A deep learning-based groundwater pollution source tracing and recovery period evaluation method

Abstract

The application relates to the field of deep learning, in particular to a groundwater pollution source tracing and recovery cycle evaluation method based on deep learning, which comprises the following steps: obtaining pollution concentration monitoring values of multiple pollution leakage points in an industrial park, and inputting the pollution concentration monitoring values into an encoder to generate an encoded pollution source intensity field vector, wherein the encoder at least comprises a convolution layer and an LSTM layer; obtaining multiple permeability coefficients on a groundwater runoff discharge path in the industrial park, and calculating a permeability coefficient field vector based on the multiple permeability coefficients; inputting the encoded pollution source intensity field vector and the permeability coefficient field vector into a physical information convolution decoder to generate a groundwater pollution concentration distribution map, and performing groundwater pollution source tracing and recovery cycle evaluation based on the groundwater pollution concentration distribution map. The application realizes coupling modeling through a physical information convolution decoder, and provides accurate basic data for groundwater pollution source tracing and recovery cycle evaluation.

Description

Technical Field

[0001] This invention relates to the field of deep learning, and in particular to a deep learning-based method for tracing the source of groundwater pollution and assessing its recovery cycle. Background Technology

[0002] With the acceleration of industrialization, industrial parks, as concentrated areas of industrial production, are facing increasingly severe groundwater pollution problems. This seriously threatens ecological and environmental security and has become a bottleneck restricting the green, low-carbon transformation and high-quality development of industrial parks. Frequent production activities within industrial parks result in various types of wastewater, hazardous waste leachate, and underground pipeline leakage, which enter the underground vadose zone through discharge outlets. Under the influence of gravity, these pollutants vertically penetrate the vadose zone and seep into the aquifer, causing groundwater pollution and forming pollution plumes of varying extents and concentrations. This poses a significant challenge to groundwater pollution control, remediation, and restoration.

[0003] Currently, the traditional methods for tracing the source of groundwater pollution and assessing the recovery cycle in industrial parks are mainly based on hydrogeological surveys and numerical simulations. These methods involve obtaining parameters such as groundwater pollution concentration and permeability coefficient through methods such as borehole sampling and pumping tests, constructing numerical models of groundwater solute transport, retrieving pollution source information, predicting the evolution trend of pollution plumes, and then assessing the recovery cycle.

[0004] However, the aforementioned traditional numerical simulation methods rely on fine grid discretization and a large number of hydrogeological parameters. Industrial parks have complex geological structures, highly heterogeneous aquifers, and significant spatial differences in parameters such as permeability coefficients. Accurately obtaining parameters for the entire region is difficult and costly. Furthermore, the numerical simulation process is cumbersome and extremely time-consuming, which cannot meet the needs of emergency response and real-time decision-making for pollution incidents.

[0005] In recent years, deep learning technology, with its powerful feature extraction and nonlinear fitting capabilities, has been gradually applied to the field of groundwater pollution simulation, providing a new technical path to address the pain points of traditional methods. Among them, the Physical Information Convolutional Decoder (PICD) has been gradually applied to the field of groundwater pollution simulation. Relying on the powerful spatial feature extraction advantages of convolutional structures, it can effectively capture the spatial distribution differences and gradient changes of groundwater pollution concentration. At the same time, combined with the physical constraints of groundwater solute transport, it breaks through the limitations of pure data-driven models that emphasize fitting but neglect physics. To a certain extent, it makes up for the shortcomings of the traditional Physical Information Neural Network (PINN) fully connected neural network architecture, which is difficult to extract spatial features and adapt to complex heterogeneous aquifer scenarios. It has become the core carrier connecting pollution source information, hydrogeological parameters, and pollution concentration distribution, providing a more targeted technical path for groundwater pollution simulation and assessment.

[0006] However, the existing physical constraint embedding method of Physical Information Convolutional Decoder (PICD) has not achieved deep integration of physical laws with the functional layers of the decoder, nor has it designed differentiated physical constraint mechanisms for the functional characteristics of each layer. As a result, the groundwater pollution concentration distribution map generated by decoding still has obvious physical distortion problems. For example, it cannot accurately reflect the diffusion rate of pollutants with groundwater runoff, it is difficult to accurately depict the dynamic migration trajectory and spatial morphology of pollution plumes, and it cannot provide reliable and realistic data support for subsequent pollution source tracing and restoration cycle assessment.

[0007] In summary, how to achieve coupled modeling of pollution sources, permeability coefficients of groundwater discharge paths, and pollution concentration distribution through physical information convolutional decoders, thereby providing accurate basic data for groundwater pollution source tracing and recovery cycle assessment and improving its computational accuracy, is a technical problem that needs to be solved. Summary of the Invention

[0008] To address this, the present invention provides a deep learning-based method for tracing the source of groundwater pollution and assessing its recovery cycle. This method overcomes the limitations of existing technologies that cannot use physical information convolutional decoders to couple the modeling of pollution sources, groundwater discharge paths, permeability coefficients, and pollution concentration distribution, thereby providing accurate basic data for tracing the source of groundwater pollution and assessing its recovery cycle, and improving the accuracy of such calculations.

[0009] To achieve the above objectives, this invention proposes a deep learning-based method for tracing the source of groundwater pollution and assessing its recovery cycle, comprising: The pollution concentration monitoring values ​​of multiple pollution leakage points in the industrial park are obtained, and the pollution concentration monitoring values ​​are passed through an encoder to generate an encoded pollution source strength field vector, wherein the encoder includes a convolutional layer and an LSTM layer. Obtain multiple permeability coefficients from multiple monitoring points along the groundwater runoff discharge path in the industrial park, and calculate the permeability coefficient field vector based on the multiple permeability coefficients; The encoded pollution source strength field vector and permeability coefficient field vector are passed through a physical information convolutional decoder to generate a groundwater pollution concentration distribution map. Based on the groundwater pollution concentration distribution map, groundwater pollution source tracing and recovery cycle assessment are performed. The physical information convolutional decoder is trained using a physical constraint loss function based on the convection diffusion equation. The physical information convolutional decoder includes a physical modulation transpose convolutional block for adjusting feature bias through modulation coefficients.

