Precise transport control method and system for remediation agents in in-situ groundwater remediation

Abstract

This invention provides a method and system for precise transport control of remediation agents in in-situ groundwater remediation, relating to the field of groundwater pollution remediation technology. The method includes: acquiring images and sensor data from borehole cores at contaminated sites; extracting features through visual attention networks and temporal attention networks; fusing these features into a conditional probability diffusion model to generate hydrogeological feature vectors; constructing a high-precision hydrogeological feature model by combining spatial attention mapping and a generative diffusion model; inputting this model into a multi-regional collaborative simulation environment for optimization calculation; fusing monitoring data through edge computing nodes; and training a lightweight control model to execute the transport control of remediation agents.

Description

Technical Field

[0001] This invention relates to the field of groundwater pollution remediation technology, and in particular to a method and system for precise transport and control of remediation agents for in-situ groundwater remediation. Background Technology

[0002] As an important means of treating groundwater pollution, in-situ groundwater remediation technology focuses on the precise transport and control of remediation agents. With the development of technology, the transport and control methods of in-situ groundwater remediation have evolved from experience-based control to intelligent control. Early transport and control mainly relied on human experience, using single-point injection or simple grid layout for remediation agent injection, and often designed fixed injection parameter schemes based on static geological conditions.

[0003] With the advancement of geological exploration and characterization technologies, methods such as borehole geophysical exploration and geophysical surveys have been gradually developed. Image recognition technology has been introduced to analyze the structural characteristics of rock cores. At the same time, multi-source sensors are used to acquire hydrogeological parameters, and numerical simulation methods are used to construct geological models. However, existing technologies still have problems such as limited feature extraction accuracy from single data sources, difficulty in comprehensively depicting complex geological conditions, lack of multi-regional collaborative mechanisms in the decision-making process, poor adaptability of optimization strategies, and difficulty in coping with complex and ever-changing geological conditions.

[0004] Therefore, a solution is urgently needed to address the problems existing in the current technology. Summary of the Invention

[0005] This invention provides a method and system for precise transport and control of remediation agents for in-situ groundwater remediation, which can at least solve some of the problems existing in the prior art.

[0006] A first aspect of this invention provides a method for precise transport control of remediation agents for in-situ groundwater remediation, comprising:

[0007] Image data and sensor measurement data corresponding to borehole cores of contaminated sites are collected. The image data is input into a visual attention neural network for feature extraction to obtain soil structure feature data. The sensor measurement data is input into a temporal attention neural network for analysis to obtain groundwater dynamic feature data. The soil structure feature data and groundwater dynamic feature data are added to a conditional probability diffusion model for feature fusion to generate hydrogeological feature vectors. These vectors are then mapped to a three-dimensional spatial grid through a spatial attention calculation module to obtain an initial hydrogeological feature model. Finally, a generative probability diffusion model is used to complete the missing areas to obtain a high-precision hydrogeological feature model.

[0008] The high-precision hydrogeological feature model is input into a multi-regional collaborative simulation environment. Multiple decision-making units perform local optimization calculations on different spatial regions. The decision-making units share optimization results through a communication network. The central evaluation network performs a global evaluation based on the shared results and coordinates the adjustment of optimization parameters of each decision-making unit. The adjusted transport strategy is verified by combining the Monte Carlo probability search method, and the optimal transport parameter scheme is output.

[0009] Edge computing control nodes are deployed at the contaminated site. The monitoring data of each area are fused and processed by a distributed model parameter aggregation algorithm to generate a global transport control model. This model is then trained with the optimal transport parameter scheme to obtain a lightweight field control model. The lightweight field control model calculates the optimal transport control scheme based on real-time monitoring data and sends it to the injection device for execution.

[0010] In one alternative implementation,

[0011] Image data and sensor measurement data corresponding to borehole cores from contaminated sites are collected. The image data is input into a visual attention neural network for feature extraction to obtain soil structure feature data. The sensor measurement data is input into a temporal attention neural network for analysis to obtain groundwater dynamic feature data. The soil structure feature data and groundwater dynamic feature data are added to a conditional probability diffusion model for feature fusion, generating hydrogeological feature vectors. These vectors are then mapped to a 3D spatial grid using a spatial attention calculation module to obtain an initial hydrogeological feature model. Finally, a generative probability diffusion model is used to complete the model for missing areas, resulting in a high-precision hydrogeological feature model, including:

[0012] A borehole sampling system was deployed at the contaminated site to collect core image data and groundwater sensor measurement data.

[0013] The core image data is input into an adaptive histogram equalization processing module to eliminate illumination interference and obtain a first processed image. The first processed image is then processed by local contrast sharpening to enhance structural details and obtain a second processed image. The second processed image is input into a visual attention neural network. The encoder of the visual attention neural network performs convolution operations to extract features, and the decoder reconstructs the soil structure feature data.

[0014] The groundwater sensor measurement data is input into the outlier detection module to remove outliers and obtain the first processed data. The first processed data is then denoised using an adaptive sliding window to obtain the second processed data. The second processed data is then input into a temporal attention neural network, and the temporal correlation is analyzed using a bidirectional long short-term memory unit to obtain the groundwater dynamic feature data.

[0015] The soil structure feature data and the groundwater dynamic feature data are concatenated and input into the conditional probability diffusion model. The hydrogeological feature vector is obtained through dynamic noise reduction iteration. The hydrogeological feature vector is input into the spatial attention calculation module. The spatial weight is calculated based on the feature similarity and the grid is densified. The initial hydrogeological feature model is obtained through feature interpolation.

[0016] The initial hydrogeological feature model is input into the generative probability diffusion model. Based on the known regional feature distribution, feature generation is performed on the missing region. A high-precision hydrogeological feature model is obtained by evaluating and optimizing it through a multi-scale discriminator.

[0017] In one alternative implementation,

[0018] The soil structure feature data and the groundwater dynamic feature data are concatenated and input into a conditional probability diffusion model. Through dynamic noise reduction iteration, a hydrogeological feature vector is obtained, including:

[0019] Soil structural feature data and groundwater dynamic feature data are concatenated to obtain a hybrid feature vector. This hybrid feature vector is then input into a two-branch structure of a conditional probability diffusion model. Latent features are obtained through the main branch, and constraints are obtained through the auxiliary branch. Differential noise injection and denoising processing are applied to the features. Combined with constraints and an attention mechanism, fused features are generated. Based on feature quality assessment, parameters are optimized, and the hydrogeological feature vector is reconstructed. Specifically, this includes:

[0020] Soil structural feature data and groundwater dynamic feature data are spliced ​​together using a hierarchical cascade structure. Features of the same scale are spliced ​​together to obtain a feature vector of the same scale. The feature vectors of the same scale are combined to form a hybrid feature vector and input into a conditional probability diffusion model with a dual-branch parallel structure. The main branch of the conditional probability diffusion model maps the hybrid feature vector to the latent space through a feature encoder to obtain latent features. The auxiliary branch extracts physical law constraint information, spatial continuity constraint information and temporal consistency constraint information through a constraint encoder to obtain a constraint condition vector.

[0021] Based on the hybrid feature vector, the feature importance distribution is calculated. Differential noise injection is performed on different feature regions according to the feature importance distribution. Strong noise is injected into feature regions with an importance of less than 0.3, and weak noise is injected into feature regions with an importance of more than 0.75. Noise features are obtained by superimposing multi-level noise to achieve feature degradation.

[0022] The noise features are input into a multi-cascaded noise reduction unit. The multi-cascaded noise reduction unit extracts feature information through a densely connected feature extraction module, predicts noise distribution through a noise prediction module, and achieves dynamic noise reduction by using a skip connection method to obtain noise reduction features. The constraint vector is input into a constraint decoder to generate feature constraints. The noise reduction features are then constrained and adjusted through a recursive feedback structure to obtain constraint features.

[0023] A multi-head attention module is constructed to calculate the feature correlation degree of the constraint features. Attention weights are generated based on the feature correlation degree. The constraint features are enhanced by the attention weights and residual connections are introduced to retain the original feature information to obtain fused features.

[0024] A parameter evaluation module is constructed to calculate the feature quality index and fusion effect index of the fused features. Based on the feature quality index and the fusion effect index, the noise reduction parameters are dynamically optimized to obtain optimized parameters. The optimized parameters are input into the feature decoder to decode and reconstruct the fused features to obtain the hydrogeological feature vector.

[0025] In one alternative implementation,

[0026] The high-precision hydrogeological feature model is input into a multi-regional collaborative simulation environment. Multiple decision-making units perform local optimization calculations for different spatial regions. These units share optimization results via a communication network. A central evaluation network performs a global evaluation based on the shared results and collaboratively adjusts the optimization parameters of each decision-making unit. The adjusted transport strategy is verified using a Monte Carlo probabilistic search method, and the optimal transport parameter scheme is output, including:

[0027] The high-precision hydrogeological feature model is divided into multiple spatial regions according to the boundaries of hydrogeological units and management zones. Decision units with data layers, computing layers and communication layers are configured in each spatial region.

[0028] The data layer of the decision-making unit extracts the geological structure characteristics, hydrological parameter distribution and boundary condition information of the spatial region, establishes a groundwater flow field numerical model, collects groundwater level, water quality index and extraction volume data to construct a dynamic database, and the calculation layer of the decision-making unit inputs the state data in the dynamic database into the optimizer, performs local optimization calculation based on the optimization objectives of water resource balance, water quality safety and system stability to obtain the optimization result;

[0029] The decision-making units interconnect through a communication network to build a hierarchical caching mechanism to store the optimization results. They adopt an incremental update strategy to transmit system status data to the central evaluation network. The central evaluation network extracts features to calculate evaluation indicators for resource regulation effectiveness, environmental impact, and economic benefit level. It uses an attention mechanism to highlight the contribution of key indicators and generate a comprehensive score.

[0030] Based on the comprehensive score, the parameter search space of each decision-making unit is expanded and the weight of the constraints is adjusted. A coordination factor is introduced to balance the optimization objective. The adjustment results are transmitted to the decision-making unit. A Monte Carlo probability search method is used to generate a verification scenario. Typical failure modes are identified through cluster analysis and weak links are optimized to obtain the verification results. Based on the verification results, water allocation coefficients, water quality control thresholds and system regulation parameters are selected. Dynamic adjustment rules for the transport strategy are constructed, and the optimal transport parameter scheme is output.

[0031] In one alternative implementation,

[0032] The adjustment results are transmitted to the decision-making unit, and a verification scenario is generated using the Monte Carlo probability search method. Typical failure modes are identified through cluster analysis, and weak links are optimized to obtain the verification results, including:

[0033] A tiered adjustment strategy is constructed based on the evaluation results of the central evaluation network. Adjustment results are generated for decision units in different score ranges. The adjustment results are transmitted to the decision units through a multi-level caching structure. Features are extracted and weak links are identified using the Monte Carlo probabilistic search method. A protection mechanism is constructed for the weak links, and evaluation results are generated. Parameter configurations are dynamically adjusted based on the evaluation results. Specifically, this includes:

[0034] Based on the evaluation results of the central evaluation network, a hierarchical adjustment strategy is constructed. By expanding the parameter search space and increasing the constraint weights, appropriately expanding the search range and fine-tuning the constraint weights, and maintaining the existing parameter configuration, the decision-making units in different scoring ranges are differentiated and adjustment results are generated.

[0035] A multi-level caching structure is constructed to transmit the adjustment results to the decision-making unit. Historical adjustment sequences are stored in the local cache, and parameter information of neighboring units is aggregated through regional cache nodes. A differential update mechanism is used to transmit data, and data integrity is ensured through validity verification and timeout retransmission.

