Groundwater intelligent monitoring system and method based on cloud edge end cooperation and dynamic optimization

Abstract

The application relates to the technical field of groundwater monitoring and discloses a groundwater intelligent monitoring system and method based on cloud edge end cooperation and dynamic optimization, which comprises a sensing layer end side, an edge layer and a cloud end layer. The sensing layer end side is integrated with intelligent monitoring terminals, target aquifer water quality parameters are obtained through multilayer nest monitoring wells, and interlayer hydraulic connection is blocked by using bentonite water stopping materials to prevent cross contamination; the edge layer receives end side data, carries out pollution diffusion simulation by using a hydrogeological model engine, and outputs evaluation results and early warning information; the cloud end layer generates optimization strategies based on early warning information, and issues configuration instructions to adjust hardware states. The application classifies and processes calculation tasks, reduces the original data transmission amount, can dynamically adjust the monitoring network density and the data acquisition frequency according to real-time pollution risks, realizes self-adaptive adjustment of the data acquisition strategy, and shortens the response time of the pollution early warning.

Description

Technical Field

[0001] This invention relates to the field of groundwater monitoring technology, specifically to a groundwater intelligent monitoring system and method based on cloud-edge-device collaboration and dynamic optimization. Background Technology

[0002] Groundwater is an important component of water resources, and groundwater pollution has a direct impact on the ecological environment and industrial and agricultural production. Currently, groundwater environmental monitoring mainly relies on manual periodic sampling and analysis, or automated monitoring schemes that deploy sensors at fixed locations and transmit data directly back to a central server.

[0003] However, existing automatic monitoring systems have several shortcomings in actual operation. These systems typically employ fixed network deployment and data acquisition frequencies. Once the equipment is installed, it becomes difficult to adjust monitoring strategies in a timely manner based on the migration paths of groundwater pollution or dynamic changes in the hydrogeological environment. This static monitoring model can lead to the omission of high-risk areas when pollution occurs, while consuming unnecessary equipment resources in low-risk areas, making it difficult to achieve real-time tracking of changes in the groundwater environment.

[0004] In terms of data processing and transmission architecture, existing systems mostly rely on a centralized processing model, directly uploading all raw data collected from the front end to cloud servers. With the increase in the number of monitoring nodes and the higher sampling frequency, the continuous reporting of massive amounts of raw data puts significant pressure on communication network bandwidth. Centralizing all computational tasks in the cloud results in a long feedback path between data analysis and command issuance, leading to delays in the generation of pollution warning information and making it difficult to meet the business needs for rapid response.

[0005] At the front-end sensing hardware level, conventional groundwater monitoring wells mostly adopt a single-layer structure, primarily targeting shallow or single-target layers. Faced with complex geological conditions involving multiple overlapping aquifers, conventional structures struggle to effectively capture the three-dimensional distribution of groundwater pollution. Attempts to conduct simultaneous monitoring at multiple depths within the same borehole fail due to a lack of effective hydraulic isolation measures, leading to groundwater mixing between upper and lower aquifers and causing cross-contamination. This results in the collected water quality parameters losing their stratified representativeness and accuracy. Summary of the Invention

[0006] The technical problem addressed by this invention is that existing groundwater monitoring systems typically employ fixed data acquisition and network deployment patterns, making it difficult to adjust monitoring strategies in a timely manner according to dynamic changes in the groundwater environment. Furthermore, existing systems largely centralize data processing and computation in the cloud, resulting in high response latency, significant communication bandwidth constraints, and insufficient monitoring capabilities and diffusion trend prediction capabilities for multi-layered aquifer pollution.

[0007] Furthermore, existing groundwater monitoring systems that utilize edge computing typically only offload general data preprocessing tasks such as data cleaning, format conversion, or outlier removal to the edge. The edge primarily handles data forwarding or lightweight preprocessing, and cannot perform dynamic parameter correction of hydrogeological models, pollution diffusion trend prediction, or regional risk index calculation locally within the monitoring area. Therefore, existing edge computing architectures struggle to address the challenges of massive edge data uploads, cloud-based centralized computing latency, and the inability to timely couple dynamic hydrogeological parameters in groundwater pollution diffusion prediction.

[0008] To address the above problems, the present invention provides the following technical solution:

[0009] A groundwater intelligent monitoring system based on cloud-edge-device collaboration and dynamic optimization includes:

[0010] The sensing layer end side includes an intelligent monitoring terminal, which acquires water quality parameters of at least one target aquifer through multi-layer nested monitoring wells and forms end-side collected data.

[0011] The edge layer includes edge computing nodes deployed at locations corresponding to the monitoring clusters. Each edge computing node contains a basic three-dimensional hydrogeological model and is configured to: receive edge-side data collected from each intelligent monitoring terminal within its monitoring cluster; dynamically correct the transport parameters of the basic three-dimensional hydrogeological model using flow velocity or direction parameters and pollutant concentration gradients from the edge-side data, generating a dynamic coupling model; run the dynamic coupling model to predict groundwater pollution diffusion trends, and output assessment results and early warning information.

[0012] The cloud layer is used to receive the assessment results and early warning information, and generate system optimization strategies based on changes in pollution risk status. The system optimization strategies are then converted into optimization configuration instructions and sent to the edge layer and the sensing layer to adjust the monitoring network density and data acquisition frequency.

[0013] This system achieves real-time perception and dynamic response to changes in the groundwater environment by hierarchically processing monitoring tasks at the cloud, edge, and sensing layers, thereby reducing data communication volume. Unlike conventional edge computing, which only decentralizes data preprocessing, caching, or format conversion tasks to the edge, this invention fully deploys the hydrogeological model engine on the edge computing nodes, performing dynamic dispersion coefficient correction, dynamic coupling model generation, pollution diffusion trend prediction, and regional risk index calculation on the edge side. Under normal operating conditions, the cloud layer does not receive all edge-side collected data, but instead receives the assessment results, early warning information, and observation feature matrices or statistical summary data output by the edge layer for use by the cloud-based time-series prediction model. It mainly performs pollution risk status comparison, system optimization strategy generation, and optimization configuration command issuance. This forms a collaborative division of labor of "edge-side model inference and risk quantification, and cloud-side strategy optimization and global scheduling," which is different from the simple data forwarding or lightweight preprocessing mode in non-domain-specific general edge computing architectures.

[0014] Preferably, the sensing layer end also includes a solar power supply module and a wireless communication module; the multi-layer nested monitoring well includes a borehole wall, inside which shallow monitoring well pipes, middle monitoring well pipes and deep monitoring well pipes are arranged side by side along the borehole axis; the shallow monitoring well pipes, middle monitoring well pipes and deep monitoring well pipes are respectively connected to screen pipe sections at positions corresponding to different target aquifers; in the annular gap between the borehole wall and each level of monitoring well pipes and in the gap between adjacent monitoring well pipes, bentonite water-stopping material is filled at the positions corresponding to the aquitard in the natural strata, and quartz sand filter material is backfilled at the positions corresponding to the aquifers, so that each screen pipe section is only connected to the corresponding target aquifer.

[0015] This structure utilizes multi-layered nested monitoring wells to divide the same borehole into multiple independent monitoring layers. Bentonite sealing material blocks the hydraulic connection between different aquifers, preventing cross-contamination of water between layers and improving the accuracy of obtaining water quality parameters for a single aquifer. The layered, independent water quality data provided by the multi-layered nested monitoring wells is the data foundation for the edge-layer dynamic coupling model to correct migration parameters layer by layer and grid by grid. Without this structure, mixed samples within the same borehole would be unable to distinguish the pollution contributions of different aquifers, reducing the vertical resolution in the dynamic coupling model and causing distortion in the pollution migration simulation. Therefore, the multi-layered nested monitoring well structure and the cloud-edge-device collaborative data processing method of this invention are mutually supportive and belong to a unified inventive concept.