[0010] Furthermore, the process of generating the encoded pollution source strength field vector through the encoder includes: After preprocessing multiple pollution concentration monitoring values, they are input into a convolutional layer to generate spatial features of pollution concentration. The spatial features of pollution concentration are passed through an LSTM layer to generate spatial temporal features of pollution concentration. The spatial-temporal features of the pollution concentration are passed through a fully connected layer to generate the encoded pollution source strength field vector; The encoder also includes a preprocessing layer and a fully connected layer.

[0011] Furthermore, the process of generating the encoded pollution source strength field vector through the encoder also includes: The measured values ​​of pollution source strength and the encoded pollution source strength field vector are used to train the initial encoder using the cross-entropy pollution source strength fitting loss function to generate the encoder.

[0012] Furthermore, the process of calculating the permeability coefficient field vector includes: Calculate the mean and covariance of multiple permeability coefficients, and substitute the mean and covariance into the KLE expansion formula to calculate the permeability coefficient field vector; The permeability coefficient is obtained through field tests or indoor undisturbed sample experiments at different strata and locations along the groundwater runoff discharge path.

[0013] Furthermore, the process of generating a groundwater pollution concentration distribution map using a physical information convolutional decoder includes: The permeability coefficient field vector and the encoded pollution source strength field vector are respectively decoded through a fully connected layer and remodeled, and then the vectors are spliced ​​together to generate an initial groundwater pollution concentration distribution map. The initial groundwater pollution concentration distribution map, permeability coefficient field vector, and spatiotemporal coordinates of monitoring points are used to generate groundwater pollution concentration mapping features by physically modulating transposed convolutional blocks. The groundwater pollution concentration mapping features are passed through the output convolutional layer to generate the groundwater pollution concentration distribution map; The physical information convolutional decoder further includes a decoding fully connected layer, a reshaping operation, and an output convolutional layer.

[0014] Furthermore, the process of generating groundwater pollution concentration mapping features through physical modulation of transposed convolutional blocks includes: The permeability coefficient field vector is passed through a feature extraction convolutional layer to generate permeability coefficient features; The spatiotemporal coordinates of the monitoring points are passed through a spatiotemporal coding projection convolutional layer to generate coded features for the monitoring points. The concatenated vector of the penetration coefficient feature and the monitoring point coding feature is passed through a convolutional mapping layer to generate the modulation coefficient; The initial groundwater pollution concentration distribution map is passed through a transposed convolutional layer to generate an upsampled groundwater pollution concentration distribution map; Based on the modulation coefficient, the channel element modulation is performed on the upsampled groundwater pollution concentration distribution map to generate a biased modulation groundwater pollution concentration distribution map; The biased modulation groundwater pollution concentration distribution map and modulation coefficient are passed through a normalization activation unit to generate the groundwater pollution concentration mapping feature. The physical modulation transposed convolutional block includes a feature extraction convolutional layer, a spatiotemporal coding projection convolutional layer, a convolutional mapping layer, a transposed convolutional layer, and a normalized activation unit.

[0015] Furthermore, the process of training the physical information convolutional decoder using a physical constraint loss function includes: Based on the mean square error loss function of the groundwater pollution concentration distribution map and the groundwater pollution concentration observation map, a data matching loss term is generated; Based on the groundwater pollution concentration distribution map, groundwater flow velocity, dispersion coefficient, source-sink flow, and pollution concentration monitoring values, physical constraint terms are generated using the mean square residual of the convection dispersion equation. The physical constraint loss function is calculated by weighted summation of the data matching loss term and the physical constraint term, and the physical information convolutional decoder is trained using the physical constraint loss function.

[0016] Furthermore, the process of generating physical constraint terms includes: The groundwater flow velocity is calculated using Darcy's law based on hydraulic head and permeability coefficient. The flow rate is calculated by multiplying the gradient values ​​in each direction at the current point in the groundwater pollution concentration distribution map with the groundwater flow velocity. The concentration change is calculated based on the partial derivative of the current location with respect to time in the groundwater pollution concentration distribution map. The amount of pollution dispersion is calculated by multiplying the gradient value in each direction of the current point in the groundwater pollution concentration distribution map with the dispersion coefficient. The source and sink pollution amounts are calculated based on the product of the groundwater pollution concentration distribution map and the pollution concentration monitoring values. Based on the aforementioned flow rate, concentration change, pollution dispersion, and source-sink pollution levels, physical constraints are generated.

[0017] Furthermore, the process of tracing the source of groundwater pollution based on groundwater pollution concentration distribution maps includes: The concentrated area of ​​pollution leakage was determined based on the groundwater pollution concentration distribution map. The pollution contribution of multiple pollution leakage concentration areas was determined by the entropy weight method. Based on the comparison and ranking of the pollution contribution of multiple leakage areas, the results of the responsibility division for groundwater pollution source tracing are generated.

[0018] Furthermore, the process of assessing the recovery cycle based on groundwater pollution concentration distribution maps includes: Based on the groundwater pollution concentration distribution map, the spatial boundary, volume, and total amount of pollutants of the pollution plume are determined. Combined with the hydrogeological conditions represented by the permeability coefficient field vector, the shrinkage trend of the pollution plume under different remediation conditions is simulated, the natural decay recovery cycle and the engineering remediation cycle are quantified, and a cycle assessment report is output.

[0019] Compared with the prior art, the beneficial effects of the present invention are that the convolutional layer of the encoder can accurately extract the spatial distribution characteristics of pollution concentration monitoring values, capture the spatial correlation between different pollution leakage points, and the bidirectional LSTM layer can efficiently capture the temporal evolution law of pollution concentration changing over time, achieving deep fusion of spatial and temporal features. Then, the fully connected layer maps the fused features into a coded pollution source strength field vector with clear physical meaning, improving the coding accuracy compared with the traditional encoder, and providing an accurate source term basis for the subsequent accurate decoding of physical information convolutional decoder, pollution source tracing and recovery cycle assessment.