[0036] The Monte Carlo probabilistic search method is used to extract the probability distribution features of the adjustment results. Validation samples are obtained through a hierarchical quantization adaptive sampling strategy. After orthogonalization of the samples to eliminate parameter correlation, a validation scenario is generated. The validation results of the validation scenario are processed by cluster analysis. Cluster centers are determined through dynamic neighborhood search to identify typical failure modes. A fault tree model is constructed to track weak links. A hierarchical early warning mechanism is established through system response feature analysis.

[0037] To address the weak points, a spatial protection circle is constructed by adjusting buffer parameters, a time response mechanism is built by configuring the control cycle, a collaborative optimization network is established to achieve regional linkage, the optimization effect is evaluated through incremental verification to obtain verification results, the verification results are evaluated at multiple levels, the compliance rate of monitoring indicators is determined through statistical analysis, the spatial distribution characteristics are evaluated through the coefficient of variation, the system stability is analyzed through disturbance response, and evaluation results are generated.

[0038] Based on the evaluation results, key indicators are calculated using a sliding time window. The calculated key indicators are then input into the hierarchical adjustment strategy to dynamically adjust the parameter search range and constraint weights of the decision-making unit, and output the optimal parameter configuration.

[0039] In one alternative implementation,

[0040] Edge computing control nodes are deployed at the contaminated site. A distributed model parameter aggregation algorithm is used to fuse monitoring data from various areas, generating a global transport control model. This model is then trained with the optimal transport parameter scheme to obtain a lightweight on-site control model. The lightweight on-site control model calculates the optimal transport control scheme based on real-time monitoring data and sends it to the injection device for execution, including:

[0041] Edge computing control nodes are deployed at the contaminated site. The edge computing control nodes acquire water level monitoring data, water quality monitoring data, and flow monitoring data through a data acquisition module. The time-series feature extraction method is used to analyze the variation patterns of the water level monitoring data, the water quality monitoring data, and the flow monitoring data to obtain the sampling period.

[0042] The edge computing control node uses an anomaly detection algorithm based on differential thresholds to identify outliers, uses local linear interpolation to repair missing data, and calculates time-series features through a sliding window to obtain feature vectors.

[0043] A distributed model parameter aggregation algorithm is used to construct a transport parameter sub-model at the edge computing control node. A node affinity matrix is ​​constructed based on the geographic location relationship. The transport parameter sub-model is iteratively aggregated according to the node affinity matrix to form a regional model. An adaptive weighting method is used to merge the regional models to generate a global transport control model.

[0044] The global transport control model is densified by a nested mesh structure. The computational task is distributed to the edge computing control node for parallel processing by a block-based iterative method. The optimal transport parameter scheme is obtained by solving in a hierarchical progressive manner.

[0045] The key features of the global transport control model are extracted by knowledge mapping, and a lightweight field control model is constructed by progressive compression strategy. The lightweight field control model receives the water level monitoring data, the water quality monitoring data and the flow monitoring data, determines the feasible solution space by piecewise optimization strategy, and generates the optimal transport control scheme by gradient search method and heuristic algorithm.

[0046] The optimal transportation control scheme is sent to the injection device for execution. The injection parameters are adjusted using a feedback correction mechanism, and potential faults are identified and the control strategy is automatically switched using a state prediction algorithm.

[0047] In one alternative implementation,

[0048] A distributed model parameter aggregation algorithm is used to construct a transport parameter sub-model at the edge computing control node. A node affinity matrix is ​​constructed based on geographic location relationships. The transport parameter sub-model is then iteratively aggregated using the node affinity matrix to form a regional model, including:

[0049] A distributed model parameter aggregation algorithm is used to construct transport parameter sub-models at edge computing control nodes. Feature sequences are generated from monitoring data through multi-scale analysis and filtering. A node affinity matrix is ​​constructed based on spatial correlation coefficients, hydrogeological similarity coefficients, and migration flux coefficients between nodes, calculated according to geographical location relationships. The transport parameter sub-models are then iteratively aggregated using the node affinity matrix to form a regional model. Specifically, this includes:

[0050] A distributed model parameter aggregation algorithm is used to construct a transport parameter sub-model at the edge computing control node. The distributed model parameter aggregation algorithm performs multi-scale analysis on water level monitoring data, water quality monitoring data and flow monitoring data through wavelet decomposition, extracts high-frequency and low-frequency feature components, sets an adaptive threshold to remove abnormal fluctuations based on the energy distribution of feature components, and uses recursive filtering to smooth the signal and generate a stable monitoring data sequence.

[0051] Spatiotemporal features are extracted from the stable monitoring data sequence, time-series statistics are calculated using a sliding window, data distribution probability is calculated using kernel density estimation, data mutation locations are determined based on change point detection, and a comprehensive feature vector is constructed by combining a multi-source data fusion algorithm.

[0052] The comprehensive feature vector is input into the gradient descent optimizer to calculate the gradient of the model parameters. The cross-validation method is used to calculate the parameter importance score. Iterative optimization is performed on the high-scoring parameters, and linear regression is performed on the low-scoring parameters to construct the transport parameter sub-model and generate the initial parameter field.

[0053] A node affinity matrix is ​​constructed based on geographical location relationships. The spatial correlation coefficient is obtained by calculating the Euclidean distance between nodes. The hydrogeological similarity coefficient is calculated using a hierarchical clustering method. The migration flux coefficient is calculated based on the pollutant concentration gradient. The affinity weight is calculated by normalizing and weighting the spatial correlation coefficient, the hydrogeological similarity coefficient, and the migration flux coefficient.

[0054] The transport parameter sub-model is iteratively aggregated based on the node affinity matrix to form a regional model. Nodes with high affinity are preferentially fused. The initial parameter field is weighted and averaged using the affinity weights. The fusion residual is calculated and the affinity weights are adjusted to update the node affinity matrix.

[0055] The region model is divided into multiple computational subtasks and assigned to the edge computing control nodes. A message queue is established between nodes to transmit data. Checkpoints are set according to the fusion residuals. The node load is allocated according to the computational load of the initial parameter field.

[0056] A second aspect of this invention provides a precise delivery control system for remediation agents in in-situ groundwater remediation, comprising:

[0057] The first unit is used to collect image data and sensor measurement data corresponding to borehole cores of contaminated sites. The image data is input into a visual attention neural network for feature extraction to obtain soil structure feature data. The sensor measurement data is input into a temporal attention neural network for analysis to obtain groundwater dynamic feature data. The soil structure feature data and groundwater dynamic feature data are added to a conditional probability diffusion model for feature fusion to generate hydrogeological feature vectors. These vectors are then mapped to a three-dimensional spatial grid through a spatial attention calculation module to obtain an initial hydrogeological feature model. Finally, a generative probability diffusion model is used to complete the missing areas to obtain a high-precision hydrogeological feature model.

[0058] The second unit is used to input the high-precision hydrogeological feature model into a multi-regional collaborative simulation environment. Multiple decision units perform local optimization calculations on different spatial regions. The decision units share optimization results through a communication network. The central evaluation network performs a global evaluation based on the shared results and coordinates the adjustment of the optimization parameters of each decision unit. The adjusted transport strategy is verified by combining the Monte Carlo probability search method, and the optimal transport parameter scheme is output.

[0059] The third unit is used to deploy edge computing control nodes at the contaminated site. It uses a distributed model parameter aggregation algorithm to fuse monitoring data from each area, generate a global transport control model, and trains it with the optimal transport parameter scheme to obtain a lightweight field control model. The lightweight field control model calculates the optimal transport control scheme based on real-time monitoring data and sends it to the injection device for execution.

[0060] This invention innovatively combines visual attention networks and temporal attention networks to accurately extract the dynamic characteristics of soil structure and groundwater. Employing a multi-regional collaborative simulation environment and a distributed decision-making unit architecture significantly enhances the adaptability of the optimization strategy to site heterogeneity. A distributed model parameter aggregation algorithm enables efficient fusion of multi-source monitoring data. A real-time optimization mechanism based on a lightweight field control model allows the control scheme to be rapidly adjusted according to site conditions. Through a multi-layered control architecture design, the stability of system operation is significantly improved. The dual attention mechanism at the feature extraction level provides a reliable data foundation, while the multi-regional collaborative mechanism at the optimization level ensures the balance of the optimization strategy. This invention provides an efficient, reliable, and precise transport control method for in-situ groundwater remediation, possessing significant engineering application value and promotional significance. Attached Figure Description

[0061] Figure 1 This is a flowchart illustrating the precise transport and control method for remediation agents in in-situ groundwater remediation according to an embodiment of the present invention.

[0062] Figure 2 This is a comparison diagram of the feature fusion effect of the precise transport and control method for remediation agents in in-situ groundwater remediation according to an embodiment of the present invention;

[0063] Figure 3 This is a comparison chart showing the system optimization effect of the precise transport control method for remediation agents in in-situ groundwater remediation according to an embodiment of the present invention;

[0064] Figure 4 This is a comparison chart showing the data quality improvement effect of the precise transport and control method for remediation agents in in-situ groundwater remediation according to an embodiment of the present invention. Detailed Implementation

[0065] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0067] Figure 1 This is a flowchart illustrating the precise transport and control method for remediation agents in in-situ groundwater remediation according to an embodiment of the present invention. Figure 1As shown, the method includes:

[0068] Image data and sensor measurement data corresponding to borehole cores of contaminated sites are collected. The image data is input into a visual attention neural network for feature extraction to obtain soil structure feature data. The sensor measurement data is input into a temporal attention neural network for analysis to obtain groundwater dynamic feature data. The soil structure feature data and groundwater dynamic feature data are added to a conditional probability diffusion model for feature fusion to generate hydrogeological feature vectors. These vectors are then mapped to a three-dimensional spatial grid through a spatial attention calculation module to obtain an initial hydrogeological feature model. Finally, a generative probability diffusion model is used to complete the missing areas to obtain a high-precision hydrogeological feature model.

[0069] In one alternative implementation,

[0070] Image data and sensor measurement data corresponding to borehole cores from contaminated sites are collected. The image data is input into a visual attention neural network for feature extraction to obtain soil structure feature data. The sensor measurement data is input into a temporal attention neural network for analysis to obtain groundwater dynamic feature data. The soil structure feature data and groundwater dynamic feature data are added to a conditional probability diffusion model for feature fusion, generating hydrogeological feature vectors. These vectors are then mapped to a 3D spatial grid using a spatial attention calculation module to obtain an initial hydrogeological feature model. Finally, a generative probability diffusion model is used to complete the model for missing areas, resulting in a high-precision hydrogeological feature model, including:

[0071] A borehole sampling system was deployed at the contaminated site to collect core image data and groundwater sensor measurement data.

[0072] The core image data is input into an adaptive histogram equalization processing module to eliminate illumination interference and obtain a first processed image. The first processed image is then processed by local contrast sharpening to enhance structural details and obtain a second processed image. The second processed image is input into a visual attention neural network. The encoder of the visual attention neural network performs convolution operations to extract features, and the decoder reconstructs the soil structure feature data.

[0073] The groundwater sensor measurement data is input into the outlier detection module to remove outliers and obtain the first processed data. The first processed data is then denoised using an adaptive sliding window to obtain the second processed data. The second processed data is then input into a temporal attention neural network, and the temporal correlation is analyzed using a bidirectional long short-term memory unit to obtain the groundwater dynamic feature data.