[0016] Furthermore, the edge layer includes edge computing nodes, local data storage units, and a hydrogeological model engine. The local data storage unit stores the basic three-dimensional hydrogeological model. The edge computing nodes complete, clean, and spatially map the edge-acquired data, and input the processed edge-acquired data into the basic three-dimensional hydrogeological model to perform grid-by-grid correction on the migration parameters of the basic three-dimensional hydrogeological model, obtaining the dynamic coupling model. The hydrogeological model engine runs the dynamic coupling model at the edge and performs groundwater pollution diffusion trend prediction and edge-side risk index calculation at the edge, generating and outputting the assessment results and early warning information. Introducing a geological model engine at the edge for local processing of high-frequency monitoring data shortens the data processing path and improves the efficiency of early warning information generation.

[0017] Furthermore, the cloud layer includes a cloud server cluster, a long-term historical database, and a decision support system. The decision support system receives the assessment results and early warning information, and combines them with historical geological data, annual water quality monitoring records, and observation feature matrices or statistical summary data uploaded from the edge layer in the long-term historical database to generate the pollution risk status for the current period. The decision support system also compares the pollution risk status for the current period with the pollution risk status stored in the previous monitoring period, and generates the system optimization strategy when it is determined that the pollution risk status has changed. The cloud server cluster converts the system optimization strategy into optimization configuration instructions and issues them.

[0018] A second aspect of this invention provides a groundwater intelligent monitoring method based on cloud-edge-device collaboration and dynamic optimization, applied to the aforementioned groundwater intelligent monitoring system based on cloud-edge-device collaboration and dynamic optimization, comprising the following steps:

[0019] The cloud server cluster generates an initial network deployment plan, and the decision support system determines the deployment location and issues data collection instructions based on the initial network deployment plan.

[0020] According to the data acquisition instructions, the intelligent monitoring terminal acquires water quality parameters of at least one target aquifer through a multi-layer nested monitoring well and generates end-side acquisition data at the end side.

[0021] Edge computing nodes receive the data collected from the end side and use the flow velocity or flow direction parameters and pollutant concentration gradient in the data collected from the end side to dynamically correct the migration parameters of the basic three-dimensional hydrogeological model, thereby generating a dynamic coupling model.

[0022] The hydrogeological model engine runs the dynamic coupling model on the edge side to predict the spread trend of groundwater pollution, calculates the edge side risk index, and outputs assessment results and early warning information.

[0023] The decision support system generates a pollution risk status based on the assessment results and early warning information, and executes scheduling tasks in parallel when it determines that the pollution risk status has changed, and summarizes and generates a system optimization strategy.

[0024] The cloud server cluster translates the system optimization strategy into optimization configuration instructions and sends them to the edge layer and sensing layer to adjust the monitoring network density and data acquisition frequency, and then returns to execute the step of obtaining water quality parameters.

[0025] This method enables closed-loop control that adjusts the monitoring network parameters in reverse based on real-time monitoring data.

[0026] Preferably, the steps of generating an initial network deployment scheme by the cloud server cluster, determining deployment locations based on the initial network deployment scheme by the decision support system, and issuing the data acquisition command include: extracting historical geological data, pollution source distribution coordinates, and groundwater sensitive protection target distribution information of the target monitoring area, and discretizing the continuous geographic space into candidate grids; calculating the deployment fitness value of each candidate grid as a monitoring node using the initial deployment fitness evaluation formula; selecting candidate grids to generate the initial network deployment scheme based on the descendingly ordered deployment fitness values, under the condition of satisfying the minimum point spacing constraint; and determining the deployment location and monitoring density of the intelligent monitoring terminal based on the initial network deployment scheme by the decision support system, and issuing the data acquisition command.

[0027] Preferably, the step of generating a dynamically coupled model by dynamically correcting the transport parameters of the basic three-dimensional hydrogeological model using the flow velocity or direction parameters and pollutant concentration gradient in the edge computing node based on the end-side acquired data includes: calling the basic three-dimensional hydrogeological model from the local data storage unit; extracting the absolute value of groundwater flow velocity from the end-side acquired data or the estimated flow velocity determined based on the groundwater level difference and spatial distance between adjacent monitoring points, and extracting the absolute value of the spatial gradient of pollutant concentration; using the dynamic dispersion coefficient evaluation formula to correct the transport parameters in the model grid of the basic three-dimensional hydrogeological model point by point, and calculating the normalized dynamic dispersion coefficient; performing denormalization processing on the normalized dynamic dispersion coefficient, and writing the denormalized dispersion coefficient into the corresponding grid node of the basic three-dimensional hydrogeological model, so that the static geological skeleton is integrated with the real-time updated hydrodynamic parameters to obtain the dynamically coupled model. This step introduces a dynamic parameter correction mechanism, which improves the model's fit to the current hydrogeological conditions.

[0028] Furthermore, the steps of the hydrogeological model engine running the dynamic coupling model on the edge side to predict the groundwater pollution diffusion trend and calculating the edge-side risk index, and outputting assessment results and early warning information include: inputting the groundwater seepage velocity vector, dynamic dispersion coefficient, source-sink term, and boundary conditions in the dynamic coupling model as physical constraint parameters into the pollution migration calculation process to obtain the predicted pollutant concentration distribution field; using a multi-dimensional pollution risk assessment formula, calculating the edge-side risk index of the target grid based on the predicted concentration of the target pollutant and the migration rate of the pollution plume front; comparing the edge-side risk index with the risk warning threshold; when the edge-side risk index is greater than or equal to the risk warning threshold, defining the data matrix containing the risk index of each grid as the assessment result, and encapsulating the area coordinates and the time of exceedance exceeding the risk warning threshold as early warning information. The risk warning threshold is determined by the upper limit of the 95% confidence interval of pollution-free normal fluctuation data in a long-term historical database.

[0029] Furthermore, the decision support system generates a pollution risk status based on the assessment results and early warning information, and executes scheduling tasks in parallel when it determines that the pollution risk status has changed. The steps of summarizing and generating system optimization strategies include: parsing the assessment results and early warning information to generate the pollution risk status for the current period, and comparing the pollution risk status for the current period with the pollution risk status stored in the previous monitoring period; determining that the pollution risk status has changed when the risk level of the target grid increases or decreases, or when the number of target grids exceeding the risk early warning threshold changes; and triggering scheduling tasks in parallel to increase the monitoring network density and collection frequency in high-risk areas, decrease the data collection frequency in low-risk areas, and generate suggestions for deploying new temporary monitoring points when the pollution risk status has changed.

[0030] The scheduling task of increasing the monitoring network density and acquisition frequency in high-risk areas includes: using an adaptive acquisition frequency adjustment formula, calculating the update acquisition frequency of the intelligent monitoring terminal corresponding to the target grid based on the risk difference between the cloud fusion risk index and the risk warning threshold of the target grid node at the current moment, and using the nonlinear smoothing characteristics of the hyperbolic tangent function to avoid numerical abrupt changes during the frequency adjustment process.

[0031] The scheduling task for generating deployment suggestions for new temporary monitoring points includes: using a blind zone deployment benefit evaluation formula, calculating the deployment benefit value of candidate points based on the Kriging interpolation variance of the candidate points and the predicted exceedance of pollutant concentrations relative to national standard limits, and using spatial coordinate points that meet the deployment benefit threshold, minimum spacing between points, drilling conditions, and communication coverage conditions as deployment suggestions for new temporary monitoring points (the deployment benefit threshold is set to 1.5 times the mathematical expectation of the deployment benefit values ​​of all candidate points in the entire area).

[0032] The system optimization strategy is generated by summarizing the updated collection frequency, low-risk area dormancy adjustment parameters, and new temporary monitoring point deployment suggestions.

[0033] Preferably, the step of the intelligent monitoring terminal acquiring water quality parameters of at least one target aquifer through multi-layer nested monitoring wells and forming end-side acquired data according to the data acquisition instruction includes: waking up the corresponding sensor channel according to the data acquisition instruction, continuously sampling the same environmental indicator to extract water quality parameters, calculating the arithmetic mean of the remaining sampled values ​​after removing the maximum and minimum values ​​in the continuous sampling sequence; combining the arithmetic mean with the standard timestamp, the device's unique identifier, and the battery power status data of the solar power module into a structure and serializing it to form a standardized serialized text; encapsulating the serialized text to form the end-side acquired data; and writing the end-side acquired data into a non-volatile storage unit and marking it as pending transmission when the network recovers in the next communication cycle, when the wireless communication module does not receive a successful upload response signal within a preset timeout period.