[0020] In particular, the physical information convolutional decoder of this invention features a standardized decoding process with extremely high accuracy and outstanding physical consistency. It effectively avoids prediction biases caused by the fuzzy architecture and non-standard decoding process of traditional decoders. The decoder also optimizes the residual backpropagation through a physical constraint loss function containing Darcy's law and convection diffusion equations. The optimization process is standardized and controllable, ensuring that the parameters of each stage of the decoder always adapt to actual needs. The generated pollution concentration distribution map strictly follows the physical laws of groundwater solute transport, completely solving the fatal defect of physical distortion in pure data-driven models. It can accurately depict the spatial boundary, concentration gradient and diffusion trend of pollution plumes, providing core and reliable data support for pollution source tracing and recovery cycle assessment. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the deep learning-based groundwater pollution source tracing and restoration cycle assessment method according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating the encoder of the deep learning-based groundwater pollution source tracing and recovery cycle assessment method according to an embodiment of the present invention. Figure 3 This is a schematic diagram of the forward propagation process of the physical information convolutional decoder in the deep learning-based groundwater pollution source tracing and recovery cycle assessment method according to an embodiment of the present invention. Figure 4 This is a schematic diagram of the backpropagation process of the physical information convolutional decoder in the deep learning-based groundwater pollution source tracing and recovery cycle assessment method according to an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0023] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0024] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0025] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0026] like Figures 1 to 4 As shown, this invention provides a deep learning-based method for tracing the source of groundwater pollution and assessing its recovery cycle. This method overcomes the problem in existing technologies that cannot achieve coupled modeling of pollution sources, groundwater discharge path permeability coefficients, and pollution concentration distribution through physical information convolutional decoders, thereby providing accurate basic data for tracing the source of groundwater pollution and assessing its recovery cycle, and improving its computational accuracy.

[0027] like Figure 1 As shown in the figure, this embodiment proposes a deep learning-based method for tracing the source of groundwater pollution and assessing its recovery cycle, including: The pollution concentration monitoring values ​​of multiple pollution leakage points in the industrial park are obtained, and the pollution concentration monitoring values ​​are passed through an encoder to generate an encoded pollution source strength field vector, wherein the encoder includes a convolutional layer and an LSTM layer. Obtain multiple permeability coefficients from multiple monitoring points along the groundwater runoff discharge path in the industrial park, and calculate the permeability coefficient field vector based on the multiple permeability coefficients; The encoded pollution source strength field vector and permeability coefficient field vector are passed through a physical information convolutional decoder to generate a groundwater pollution concentration distribution map. Based on the groundwater pollution concentration distribution map, groundwater pollution source tracing and recovery cycle assessment are performed. The physical information convolutional decoder is trained using a physical constraint loss function based on the convection diffusion equation. The physical information convolutional decoder includes a physical modulation transpose convolutional block for adjusting feature bias through modulation coefficients.

[0028] Understandably, by embedding the physical laws of pollutant transport into a deep network through a Physical Information Convolutional Decoder (PICD), pollution sources can be directly retrieved from monitoring data without the need for thousands of numerical simulation calls in traditional PINN convolutional neural networks.

[0029] Specifically, the pollution leakage points refer to groundwater monitoring wells and monitoring points within the industrial park where online monitoring equipment has been deployed or where regular manual sampling is performed. The encoded pollution source strength field vector is a low-dimensional vector output by the encoder after compressing and extracting features from the time-series data of multiple monitoring wells. This vector contains all the potential information needed to reconstruct the complete spatial source strength field while retaining only the information most fundamentally related to the spatial distribution of the pollution source.

[0030] Specifically, multiple drawdowns are obtained by analyzing the main seepage aquifers along the main flow path of the industrial park, namely sand, gravel, and fractured rock layers. For steady-state flow at these drawdowns, the Dupuit formula is used to calculate the permeability coefficient. For unsteady-state flow at these drawdowns, the Theis formula is used. The Dupuit formula is a fundamental calculation formula derived from steady-state well flow theory, while the Theis formula is the theoretical basis for unsteady-state pumping tests, which do not require achieving steady-state flow. This perfectly suits the needs of low-permeability formations and short-term pumping in contaminated areas within industrial parks.

[0031] like Figure 2 As shown, the process of generating the encoded pollution source strength field vector through the encoder further includes: After preprocessing multiple pollution concentration monitoring values, they are input into a convolutional layer to generate spatial features of pollution concentration. The spatial features of pollution concentration are passed through an LSTM layer to generate spatial temporal features of pollution concentration. The spatial-temporal features of the pollution concentration are passed through a fully connected layer to generate the encoded pollution source strength field vector; The encoder also includes a preprocessing layer and a fully connected layer.

[0032] Specifically, the monitoring points are located downstream and laterally along the groundwater runoff discharge path. The sampling frequency for pollution concentration monitoring values ​​is once a day, and multiple pollution concentration monitoring values ​​cover at least one full quarterly monitoring period to capture the rising, peak, and decay phases of pollution events. The preprocessing includes Z-score normalization to eliminate differences in concentration magnitudes between different monitoring points, thereby stabilizing network training.

[0033] Specifically, the convolutional layer consists of two layers, comprising a first sub-convolutional layer and a second sub-convolutional layer. The first sub-convolutional layer has 64 kernels, a kernel length of 7, and a stride of 2. Because the pollution concentration monitoring data in the industrial park exhibits significant weekly fluctuations, the first sub-convolutional layer can capture short-term patterns of sudden increases and small oscillations in concentration. The larger kernel length captures weekly trends, and the stride of 2 enables time-dimensional downsampling, compressing the sequence length and increasing the receptive field. The second sub-convolutional layer has 128 kernels, a kernel length of 5, and a stride of 2. This is to extract more abstract composite patterns, such as the synchronous response of monitoring points. The smaller kernel length finely characterizes daily fluctuations, and the stride of 2 enables time-dimensional downsampling, compressing the sequence length. Both the first and second sub-convolutional layers employ the LeakyReLU activation function with a negative slope of 0.2 to preserve negative gradients, prevent neuron death, and is more suitable for noisy monitoring data than the ReLU activation function.

[0034] Specifically, the LSTM layer adopts a bidirectional LSTM structure to capture the temporal evolution characteristics of pollution concentration monitoring values ​​at different time points, so as to utilize past and future information simultaneously to more accurately identify the start, end and peak of pollution events. The number of hidden units in the LSTM layer is 128 to avoid overfitting if too large and information bottlenecks if too small.