[0074] The soil structure feature data and the groundwater dynamic feature data are concatenated and input into the conditional probability diffusion model. The hydrogeological feature vector is obtained through dynamic noise reduction iteration. The hydrogeological feature vector is input into the spatial attention calculation module. The spatial weight is calculated based on the feature similarity and the grid is densified. The initial hydrogeological feature model is obtained through feature interpolation.

[0075] The initial hydrogeological feature model is input into the generative probability diffusion model. Based on the known regional feature distribution, feature generation is performed on the missing region. A high-precision hydrogeological feature model is obtained by evaluating and optimizing it through a multi-scale discriminator.

[0076] First, a borehole sampling system was deployed inside the contaminated site. This system includes a core sampling device and a groundwater sensor system. The core sampling device is equipped with high-resolution imaging equipment, which can continuously acquire core image data. The groundwater sensor system integrates multiple parameter sensors, which can monitor key parameters such as groundwater level, temperature, and conductivity in real time.

[0077] The acquired core image data is preprocessed by an adaptive histogram equalization module. This module calculates the cumulative distribution function of local image regions and adaptively adjusts the pixel distribution to effectively eliminate interference from uneven illumination, resulting in a first processed image with balanced brightness distribution. Local contrast sharpening is then applied to the first processed image. By calculating gradient information in local image regions, structural details such as edges and textures are enhanced, generating a second processed image with clearer details. The processed second image is then input into a pre-trained visual attention neural network. The network's encoder extracts image features step-by-step through multi-layer convolution operations, with an attention mechanism highlighting key region features. The decoder reconstructs soil structural feature data, including key parameters such as porosity and permeability, through deconvolution operations.

[0078] The measurement data collected by the sensors is preprocessed by the outlier detection module. Data points that significantly deviate from the normal range are identified and removed using statistical methods, generating the first processed data. An adaptive sliding window algorithm is then applied to the first processed data for noise reduction; the window size dynamically adjusts according to the data fluctuation, resulting in smoother second processed data. This second processed data is then input into a temporal attention neural network. The bidirectional long short-term memory units in the network analyze the forward and backward temporal correlations of the data, and, combined with the attention mechanism, capture key temporal features, ultimately obtaining data characterizing the dynamic changes in groundwater.

[0079] After obtaining soil structure characteristic data and groundwater dynamic characteristic data, the two types of data are concatenated according to a preset format and used as input for a conditional probability diffusion model. The conditional probability diffusion model gradually extracts and fuses the correlation information between the two types of features through a multi-step denoising iterative process, generating a comprehensive hydrogeological feature vector. The obtained feature vector is input into a spatial attention calculation module to calculate the similarity of features at different locations in the spatial domain, generating a spatial weight matrix. Based on the generated weight matrix, a grid refinement process is guided, with grid subdivision performed in key areas. Grid node values ​​are then filled using feature interpolation to construct an initial hydrogeological feature model.

[0080] An initial hydrogeological feature model is input into a generative probabilistic diffusion model to analyze the feature distribution patterns of known areas. Based on the analyzed patterns, features are generated for missing areas in the model through an iterative diffusion process. During the generation process, a multi-scale discriminator is used to evaluate the rationality of the generated features at different scales and returns an optimization signal. By repeatedly executing the generation-evaluation-optimization process, a complete and high-precision hydrogeological feature model is obtained.

[0081] For example, taking a contaminated site remediation project as an example, 25 borehole sampling points were set up within a 100m × 100m area. A 12-megapixel industrial camera was used in each borehole to capture one core image every 10cm, with an image resolution of 4000 × 3000 pixels. Adaptive histogram equalization was used to adjust the image brightness range from the original [50, 180] to [30, 220], improving local contrast by 30%. The soil structural features extracted by the visual attention network included: average porosity of 0.35 and permeability coefficient of 1.2 × 10⁻⁻⁻⁶. 5 For sensor data acquisition, water level, temperature, and conductivity data are collected hourly. Outlier detection removes data points with deviations exceeding three standard deviations (approximately 0.5%). An adaptive window of 15 data points is used for noise reduction, and temporal attention network analysis yields dynamic features such as groundwater flow velocity (0.5 m / d) and flow direction (45°). In the feature fusion and modeling stage, the extracted features are organized into 256-dimensional vectors, and fused features are obtained through 1000 steps of noise reduction iterations. Spatial attention calculations show that the feature similarity within a 20m radius exceeds 0.8; therefore, the original 1m grid is densified to 0.2m in high-similarity areas. The generative model completes approximately 20% of the missing areas in five locations. A multi-scale discriminator evaluates the generation quality at three scales: 0.2m, 1m, and 5m, achieving an average accuracy of 92%. The final hydrogeological feature model contains 500×500×50 grid nodes, each node containing 25 feature parameters, which can be used to guide subsequent remediation scheme optimization.

[0082] In this embodiment, a technical solution combining multi-source data processing and deep learning effectively solves the problem of inaccurate hydrogeological feature representation during in-situ groundwater remediation. Existing technologies typically employ a single data source for geological feature analysis in in-situ groundwater remediation, such as relying solely on borehole core images for soil structure analysis or solely on sensor data to study groundwater dynamics, which fails to comprehensively reflect the site's hydrogeological characteristics. Regarding data processing, traditional methods often employ image enhancement algorithms with fixed parameters and simple data filtering methods, which cannot adapt to the differences in data quality under different site conditions. Feature extraction mainly relies on manually designed feature extraction rules, resulting in features that are often one-sided and highly subjective. To address these issues, this embodiment simultaneously deploys core image acquisition and groundwater parameter monitoring systems at the data acquisition end to obtain multi-source data. In the data processing stage, adaptive histogram equalization and local contrast enhancement techniques were introduced to achieve dynamic optimization of image quality. Outlier detection and adaptive sliding window algorithms were adopted to improve the reliability of sensor data. Visual attention neural networks and temporal attention neural networks were innovatively applied to extract features from spatial and temporal dimensions, respectively, overcoming the limitations of traditional feature extraction methods. Deep fusion of multi-source features was achieved through a conditional probability diffusion model, and a generative probability diffusion model was used to intelligently complete missing regions, significantly improving the completeness and accuracy of the model.

[0083] In one alternative implementation,

[0084] The soil structure feature data and the groundwater dynamic feature data are concatenated and input into a conditional probability diffusion model. Through dynamic noise reduction iteration, a hydrogeological feature vector is obtained, including:

[0085] Soil structural feature data and groundwater dynamic feature data are concatenated to obtain a hybrid feature vector. This hybrid feature vector is then input into a two-branch structure of a conditional probability diffusion model. Latent features are obtained through the main branch, and constraints are obtained through the auxiliary branch. Differential noise injection and denoising processing are applied to the features. Combined with constraints and an attention mechanism, fused features are generated. Based on feature quality assessment, parameters are optimized, and the hydrogeological feature vector is reconstructed. Specifically, this includes:

[0086] Soil structural feature data and groundwater dynamic feature data are spliced ​​together using a hierarchical cascade structure. Features of the same scale are spliced ​​together to obtain a feature vector of the same scale. The feature vectors of the same scale are combined to form a hybrid feature vector and input into a conditional probability diffusion model with a dual-branch parallel structure. The main branch of the conditional probability diffusion model maps the hybrid feature vector to the latent space through a feature encoder to obtain latent features. The auxiliary branch extracts physical law constraint information, spatial continuity constraint information and temporal consistency constraint information through a constraint encoder to obtain a constraint condition vector.

[0087] Based on the hybrid feature vector, the feature importance distribution is calculated. Differential noise injection is performed on different feature regions according to the feature importance distribution. Strong noise is injected into feature regions with an importance of less than 0.3, and weak noise is injected into feature regions with an importance of more than 0.75. Noise features are obtained by superimposing multi-level noise to achieve feature degradation.

[0088] The noise features are input into a multi-cascaded noise reduction unit. The multi-cascaded noise reduction unit extracts feature information through a densely connected feature extraction module, predicts noise distribution through a noise prediction module, and achieves dynamic noise reduction by using a skip connection method to obtain noise reduction features. The constraint vector is input into a constraint decoder to generate feature constraints. The noise reduction features are then constrained and adjusted through a recursive feedback structure to obtain constraint features.

[0089] A multi-head attention module is constructed to calculate the feature correlation degree of the constraint features. Attention weights are generated based on the feature correlation degree. The constraint features are enhanced by the attention weights and residual connections are introduced to retain the original feature information to obtain fused features.

[0090] A parameter evaluation module is constructed to calculate the feature quality index and fusion effect index of the fused features. Based on the feature quality index and the fusion effect index, the noise reduction parameters are dynamically optimized to obtain optimized parameters. The optimized parameters are input into the feature decoder to decode and reconstruct the fused features to obtain the hydrogeological feature vector.

[0091] Feature stitching preprocessing is performed. The acquired soil structure feature data and groundwater dynamic feature data are organized using a hierarchical cascade structure. Features with the same resolution and scale from both types of data are stitched together to obtain multiple feature vectors of the same scale. These feature vectors are then combined according to preset rules to form a hybrid feature vector containing multi-scale information. This hybrid feature vector is input into a designed dual-branch parallel structure conditional probability diffusion model. The main branch contains a feature encoder that maps the hybrid feature vector to the latent space through a multi-layer neural network to obtain latent feature representations. The auxiliary branch contains a constraint encoder that extracts physical constraint information (such as the relationship between permeability and porosity), spatial continuity constraint information (such as the spatial gradient of features), and temporal consistency constraint information (such as the dynamic change pattern of groundwater), ultimately obtaining a constraint condition vector.

[0092] Differentiated noise processing for features. The importance distribution of each feature region is calculated based on the mixed feature vector, and a gradient-based feature importance evaluation method is used. Based on the calculated feature importance distribution, a differentiated noise injection strategy is applied to feature regions of different importance: feature regions with importance scores below 0.3 are injected with strong Gaussian noise; feature regions with importance scores above 0.75 are injected with only slight noise; and feature regions in between are injected with noise intensity linearly according to their importance. By superimposing multiple noises of different scales, multi-level feature degradation is achieved, resulting in noise features.

[0093] The noise features are input into a multi-tiered denoising unit for processing. This unit first extracts feature information using a densely connected feature extraction module with a multi-layer convolutional network, and simultaneously predicts the noise distribution in each region using a dedicated noise prediction module. Based on this, a skip connection structure is used to fuse feature information from different levels, achieving a dynamic denoising process and obtaining preliminary denoised features. Subsequently, the previously obtained constraint vector is input into a constraint decoder to generate specific feature constraints. Through a designed recursive feedback structure, the constraints and denoised features are iteratively adjusted in multiple rounds to ensure that the features satisfy the physical constraints, ultimately yielding the constrained features.

[0094] A multi-head attention module is constructed to calculate the correlation between constrained features. This module divides the features into multiple attention heads, each of which independently calculates the correlation between features and generates corresponding attention weights. The attention weights are used to selectively enhance the constrained features, while residual connections preserve the original feature information, ultimately resulting in a fully fused feature representation.

[0095] A parameter evaluation module is constructed to quantitatively evaluate the feature fusion effect using designed evaluation metrics. This module calculates feature quality metrics (such as signal-to-noise ratio and sharpness) and fusion effect metrics (such as feature consistency and spatial continuity). Based on these evaluation metrics, the gradient descent method is used to dynamically optimize the key parameters of the noise reduction process, resulting in an optimized parameter set. The optimized parameters are then input into the feature decoder to decode and reconstruct the fused features, ultimately yielding a high-quality hydrogeological feature vector.