[0034] In a preferred embodiment of the present invention, the decision support system generates a pollution risk status based on the assessment results and early warning information. The process also incorporates a long short-term memory neural network model to assist in the determination: the time-series predicted concentration output by the long short-term memory neural network model is weighted and calculated with the diffusion simulation concentration obtained by analyzing the assessment results and early warning information to obtain a fused predicted concentration. Based on this fused predicted concentration, a cloud-based fused risk index is calculated to quantitatively generate the pollution risk status. This combination of a data-driven model and a physical mechanism model reduces the prediction error of a single model.

[0035] In a preferred embodiment of the present invention, after the edge computing node receives the edge-collected data, it further includes an execution step of performing multi-source time-series preprocessing on the edge-collected data: identifying the time missing span of the edge-collected data; when the time missing span is less than the maximum missing time tolerance threshold, performing linear interpolation to complete the edge-collected data; when the time missing span is greater than or equal to the maximum missing time tolerance threshold, calling local historical average data for replacement and marking downgraded trust level; wherein, the maximum missing time tolerance threshold is set internally by the system.

[0036] In a preferred embodiment of the present invention, after the edge computing node receives the end-side collected data, it further includes an edge fault self-diagnosis execution step: based on the consistency of the end-side collected data uploaded by each intelligent monitoring terminal within the same cluster, calculations are performed, and when the readings of one or more sensors continuously deviate from the normal fluctuation range of other terminals within the same cluster and exceed the set standard deviation threshold, a fault alarm is automatically uploaded to the cloud server cluster to trigger a maintenance work order; wherein, the standard deviation threshold is set to 3 standard deviations.

[0037] In a preferred embodiment of the present invention, before the hydrogeological model engine runs the dynamically coupled model to predict the diffusion trend of groundwater pollution, it further includes a training and solidification execution step of a physical information neural network proxy model: extracting historical geological data and historical measured concentration values ​​to construct a supervised sample dataset; adding the residual of the groundwater convection dispersion equation as a physical constraint penalty term to the loss function containing the supervised sample dataset for iterative training until the loss function converges to obtain a solidified weight model; and distributing the solidified weight model to the hydrogeological model engine as the basic architecture for prediction inference. Introducing physical laws as constraints into the neural network improves the model's generalization ability and physical rationality under conditions of scarce samples.

[0038] This invention provides a groundwater intelligent monitoring system and method based on cloud-edge-device collaboration and dynamic optimization. It has the following beneficial effects:

[0039] 1. This invention adopts a cloud-edge-device collaborative architecture, deploying the hydrogeological model engine entirely on the computing nodes of the edge layer. Dynamic correction of hydrogeological transport parameters, model coupling, and pollution diffusion trend prediction are completed locally at the edge. Under normal circumstances, the cloud layer no longer receives all the data collected from the edge layer, but instead receives the evaluation results and early warning information output by the edge layer. Compared to traditional architectures that only use edge nodes for data forwarding or lightweight preprocessing, this collaborative division of labor mechanism effectively alleviates the latency problem caused by centralized cloud computing, reduces communication bandwidth pressure, and improves the system's early warning response speed to sudden groundwater pollution events.

[0040] 2. This invention receives assessment results and early warning information from the edge layer through a decision support system, generates system optimization strategies based on changes in groundwater pollution risk status, and dynamically adjusts the data acquisition frequency and monitoring network density at the sensing layer end. Through an adaptive acquisition frequency adjustment mechanism and blind zone deployment benefit assessment, the system can automatically increase the monitoring frequency in high-risk areas and generate temporary monitoring point deployment suggestions, while reducing the acquisition frequency in low-risk areas. Compared to traditional fixed data acquisition and network deployment modes, this mechanism ensures data capture rates in high-risk areas and monitoring blind zones while reducing monitoring resource consumption and equipment power consumption in low-risk areas, achieving dynamic closed-loop control of the monitoring strategy.

[0041] 3. This invention introduces a multi-layered nested monitoring well structure at the sensing layer end, arranging shallow, intermediate, and deep monitoring well pipes side-by-side along the same borehole axis. Bentonite sealing material is used to fill the corresponding natural aquifer locations, ensuring that each screen pipe section communicates only with a single target aquifer. This structure blocks the hydraulic connection between different aquifers, avoiding cross-contamination caused by groundwater mixing between layers, and providing accurate, layered, independent water quality data for the edge layer model. This data supports the dynamic coupling model in layer-by-layer, grid-by-grid correction of transport parameters, achieving coordinated prediction of horizontal and vertical migration of groundwater pollutants, and solving the problem of insufficient monitoring and prediction capabilities for multi-layered, three-dimensional pollution in existing systems. Attached Figure Description

[0042] Figure 1 This is a diagram illustrating the overall architecture of the intelligent groundwater environment monitoring system based on cloud-edge-device collaboration according to the present invention.

[0043] Figure 2 This is a flowchart of the adaptive monitoring network optimization process of the present invention;

[0044] Figure 3 This is a schematic diagram of the spatial topology between the sensing layer end and the edge layer of the present invention;

[0045] Figure 4 This is a diagram illustrating the architecture of the multi-level intelligent early warning system for groundwater pollution according to the present invention.

[0046] Figure 5 This is a schematic diagram of the single-hole multi-layer nested monitoring well structure of the present invention.

[0047] Among them, 101, intelligent monitoring terminal; 102, solar power supply module; 103, wireless communication module; 201, edge computing node; 202, local data storage unit; 203, hydrogeological model engine; 301, cloud server cluster; 302, long-term historical database; 303, decision support system; 501, borehole wall; 502, shallow monitoring well casing; 503, middle-layer monitoring well casing; 504, deep monitoring well casing; 505, bentonite water-stopping material; 506, quartz sand filter media; 507, screen pipe section; 508, aquitard; 509, aquifer. Detailed Implementation

[0048] The technical solutions in 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.

[0049] See attached document Figure 1 This invention provides a groundwater intelligent monitoring system based on cloud-edge-device collaboration and dynamic optimization, comprising: a sensing layer, an edge layer, and a cloud layer.

[0050] The sensing layer integrates an intelligent monitoring terminal 101, a solar power supply module 102, and a wireless communication module 103 on the end side. The intelligent monitoring terminal 101 includes a multi-parameter water quality sensor, a water level sensor, a flow velocity or flow rate detection unit, a positioning module, a non-volatile storage unit, and an end-side main control unit. The multi-parameter water quality sensor collects at least one water quality indicator from the following: pH value, dissolved oxygen, conductivity, turbidity, water temperature, ammonia nitrogen, target heavy metal ion concentration, and target organic matter concentration. The water level sensor collects groundwater level data. The flow velocity or flow rate detection unit obtains groundwater flow velocity or assists in calculating groundwater flow direction. The positioning module records the spatial location of the intelligent monitoring terminal 101. The non-volatile storage unit caches end-side collected data that was not successfully uploaded due to communication failure. After performing timing filtering and message encapsulation on the end side, the intelligent monitoring terminal 101 generates end-side collected data and uploads it via the wireless communication module 103.

[0051] The edge layer is configured with edge computing nodes 201, local data storage units 202 and hydrogeological model engine 203. It receives data collected from the edge through edge computing nodes 201 and performs data preprocessing. The hydrogeological model engine 203 performs simulation calculations to output evaluation results and early warning information.

[0052] The cloud layer includes a cloud server cluster 301, a long-term historical database 302, and a decision support system 303. It receives evaluation results and early warning information through the cloud server cluster 301, and the decision support system 303 generates optimization configuration instructions and sends the optimization configuration instructions to the edge layer and the perception layer.

[0053] It should be noted that conventional edge computing architectures typically only offload common tasks such as data preprocessing, caching, protocol conversion, or outlier removal to the edge, and their edge nodes do not undertake the complete reasoning and risk quantification of groundwater pollution diffusion models.