[0035] Specifically, the fully connected layer does not use an activation function to achieve a linear output for generating the latent space distribution. Its output dimension is consistent with the feature dimension of the encoded pollution source strength field vector, and is used to map the extracted spatial-temporal fusion features into an encoded pollution source strength field vector with clear physical meaning.

[0036] Furthermore, the process of generating the encoded pollution source strength field vector through the encoder also includes: The measured values ​​of pollution source strength and the encoded pollution source strength field vector are used to train the initial encoder using the cross-entropy pollution source strength fitting loss function to generate the encoder.

[0037] Specifically, the cross-entropy pollution source strength fitting loss function is used to minimize the deviation between the predicted and measured values ​​of the encoded pollution source strength, ensuring that the generated encoded pollution source strength field vector can accurately characterize the magnitude and distribution characteristics of the pollution source strength at each pollution leak point.

[0038] Specifically, the encoder training process is as follows: Collect measured values ​​of pollution source strength and match them with monitored values ​​of pollution concentration; input the monitored values ​​of pollution concentration into the initial encoder to generate an encoded pollution source strength field vector; calculate the cross-entropy loss between the encoded pollution source strength field vector and the measured values ​​of pollution source strength; update the parameters of the encoder's convolutional layer, LSTM layer, and fully connected layer through backpropagation using the Adam optimizer (learning rate 0.001); repeat the above steps until the cross-entropy loss converges (convergence threshold is set to 0.05), and obtain the trained encoder.

[0039] Furthermore, the process of calculating the permeability coefficient field vector includes: Calculate the mean and covariance of multiple permeability coefficients, and substitute the mean and covariance into the KLE expansion formula to calculate the permeability coefficient field vector; The permeability coefficient is obtained through field tests or indoor undisturbed sample experiments at different strata and locations along the groundwater runoff discharge path.

[0040] Specifically, field tests include borehole pumping tests, injection tests, and pit seepage tests. Borehole pumping tests are suitable for Quaternary loose sedimentary layers or bedrock fractured aquifers with large aquifer thickness and good water-bearing capacity, and can obtain representative permeability and hydraulic conductivity coefficients over a wide range of aquifers. Injection tests are suitable for areas with large groundwater levels, weak aquifer water-bearing capacity, or difficult pumping. Pit seepage tests are suitable for the vadose zone in the shallow part of the surface and are used to evaluate the vertical permeability performance of surface soil or strongly weathered layers. The indoor undisturbed sample experiments use undisturbed soil samples or rock cores taken from boreholes to accurately determine the permeability coefficients of different lithological sections under controlled laboratory conditions. Therefore, the permeability coefficients can be accurately obtained according to the environmental conditions of the industrial park's groundwater, providing a solid physical basis for the permeability coefficient field vector. This ensures that the generation of the permeability coefficient field during the source tracing and inversion process of the physical information convolutional decoder conforms to the statistical laws of field geology and can finely characterize the differences in hydraulic conductivity of different strata, thereby significantly improving the accuracy and reliability of pollution plume evolution prediction and recovery cycle assessment.

[0041] Specifically, the KLE expansion formula is:

[0042] in, Represents the logarithmic permeability field, i.e. , This represents the permeability coefficient at grid location x, where x represents the grid location in three-dimensional coordinates. Represents the permeability coefficient field vector. This represents the mean permeability coefficient at grid location x. This indicates the truncation order, which is determined to be 15 based on the cumulative contribution rate of the eigenvalues. Represents covariance The i-th largest eigenvalue, Represents covariance The orthogonal basis functions, with covariance as ,in This represents the variance of multiple permeability coefficients at this grid location, where h and l represent the spatial distance and horizontal correlation length between two points, respectively. This represents the i-th element of the permeability coefficient field vector. This indicates that the i-th element of the permeability coefficient field vector follows a standard normal distribution. .

[0043] The calculation process for the truncation order is as follows: First, based on the measured permeability coefficient values ​​at multiple monitoring points along the groundwater runoff discharge path, a covariance matrix of the permeability coefficient field is constructed. Second, the covariance matrix is ​​decomposed into eigenvalues ​​to obtain a series of eigenvalues ​​arranged in descending order and their corresponding orthogonal basis functions. Then, the proportion of the cumulative eigenvalues ​​to the sum of the total eigenvalues ​​is calculated, and the eigenvalues ​​are accumulated sequentially. When the cumulative contribution rate reaches a preset accuracy threshold of 95%, the accumulation stops. The number of accumulated eigenvalues ​​at this point is the truncation order 15. The top 15 largest eigenvalues ​​and their corresponding orthogonal basis functions are selected and substituted into the KLE expansion formula to transform the high-dimensional permeability coefficient field into a permeability coefficient field vector composed of 15 independent Gaussian variables. This effectively reduces the input dimension and computational load of the subsequent physical information convolutional decoder while ensuring accurate representation of the spatial distribution characteristics of the permeability coefficient, thus balancing accuracy and computational efficiency.

[0044] Therefore, by substitution , The permeability coefficient field vector can be solved; the KLE expansion formula originates from a stochastic process. Theorem states that any second-order stationary (or weakly stationary) random field can be represented as an infinite series expansion. The original permeability coefficient field has multiple degrees of freedom in the grid, while KLE compresses it into M independent Gaussian variables (e.g., 15). In the inversion of the physical information convolutional decoder, only these M variables need to be optimized, greatly reducing computational complexity while preserving the geostatistical characteristics of the permeability coefficient.

[0045] Specifically, in the pollution source tracing and evolution prediction framework of the Physical Information Convolutional Decoder (PICD), the role of KLE expansion is to compress the high-dimensional, spatially heterogeneous hydrogeological parameter field with geostatistical priors into a low-dimensional random vector, which serves as a trainable input variable for the convolutional decoder. KLE expansion uses a second-order stationary Gaussian random field to describe the spatial continuity of the permeability coefficient.