[0096] For example, consider the feature fusion process of a contaminated site. The initial data includes 32×32×64-dimensional soil structure features and 32×32×32-dimensional groundwater dynamic features. Features with the same spatial resolution (32×32) are stitched together using a hierarchical cascade structure to obtain a 96-dimensional feature vector of the same scale. Feature importance assessment shows that approximately 20% of the feature regions have an importance below 0.3, mainly distributed in the boundary areas, where Gaussian noise with a standard deviation of 0.5 is injected; approximately 30% of the feature regions have an importance above 0.75, concentrated in the core contaminated area, where only weak noise with a standard deviation of 0.1 is injected.

[0097] During the noise reduction process, a 4-tiered noise reduction unit is used, with each tier containing 3 densely connected convolutional layers (with a kernel size of 3×3). Features at different scales are fused through skip connections. The physical constraints extracted by the constraint encoder include: the power-law relationship between permeability and porosity (R²>0.9), the continuity constraint of spatial gradient (gradient change <5%), and the consistency constraint of temporal changes (change between adjacent time steps <10%).

[0098] The multi-head attention module employs eight attention heads, each independently calculating the correlation matrix for 32×32 spatial locations. Through the attention mechanism, the weights of approximately 90% of the feature locations are enhanced to varying degrees, with an average enhancement of 35%. Final parameter evaluations show that the signal-to-noise ratio (SNR) is improved by 40% and the sharpness by 25% in the feature quality metrics; the feature consistency in the fusion performance metrics reaches 0.85, and the spatial continuity error is less than 3%. After reconstruction by the feature decoder, a 32×32×128 hydrogeological feature vector is obtained, which retains the key information of the original features while satisfying the physical constraints.

[0099] In this embodiment, through innovative model structure design and processing strategies, the shortcomings of traditional methods in feature fusion, noise processing, and constraint introduction are effectively addressed. This provides more reliable technical support for the hydrogeological feature characterization in the in-situ groundwater remediation process, promoting the advancement of technology in this field. Traditional hydrogeological feature extraction methods typically employ simple feature splicing or weighted averaging methods during feature fusion, which are insufficient to effectively handle the fusion of multi-source heterogeneous data. In the feature optimization stage, uniform denoising parameters and processing strategies are often used, failing to consider the differences in importance among different feature regions. Regarding constraint introduction, simple penalty or regularization terms are often used, making it difficult to fully utilize known physical laws and spatiotemporal constraints. In the feature enhancement stage, commonly used feature selection methods lack in-depth analysis of the correlation between features, easily leading to the loss of useful information. These problems result in insufficient accuracy of the final hydrogeological feature model, making it difficult to meet the refined requirements of in-situ groundwater remediation.

[0100] This embodiment proposes a feature fusion method based on a conditional probability diffusion model. It employs a hierarchical cascaded structure for feature organization, laying the foundation for subsequent deep fusion through precise alignment and combination of features at the same scale. In terms of model architecture, an innovative dual-branch parallel structure is designed. The main branch is responsible for feature extraction and mapping, while the auxiliary branch focuses on extracting multi-dimensional constraint information, achieving an organic combination of feature extraction and constraint introduction. An importance-based differentiated noise injection strategy is introduced. By accurately calculating the feature importance distribution, differentiated noise processing is implemented for different regions, effectively protecting key feature information. Simultaneously, a multi-cascaded denoising unit is designed. Through a structure design of dense connections and skip connections, multi-scale feature extraction and dynamic denoising are achieved, significantly improving feature quality.

[0101] The high-precision hydrogeological feature model is input into a multi-regional collaborative simulation environment. Multiple decision-making units perform local optimization calculations on different spatial regions. The decision-making units share optimization results through a communication network. The central evaluation network performs a global evaluation based on the shared results and coordinates the adjustment of optimization parameters of each decision-making unit. The adjusted transport strategy is verified by combining the Monte Carlo probability search method, and the optimal transport parameter scheme is output.

[0102] In one alternative implementation,

[0103] The high-precision hydrogeological feature model is input into a multi-regional collaborative simulation environment. Multiple decision-making units perform local optimization calculations for different spatial regions. These units share optimization results via a communication network. A central evaluation network performs a global evaluation based on the shared results and collaboratively adjusts the optimization parameters of each decision-making unit. The adjusted transport strategy is verified using a Monte Carlo probabilistic search method, and the optimal transport parameter scheme is output, including:

[0104] The high-precision hydrogeological feature model is divided into multiple spatial regions according to the boundaries of hydrogeological units and management zones. Decision units with data layers, computing layers and communication layers are configured in each spatial region.

[0105] The data layer of the decision-making unit extracts the geological structure characteristics, hydrological parameter distribution and boundary condition information of the spatial region, establishes a groundwater flow field numerical model, collects groundwater level, water quality index and extraction volume data to construct a dynamic database, and the calculation layer of the decision-making unit inputs the state data in the dynamic database into the optimizer, performs local optimization calculation based on the optimization objectives of water resource balance, water quality safety and system stability to obtain the optimization result;

[0106] The decision-making units interconnect through a communication network to build a hierarchical caching mechanism to store the optimization results. They adopt an incremental update strategy to transmit system status data to the central evaluation network. The central evaluation network extracts features to calculate evaluation indicators for resource regulation effectiveness, environmental impact, and economic benefit level. It uses an attention mechanism to highlight the contribution of key indicators and generate a comprehensive score.

[0107] Based on the comprehensive score, the parameter search space of each decision-making unit is expanded and the weight of the constraints is adjusted. A coordination factor is introduced to balance the optimization objective. The adjustment results are transmitted to the decision-making unit. A Monte Carlo probability search method is used to generate a verification scenario. Typical failure modes are identified through cluster analysis and weak links are optimized to obtain the verification results. Based on the verification results, water allocation coefficients, water quality control thresholds and system regulation parameters are selected. Dynamic adjustment rules for the transport strategy are constructed, and the optimal transport parameter scheme is output.

[0108] A high-precision hydrogeological feature model was used to divide the site into regions. Based on the boundary characteristics of the hydrogeological units and management zoning requirements, the entire site was divided into several relatively independent spatial regions. Within each divided spatial region, a decision-making unit with a three-layer architecture was configured, including a data layer, a computing layer, and a communication layer, forming a distributed decision-making network.

[0109] At the data layer of the decision-making unit, geological structural feature data of the spatial region are systematically extracted, including stratigraphic distribution and fault structures; simultaneously, hydrological parameter distribution information, such as permeability coefficient and storage coefficient, is acquired; and boundary condition information of the region, including head boundary and flow boundary, is determined. Based on these fundamental data, a numerical model of the groundwater flow field in the region is established. Simultaneously, dynamic data on groundwater level, water quality indicators (such as pH value, conductivity, and pollutant concentration), and groundwater extraction volume are continuously collected to construct a real-time updated dynamic database.

[0110] At the computational layer, the decision-making unit inputs the state data from the dynamic database into the optimizer for processing. The optimizer performs calculations based on three core optimization objectives: water resource balance to ensure regional water quantity balance, water quality safety to ensure water quality meets standards, and system stability to maintain the dynamic balance of the groundwater system. Through a multi-objective optimization algorithm, local optimization results that satisfy the current regional characteristics are obtained.

[0111] The decision-making units are interconnected via a communication network, and a hierarchical caching mechanism is established to store optimization results. An incremental update strategy is adopted to periodically transmit system status data to the central evaluation network. The central evaluation network is responsible for extracting features and calculating multiple evaluation indicators: the resource regulation effectiveness indicator reflects water resource utilization efficiency, the environmental impact indicator assesses ecological and environmental impact, and the economic benefit level indicator measures economic rationality. These indicators are weighted using an attention mechanism to highlight the contribution of key indicators, ultimately generating a comprehensive score.

[0112] Based on the comprehensive scoring results, the system adjusts the optimization process of each decision-making unit: expanding the parameter search space to find better solutions, adjusting the constraint weights to balance multiple objectives, and introducing coordination factors to ensure the coordination of decisions across regions. These adjustments are transmitted back to each decision-making unit, and then multiple verification scenarios are generated using the Monte Carlo probabilistic search method. Typical failure modes are identified through cluster analysis, and weak links in the system are optimized in a targeted manner to obtain verification results. Finally, based on the verification results, key control parameters are determined, including water allocation coefficients, water quality control thresholds, and system regulation parameters. Dynamic adjustment rules for the transport strategy are constructed, and the optimal transport parameter scheme is ultimately output.

[0113] For example, taking a contaminated site remediation project as an example, the site area is 1 square kilometer, which is divided into 5 spatial areas according to hydrogeological conditions, and each area is equipped with a decision-making unit.

[0114] At the data layer, taking decision unit 1 as an example: the extracted geological structural features show that the area is mainly composed of gravel aquifers with an average thickness of 15 meters; hydrological parameters include a permeability coefficient of 1.2 × 10⁻⁻⁻⁶. 4 The groundwater flow field model has a flow rate of m / s and a storage coefficient of 0.15. Boundary conditions are a constant head boundary (20 meters) on the west side and a constant flow rate boundary (100 m³ / d) on the east side. The established groundwater flow field model has a grid size of 5 meters × 5 meters and a time step of 1 day. The dynamic database is updated hourly and includes water level data from 10 monitoring wells, water quality data (COD, ammonia nitrogen, etc.) from 5 sampling points, and extraction volume data from 3 pumping wells.

[0115] At the computational layer, the optimizer's objective function weights are set as follows: water resource balance 0.4, water quality safety 0.4, and system stability 0.2. Local optimization results show that the optimal pumping rate for this region is 150 m³ / d, and the injection rate is 120 m³ / d.

[0116] The communication layer employs a three-level caching mechanism: a local cache stores nearly 24 hours of data, a regional cache stores nearly 7 days of data, and a central cache stores nearly 30 days of data. Status data is incrementally updated and transmitted to the central evaluation network every 30 minutes. The central evaluation network calculates the following index values: resource regulation effectiveness 0.85 (out of 1.0), environmental impact degree 0.15 (out of 0.3), and economic benefit level 0.75 (out of 1.0). The overall score calculated using the attention mechanism is 0.82.

[0117] Based on this score, the system expanded the parameter search range of decision unit 1 by 20%, and adjusted the constraint weights to: water quality constraint 0.5, water quantity constraint 0.3, and cost constraint 0.2. A Monte Carlo method was used to generate 1000 verification scenarios, and cluster analysis identified three typical failure modes: water level exceeding limits, pollutant concentration exceeding standards, and system response delay. To address these issues, the optimized and adjusted key parameters included: water distribution coefficient 0.8, COD control threshold 40 mg / L, and system response time threshold 4 hours. The final output transport parameter scheme included: injection well spacing of 30 meters, injection pressure of 0.2 MPa, and injection cycle of 6 hours.

[0118] In this embodiment, the distributed decision-making architecture significantly improves the system's data processing speed and optimization efficiency, reduces response latency, and enables rapid response to changes in site conditions. The multi-objective collaborative optimization framework makes water resource regulation more balanced, significantly improves water quality compliance rate, makes system operation more stable, and significantly improves overall remediation efficiency. The probabilistic verification method significantly improves the adaptability and robustness of the optimization scheme, significantly reduces system failure rate, and greatly reduces maintenance costs. Existing groundwater in-situ remediation transport control technologies typically adopt a centralized architecture in decision system design, where data and computational tasks from all areas are concentrated in a single control center, resulting in slow system response speed and high computational load. In terms of optimization objective setting, there is often an overemphasis on optimizing a single indicator, such as only considering water balance or water quality compliance, lacking a systematic consideration of multi-objective collaborative optimization. In the decision verification stage, deterministic verification methods are usually used, which are difficult to cope with uncertainties in actual sites. In terms of parameter adjustment, fixed control rules are often used, lacking dynamic adaptability and making it difficult to respond to changes in site conditions in a timely manner. These problems seriously restrict the effectiveness of groundwater in-situ remediation.