[0054] The difference between this invention and the conventional edge computing architecture described above lies in the following: This invention fully deploys the hydrogeological model engine 203 on the edge computing node 201, and stores the basic three-dimensional hydrogeological model in the local data storage unit 202. The edge computing node 201 uses the real-time flow velocity, flow direction parameters, and pollutant concentration gradient uploaded by the intelligent monitoring terminal 101 to dynamically correct the migration parameters in the basic three-dimensional hydrogeological model grid by grid, generating a dynamically coupled model. The hydrogeological model engine 203 independently completes the prediction of pollution diffusion trends and the calculation of edge-side risk indices on the edge side. Under normal operating conditions, the cloud server cluster 301 does not receive all edge-side collected data, but instead receives assessment results, early warning information, and observation feature matrices or statistical summary data for use by the cloud-based time-series prediction model. Therefore, this invention forms a division of labor mode of "edge-side model inference and risk quantification, cloud-side strategy optimization and global scheduling," which can avoid transmission congestion and early warning delays caused by massive edge-side data uploads and centralized cloud computing.

[0055] See attached document Figure 2 This invention provides a groundwater intelligent monitoring method based on cloud-edge-device collaboration and dynamic optimization, comprising the following steps:

[0056] S101, Initial monitoring network construction: The cloud server cluster 301 extracts historical geological data, pollution source distribution coordinates, and groundwater sensitive protection target distribution information from the long-term historical database 302 to plan the basic layout array and generate the initial network deployment scheme.

[0057] S102, Initial network configuration: The decision support system 303 determines the deployment location and monitoring density of the intelligent monitoring terminal 101 based on the initial network deployment plan, and issues data acquisition instructions.

[0058] S201, Real-time hydrogeological parameter acquisition: The intelligent monitoring terminal 101 acquires groundwater level, water quality indicators and flow velocity or flow direction parameters used to characterize the flow state of groundwater according to the data acquisition instructions, and forms the end-side acquired data after completing time-series filtering and message encapsulation at the end side.

[0059] S301, Dynamic hydrodynamic model coupling: Edge computing node 201 inputs the data collected from the end side into the basic model stored in the local data storage unit 202 for computation, and obtains the dynamic coupling model.

[0060] S302, Pollution migration simulation, Hydrogeological model engine 203 runs a dynamic coupled model to predict the diffusion trend of groundwater pollution, and outputs assessment results and early warning information;

[0061] The decision support system 303 generates the pollution risk status for the current period based on the assessment results and early warning information, and compares the current pollution risk status with the pollution risk status stored in the previous monitoring period. If the current period is the initial operating period, the baseline risk status generated based on historical data during the initial network deployment is used as the pollution risk status for the previous monitoring period. When the risk level of the target grid increases or decreases, or exceeds the risk warning threshold... When the number of target grids changes, the pollution risk status is determined to change.

[0062] When the pollution risk status is determined to change, the following steps are executed in parallel: step S401, increasing the monitoring network density and collection frequency in high-risk areas; step S402, reducing the data collection frequency in low-risk areas; and step S403, generating deployment suggestions for new temporary monitoring points, in order to summarize and generate a system optimization strategy.

[0063] The configuration instructions are sent to the terminal and edge nodes. The cloud server cluster 301 converts the system optimization strategy into optimization configuration instructions and sends them to adjust the hardware operating status. Then, it returns to the execution step S201.

[0064] If the pollution risk status is determined to have not changed, return directly to step S201.

[0065] To achieve the complete technical objective of the intelligent groundwater monitoring system and method based on cloud-edge-device collaboration and dynamic optimization of this invention, the specific execution logic and parameter calculation methods of each of the above steps will be described in detail below. In this embodiment, MQTT refers to Message Queuing Telemetry Transport Protocol, CoAP refers to Restricted Application Protocol, JSON refers to JavaScript Object Notation, UTC refers to Coordinated Universal Time, GPS refers to Global Positioning System, LoRa refers to Long-Range Low-Power Wireless Communication Technology, APP refers to Mobile Application, Web platform refers to the management platform accessed through a web browser, LSTM refers to Long Short-Term Memory Neural Network, GRU refers to Gated Recurrent Unit, and PINN refers to Physical Information Neural Network.

[0066] After the system completes the physical deployment at each level, to ensure the effectiveness and coverage of the groundwater monitoring network, the cloud server cluster 301 and the decision support system 303 collaboratively execute steps S101 and S102. As a preferred implementation, the specific execution logic of the above process is divided into the following sub-steps:

[0067] Step S1011: Data Retrieval and Spatial Discretization. The cloud server cluster 301 extracts historical geological data, pollution source distribution coordinates, and groundwater sensitive protection target distribution information of the target monitoring area from the long-term historical database 302 to plan the basic deployment array. In this embodiment, the long-term historical database 302 can be deployed based on a PostgreSQL database combined with the PostGIS spatial extension module to support the storage and spatial computation of large-scale spatial data. The historical geological data extracted here includes three-dimensional geological modeling data of the target area, aquifer permeability parameters, and historical water quality sampling records.

[0068] Using the above data, the cloud server cluster 301 performs orthogonal grid division on the target monitoring area, discretizing the continuous geographic space into multiple candidate grids with independent three-dimensional coordinates. To avoid dimensional conflicts of different physical parameters in subsequent calculations, the cloud server cluster 301 adopts the maximum-minimum normalization method to uniformly map the extracted concentration, permeability, and distance values ​​to the dimensionless [0,1] interval.

[0069] Step S1012: Calculation of candidate grid placement fitness. To address the technical challenges of traditional grid placement methods that rely heavily on manual experience and struggle to simultaneously consider pollution intensity, pollution gradient, formation permeability, and the impact of pollution source distance, the cloud server cluster 301 uses an initial grid placement fitness evaluation formula to calculate the placement fitness value for each candidate grid. This evaluation process comprehensively characterizes the pollutant concentration level, spatial gradient of pollutant concentration, and geological media permeability within the target area to screen monitoring nodes that reflect the evolution characteristics of groundwater pollution. The initial grid placement fitness evaluation formula is:

[0070] ;

[0071] in, This represents the fitness value of the candidate grid points. Indicates the candidate grid number; This represents the concentration weighting coefficient, and its value ranges from [0,1]. This represents the average historical pollutant concentration after normalization of the candidate grid. This represents the gradient weight coefficient, and its value ranges from [0,1]. This represents the spatial gradient of historical pollutant concentrations after normalization of the candidate grid, where Represents the gradient operator; Let represent the penetration weight coefficient, which takes values ​​in the range [0,1] and satisfies . ; This represents the aquifer permeability coefficient after the candidate grid is normalized; Represents the natural constant; This represents the normalized linear Euclidean distance from the candidate grid center to the nearest known pollution source; This represents the distance attenuation factor. Based on the above principle, after calculation, the cloud server cluster 301 sorts all candidate grids in descending order of their fitness values, and determines the theoretical number of points based on the target monitoring area, the effective monitoring radius of a single intelligent monitoring terminal 101, the number of available terminals, and the minimum point spacing constraint. Under the condition of satisfying the minimum point spacing constraint, the cloud server cluster 301 sequentially selects candidate grids with higher fitness values ​​as theoretical points in the basic point array, and then summarizes these to generate the initial network deployment scheme.

[0072] Step S1021: Monitoring node physical coordinate mapping. The decision support system 303 determines the deployment location of the intelligent monitoring terminal 101 based on the initial network deployment scheme. In actual engineering implementation, theoretical locations often face terrain constraints. If the theoretical location is obstructed by surface buildings or exposed bedrock, the decision support system 303 will search for coordinates closest to the origin and meeting drilling conditions within a circular area centered on the theoretical location and bounded by a preset offset radius. To prevent the algorithm from getting stuck in an unsolvable dead zone, if no usable coordinates are found within the preset offset radius, the decision support system 303 will gradually expand the search radius by a fixed step size. When the search range is expanded to the set maximum search boundary and no effective coordinates are found, the system marks the candidate grid as unusable and extracts a sequential replacement node from the candidate list of the basic deployment array, thereby finally determining the actual deployment location of each intelligent monitoring terminal 101.