[0046] like Figure 3 As shown, the process of generating a groundwater pollution concentration distribution map using a physical information convolutional decoder further includes: The permeability coefficient field vector and the encoded pollution source strength field vector are respectively decoded through a fully connected layer and remodeled, and then the vectors are spliced ​​together to generate an initial groundwater pollution concentration distribution map. The initial groundwater pollution concentration distribution map, permeability coefficient field vector, and spatiotemporal coordinates of monitoring points are used to generate groundwater pollution concentration mapping features by physically modulating transposed convolutional blocks. The groundwater pollution concentration mapping features are passed through the output convolutional layer to generate the groundwater pollution concentration distribution map; The physical information convolutional decoder further includes a decoding fully connected layer, a reshaping operation, and an output convolutional layer.

[0047] Specifically, the permeability coefficient field vector is mapped to a feature map via the first fully connected layer, and the reshaping operation results in the first tensor. The encoded pollution source strength field is mapped to a feature map via the second fully connected layer, and the reshaping operation results in the second tensor. The spatial and temporal dimensions of the first and second tensors are aligned with the spatiotemporal coordinates of the monitoring points. The concatenated vector of the first and second tensors forms the initial groundwater pollution concentration distribution map. Both the first and second fully connected layers employ the LeakyReLU activation function to provide nonlinearity. Therefore, through the fully connected layers and the reshaping operation, the dimensionality expansion from the latent space to the image space is achieved, transforming the abstract encoding into a tensor form of a distribution map that can be convolutionally processed. The independent branch design decouples the generation processes of the source strength and permeability coefficient, avoiding mutual interference between the two types of heterogeneous parameters in the initial stage, which is beneficial for the conditional injection of subsequent physically modulated transposed convolutional blocks.

[0048] Specifically, the output convolutional layer has a kernel size of 3×3, a stride of 1, edge padding of 1, and an activation function of Softplus. This compresses the high-dimensional abstract feature map into a concentration field corresponding to each feature channel at a given prediction time step, generating a pollution plume evolution sequence across multiple time steps, providing efficient prediction for recovery cycle assessment. The Softplus activation function, while ensuring non-negativity, avoids the gradient breakage problem of ReLU at zero, making it suitable for concentration, a non-negative continuous physical quantity.

[0049] Furthermore, the process of generating groundwater pollution concentration mapping features through physical modulation of transposed convolutional blocks includes: The permeability coefficient field vector is passed through a feature extraction convolutional layer to generate permeability coefficient features; The spatiotemporal coordinates of the monitoring points are passed through a spatiotemporal coding projection convolutional layer to generate coded features for the monitoring points. The concatenated vector of the penetration coefficient feature and the monitoring point coding feature is passed through a convolutional mapping layer to generate the modulation coefficient; The initial groundwater pollution concentration distribution map is passed through a transposed convolutional layer to generate an upsampled groundwater pollution concentration distribution map; Based on the modulation coefficient, the channel element modulation is performed on the upsampled groundwater pollution concentration distribution map to generate a biased modulation groundwater pollution concentration distribution map; The biased modulation groundwater pollution concentration distribution map and modulation coefficient are passed through a normalization activation unit to generate the groundwater pollution concentration mapping feature. The physical modulation transposed convolutional block includes a feature extraction convolutional layer, a spatiotemporal coding projection convolutional layer, a convolutional mapping layer, a transposed convolutional layer, and a normalized activation unit.

[0050] Specifically, the kernel size of the feature extraction convolutional layer is 3×3, the stride is 1, the edge padding is 1, and the activation function is LeakyReLU. This can transform the permeation coefficient field vector into a high-level semantic feature suitable for generating modulation coefficients, enabling the decoder network to perceive the strength of the permeation ability at the current spatial location and its local gradient.

[0051] Specifically, the convolutional mapping layer adopts a linear fully connected layer architecture without activation functions to maintain the linear additivity of coordinate information, which is beneficial to the stability of the modulation signal. It can transform global spatiotemporal coordinates into conditional tensors with spatial constants but different channels, so that the decoder knows the absolute spatial position and prediction time of the currently generated content.

[0052] Specifically, the convolutional mapping layer uses a 3×3 kernel size, a stride of 1, an edge padding of 1, no activation function, and normalization calculation to generate a spatially adaptive modulation signal based on the fusion conditions, ensuring that the modulation coefficients α and β are in the range of [0,1], thus achieving conditional control at the feature map level.

[0053] Specifically, the transposed convolutional layer has a kernel size of 4×4, a stride of 2, and an edge padding of 1. The combination of an even kernel and a stride of 2 can achieve precise size doubling without overlap. Therefore, the transposed convolutional layer performs spatial upsampling, doubling the resolution of the feature map and providing a structural skeleton for generating high-resolution contamination plumes.

[0054] Specifically, the process of channel element modulation can be represented as: ,in This indicates a bias towards modifying the groundwater pollution concentration distribution map. , Both represent modulation coefficients. This indicates element-wise multiplication. This represents the distribution map of pollution concentration in the sampled groundwater. Therefore, by modulating the tensor... The values ​​of convection-related characteristic channels in the upsampled groundwater pollution concentration distribution map can be amplified, resulting in a higher concentration gradient and faster plume expansion in that area, through modulation coefficients. A negative bias can be applied to weaken non-physical concentration penetration, forcing the pollution plume to flow around it. The resulting pollution plume, which modulates the groundwater pollution concentration distribution map, exhibits a natural frontal bending and flow around the permeability abrupt interface at the lithological boundary, rather than a simple isotropic diffusion circle. This allows the decoder network to amplify the activation of convection characteristics along high-permeability strips, clearly reflecting the dominant flow path and accurately characterizing the preferential migration direction of pollutants in heterogeneous media.

[0055] Therefore, the physical modulation transposed convolution block adjusts the feature bias through the modulation coefficient, so that the generated pollution concentration distribution map conforms to the flow around and transport law of pollutants in heterogeneous strata.

[0056] Please see Figure 4 This is a schematic diagram of the backpropagation process of the physical information convolutional decoder, which is the training process. Furthermore, the training process of the physical information convolutional decoder using the physical constraint loss function includes: Based on the mean square error loss function of the groundwater pollution concentration distribution map and the groundwater pollution concentration observation map, a data matching loss term is generated; Based on the groundwater pollution concentration distribution map, groundwater flow velocity, dispersion coefficient, source-sink flow, and pollution concentration monitoring values, physical constraint terms are generated using the mean square residual of the convection dispersion equation. The physical constraint loss function is calculated by weighted summation of the data matching loss term and the physical constraint term, and the physical information convolutional decoder is trained using the physical constraint loss function.