[0119] This embodiment employs a distributed decision-making architecture, dividing the site into multiple regions based on hydrogeological unit characteristics. Each region is configured with an independent three-layer architecture decision-making unit, significantly improving the system's parallel processing capability and response speed while reducing the computational load of individual decision-making units, thus constructing a multi-layered data acquisition and processing system. The data layer system of the decision-making units collects information such as geological structure, hydrological parameters, and boundary conditions to establish a refined groundwater flow field model, and implements real-time data updates through a dynamic database. This design provides comprehensive and reliable data support for subsequent optimization decisions, and a multi-objective collaborative optimization framework is designed. By simultaneously considering the three core objectives of water resource balance, water quality safety, and system stability at the computational layer, comprehensive optimization of the remediation process is achieved.

[0120] In one alternative implementation,

[0121] The adjustment results are transmitted to the decision-making unit, and a verification scenario is generated using the Monte Carlo probability search method. Typical failure modes are identified through cluster analysis, and weak links are optimized to obtain the verification results, including:

[0122] A tiered adjustment strategy is constructed based on the evaluation results of the central evaluation network. Adjustment results are generated for decision units in different score ranges. The adjustment results are transmitted to the decision units through a multi-level caching structure. Features are extracted and weak links are identified using the Monte Carlo probabilistic search method. A protection mechanism is constructed for the weak links, and evaluation results are generated. Parameter configurations are dynamically adjusted based on the evaluation results. Specifically, this includes:

[0123] Based on the evaluation results of the central evaluation network, a hierarchical adjustment strategy is constructed. By expanding the parameter search space and increasing the constraint weights, appropriately expanding the search range and fine-tuning the constraint weights, and maintaining the existing parameter configuration, the decision-making units in different scoring ranges are differentiated and adjustment results are generated.

[0124] A multi-level caching structure is constructed to transmit the adjustment results to the decision-making unit. Historical adjustment sequences are stored in the local cache, and parameter information of neighboring units is aggregated through regional cache nodes. A differential update mechanism is used to transmit data, and data integrity is ensured through validity verification and timeout retransmission.

[0125] The Monte Carlo probabilistic search method is used to extract the probability distribution features of the adjustment results. Validation samples are obtained through a hierarchical quantization adaptive sampling strategy. After orthogonalization of the samples to eliminate parameter correlation, a validation scenario is generated. The validation results of the validation scenario are processed by cluster analysis. Cluster centers are determined through dynamic neighborhood search to identify typical failure modes. A fault tree model is constructed to track weak links. A hierarchical early warning mechanism is established through system response feature analysis.

[0126] To address the weak points, a spatial protection circle is constructed by adjusting buffer parameters, a time response mechanism is built by configuring the control cycle, a collaborative optimization network is established to achieve regional linkage, the optimization effect is evaluated through incremental verification to obtain verification results, the verification results are evaluated at multiple levels, the compliance rate of monitoring indicators is determined through statistical analysis, the spatial distribution characteristics are evaluated through the coefficient of variation, the system stability is analyzed through disturbance response, and evaluation results are generated.

[0127] Based on the evaluation results, key indicators are calculated using a sliding time window. The calculated key indicators are then input into the hierarchical adjustment strategy to dynamically adjust the parameter search range and constraint weights of the decision-making unit, and output the optimal parameter configuration.

[0128] Based on the evaluation results of the central evaluation network, a tiered adjustment strategy is constructed, employing differentiated adjustment methods for decision-making units in different score ranges: for decision-making units with low scores, a significant adjustment is made by expanding the parameter search space and increasing constraint weights; for decision-making units with medium scores, a moderate adjustment is made by appropriately expanding the search range and fine-tuning the constraint weights; and for decision-making units with high scores, the current parameter configuration is maintained. This tiered adjustment strategy generates targeted adjustment results.

[0129] To ensure efficient transmission and storage of adjustment results, a multi-level caching structure is constructed. The historical adjustment sequences of each decision-making unit are stored at the local cache level to track parameter change trends. Parameter information from neighboring decision-making units is aggregated at regional cache nodes to facilitate regional collaborative optimization. Data transmission employs a differential update mechanism, transmitting only changed parameter information, while validity checks and timeout retransmission mechanisms ensure the integrity and reliability of data transmission.

[0130] The Monte Carlo probabilistic search method is used to conduct in-depth analysis of the adjustment results, extracting the probability distribution characteristics of the parameters. Validation samples are obtained through a hierarchical quantization adaptive sampling strategy, with the sampling density dynamically adjusted according to parameter importance. The obtained samples are orthogonalized to eliminate correlations between parameters, generating independent validation scenarios. Cluster analysis is used to process the validation scenario results, determining cluster centers through dynamic neighborhood search, identifying typical failure modes of the system, constructing a fault tree model, systematically tracking and analyzing each weak link, and establishing a hierarchical early warning mechanism by analyzing system response characteristics.

[0131] To address the identified weaknesses, multi-level optimization measures were implemented: a spatial protection zone was constructed by adjusting buffer parameters to prevent the spread of adverse effects; a time response mechanism was established by configuring the control cycle to ensure the system could respond to changes in a timely manner; and a collaborative optimization network was established to achieve regional linkage and improve the overall control effect. The optimization effect was evaluated through incremental verification to obtain verification results. The verification results were then evaluated at multiple levels: the compliance rate of monitoring indicators was determined through statistical analysis to assess the optimization effect; the spatial distribution characteristics of parameters were assessed through the coefficient of variation to verify the uniformity of control; and system stability was evaluated through disturbance response analysis to verify the reliability of the optimization scheme.

[0132] Based on the evaluation results, a sliding time window method is used to calculate key indicators, such as compliance rate, uniformity, and stability. These indicators are then input into a tiered adjustment strategy to dynamically adjust the parameter search range and constraint weights of the decision-making unit, ultimately outputting the optimal parameter configuration scheme.

[0133] For example, taking a contaminated site remediation project as an example, the site comprises 5 decision-making units. According to the assessment results of the central assessment network, Unit 1 scored below the threshold, Units 2 and 3 scored moderately, and Units 4 and 5 scored highly. A significant adjustment strategy was adopted for Unit 1: the parameter search space was doubled, and the water quality constraint weight was increased from 0.3 to 0.5; a moderate adjustment was adopted for Units 2 and 3: the search range was expanded by 20%, and the constraint weights were fine-tuned; Units 4 and 5 maintained their existing configuration. In the multi-level caching structure, the local cache stores the 100 most recent adjustment records, and the regional cache nodes summarize the parameter information of the five adjacent units. Data transmission uses differential updates every 30 minutes, achieving a transmission success rate of 99.9%.

[0134] The Monte Carlo method generated 1000 validation scenarios, and cluster analysis identified three typical failure modes: water quality exceeding standards, abnormal water levels, and system response delay. The constructed fault tree showed that the main weaknesses included: injection pressure control, water quality monitoring frequency, and system response time. To address these weaknesses, a 50-meter spatial protection zone was established, with a control cycle set at 4 hours, and a collaborative optimization network covering adjacent units was created. Incremental validation and evaluation showed that the compliance rate of monitoring indicators significantly improved, the spatial distribution became more uniform, and the system's response to disturbances remained stable. A 24-hour sliding time window was used to calculate key indicators, and parameter configurations were dynamically optimized based on indicator changes to determine optimal parameters, including injection pressure, injection cycle, and monitoring frequency.

[0135] In this embodiment, a hierarchical adjustment strategy is constructed, and differentiated processing is implemented for different decision-making units based on the evaluation results. Units with poor performance are subject to significant adjustments, those with average performance are subject to moderate adjustments, and those with excellent performance are kept stable. This differentiated adjustment approach significantly improves the accuracy and efficiency of optimization. A multi-level caching structure is designed, storing historical adjustment sequences in a local cache, aggregating information from neighboring units using regional cache nodes, and transmitting data using a differential update mechanism. This hierarchical data management approach significantly improves the system's data processing efficiency and reliability. Validation samples are obtained through a hierarchical quantification adaptive sampling strategy, orthogonalization is used to eliminate parameter correlations, and cluster analysis is combined to identify typical failure modes, significantly improving the comprehensiveness and reliability of the validation results.

[0136] Some in-situ groundwater remediation and optimization control technologies often employ uniform adjustment methods for parameters, failing to implement differentiated treatment based on the characteristics of different regions, resulting in unsatisfactory optimization effects. Regarding data transmission and storage, they frequently utilize simple centralized storage structures, easily creating data transmission bottlenecks and storage pressure. In the scheme verification phase, deterministic verification methods are commonly used, making it difficult to comprehensively evaluate the system's performance under complex conditions. In system optimization, there is a lack of systematic analysis and targeted optimization of weak links, leading to insufficient system reliability. These problems seriously affect the overall effectiveness of remediation projects.

[0137] Edge computing control nodes are deployed at the contaminated site. The monitoring data of each area are fused and processed by a distributed model parameter aggregation algorithm to generate a global transport control model. This model is then trained with the optimal transport parameter scheme to obtain a lightweight field control model. The lightweight field control model calculates the optimal transport control scheme based on real-time monitoring data and sends it to the injection device for execution.

[0138] In one alternative implementation,

[0139] Edge computing control nodes are deployed at the contaminated site. A distributed model parameter aggregation algorithm is used to fuse monitoring data from various areas, generating a global transport control model. This model is then trained with the optimal transport parameter scheme to obtain a lightweight on-site control model. The lightweight on-site control model calculates the optimal transport control scheme based on real-time monitoring data and sends it to the injection device for execution, including:

[0140] Edge computing control nodes are deployed at the contaminated site. The edge computing control nodes acquire water level monitoring data, water quality monitoring data, and flow monitoring data through a data acquisition module. The time-series feature extraction method is used to analyze the variation patterns of the water level monitoring data, the water quality monitoring data, and the flow monitoring data to obtain the sampling period.

[0141] The edge computing control node uses an anomaly detection algorithm based on differential thresholds to identify outliers, uses local linear interpolation to repair missing data, and calculates time-series features through a sliding window to obtain feature vectors.

[0142] A distributed model parameter aggregation algorithm is used to construct a transport parameter sub-model at the edge computing control node. A node affinity matrix is ​​constructed based on the geographic location relationship. The transport parameter sub-model is iteratively aggregated according to the node affinity matrix to form a regional model. An adaptive weighting method is used to merge the regional models to generate a global transport control model.

[0143] The global transport control model is densified by a nested mesh structure. The computational task is distributed to the edge computing control node for parallel processing by a block-based iterative method. The optimal transport parameter scheme is obtained by solving in a hierarchical progressive manner.

[0144] The key features of the global transport control model are extracted by knowledge mapping, and a lightweight field control model is constructed by progressive compression strategy. The lightweight field control model receives the water level monitoring data, the water quality monitoring data and the flow monitoring data, determines the feasible solution space by piecewise optimization strategy, and generates the optimal transport control scheme by gradient search method and heuristic algorithm.

[0145] The optimal transportation control scheme is sent to the injection device for execution. The injection parameters are adjusted using a feedback correction mechanism, and potential faults are identified and the control strategy is automatically switched using a state prediction algorithm.

[0146] Edge computing control nodes are deployed at key locations within the contaminated site. Each node is equipped with a data acquisition module to continuously collect water level, water quality, and flow monitoring data. The changing patterns of these monitoring data are analyzed using time-series feature extraction methods, including periodic changes, trend changes, and abrupt changes, to determine the optimal sampling period and achieve dynamic adjustment of data acquisition.