[0073] Step S1022: Initial monitoring parameters and reference frequency allocation. After determining the deployment locations, the decision support system 303 further determines the monitoring density of each area. By calculating the number of adjacent nodes within a preset range around each intelligent monitoring terminal 101, if this number is greater than or equal to the density determination threshold, the corresponding area is classified as a high-density clustered area. The intelligent monitoring terminals 101 in this area are configured to collect multiple water quality indicators, including heavy metal ions and organic matter concentration, and are matched with a high-frequency reference acquisition frequency; conversely, if the number of adjacent nodes is less than the threshold, it is classified as a low-density sparse area, where the equipment is only configured to collect basic indicators such as water level, water temperature, and conductivity, and is matched with a low-frequency reference acquisition frequency.

[0074] Step S1023: Control command encapsulation and issuance. Combining the determined deployment locations and monitoring density with the corresponding hardware and software operating parameters, the decision support system 303 encapsulates the data acquisition commands as a whole. Subsequently, the cloud server cluster 301, using a wireless network architecture, issues the data acquisition commands to each intelligent monitoring terminal 101 deployed on-site. Regarding the specific networking and communication methods, those skilled in the art can typically utilize the MQTT protocol based on a publish / subscribe mechanism or the CoAP protocol for IoT-constrained environments to ensure reliable message delivery. Upon receiving the command, the device-side wireless communication module 103 drives the sensor probe into a normalized monitoring state. The underlying network connection establishment and heartbeat maintenance methods are conventional techniques in this field and will not be elaborated upon here.

[0075] After the monitoring network configuration is completed in the cloud, the intelligent monitoring terminal 101 deployed on-site enters the routine operation phase. As a preferred implementation, the intelligent monitoring terminal 101 acquires groundwater level, water quality indicators, and flow velocity or direction parameters characterizing groundwater flow based on received data acquisition instructions. After performing time-series filtering and message encapsulation at the terminal side, it forms the acquired data. This data acquisition and processing phase specifically includes the following sub-steps:

[0076] Step S2011: Data Acquisition Task Parsing and Hardware Channel Wake-up. In this embodiment, the main control unit inside the intelligent monitoring terminal 101 (e.g., an STM32L4 series low-power microcontroller based on the ARM Cortex-M4 core) performs periodic parsing of the data acquisition instructions residing in local memory, extracting the sampling frequency parameters and sensor activation channel configuration table contained therein. To control the overall power consumption of the independent field device, the intelligent monitoring terminal 101 maintains a sleep state during non-sampling periods. When the built-in real-time clock triggers the set sampling time node, the main control unit issues a power management command, instructing the solar power module 102 to provide operating voltage (such as standard 12V or 24V DC power) to the sensor bus, thereby waking up the water level probes and multi-parameter water quality electrodes connected to each physical channel.

[0077] Step S2012: Environmental parameter reading and time-series filtering and cleaning. The awakened sensor array begins measurement operations to acquire groundwater level, water quality indicators, and flow velocity or direction parameters characterizing the groundwater flow state. In actual groundwater monitoring environments, physical disturbances from suspended particles within the water body or transient noise from circuits can easily cause abrupt changes in single-sample values. To smooth out such data fluctuations, the intelligent monitoring terminal 101 performs at least three array-style samplings of the same environmental indicator within a set time window. The maximum and minimum values ​​in the continuous sampling sequence are removed using a sorting algorithm, and then the arithmetic mean of the remaining sampled values ​​is calculated and used as the parameter value for the current period. As a preferred method, the time window is set to 3 to 5 minutes, and the number of continuous samplings is set to 5 to 15 times. The specific code writing methods for the underlying analog-to-digital conversion logic are conventional techniques in this field and will not be elaborated here.

[0078] Step S2013: Data structure serialization and communication message encapsulation. After obtaining the filtered and cleaned parameters, the intelligent monitoring terminal 101 needs to format them to meet the requirements of cross-level network transmission. The main control unit combines the extracted groundwater level, water quality indicators, and flow parameters with the current standard timestamp (usually recorded in UTC time format), the device's unique identifier, and the current battery power status data of the solar power module 102 into a structure. According to the pre-set communication protocol, the system serializes the above structured data stream into JSON format text or a custom binary byte order, thereby forming standardized end-side acquired data.

[0079] Step S2014: Data upload and status reset on the terminal side. After data encapsulation is completed, the intelligent monitoring terminal 101 activates the wireless communication module 103 to establish a communication link.

[0080] As a preferred approach, the wireless communication module 103 can employ a 4G, 5G, NB-IoT, or LoRa communication module. To address the risk of program deadlock and data loss caused by network fluctuations in the field, the intelligent monitoring terminal 101 internally activates a timeout timer. If the wireless communication module 103 receives a successful upload response signal from the edge computing node 201 within the preset timeout period, the intelligent monitoring terminal 101 records the current successful transmission status; conversely, if no response signal is received within the preset timeout period, the intelligent monitoring terminal 101 determines that the communication has failed, writes the data collected at the terminal side into a local non-volatile storage medium, and marks it as pending transmission, so that it can be merged and retransmitted when the network recovers in the next communication cycle. After completing the above status processing, the intelligent monitoring terminal 101 disconnects the power supply circuit between the sensor probe and the wireless communication module 103, re-enters a low-power sleep mode, and waits for the arrival of the next physical sampling cycle specified by the data acquisition command.

[0081] See attached document Figure 3 Multiple intelligent monitoring terminals 101 deployed in the field are divided into several monitoring clusters based on data similarity and spatial distribution. Specifically, the system uses hydrogeological data from all monitoring locations in the initial monitoring network as sample data, extracting groundwater level, target pollutant concentration, aquifer permeability coefficient, and flow velocity or direction parameters corresponding to each monitoring location to form a hydrogeological feature vector. Edge computing nodes 201 or cloud server clusters 301 calculate the hydrogeological feature distance between any two monitoring locations and combine it with the actual geographical distance between them to form a comprehensive clustering distance, thereby clustering the intelligent monitoring terminals 101. After clustering, the system deploys at least one edge computing node 201 at the geographical center of each cluster, ensuring high data comparability among the intelligent monitoring terminals 101 managed by each edge computing node 201. The edge computing node 201 is used to receive edge-side collected data uploaded by each terminal within the monitoring cluster.

[0082] In this embodiment, the edge computing node 201, in collaboration with its internal local data storage unit 202 and hydrogeological model engine 203, jointly executes the dynamic hydrodynamic model coupling step S301 and the pollution migration simulation step S302. Simultaneously, the edge computing node 201 calculates dynamic health indicators of groundwater resources based on groundwater level, water quality indicators, flow velocity, or flow direction parameters uploaded by the managed intelligent monitoring terminal 101. These dynamic health indicators include the deviation from the historical average, the rate of change between adjacent monitoring periods, and the consistency deviation between nodes in the same cluster. The edge computing node 201 identifies the trend changes, periodic fluctuations, and abnormal patterns of groundwater status in the target area based on these dynamic health indicators and uploads the health dynamic assessment results as part of the assessment results to the cloud server cluster 301. The specific execution logic is divided into the following sub-steps:

[0083] Step S3011: Multi-source time-series data preprocessing and spatial mapping. Edge computing node 201 receives end-side collected data uploaded by terminals within its cluster via wireless LAN. To address packet loss or out-of-order delivery during field transmission, edge computing node 201 uses a linear interpolation algorithm to fill in missing timestamp data. To avoid interpolation distortion caused by prolonged disconnection, the system internally sets a maximum missing time tolerance threshold (preferably set to 2h to 6h). If the time span of missing data for a node is less than this maximum missing time tolerance threshold, linear interpolation is performed directly; if the missing time span is greater than or equal to this maximum missing time tolerance threshold, edge computing node 201 stops the interpolation operation for that node and instead calls the historical average data of the same intelligent monitoring terminal 101 in the same season and time period stored in the local data storage unit 202; when the historical data of the same intelligent monitoring terminal 101 is insufficient, the historical average data of adjacent grid nodes within the same time window is called to replace it, and the trust level of the replaced data points is downgraded.

[0084] During this process, to ensure the reliability of the sensing layer hardware, the edge computing node 201 also simultaneously performs edge-end fault self-diagnosis tasks. The edge computing node 201 performs calculations based on the consistency of data uploaded by each intelligent monitoring terminal 101 within its cluster. When the readings of one or more sensors of a certain intelligent monitoring terminal 101 continuously deviate from the normal fluctuation range of other terminals within the same cluster and exceed the set standard deviation threshold (in a preferred implementation, the standard deviation threshold can be set to 3 standard deviations), the edge computing node 201 determines that the sensor may have a hardware fault or probe contamination, and automatically uploads a fault alarm to the cloud server cluster 301, triggering the cloud platform to automatically generate a maintenance work order including the faulty terminal number, location coordinates, and suggested maintenance measures.