[0057] Specifically, the measured values ​​of pollution concentration and their coordinates at all monitoring points are collected, and spatial kriging interpolation is used to extrapolate the concentration at discrete points to the entire model grid to generate a groundwater pollution concentration observation map.

[0058] Specifically, the training process of the physical information convolutional decoder is as follows: input the encoded pollution source strength field vector and permeability coefficient field vector to generate an initial groundwater pollution concentration distribution map; calculate the data matching loss term and physical constraint term; calculate the weighted summation of the physical constraint loss function; update the parameters of the fully connected layer, the physical modulation transposed convolutional block, and the output convolutional layer of the decoder through backpropagation using the Adam optimizer; repeat steps 1-4 above for 1000 iterations until the loss function converges, and obtain the trained physical information convolutional decoder.

[0059] Specifically, the weighting coefficients for the data matching loss term and the physical constraint term in the physical constraint loss function are 1.0 and 0.1, respectively. After multiple experiments, it has been verified that these weighting coefficients can ensure that the network output approximates the source accuracy of the monitoring data, and can also force the output to satisfy the convection diffusion equation, ensuring physical consistency and avoiding overfitting or physical distortion.

[0060] Furthermore, the process of generating physical constraint terms includes: The groundwater flow velocity is calculated using Darcy's law based on hydraulic head and permeability coefficient. The flow rate is calculated by multiplying the gradient values ​​in each direction at the current point in the groundwater pollution concentration distribution map with the groundwater flow velocity. The concentration change is calculated based on the partial derivative of the current location with respect to time in the groundwater pollution concentration distribution map. The amount of pollution dispersion is calculated by multiplying the gradient value in each direction of the current point in the groundwater pollution concentration distribution map with the dispersion coefficient. The source and sink pollution amounts are calculated based on the product of the groundwater pollution concentration distribution map and the pollution concentration monitoring values. Based on the aforementioned flow rate, concentration change, pollution dispersion, and source-sink pollution levels, physical constraints are generated.

[0061] Specifically, the physical constraint term can be expressed as:

[0062] in, Represents physical constraint terms. The total number of grid points is represented by C, which represents the groundwater pollution concentration distribution map, and t represents time. Indicates the amount of change in concentration. , , This represents the dispersion coefficient in the x, y, and xy directions, and is determined based on the type of pollutant being detected. , , The second-order partial derivative and mixed partial derivative, representing concentration, are obtained by automatic differentiation based on the first-order derivative. They describe the concentration curvature and are used to calculate hydrodynamic dispersion. Indicates the amount of pollution dispersed. , The components of the groundwater flow velocity in the x and y directions are represented and calculated using Darcy's law: Where K represents the permeability coefficient. The effective porosity is represented by h, and the spatial distance between the two points is represented by h. , This represents the gradient values ​​in each direction at the current point in the groundwater pollution concentration distribution map. S represents the pollution concentration monitoring value after Z-score normalization, representing the flow rate. The source characteristic concentration is represented in the groundwater pollution concentration distribution map. Indicates effective porosity. This indicates the amount of pollution at the source and sink. Effective porosity, determined by water immersion testing (WIP) on borehole cores, is the ratio of the volume of interconnected pores through which fluid can pass to the total volume of the rock.

[0063] Therefore, the above physical constraint term, namely the change in concentration minus the amount of pollution dispersion plus the amount of pollution from the source and sink, is used to make the change in concentration subject to the physical constraints of the flow rate, the amount of pollution dispersion, and the amount of pollution from the source and sink. The amount of pollution from the source and sink can be understood as the mass of pollutants injected into the aquifer per unit volume per unit time.

[0064] Understandably, the monitoring data input to the model often contains random errors and system noise. If only data matching loss terms are used for training, the network will overfit the noise, inverting false pollution sources or infiltration anomalies. Physical constraints, as smoothing priors, filter out short-wavelength disturbances that violate mass conservation, making the inversion results more robust.

[0065] Specifically, in the residual backpropagation optimization of the physical information convolutional decoder, the Adam optimizer is used to calculate the partial derivatives of the physical constraint loss function with respect to the decoding fully connected layer, the reshaping operation, the transposed convolutional block, and the output convolutional layer, in order to perform residual backpropagation optimization.

[0066] Furthermore, the process of tracing the source of groundwater pollution based on groundwater pollution concentration distribution maps includes: The concentrated area of ​​pollution leakage was determined based on the groundwater pollution concentration distribution map. The pollution contribution of multiple pollution leakage concentration areas was determined by the entropy weight method. Based on the comparison and ranking of the pollution contribution of multiple leakage areas, the results of the responsibility division for groundwater pollution source tracing are generated.

[0067] Specifically, 0.3 of the maximum concentration in the groundwater pollution concentration distribution map is used as the concentration threshold, and grids exceeding the concentration threshold are marked as pollution leakage concentration areas, which include multiple grid points.

[0068] Specifically, the process of determining the pollution contribution of the leak area is as follows: Obtain multiple concentration values ​​from the concentrated area of ​​the pollution leak on the groundwater pollution concentration distribution map. Calculate the regional concentration integral and regional concentration peak of these multiple concentration values. After standardizing the regional concentration integral and regional concentration peak, divide the standardized score of the regional concentration integral at that point by the sum of the standardized scores of the concentration integrals at all points in the concentrated pollution leak area to calculate the weight of this indicator. Similarly, divide the standardized score of the regional concentration peak at that point by the sum of the standardized scores of the regional concentration peaks at all points in the concentrated pollution leak area to calculate the weight of this indicator. For each indicator, using the regional weight values ​​obtained in the previous step, calculate according to the definition of information entropy: multiply the weight value of each region by its natural logarithm, sum over all regions, take the opposite number, and divide by the natural logarithm of the number of regions for normalization correction. The result is the entropy value of this indicator. Based on the entropy value obtained in the previous step, for each indicator, subtract its entropy value from a constant to obtain the difference coefficient of this indicator. Sum the difference coefficients of all indicators to obtain a total value. Then, for each indicator, its own difference coefficient is divided by the sum value, and the result is the final weight of the indicator in the comprehensive evaluation system. Based on the final weight, the indicators (regional concentration integral and regional concentration peak) are weighted and summed to obtain the pollution contribution of the leak area.