[0147] In the data preprocessing stage, the edge computing control node first performs anomaly detection. An algorithm based on differential thresholding is used to identify potential outliers by calculating the difference between adjacent data points and comparing it to a preset threshold. For identified outliers, the system marks and temporarily removes them. Subsequently, for missing values ​​in the dataset, a local linear interpolation method is used for repair. This method considers the changing trends of neighboring data points to ensure the reasonableness of the interpolation results. After data repair, time-series features, including statistical features such as mean, variance, and trend, are calculated using a sliding window method to form a feature vector.

[0148] In the model building phase, a distributed model parameter aggregation algorithm is employed. First, a transport parameter sub-model is constructed at each edge computing control node, reflecting the characteristics of a local area. Then, based on the geographical relationships between nodes, a node affinity matrix is ​​constructed, describing the degree of association between different nodes. According to the affinity matrix, the system iteratively aggregates the transport parameter sub-models to form a regional model capable of representing the characteristics of a larger area. Finally, an adaptive weighting method is used to merge the regional models into a global transport control model based on their reliability and representativeness.

[0149] To improve computational efficiency, the system employs a nested mesh structure to refine the mesh of the global transport control model. A finer mesh is used in critical regions, while a coarser mesh is maintained in general areas. A block-based iterative method is used to distribute large-scale computational tasks to various edge computing control nodes, achieving parallel processing. A hierarchical, progressive solution strategy is adopted: first, an initial solution is obtained by solving the coarse mesh, then the mesh is gradually refined to obtain a high-precision solution, ultimately yielding the optimal transport parameter scheme.

[0150] To achieve rapid on-site control, the system extracts key features of the global transport control model through knowledge mapping, including the relationships between major parameters and control rules. A progressive compression strategy is employed to remove secondary features while retaining the core control logic, constructing a lightweight on-site control model. This lightweight model receives real-time monitoring data, uses a piecewise optimization strategy to determine the feasible solution space under current conditions, and then combines gradient search and heuristic algorithms for rapid optimization to generate the optimal transport control scheme.

[0151] The system sends the optimal transport control scheme to the injection device for execution. During execution, a feedback correction mechanism is used to adjust the injection parameters in real time to ensure control effectiveness. Simultaneously, a state prediction algorithm continuously analyzes the system's operating status, and when a potential fault is identified, it automatically switches to a backup control strategy to ensure the system's continuous and stable operation.

[0152] For example, consider a remediation project of a contaminated site at a chemical plant. Twelve edge computing control nodes are deployed within the site, each equipped with monitoring devices for water level, water quality (pH, conductivity, COD, etc.), and flow rate. Time-series analysis determines that water level data is sampled hourly, water quality data every four hours, and flow rate data every half hour.

[0153] In data preprocessing, the system set the difference threshold to three times the standard deviation, successfully identifying outliers accounting for 2% of the total data. Local linear interpolation was used to repair approximately 5% of the missing data. A 20-dimensional feature vector, including mean, variance, and trend, was calculated using a sliding window of 30 data points.

[0154] During the model building process, each node first constructs a local sub-model containing parameters such as permeability coefficient and water storage coefficient. Based on the distance relationship between nodes, an affinity matrix is ​​constructed. The 12 sub-models are aggregated through 5 rounds of iteration to form 3 regional models, and finally merged into a global model covering the entire site.

[0155] During the computational optimization phase, a three-layer nested mesh structure was adopted, with the finest mesh spacing being 1 meter and the coarsest mesh spacing being 10 meters. The computational task was distributed to 12 nodes for parallel processing, and the optimal solution was obtained through three rounds of progressive solving.

[0156] The lightweight model retains 30% of the original model's feature parameters and improves response speed by three times. This model employs a five-segment optimization strategy to divide the solution space and combines it with an improved particle swarm optimization algorithm for optimization, generating a control scheme on average within a few seconds.

[0157] During the implementation of the scheme, the feedback correction mechanism adjusts the injection parameters every minute, the state prediction algorithm provides early warning of potential faults, and the system can switch control strategies within seconds when an anomaly is detected.

[0158] In this embodiment, edge computing control nodes are deployed at the contaminated site. The changing patterns of monitoring data are analyzed using a time-series feature extraction method, enabling dynamic adjustment of the sampling period. This significantly improves the targeting and efficiency of data collection and ensures data quality. An anomaly detection algorithm based on differential thresholding and a local linear interpolation method are employed to accurately identify abnormal data and reasonably repair missing data. A vector representation comprehensively reflecting data characteristics is constructed by calculating time-series features through a sliding window, providing a reliable foundation for subsequent modeling. By constructing sub-models at edge nodes and iteratively aggregating them based on geographical location relationships, a global transport control model is ultimately formed, improving the efficiency and accuracy of model construction. In particular, the use of a nested mesh structure and a block-based iterative method enables efficient parallel processing of computational tasks.

[0159] Existing in-situ groundwater remediation control technologies typically employ fixed sampling periods and uniform data processing methods, which are ill-suited to adapting to dynamic changes in site conditions. Regarding anomaly handling, traditional methods are simplistic in identifying outliers and repairing missing data, easily leading to data distortion. In model building, centralized computing architectures are commonly used, resulting in low computational efficiency and difficulty in handling large-scale data. Furthermore, commonly used control models are complex in structure and slow in response, failing to meet real-time control requirements. These problems severely restrict the improvement of remediation effectiveness. This embodiment, through an innovative edge computing architecture, intelligent data processing methods, distributed modeling strategies, and lightweight control models, effectively addresses the shortcomings of traditional technologies in data processing, computational efficiency, control performance, and system reliability. It significantly improves the control accuracy and efficiency in the in-situ groundwater remediation process, reduces system operating costs, and enhances remediation results, providing a new technical path for the development of in-situ groundwater remediation technology and possessing significant engineering application value.

[0160] In one alternative implementation,

[0161] A distributed model parameter aggregation algorithm is used to construct a transport parameter sub-model at the edge computing control node. A node affinity matrix is ​​constructed based on geographic location relationships. The transport parameter sub-model is then iteratively aggregated using the node affinity matrix to form a regional model, including:

[0162] A distributed model parameter aggregation algorithm is used to construct transport parameter sub-models at edge computing control nodes. Feature sequences are generated from monitoring data through multi-scale analysis and filtering. A node affinity matrix is ​​constructed based on spatial correlation coefficients, hydrogeological similarity coefficients, and migration flux coefficients between nodes, calculated according to geographical location relationships. The transport parameter sub-models are then iteratively aggregated using the node affinity matrix to form a regional model. Specifically, this includes:

[0163] A distributed model parameter aggregation algorithm is used to construct a transport parameter sub-model at the edge computing control node. The distributed model parameter aggregation algorithm performs multi-scale analysis on water level monitoring data, water quality monitoring data and flow monitoring data through wavelet decomposition, extracts high-frequency and low-frequency feature components, sets an adaptive threshold to remove abnormal fluctuations based on the energy distribution of feature components, and uses recursive filtering to smooth the signal and generate a stable monitoring data sequence.

[0164] Spatiotemporal features are extracted from the stable monitoring data sequence, time-series statistics are calculated using a sliding window, data distribution probability is calculated using kernel density estimation, data mutation locations are determined based on change point detection, and a comprehensive feature vector is constructed by combining a multi-source data fusion algorithm.

[0165] The comprehensive feature vector is input into the gradient descent optimizer to calculate the gradient of the model parameters. The cross-validation method is used to calculate the parameter importance score. Iterative optimization is performed on the high-scoring parameters, and linear regression is performed on the low-scoring parameters to construct the transport parameter sub-model and generate the initial parameter field.

[0166] A node affinity matrix is ​​constructed based on geographical location relationships. The spatial correlation coefficient is obtained by calculating the Euclidean distance between nodes. The hydrogeological similarity coefficient is calculated using a hierarchical clustering method. The migration flux coefficient is calculated based on the pollutant concentration gradient. The affinity weight is calculated by normalizing and weighting the spatial correlation coefficient, the hydrogeological similarity coefficient, and the migration flux coefficient.

[0167] The transport parameter sub-model is iteratively aggregated based on the node affinity matrix to form a regional model. Nodes with high affinity are preferentially fused. The initial parameter field is weighted and averaged using the affinity weights. The fusion residual is calculated and the affinity weights are adjusted to update the node affinity matrix.

[0168] The region model is divided into multiple computational subtasks and assigned to the edge computing control nodes. A message queue is established between nodes to transmit data. Checkpoints are set according to the fusion residuals. The node load is allocated according to the computational load of the initial parameter field.

[0169] A distributed model parameter aggregation algorithm is employed to construct a transport parameter sub-model at the edge computing control node. This algorithm performs wavelet decomposition on the collected water level, water quality, and flow monitoring data to achieve multi-scale analysis. Wavelet transform decomposes the original signal into feature components of different frequencies, extracting high-frequency components (reflecting short-term fluctuations) and low-frequency components (reflecting long-term trends). Based on the energy distribution of each feature component, an adaptive threshold is set to identify and remove abnormal fluctuations. Subsequently, a recursive filtering method is used to smooth the signal, eliminate residual noise, and generate a stable monitoring data sequence.

[0170] Spatiotemporal features were extracted from the processed stable monitoring data sequences. First, time-series statistics, including mean, variance, skewness, and kurtosis, were calculated using the sliding window method. Then, the probability distribution of the data was calculated using kernel density estimation to obtain the probability density features. Next, a change point detection algorithm was used to determine the locations of abrupt changes and identify key change points. Finally, a multi-source data fusion algorithm was combined to integrate different types of features and construct a comprehensive feature vector that fully reflects the system characteristics.

[0171] The constructed integrated feature vector is input into the gradient descent optimizer to calculate the gradient information of each parameter of the model. Cross-validation is used to evaluate the importance of different parameters, generating parameter importance scores. Parameters with high importance scores are fine-tuned using iterative optimization methods, while parameters with low importance scores are quickly estimated using linear regression. Based on the optimized parameters, a transport parameter sub-model is constructed, and an initial parameter field describing the characteristics of the current region is generated.

[0172] A node affinity matrix is ​​constructed based on geographical location relationships, and the Euclidean distance between nodes is calculated to obtain spatial correlation coefficients characterizing the degree of spatial association. Hierarchical clustering is used to analyze the hydrogeological characteristics of each node region, and hydrogeological similarity coefficients are calculated. Migration flux coefficients are calculated based on the spatial distribution of pollutant concentrations to characterize the correlation of pollutant migration. These three types of coefficients are normalized and then weighted to obtain a comprehensive affinity weight.

[0173] The transport parameter sub-models are iteratively aggregated based on the node affinity matrix to form a regional model. During the aggregation process, nodes with higher affinity are prioritized for fusion, and the initial parameter field is weighted and averaged using the calculated affinity weights. The residuals generated during the fusion process are calculated, and the affinity weights are dynamically adjusted based on the magnitude of the residuals, thereby updating the node affinity matrix to ensure the accuracy of the aggregation results.

[0174] The regional model is divided into multiple computational subtasks, which are then distributed to various edge computing control nodes for parallel processing. A message queue system is established between nodes to achieve efficient data transmission. Checkpoints are set based on the fused residuals to monitor the convergence of the computation process. Simultaneously, node load is rationally allocated according to the computational load of the initial parameter field to ensure balanced utilization of computing resources.