[0085] After completing the data supplementation, cleaning, and fault screening, the edge computing node 201 maps the valid groundwater level, water quality indicators, and flow velocity or flow direction parameters to the three-dimensional coordinate nodes of the pre-built geological grid based on the device's unique identification code, forming the observation data matrix for the current period.

[0086] Step S3012: Basic model reading and dynamic parameter coupling calculation. Edge computing node 201 retrieves the static basic three-dimensional hydrogeological model of the area from local data storage unit 202. To endow the static model with the ability to reflect changes in dynamic physical flow field, the system introduces a dynamic boundary condition update mechanism. Since groundwater velocity and pollutant concentration gradient directly affect the mechanical dispersion and diffusion driving force of pollutants, traditional simulation methods using constant dispersion coefficients are difficult to track changes in the pollution plume diffusion rate in a timely manner, easily causing prediction results to lag or deviate. Therefore, this invention introduces a dynamic dispersion coefficient evaluation formula at the edge side, incorporating the absolute value of groundwater velocity and the absolute value of the spatial gradient of pollutant concentration obtained in real time or estimated at the edge side into the migration parameter correction process of the model grid, in order to solve the technical problem that the static model cannot track changes in the pollution diffusion rate.

[0087] A static, basic 3D hydrogeological model is embedded in the local data storage unit 202 of the edge computing node 201. Flow velocity or direction parameters, pollutant concentrations, and the concentration gradients calculated from them, collected in real time at the edge, are used to correct the model parameters grid-by-grid, upgrading the basic 3D hydrogeological model from a static geological skeleton to a dynamically coupled model. This correction step is completed only at the edge; the cloud server cluster 301 does not participate in the point-by-point correction, thus avoiding massive data transmission at the edge and latency in centralized computing.

[0088] To eliminate differences in physical dimensions, all parameters input into this formula have been pre-normalized. The result is obtained from the following formula. The normalized dynamic dispersion coefficient of the target mesh node at the current moment:

[0089] ;

[0090] in, This represents the normalized dynamic dispersion coefficient of the target mesh node at the current moment; Indicates the current time; Indicates the target grid node number; Indicates the basic diffusion weight; This represents the basic molecular diffusion coefficient after normalization of the target grid nodes; This indicates the weighting factor for the influence of flow velocity; This represents the absolute value of the groundwater flow velocity at the target grid node at the current moment. The groundwater flow velocity can be directly acquired by the velocity or flow rate detection unit configured in the intelligent monitoring terminal 101. When the groundwater flow velocity is not directly acquired, the edge computing node 201 calculates the hydraulic gradient based on the groundwater level difference and spatial distance between adjacent monitoring points, and combines this with the aquifer permeability coefficient to calculate the estimated flow velocity of the target grid node. As a preferred method, the estimated flow velocity is determined according to the following relationship:

[0091] ;

[0092] in, This represents the aquifer permeability coefficient of the target grid node. The hydraulic gradient is determined by the difference in groundwater levels and the spatial distance between adjacent monitoring points. Indicates effective porosity; Indicates the concentration gradient weighting coefficient; This represents the pollutant concentration at the target grid node at the current moment. The absolute value of the spatial gradient, where This represents the gradient operator.

[0093] Edge computing node 201 performs denormalization on the normalized dynamic dispersion coefficient based on the preset dispersion coefficient range in the basic three-dimensional hydrogeological model, and writes the denormalized dispersion coefficient into the corresponding grid node of the dynamic coupling model. After the above correction process is completed, the static geological skeleton is integrated with the real-time updated hydrodynamic parameters to obtain the dynamic coupling model.

[0094] Step S3021: Prediction of pollution migration trends under physical constraints. The hydrogeological model engine 203 runs the aforementioned dynamically coupled model to predict the diffusion trend of groundwater pollution. As a preferred option, the hydrogeological model engine 203 integrates a physical information neural network proxy model built on the PyTorch deep learning framework. The specific structure of this neural network includes one input layer, four fully connected hidden layers (each layer configured with 128 neurons and using the Tanh activation function), and one output layer. The basic input data of the model are the grid three-dimensional spatial coordinates (x, y, z) and the prediction time step. The preprocessing logic is to normalize the coordinates and time to the [-1,1] interval; when the hydrogeological model engine 203 is in the model inference, it uses the groundwater seepage velocity vector, dynamic dispersion coefficient, source and sink terms and boundary conditions provided by the dynamic coupling model as physical constraint parameters to input the pollution migration calculation process; the output result is the predicted pollutant concentration value of the corresponding spatiotemporal node, and its physical meaning is the pollutant concentration level of the location in the future predicted time step.

[0095] For the model construction and training process, the cloud server cluster 301 extracts historical geological data and historical measured concentration values ​​from the long-term historical database 302 to construct a supervised sample dataset, where the sample label is defined as the corresponding measured concentration. During the training phase, the system incorporates the residual of the groundwater convection dispersion equation as a physical constraint penalty term into the loss function to address the problem that pure data-driven models are prone to violating physical laws under conditions of limited groundwater monitoring samples and scarcity of extreme pollution diffusion samples. In this embodiment, the residual of the groundwater convection dispersion equation is calculated using the formula for calculating the residual of the groundwater convection dispersion equation. The formula for calculating the residual of the groundwater convection dispersion equation is as follows:

[0096] ;

[0097] in, Represents the residuals of the equation; Indicates pollutant concentration; Indicates the current time; Represents the partial derivative operator; This represents the partial derivative of the pollutant concentration with respect to the current moment; This represents the groundwater seepage velocity vector; Represents the gradient operator; The spatial gradient vector representing the pollutant concentration; This represents the dot product operator; This represents the convection term, which is the dot product of the groundwater seepage velocity vector and the spatial gradient vector of the pollutant concentration. Indicates the dynamic dispersion coefficient; This represents the diffusion flux vector, which is the product of the dynamic diffusion coefficient and the spatial gradient vector of the pollutant concentration. Represents the divergence operator; This represents the dispersion term, which is the divergence of the product of the dynamic dispersion coefficient and the spatial gradient vector of the pollutant concentration. Indicates source and sink terms.

[0098] The overall training loss function is a weighted average of the supervised data error, the stack error, and the boundary condition error. The overall training loss function is:

[0099] ;

[0100] in, This represents the overall training loss function; Indicates the error weights of the supervised data; This represents the mean square error between the predicted pollutant concentration and the measured pollutant concentration. Indicates the weight of physical constraint error; This represents the mean square error of the equation residuals obtained from the formula for calculating the residuals of the groundwater convection dispersion equation; Indicates the boundary condition error weights; This represents the boundary condition error. As a preferred method, , , Boundary and initial conditions are provided by the basic three-dimensional hydrogeological model stored in local data storage unit 202. As a verification example, the training samples are constructed using monthly water quality monitoring data from 2015 to 2020, historical geological data, and historical measured concentration values ​​at corresponding monitoring points. The training epochs are set to 2000, and the Adam optimization algorithm is used for iterative weight updates. The initial learning rate is set to 0.001, and an exponential decay strategy is used to gradually reduce the learning rate. During training, when the validation set loss decreases below the preset allowable error range within 50 consecutive epochs, the model is considered converged. Subsequently, the model with fixed weights is deployed and run in the hydrogeological model engine 203.