[0069] Furthermore, the process of assessing the recovery cycle based on groundwater pollution concentration distribution maps includes: Based on the groundwater pollution concentration distribution map, the spatial boundary, volume, and total amount of pollutants of the pollution plume are determined. Combined with the hydrogeological conditions represented by the permeability coefficient field vector, the shrinkage trend of the pollution plume under different remediation conditions is simulated, the natural decay recovery cycle and the engineering remediation cycle are quantified, and a cycle assessment report is output.

[0070] Specifically, the groundwater pollution concentration distribution map delineates the spatial boundary of the pollution plume using the concentration threshold, identifying the plume's planar distribution morphology, maximum migration distance, and mainstream direction. Based on this, the area and average vertical thickness of the pollution plume are calculated using the model mesh size, further estimating the total volume of the contaminated aquifer. Simultaneously, the concentration values ​​within the plume boundary are spatially integrated, and combined with effective porosity parameters, the total mass of pollutants currently remaining in the groundwater is calculated. Before the assessment, several alternative remediation scenarios need to be set. The natural decay scenario refers to the gradual dissipation of the pollution plume through groundwater convection, dispersion, and natural degradation of pollutants without any human intervention. Engineering remediation scenarios include hydraulic control via pumping wells at specific locations and in-situ remediation via the injection of chemical oxidants in specific areas. The permeability coefficient input to the KLE expansion formula is determined based on the engineering remediation scenario, and then the shrinkage trend of the pollution plume under different remediation scenarios is simulated using an encoder and decoder to quantify the recovery period.

[0071] Specifically, a chemical industrial park was selected as the experimental site. This park has five known pollution leakage points. Twenty groundwater monitoring wells were deployed (10 located around the pollution leakage points, 5 downstream of the groundwater runoff, and 5 laterally), collecting pollution concentration data for 30 consecutive days. The chemical industrial park covers an area of ​​5 km². The groundwater type is Quaternary loose rock pore water, with the aquifer mainly consisting of sand and gravel layers. The permeability coefficient ranges from 1.2 to 8.5 m / d, and the main pollutant is the heavy metal cadmium. The pollution leak points are located in the chemical production area on the west side of the industrial park. Through borehole pumping tests and core permeability tests, measured permeability coefficients were obtained at 30 points along the groundwater runoff discharge path in the park. Kriging interpolation was used to supplement the data for the entire park, resulting in a permeability coefficient dataset. Simultaneously, through on-site sampling and detection, measured pollution source strength values ​​for each leak point and a measured distribution map of groundwater pollution concentration across the entire park were obtained as baseline data for experimental verification. The encoder and physical information convolutional decoder models of this embodiment were constructed. The convolutional layers consist of 2 layers with 3×3 convolutional kernels, the bidirectional LSTM layer has 128 hidden units, and the fully connected layer has an output dimension of 5. The transposed convolutional block of the physical information convolutional decoder model contains 2 transposed convolutional layers (3×3 convolutional kernels, stride 2) and 2 convolutional batch normalization layers. The output convolutional layer uses a 1×1 convolutional kernel. The comparison groups consisted of a traditional CNN encoder, a standard convolutional decoder, and a PINN model. All experimental parameters were kept consistent across groups, with 1000 training iterations, a learning rate of 0.001, and a preset residual threshold of 10. -4 The fitting error threshold is 5%.

[0072] The encoder's pollution source strength fitting loss value converged to 0.023, and the average relative error between the predicted and measured values ​​of the encoded pollution source strength was 3.8%. In contrast, the traditional CNN encoder in the comparison group had a fitting loss value of 0.087 and an average relative error of 12.5%, indicating that the encoder can encode pollution source strength information more accurately.

[0073] The mean absolute error (MAE) between the groundwater pollution concentration distribution map generated by the physical information convolutional decoder and the measured distribution map is 0.056 mg / L, and the coefficient of determination is... The MAE was 0.962; the MAE of the ordinary convolutional decoder in the control group was 0.189 mg / L. The MAE of the PINN model was 0.783, and the MAE of the PINN model was 0.112 mg / L. The value of 0.875 indicates that the pollution concentration distribution map generated by the decoder is significantly more accurate than that of the control group, and has better physical consistency.

[0074] In this embodiment, the spatial distribution characteristics of pollution concentration monitoring values ​​can be accurately extracted through the convolutional layer of the encoder, capturing the spatial correlation between different pollution leakage points. The bidirectional LSTM layer can efficiently capture the temporal evolution law of pollution concentration changing over time, achieving deep fusion of spatial and temporal features. Then, the fused features are mapped into encoded pollution source strength field vectors with clear physical meaning through the fully connected layer. The encoding accuracy is improved compared with traditional encoders, providing an accurate source term basis for the accurate decoding of subsequent physical information convolutional decoders, pollution source tracing, and recovery cycle assessment. The physical information convolutional decoder boasts a standardized decoding process, extremely high accuracy, and outstanding physical consistency. It effectively avoids prediction biases caused by the fuzzy architecture and non-standard decoding process of traditional decoders. The decoder also optimizes residual backpropagation through a physical constraint loss function containing Darcy's law and convection diffusion equations. The optimization process is standardized and controllable, ensuring that the parameters of each stage of the decoder always adapt to actual needs. The generated pollution concentration distribution map strictly follows the physical laws of groundwater solute transport, completely solving the fatal flaw of physical distortion in pure data-driven models. It can accurately depict the spatial boundary, concentration gradient, and diffusion trend of pollution plumes, providing core and reliable data support for pollution source tracing and recovery cycle assessment.