[0175] For example, taking a groundwater contaminated site as an example, 15 edge computing control nodes are deployed within the site. The raw data collected by each node includes: water level data (once per hour), water quality data (pH, COD, etc., once every four hours), and flow rate data (once every half hour).

[0176] Four-level wavelet decomposition was used to decompose various monitoring data into components of different frequencies. Energy distribution analysis showed that high-frequency components accounted for approximately 20%, and an adaptive threshold was set to remove abnormal fluctuations with amplitudes exceeding three times the mean. Kalman filtering was used for signal smoothing to obtain a stable monitoring data sequence. A 24-hour sliding window was used to calculate time-series statistical characteristics, and a Gaussian kernel function was used for kernel density estimation. The CUSUM algorithm was used to detect data abrupt changes. After fusing the multi-source data, a 128-dimensional comprehensive feature vector was obtained.

[0177] During gradient descent optimization, parameter importance is calculated using 5-fold cross-validation. Parameters with importance scores exceeding 0.7 undergo 100 rounds of iterative optimization, while other parameters are estimated using linear regression. A sub-model containing 20 key parameters is constructed. In node affinity calculation, the spatial correlation coefficient has a weight of 0.4, the hydrogeological similarity coefficient has a weight of 0.3, and the migration flux coefficient has a weight of 0.3. Through 5 rounds of iterative aggregation, the 15 sub-models are merged into 3 regional models. Finally, the computational task is divided into 60 sub-tasks, distributed across 15 nodes for parallel processing. A checkpoint is set every 20% of the data volume processed, and 3-5 sub-tasks are assigned to each node based on computational complexity.

[0178] In this embodiment, wavelet decomposition is used to achieve multi-scale data analysis. By extracting high-frequency and low-frequency feature components and combining adaptive thresholding and recursive filtering, the monitoring data is refined, significantly improving data quality and providing a reliable foundation for subsequent modeling. By combining various methods such as time series statistics, kernel density estimation, and change point detection, a comprehensive feature extraction system is constructed. Multi-scale analysis and adaptive filtering significantly improve the signal-to-noise ratio of the monitoring data, effectively eliminate abnormal fluctuations, and ensure data reliability. The multi-dimensional feature extraction and fusion mechanism enables the model to more comprehensively characterize the site features, significantly improving the richness and accuracy of feature expression. The differentiated optimization strategy significantly improves computational efficiency while ensuring the optimization accuracy of key parameters, achieving a balance between optimization efficiency and accuracy.

[0179] Existing groundwater in-situ remediation model construction technologies typically employ single filtering methods for monitoring data processing, failing to effectively identify and eliminate multi-scale noise. In feature extraction, they are often limited to simple statistical features, lacking in-depth analysis of the spatiotemporal evolution of the data. Regarding parameter optimization, a uniform optimization strategy is used for all parameters, ignoring differences in parameter importance. In model aggregation, simple spatial distance is often used as the fusion basis, failing to fully consider the influence of hydrogeological characteristics and pollutant migration patterns. These problems lead to insufficient model accuracy and difficulty in accurately describing site characteristics. This embodiment effectively addresses the shortcomings of traditional technologies in data processing, feature extraction, parameter optimization, and model aggregation through innovative data processing methods, a comprehensive feature extraction system, differentiated optimization strategies, and a multi-factor coupled aggregation mechanism. It significantly improves the construction efficiency and accuracy of groundwater in-situ remediation models, providing more reliable support for the optimization of remediation schemes.

[0180] Figure 2This diagram compares the feature fusion performance of the precise transport control method for in-situ groundwater remediation agents according to an embodiment of the present invention. It shows the trend of feature consistency changing with the number of iterations during feature fusion for different technical solutions. The embodiment of the present invention, through an innovative dual-branch parallel structure and multi-cascade noise reduction mechanism, achieved a high fusion accuracy (0.68) in the initial iteration (5 iterations) and continued to improve in subsequent iterations, ultimately reaching a high accuracy of 0.87 after 35 iterations. In contrast, the traditional feature fusion method and the single-constraint feature fusion method achieved fusion accuracies of 0.58 and 0.62 respectively in the initial iteration, eventually reaching only 0.70 and 0.76. Particularly in the critical stage of 10-20 iterations, the accuracy improvement rate of the embodiment of the present invention is significantly higher than other schemes, thanks to the effective combination of differentiated noise processing strategies and multi-head attention mechanisms. After 25 iterations, the accuracy curve of the embodiment of the present invention tends to smooth out, indicating that the fusion process has reached a stable state. The smoothness of the curve is also better than other schemes, reflecting the stability and reliability of the fusion process. Furthermore, by preserving the original feature information through residual connections, the embodiments of the present invention maintain high fusion accuracy while also preserving the physical meaning of the features. These data fully demonstrate the significant advantages of the embodiments of the present invention in terms of feature fusion effect and efficiency.

[0181] Figure 3 This diagram compares the system optimization effects of the precise transport control method for remediation agents in in-situ groundwater remediation according to an embodiment of the present invention. It shows the trend of system compliance rate changing with the optimization cycle under different optimization strategies. The embodiment of the present invention, through a hierarchical adjustment strategy and a multi-level buffer structure, achieved a high compliance rate (82%) in the initial optimization stage (4 hours), and continued to improve in subsequent optimization processes, ultimately reaching a high compliance rate of 94% at 28 hours. In contrast, the compliance rates of the traditional fixed strategy and the simple adaptive strategy were 75% and 78% respectively in the initial optimization stage, and ultimately only reached 85% and 88%. Especially in the critical optimization stage of 8-16 hours, the compliance rate improvement speed of the embodiment of the present invention was significantly faster than other schemes. After 20 hours, the compliance rate curve of the embodiment of the present invention tended to stabilize, indicating that the system reached a stable optimization state. At the same time, the smoothness of the curve was also better than other schemes, reflecting the stability and reliability of the system optimization process. These data fully demonstrate the significant advantages of the embodiment of the present invention in terms of system optimization effect and efficiency.

[0182] Figure 4This chart compares the data quality improvement effects of the precise transport control method for remediation agents in in-situ groundwater remediation according to an embodiment of the present invention. It shows the signal-to-noise ratio (SNR) trends of different technical solutions during data processing. The embodiment of the present invention, through multi-scale wavelet decomposition and adaptive threshold filtering, achieved a high SNR (38dB) in the initial processing stage (20 min), and continued to improve with increasing processing time, ultimately reaching a high SNR of 85dB at 100 min. In contrast, traditional wavelet threshold denoising and empirical mode decomposition achieved SNRs of 31dB and 33dB respectively in the initial processing stage, ultimately only reaching 55dB and 60dB. Particularly in the critical processing stage of 40-60 min, the SNR improvement rate of the embodiment of the present invention was significantly higher than other solutions, thanks to the effective elimination of residual noise by the recursive filtering method. Simultaneously, the SNR curve of this technical solution is smoother, indicating a more stable and reliable processing process. These data fully demonstrate the significant advantages of this technical solution in improving data quality.

[0183] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0184] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for precise delivery control of a remediation agent for in-situ groundwater remediation, characterized by, The method comprises the following steps: Collecting image data and sensor measurement data corresponding to the drilling core of the contaminated site, inputting the image data into a visual attention neural network to extract soil structure feature data, inputting the sensor measurement data into a time series attention neural network to analyze groundwater dynamic feature data, adding the soil structure feature data and the groundwater dynamic feature data to a conditional probability diffusion model for feature fusion, generating a hydrogeological feature vector and mapping it to a three-dimensional spatial grid through a spatial attention calculation module to obtain an initial hydrogeological feature model, and combining a generative probability diffusion model to complete the missing area to obtain a high-precision hydrogeological feature model; Inputting the high-precision hydrogeological feature model into a multi-region collaborative simulation environment, performing local optimization calculation on different spatial regions through multiple decision units, sharing optimization results through a communication network, performing global evaluation and cooperatively adjusting optimization parameters of each decision unit based on the shared results through a central evaluation network, verifying the adjusted transport strategy through a Monte Carlo probability search method, and outputting an optimal transport parameter scheme; Laying out edge computing control nodes at the contaminated site, fusing regional monitoring data through a distributed model parameter aggregation algorithm to generate a global transport control model and train a lightweight on-site control model with the optimal transport parameter scheme, and calculating an optimal transport control scheme from real-time monitoring data through the lightweight on-site control model and sending it to an injection device for execution.

2. The method of claim 1, wherein, Collecting image data and sensor measurement data corresponding to the drilling core of the contaminated site, inputting the image data into a visual attention neural network to extract soil structure feature data, inputting the sensor measurement data into a time series attention neural network to analyze groundwater dynamic feature data, adding the soil structure feature data and the groundwater dynamic feature data to a conditional probability diffusion model for feature fusion, generating a hydrogeological feature vector and mapping it to a three-dimensional spatial grid through a spatial attention calculation module to obtain an initial hydrogeological feature model, and combining a generative probability diffusion model to complete the missing area to obtain a high-precision hydrogeological feature model comprises: Laying out a drilling sampling system at the contaminated site to collect core image data and groundwater sensor measurement data; Inputting the core image data into an adaptive histogram equalization processing module to eliminate light interference and obtain a first processed image, inputting the first processed image into a local contrast sharpening processing module to enhance structural details and obtain a second processed image, inputting the second processed image into a visual attention neural network, performing convolution operation through an encoder of the visual attention neural network to extract features, and reconstructing the features through a decoder to obtain soil structure feature data; Inputting the groundwater sensor measurement data into an outlier detection module to remove outliers and obtain first processed data, inputting the first processed data into an adaptive sliding window noise reduction module to obtain second processed data, inputting the second processed data into a time series attention neural network, analyzing time series correlation through a bidirectional long short-term memory unit to obtain groundwater dynamic feature data; The soil structure feature data and the underground water dynamic feature data are spliced to input a conditional probability diffusion model, a hydrogeological feature vector is obtained through dynamic noise reduction iteration, the hydrogeological feature vector is input into a spatial attention calculation module, a spatial weight is calculated according to feature similarity, and grid encryption is performed, and an initial hydrogeological feature model is obtained through feature interpolation; The initial hydrogeological feature model is input into a generative probability diffusion model, feature generation is performed on a missing area based on a known regional feature distribution, and a high-precision hydrogeological feature model is obtained through multi-scale discriminator evaluation and optimization.

