based on the features extracted by the first convolutional layer, it further captures local features. The third convolutional layer uses a 7×7 kernel with a stride of 1 and 128 channels. This layer aims to obtain deep-level global feature information. After the convolutional layers, two pooling layers are connected: The first pooling layer uses max pooling with a 2×2 window and a stride of 2. By taking the maximum value within the window, the dimensionality of the feature map is reduced while retaining the most important feature information. The second pooling layer uses max pooling with a 3×3 window and a stride of 2, further compressing the data dimensionality. The Long Short-Term Memory (LSTM) network consists of two memory units, each containing an input gate, a forget gate, and an output gate: The first memory unit has 256 neurons and is used to process time-series data. It controls the input of new information through the input gate, the forget gate determines whether to retain or discard the memory information of the previous moment, and the output gate determines the output of the current moment in order to capture the short-term dynamic changes of the data in the time dimension. The second memory unit has 512 neurons and is used to capture long-term dynamic change patterns.

8. The groundwater pollution source tracing method based on deep learning and multi-source data fusion according to claim 1, characterized in that, In step 9, the step of establishing a potential pollution source contribution ratio model by combining Monte Carlo simulation and the conditional probability distribution to obtain the contribution ratio of each potential pollution source includes: Random samples are generated using Monte Carlo simulation: the number of simulations is set to [number]. N For each simulation, a sample is randomly selected from the conditional probability distribution, and the probability that the sample belongs to each pollution source is calculated based on the selected sample; the above simulation process is repeated. N The number of times each pollution source was selected was counted, and the probability distribution of each pollution source was calculated. The contribution ratio of each pollution source is calculated based on the probability distribution of each pollution source, where for the th i The formula for calculating the contribution ratio of each pollution source is as follows: ,in P i avg It is the first i The average probability of each pollution source is calculated by averaging the probability distribution of the pollution source obtained from Monte Carlo simulation. It is the sum of the average probabilities of all pollution sources.

9. A computer system, characterized in that, include: One or more processors; A memory that stores operable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, including the steps of the method as described in any one of claims 1-8.

Images