[0075] Those skilled in the art will recognize that the modules and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0076] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0077] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A deep learning-based method for tracing the source of groundwater pollution and assessing its recovery cycle, characterized in that, include: The pollution concentration monitoring values ​​of multiple pollution leakage points in the industrial park are obtained, and the pollution concentration monitoring values ​​are passed through an encoder to generate an encoded pollution source strength field vector, wherein the encoder includes a convolutional layer and an LSTM layer. Obtain multiple permeability coefficients from multiple monitoring points along the groundwater runoff discharge path in the industrial park, and calculate the permeability coefficient field vector based on the multiple permeability coefficients; The encoded pollution source strength field vector and permeability coefficient field vector are passed through a physical information convolutional decoder to generate a groundwater pollution concentration distribution map. Based on the groundwater pollution concentration distribution map, groundwater pollution source tracing and recovery cycle assessment are performed. The physical information convolutional decoder is trained using a physical constraint loss function based on the convection diffusion equation. The physical information convolutional decoder includes a physical modulation transpose convolutional block for adjusting feature bias through modulation coefficients. The process of generating encoded pollution source strength field vectors through an encoder includes: After preprocessing multiple pollution concentration monitoring values, they are input into a convolutional layer to generate spatial features of pollution concentration. The spatial features of pollution concentration are passed through an LSTM layer to generate spatial temporal features of pollution concentration. The spatial-temporal features of the pollution concentration are passed through a fully connected layer to generate the encoded pollution source strength field vector; The encoder further includes a preprocessing layer and a fully connected layer; The process of generating encoded pollution source strength field vectors through an encoder also includes: The measured values ​​of pollution source strength and the encoded pollution source strength field vector are used to train the initial encoder using the cross-entropy pollution source strength fitting loss function to generate the encoder. The process of calculating the permeability coefficient field vector includes: Calculate the mean and covariance of multiple permeability coefficients, and substitute the mean and covariance into the KLE expansion formula to calculate the permeability coefficient field vector; The permeability coefficient was obtained through field tests or indoor undisturbed sample experiments at different strata and locations along the groundwater runoff discharge path. The training process of the physical information convolutional decoder using a physical constraint loss function includes: Based on the mean square error loss function of the groundwater pollution concentration distribution map and the groundwater pollution concentration observation map, a data matching loss term is generated; Based on the groundwater pollution concentration distribution map, groundwater flow velocity, dispersion coefficient, source-sink flow, and pollution concentration monitoring values, physical constraint terms are generated using the mean square residual of the convection dispersion equation. The physical constraint loss function is calculated by weighted summation of the data matching loss term and the physical constraint term, and the physical information convolutional decoder is trained using the physical constraint loss function. The process of generating physical constraint terms includes: The groundwater flow velocity is calculated using Darcy's law based on hydraulic head and permeability coefficient. The flow rate is calculated by multiplying the gradient values ​​in each direction at the current point in the groundwater pollution concentration distribution map with the groundwater flow velocity. The concentration change is calculated based on the partial derivative of the current location with respect to time in the groundwater pollution concentration distribution map. The amount of pollution dispersion is calculated by multiplying the gradient value in each direction of the current point in the groundwater pollution concentration distribution map with the dispersion coefficient. The source and sink pollution amounts are calculated based on the product of the groundwater pollution concentration distribution map and the pollution concentration monitoring values. Based on the aforementioned flow rate, concentration change, pollution dispersion, and source-sink pollution levels, physical constraints are generated.

2. The method for tracing the source of groundwater pollution and assessing its recovery cycle based on deep learning as described in claim 1, characterized in that, The process of generating a groundwater pollution concentration distribution map using a physical information convolutional decoder includes: The permeability coefficient field vector and the encoded pollution source strength field vector are respectively decoded through a fully connected layer and remodeled, and then the vectors are spliced ​​together to generate an initial groundwater pollution concentration distribution map. The initial groundwater pollution concentration distribution map, permeability coefficient field vector, and spatiotemporal coordinates of monitoring points are used to generate groundwater pollution concentration mapping features by physically modulating transposed convolutional blocks. The groundwater pollution concentration mapping features are passed through the output convolutional layer to generate the groundwater pollution concentration distribution map; The physical information convolutional decoder further includes a decoding fully connected layer, a reshaping operation, and an output convolutional layer.

3. The method for tracing groundwater pollution sources and assessing restoration cycles based on deep learning according to claim 2, characterized in that, The process of generating groundwater pollution concentration mapping features by physically modulating transposed convolutional blocks includes: The permeability coefficient field vector is passed through a feature extraction convolutional layer to generate permeability coefficient features; The spatiotemporal coordinates of the monitoring points are passed through a spatiotemporal coding projection convolutional layer to generate coded features for the monitoring points. The concatenated vector of the penetration coefficient feature and the monitoring point coding feature is passed through a convolutional mapping layer to generate the modulation coefficient; The initial groundwater pollution concentration distribution map is passed through a transposed convolutional layer to generate an upsampled groundwater pollution concentration distribution map; Based on the modulation coefficient, the channel element modulation is performed on the upsampled groundwater pollution concentration distribution map to generate a biased modulation groundwater pollution concentration distribution map; The biased modulation groundwater pollution concentration distribution map and modulation coefficient are passed through a normalization activation unit to generate the groundwater pollution concentration mapping feature. The physical modulation transposed convolutional block includes a feature extraction convolutional layer, a spatiotemporal coding projection convolutional layer, a convolutional mapping layer, a transposed convolutional layer, and a normalized activation unit.

4. The method for tracing the source and assessing the recovery cycle of groundwater pollution based on deep learning according to any one of claims 1 to 3, characterized in that, The process of tracing the source of groundwater pollution based on groundwater pollution concentration distribution maps includes: The concentrated area of ​​pollution leakage was determined based on the groundwater pollution concentration distribution map. The pollution contribution of multiple pollution leakage concentration areas was determined by the entropy weight method. Based on the comparison and ranking of the pollution contribution of multiple leakage areas, the results of the responsibility division for groundwater pollution source tracing are generated.

5. The method for tracing the source of groundwater pollution and assessing the recovery cycle based on deep learning according to any one of claims 1 to 3, characterized in that, The process of assessing remediation cycles based on groundwater pollution concentration distribution maps includes: Based on the groundwater pollution concentration distribution map, the spatial boundary, volume, and total amount of pollutants of the pollution plume are determined. Combined with the hydrogeological conditions represented by the permeability coefficient field vector, the shrinkage trend of the pollution plume under different remediation conditions is simulated. Based on the shrinkage trend, natural decay recovery period, and engineering remediation period, a period assessment report is output.

Images