3. The method of claim 2, wherein, The soil structure feature data and the underground water dynamic feature data are spliced to input a conditional probability diffusion model, a hydrogeological feature vector is obtained through dynamic noise reduction iteration, the hydrogeological feature vector is input into a spatial attention calculation module, a spatial weight is calculated according to feature similarity, and grid encryption is performed, and an initial hydrogeological feature model is obtained through feature interpolation; The soil structure feature data and the underground water dynamic feature data are spliced to input a conditional probability diffusion model, a hydrogeological feature vector is obtained through dynamic noise reduction iteration, the hydrogeological feature vector is input into a spatial attention calculation module, a spatial weight is calculated according to feature similarity, and grid encryption is performed, and an initial hydrogeological feature model is obtained through feature interpolation; The soil structure feature data and the underground water dynamic feature data are spliced to input a conditional probability diffusion model, a hydrogeological feature vector is obtained through dynamic noise reduction iteration, the hydrogeological feature vector is input into a spatial attention calculation module, a spatial weight is calculated according to feature similarity, and grid encryption is performed, and an initial hydrogeological feature model is obtained through feature interpolation; The soil structure feature data and the underground water dynamic feature data are spliced to input a conditional probability diffusion model, a hydrogeological feature vector is obtained through dynamic noise reduction iteration, the hydrogeological feature vector is input into a spatial attention calculation module, a spatial weight is calculated according to feature similarity, and grid encryption is performed, and an initial hydrogeological feature model is obtained through feature interpolation; The soil structure feature data and the underground water dynamic feature data are spliced to input a conditional probability diffusion model, a hydrogeological feature vector is obtained through dynamic noise reduction iteration, the hydrogeological feature vector is input into a spatial attention calculation module, a spatial weight is calculated according to feature similarity, and grid encryption is performed, and an initial hydrogeological feature model is obtained through feature interpolation; The soil structure feature data and the underground water dynamic feature data are spliced to input a conditional probability diffusion model, a hydrogeological feature vector is obtained through dynamic noise reduction iteration, the hydrogeological feature vector is input into a spatial attention calculation module, a spatial weight is calculated according to feature similarity, and grid encryption is performed, and an initial hydrogeological feature model is obtained through feature interpolation; A multi-head attention module is constructed to calculate the feature correlation degree of the constraint feature, an attention weight is generated based on the feature correlation degree, the constraint feature is enhanced through the attention weight and residual connection is introduced to reserve original feature information to obtain a fusion feature. The construction parameter evaluation module calculates a feature quality index and a fusion effect index of the fusion feature, dynamically optimizes a noise reduction parameter based on the feature quality index and the fusion effect index to obtain an optimized parameter, inputs the optimized parameter into a feature decoder, and decodes and reconstructs the fusion feature to obtain a hydrogeological feature vector.

4. The method of claim 1, wherein, The high-precision hydrogeological feature model is input into a multi-region collaborative simulation environment, and local optimization calculation is performed on different spatial regions by multiple decision units. The decision units share optimization results through a communication network, a central evaluation network performs global evaluation and cooperatively adjusts optimization parameters of each decision unit based on the shared results, and a Monte Carlo probability search method is used to verify the adjusted transport strategy, and an optimal transport parameter scheme is output, including: The high-precision hydrogeological feature model is divided into multiple spatial regions according to the hydrogeological unit boundary and the management partition, and a decision unit with a data layer, a calculation layer and a communication layer is configured in each spatial region; The data layer of the decision unit extracts geological structure features, hydrological parameter distribution and boundary condition information of the spatial region, establishes a groundwater flow field numerical model, collects groundwater level, water quality index and exploitation amount data to construct a dynamic database, and the calculation layer of the decision unit inputs state data in the dynamic database into an optimizer to perform local optimization calculation based on optimization objectives of water resource balance, water quality safety and system stability to obtain optimization results; The decision units are interconnected through a communication network to build a hierarchical caching mechanism to store the optimization results, and an incremental update strategy is used to transmit system state data to a central evaluation network. The central evaluation network extracts evaluation indexes of feature calculation resource regulation effect, environmental impact degree and economic benefit level, highlights key index contributions through an attention mechanism to generate a comprehensive score; Based on the comprehensive score, the parameter search space of each decision unit is expanded and the constraint condition weight is adjusted, a coordination factor is introduced to balance the optimization objectives, the adjustment results are transmitted to the decision units, a Monte Carlo probability search method is used to generate a verification scenario, typical failure modes are identified through cluster analysis and weak links are optimized to obtain verification results, water quantity distribution coefficients, water quality control thresholds and system regulation parameters are selected based on the verification results, dynamic adjustment rules of the transport strategy are constructed, and an optimal transport parameter scheme is output.

5. The method of claim 1, wherein, The adjustment results are transmitted to the decision units, a Monte Carlo probability search method is used to generate a verification scenario, typical failure modes are identified through cluster analysis and weak links are optimized to obtain verification results, including: Based on the evaluation results of the central evaluation network, a hierarchical adjustment strategy is constructed, adjustment results are generated for decision units in different score intervals, the adjustment results are transmitted to the decision units through a multi-level cache structure, a Monte Carlo probability search method is used to extract features and identify weak links, a protection mechanism is constructed for the weak links and evaluation results are generated, and parameter configuration is dynamically adjusted based on the evaluation results, including: Based on the evaluation results of the central evaluation network, a hierarchical adjustment strategy is constructed, the decision units in different score intervals are processed differently by expanding the parameter search space and increasing the constraint weight, moderately expanding the search range and fine-tuning the constraint weight, and maintaining the existing parameter configuration adjustment mode, and an adjustment result is generated; The adjustment result is transmitted to the decision unit by constructing a multi-level cache structure, the historical adjustment sequence is stored by local cache, the parameter information of adjacent units is summarized by regional cache nodes, the data is transmitted by differential update mechanism and the data integrity is ensured by validity check and timeout retransmission; The probability distribution characteristics of the adjustment result are extracted by using the Monte Carlo probability search method, the verification samples are obtained by using the adaptive sampling strategy of hierarchical quantization, the verification scenes are generated after orthogonalization processing of the samples to eliminate parameter correlation, the verification results of the verification scenes are processed by cluster analysis, the typical failure modes are identified by determining the cluster centers through dynamic neighborhood search, the fault tree model is constructed to track the weak links, and the hierarchical early warning mechanism is established by analyzing the system response characteristics; A spatial protection ring is constructed by adjusting the buffer parameters of the weak links, a time response mechanism is constructed by configuring the control period, a regional linkage is realized by establishing a collaborative optimization network, and the optimization effect is verified by incremental verification to obtain a verification result, the monitoring index compliance rate is determined by statistical analysis, the spatial distribution characteristics are evaluated by dispersion coefficient, and the system stability is analyzed by disturbance response to generate an evaluation result; Based on the evaluation results, the key indicators are calculated by using a sliding time window, the calculated key indicators are input into the hierarchical adjustment strategy, the parameter search range and constraint weight of the decision unit are dynamically adjusted, and the optimal parameter configuration is output.

6. The method of claim 1, wherein, Edge computing control nodes are arranged at the pollution site, regional monitoring data is fused by a distributed model parameter aggregation algorithm to generate a global transport control model, and a lightweight field control model is trained with the optimal transport parameter scheme, the lightweight field control model calculates the optimal transport control scheme according to real-time monitoring data and sends it to the injection device for execution, including: Edge computing control nodes are arranged at the pollution site, the edge computing control nodes acquire water level monitoring data, water quality monitoring data and flow monitoring data through a data acquisition module, analyze the change law of the water level monitoring data, the water quality monitoring data and the flow monitoring data by using a time sequence feature extraction method, and obtain a sampling period; The edge computing control nodes identify outliers by using a differential threshold-based anomaly detection algorithm, repair missing data by using local linear interpolation, and obtain a feature vector by calculating time sequence features through a sliding window; A transport parameter sub-model is constructed at the edge computing control nodes by using a distributed model parameter aggregation algorithm, a node affinity matrix is constructed based on geographical location relationship, the transport parameter sub-model is iteratively aggregated based on the node affinity matrix to form a regional model, and the regional model is merged by using an adaptive weight method to generate a global transport control model; The global transport control model is meshed by using a nested grid structure, the calculation task is distributed to the edge computing control nodes for parallel processing by using a block iteration method, and an optimal transport parameter scheme is obtained by using a hierarchical progressive solution method; The key features of the global transport control model are extracted by using a knowledge mapping method, a lightweight field control model is constructed by using a progressive compression strategy, the lightweight field control model receives the water level monitoring data, the water quality monitoring data and the flow monitoring data, determines a feasible solution space by using a piecewise optimization strategy, and generates an optimal transport control scheme by using a gradient search method and a heuristic algorithm for optimization; The optimal transport control scheme is sent to the injection device for execution, the injection parameters are adjusted by using a feedback correction mechanism, and potential faults are identified by using a state prediction algorithm and the control strategy is automatically switched.

7. The method of claim 6, wherein, A distributed model parameter aggregation algorithm is used to construct a transport parameter sub-model on the edge computing control nodes, a node affinity matrix is constructed based on geographical location relationships, and the transport parameter sub-model is iteratively aggregated to form a regional model based on the node affinity matrix, including: A distributed model parameter aggregation algorithm is used to construct a transport parameter sub-model on the edge computing control nodes, a node affinity matrix is constructed based on geographical location relationships, and the transport parameter sub-model is iteratively aggregated to form a regional model based on the node affinity matrix, including: A distributed model parameter aggregation algorithm is used to construct a transport parameter sub-model on the edge computing control nodes, the distributed model parameter aggregation algorithm performs multi-scale analysis on the water level monitoring data, the water quality monitoring data and the flow monitoring data by wavelet decomposition, extracts high-frequency and low-frequency feature components, sets an adaptive threshold based on the energy distribution of the feature components to remove abnormal fluctuations, and performs smoothing processing on the signal by using recursive filtering to generate a stable monitoring data sequence; Temporal and spatial features are extracted from the stable monitoring data sequence, time series statistics are calculated by using a sliding window, data distribution probabilities are calculated by using kernel density estimation, data mutation positions are determined based on change point detection, and a comprehensive feature vector is constructed by using a multi-source data fusion algorithm; The comprehensive feature vector is input into a gradient descent optimizer, the model parameter gradient is calculated, the parameter importance score is calculated by using a cross-validation method, the high-score parameters are iteratively optimized, the low-score parameters are linearly regressed, the transport parameter sub-model is constructed, and an initial parameter field is generated; A node affinity matrix is constructed based on geographical location relationships, the spatial correlation coefficient is calculated by calculating the Euclidean distance between nodes, the hydrogeological similarity coefficient is calculated by using hierarchical clustering, the migration flux coefficient is calculated based on the concentration gradient of the pollutant, and the spatial correlation coefficient, the hydrogeological similarity coefficient and the migration flux coefficient are normalized and weighted to calculate the affinity weight. According to the node affinity matrix, the transport parameter sub-model is iteratively aggregated to form a regional model, high-affinity nodes are preferentially fused, the initial parameter field is weighted and averaged using the affinity weight, a fusion residual is calculated and the affinity weight is adjusted, and the node affinity matrix is updated; The regional model is divided into multiple calculation sub-tasks and distributed to the edge computing control nodes, a message queue is established between nodes to transmit data, a checkpoint is set according to the fusion residual, and node load is distributed according to the calculation amount of the initial parameter field.

8. A precise delivery control system for a remediation agent for in-situ groundwater remediation for implementing the method according to any one of the preceding claims 1 to 7, characterized in that, Comprise: The first unit is used for collecting image data and sensor measurement data corresponding to the drilling core of the contaminated site, inputting the image data into a visual attention neural network to extract soil structure feature data, inputting the sensor measurement data into a time series attention neural network to analyze underground water dynamic feature data, adding the soil structure feature data and the underground water dynamic feature data to a conditional probability diffusion model for feature fusion, generating a hydrogeological feature vector and mapping it to a three-dimensional space grid through a spatial attention calculation module to obtain an initial hydrogeological feature model, and combining a generative probability diffusion model to complete the missing area to obtain a high-precision hydrogeological feature model; The second unit is used for inputting the high-precision hydrogeological feature model into a multi-region collaborative simulation environment, performing local optimization calculation on different spatial regions through multiple decision units, sharing optimization results through a communication network, performing global evaluation based on the shared results and cooperatively adjusting the optimization parameters of each decision unit, verifying the adjusted transport strategy by combining a Monte Carlo probability search method, and outputting an optimal transport parameter scheme; The third unit is used for arranging edge computing control nodes at the contaminated site, fusing regional monitoring data through a distributed model parameter aggregation algorithm, generating a global transport control model, and training a lightweight on-site control model with the optimal transport parameter scheme, so as to calculate an optimal transport control scheme according to real-time monitoring data and send it to an injection device for execution.

Images