[0101] Step S3022: Quantification and Status Determination of Multidimensional Risk Indicators. After obtaining the predicted pollutant concentration distribution field, to address the problem that simply relying on whether pollutant concentration exceeds the standard is insufficient to characterize the dynamic expansion threat of pollution plumes, the hydrogeological model engine 203 calculates the edge-side risk index for each region using a multidimensional pollution risk assessment formula. The assessment logic is as follows: the proportion of concentration exceeding the standard reflects the static environmental hazard level of the region, while the migration rate represents the potential threat of dynamic expansion of the pollution range. By incorporating both into the edge-side risk index, it is possible to avoid missing low-concentration but rapidly migrating pollution plumes and to avoid over-warning of high-concentration but slowly migrating localized pollution. The multidimensional pollution risk assessment formula is:

[0102] ;

[0103] in, This indicates the edge-side risk index of the target grid; This indicates the weighting factor for concentration exceeding the standard; This indicates the predicted concentration of the target pollutant within the target grid. This indicates the national standard limit concentration corresponding to the target pollutant. When multiple target pollutants exist in the same target grid, the concentration exceedance ratio for each target pollutant is calculated separately, and the maximum value or weighted average value is used as the concentration risk item for that target grid. Denotes the migration rate weighting coefficient, and satisfies... ; This represents the migration rate of the leading edge of the contamination plume in the target grid. The migration rate of the leading edge of the contamination plume is determined by the ratio of the spatial displacement of the leading edge position of the preset concentration isotope to the prediction time interval between two adjacent prediction periods. Indicates the baseline migration rate, and and All are preset parameters that are greater than zero.

[0104] The hydrogeological model engine 203 calculates the edge risk index of the target grid and compares it with the system's built-in risk warning threshold. A comparison was performed. This risk warning threshold... The risk index is determined empirically by technical personnel based on historical environmental damage assessment reports and safety redundancy criteria, with a standard range of 0.7 to 0.85. If the risk index at the edge of the target grid is greater than or equal to this risk warning threshold... If the current area shows a trend of spreading and deterioration, then the assessment results and early warning information will be output.

[0105] Step S3023: Package the assessment results and report the early warning information. The hydrogeological model engine 203 defines the data matrix containing the risk index of each grid as the assessment result, and reports data exceeding the risk warning threshold. The regional coordinates and the time of exceeding the standard are packaged into early warning information.

[0106] Finally, edge computing node 201 packages and transmits the aforementioned evaluation results, anomaly detection records, fault alarms, and early warning information to cloud server cluster 301 via a wide area network link. For edge-side collected data that has completed local preprocessing, model computation, and risk quantification, edge computing node 201 only uploads detailed data within the corresponding time window when cloud server cluster 301 issues a traceability request, model retraining request, or anomaly review request. Under normal operating conditions, edge computing node 201 does not upload the full amount of edge-side collected data to the cloud, but instead uploads evaluation results, early warning information, and observation feature matrices or statistical summary data for use by cloud time series prediction models, in order to reduce wide area network transmission volume and centralized computing pressure on the cloud. This data is then used by the upstream decision support system 303 for risk status assessment and network optimization scheduling.

[0107] See attached document Figure 4 Under this system architecture, after receiving assessment results and early warning information from the edge layer, the cloud server cluster 301 transmits them to the internally mounted decision support system 303. The decision support system 303 then performs a global pollution risk state evolution determination based on preset multi-level tiered rules. For example... Figure 4 As shown, the system architecture, from bottom to top, covers the data acquisition layer, data analysis layer, early warning judgment layer, and information dissemination layer.

[0108] In the data analysis layer of the aforementioned architecture, a long short-term memory neural network model combined with anomaly detection algorithms is used to extrapolate and identify temporal fluctuations in multidimensional groundwater indicators. As an alternative, a gated cyclic unit model can also be used for temporal trend prediction. Both the long short-term memory neural network model and the gated cyclic unit model are used to process time-dependent sequences of groundwater level, water quality, and meteorological geological parameters, and output predicted values ​​of target pollutant concentrations for a specified future time step.

[0109] To ensure the model's applicability in hydrogeological environments, this embodiment introduces a long short-term memory neural network model built on the PyTorch 1.12 framework, and deploys either an isolation forest or autoencoder models in parallel for anomaly pattern detection. Regarding data flow and preprocessing logic, the model's input data originates from observation feature matrices or statistical summary data uploaded by edge computing node 201, and is integrated with externally accessed meteorological data and regional geological baseline data. The observation feature matrix or statistical summary data is generated from data collected by intelligent monitoring terminal 101, which is then completed, cleaned, and spatially mapped by edge computing node 201. After preprocessing the aforementioned multi-source heterogeneous data through max-min normalization and a sliding window slice with a time step of 24 hours, a multi-dimensional time series matrix is ​​constructed. This multi-dimensional time series matrix specifically covers items such as groundwater level, water temperature, pH value, target pollutant concentration, and meteorological and geological parameters.

[0110] In terms of the internal hierarchical structure, this Long Short-Term Memory (LSTM) neural network model comprises, from bottom to top, an input layer, two stacked hidden layers, and a fully connected output layer. Each hidden layer is configured with 64 computational neurons equipped with forget gates, input gates, and output gates to extract hydrodynamic nonlinear coupling features over a long time span. The data output of this network architecture is the predicted concentration value at a specified future time step, corresponding to the actual physical state of the pollutant plume's migration and diffusion trend within the aquifer.

[0111] To ensure the model's reliable predictive capabilities, historical geological data and annual water quality monitoring records were extracted from the long-term historical database 302. The time-series segments were then used as feature samples, and the actual monitoring concentrations for the next time period were used as labels to construct the training set. During training, the system used mean squared error as the loss function and employed the Adam optimization algorithm to perform backpropagation of network parameters and weight iteration until the loss function's descent gradient fell below a set allowable error range for 50 consecutive cycles, at which point training was considered complete.

[0112] The decision support system 303 fuses the time-series predicted concentration output by the long short-term memory neural network model with the diffusion simulation concentration output by the hydrogeological model engine 203 to obtain the fused predicted concentration of the target grid. As a preferred method, the fused predicted concentration is determined as follows:

[0113] ;

[0114] in, This represents the fusion prediction concentration of the target grid. This indicates the diffusion simulation concentration output by the hydrogeological model engine 203. This represents the time-series predicted concentration output by the Long Short-Term Memory (LSTM) neural network model. The value represents the fusion weight and ranges from 0 to 1. The decision support system 303 performs spatiotemporal alignment and fusion processing on the multi-source monitoring data collected by the optimized monitoring network to generate a three-dimensional boundary and concentration gradient distribution map of the pollution plume. This map is then input into the built-in migration feature analysis engine for pollution behavior analysis. The analysis includes identifying pollution sources, determining the pollution diffusion rate and direction, assessing the evolution trend of the pollution range, and ultimately generating a pollution behavior report containing pollution source location information, diffusion rate predictions, and a list of affected sensitive targets.

[0115] Meanwhile, the decision support system 303 calculates a cloud-based fusion risk index based on fusion predicted concentrations and quantitatively divides the regional safety status into four tiers: a blue alert (attention level) indicating abnormal trends in indicators but not yet exceeding limits; a yellow alert (warning level) indicating single indicator exceeding limits but no spread detected; an orange alert (warning level) indicating multiple indicators exceeding limits and a spread trend detected; and a red alert (emergency level) indicating pollution has spread to sensitive target areas. Based on this, the decision support system 303 generates the pollution risk status for the current period and releases warning information in a tiered manner through multiple channels, including SMS, mobile app push notifications, and web platforms. Furthermore, the decision support system 303 accesses the historical case library stored in the long-term historical database 302, matches historical cases with similar pollution characteristics to the current situation, and generates emergency response strategy recommendations. Pollution characteristics include at least one of the following: pollution type, geological conditions, warning level, pollution plume migration direction, and type of affected sensitive targets. Emergency response strategy recommendations include at least one of the following: pollution source control, adding emergency monitoring points, or increasing sampling frequency in key areas, thereby assisting regulatory authorities in making rapid decisions.

[0116] The decision support system 303 compares the pollution risk status of the current period with the pollution risk status stored in the previous monitoring period. If the current period is the initial operating period, the baseline risk status generated based on historical data during the initial network deployment is used as the pollution risk status of the previous monitoring period. When the risk level of the target grid increases or decreases, or exceeds the risk warning threshold... When the number of target grids changes, the pollution risk status is determined to change.

[0117] When a change in pollution risk status is detected, the decision support system 303 triggers a dynamic scheduling task for field equipment in parallel, and generates a system optimization strategy. If the fusion prediction results indicate that the direction of pollution migration has changed compared to the previous monitoring period, the decision support system 303 re-determines the high-risk path based on the new migration direction of the pollution plume leading edge, and adjusts the monitoring resources from the original high-risk path to the new high-risk path. This specifically includes the following collaborative execution steps:

[0118] Step S401: Increase the density and sampling frequency of the monitoring network in high-risk areas. For grids that exceed the warning standard in the assessment results, shorten their sampling waiting interval to obtain high-density process data. When the distance between the pollution plume front and the groundwater sensitive protection target is less than the preset safe distance, or the expected arrival time is less than the preset emergency time threshold, the decision support system 303 triggers the emergency monitoring mode, increases the sampling frequency of the existing intelligent monitoring terminal 101 upstream of the sensitive protection target, the lateral boundary, and within the expected impact range, and generates temporary monitoring point deployment suggestions.

[0119] Step S402: To reduce the data acquisition frequency in low-risk areas: For safe areas that have been below the pollution baseline for a long time, extend the sensor sleep time, or implement temporary sleep control for intelligent monitoring terminals 101 that have been in a low-risk state for multiple consecutive monitoring cycles and have sufficient coverage from neighboring nodes. Intelligent monitoring terminals 101 in temporary sleep state retain low-frequency heartbeat reporting and undervoltage alarm functions, and resume normal acquisition tasks after receiving a wake-up command from the cloud server cluster 301 or edge computing node 201. This adaptive regulation can reduce the battery power consumption of the solar power module 102 and reduce the network throughput load when the wireless communication module 103 transmits data upwards.

[0120] In this embodiment, to achieve continuous adjustment of the acquisition frequency in both spatiotemporal dimensions and avoid hardware control jitter caused by step-like hard shearing, the decision support system 303 uses an adaptive acquisition frequency adjustment formula to calculate the updated acquisition frequency of the intelligent monitoring terminal corresponding to the target grid. The engineering problem addressed by this formula is that the unattended monitoring terminal in the field is limited by the capacity of the solar power module 102, the lifespan of the sensor, and the number of times the wireless communication module 103 can be activated. If a step-like adjustment is directly performed based on the risk level, the sampling frequency may suddenly jump from a low-frequency state to an extremely high-frequency state, leading to instability in the power supply system or excessive sensor wear. Therefore, this invention utilizes the nonlinear smoothing characteristics of the hyperbolic tangent function to map the dimensionless risk difference to a fixed frequency band, enabling the adjustment mechanism to balance high-risk area tracking and low-risk area energy saving. The adaptive acquisition frequency adjustment formula is:

[0121] ;

[0122] in, This indicates the update frequency of the intelligent monitoring terminal corresponding to the target grid; Indicates the target grid node number; Indicates the update status superscript; Indicates the reference acquisition frequency; This indicates that the frequency dynamically adjusts the weight, and ; Represents the hyperbolic tangent function; This represents the cloud fusion risk index of the target grid node at the current moment; Indicates the risk warning threshold; This represents the risk difference.

[0123] In the above formula, the following is used The function maps the risk difference to The range allows for smooth adjustment of the sampling frequency and automatic limitation within a certain range. . to Within the specified range, to avoid drastic fluctuations in the sampling frequency due to sudden risk changes, such as avoiding a sudden change from once per hour to once per minute, thereby protecting the sensor's lifespan and the power supply stability of the solar power module 102. This smooth adjustment mechanism is specifically designed for unmanned groundwater monitoring stations in the field, differing from the stepped or proportional adjustment methods in general IoT. In this embodiment, to achieve continuous adjustment of the acquisition frequency in both spatiotemporal dimensions and avoid hardware control jitter caused by hard step-like shearing, the decision support system 303 uses an adaptive acquisition frequency adjustment formula to calculate the updated acquisition frequency of the intelligent monitoring terminal corresponding to the target grid. The engineering principle of this formula is to utilize the nonlinear smoothing characteristics of the hyperbolic tangent function to map the dimensionless risk difference to a fixed frequency band, enabling the adjustment mechanism to balance high-risk area tracking and low-risk area energy saving. The adaptive acquisition frequency adjustment formula is:

[0124] ;

[0125] in, This indicates the update and acquisition frequency of the intelligent monitoring terminal corresponding to the target grid; Indicates the target grid node number; Indicates the update status superscript; Indicates the reference acquisition frequency; Represents the multiplication operator; Represents the constant 1; This indicates that the frequency dynamically adjusts the weight, and ; Represents the hyperbolic tangent function; This represents the cloud fusion risk index of the target grid node at the current moment; Indicates the risk warning threshold; This represents the risk difference, which is the difference between the cloud-integrated risk index of the target grid node at the current moment and the risk warning threshold. The hyperbolic tangent mapping value represents the risk difference; This represents the adjustment factor, which is the product of the frequency dynamic adjustment weight and the hyperbolic tangent mapping value of the risk difference; This represents the frequency adjustment coefficient, which is the sum of the constant 1 and the adjustment factor.

[0126] To avoid the data acquisition frequency exceeding the hardware's allowable range, the decision support system 303 further limits the update acquisition frequency of the intelligent monitoring terminal corresponding to the target grid to between a preset minimum acquisition frequency and a preset maximum acquisition frequency. As a preferred implementation, the aforementioned risk warning threshold is determined by the upper limit of the 95% confidence interval of the pollution-free normal fluctuation data in the long-term historical database 302, and its value range is typically set to 0.7 to 0.85.

[0127] Step S403 in generating recommendations for new temporary monitoring site deployment: Due to the three-dimensional heterogeneous characteristics of the groundwater dynamic field, the initial network deployment scheme is difficult to completely capture the secondary free trajectory of the pollution plume leading edge during long-term operation. To address the problem that traditional methods of selecting new monitoring sites rely solely on predicted concentration levels and easily overlook the interpolation uncertainty of the existing monitoring network, the decision support system 303 uses a blind zone deployment benefit assessment formula to calculate the deployment value of areas without fixed terminals. This logic is based on a cross-assessment of spatial interpolation uncertainty and environmental hazard level; if the predicted concentration in the area does not exceed the standard, the maximum value function sets the deployment benefit value of the corresponding candidate site to zero, thereby excluding candidate areas with predicted concentrations below the standard limit and prioritizing spatial locations with a risk of exceeding the standard and insufficient coverage by existing nodes. The blind zone deployment benefit assessment formula is as follows:

[0128] ;

[0129] in, This represents the benefit value of arranging candidate locations; The spatial coordinate vector representing the candidate point location; This represents the Kriging interpolation variance at the spatial location of the candidate point, used to characterize the spatial interpolation uncertainty of the existing monitoring network at that location; This represents the function that takes the maximum value. Indicates the predicted pollutant concentration at the candidate site; Indicates the concentration limit specified by the national standard; It is used to characterize the severity of environmental hazards when the concentration of pollutants exceeds the national standard limit.

[0130] When the predicted concentration at a candidate site does not exceed the national standard limit, this item is set to 0, preventing the system from recommending low-risk areas simply due to high spatial uncertainty. The product of these two factors achieves a joint assessment of "uncertainty × hazard," prioritizing locations that may exceed the standard and where existing monitoring networks cannot accurately interpolate, rather than simply selecting the point with the highest predicted concentration.

[0131] The decision support system 303 calculates the candidate coordinates across the entire region based on the above formula and compares the calculated placement benefit value of the candidate points with a preset placement benefit threshold. In this embodiment, the placement benefit threshold is set to 1.5 times the mathematical expectation of the placement benefit values ​​of all candidate points across the entire region. The system extracts spatial coordinates of candidate points whose placement benefit values ​​are greater than the placement benefit threshold, and further eliminates spatial coordinates that are less than the preset minimum placement spacing from the existing intelligent monitoring terminal 101, located in an area that cannot be drilled, or do not meet the communication coverage conditions. Finally, the remaining spatial coordinates are output as deployment suggestions for new temporary monitoring points.

[0132] As a preferred approach, to support the requirement for three-dimensional tracking of the vertical dimension of the groundwater dynamic field in the above deployment recommendations, the newly deployed nodes adopt a multi-layer parallel network configuration.

[0133] See attached document Figure 5 The cross-sectional structure of this monitoring node is constructed based on the outer borehole wall 501. Inside the borehole wall 501, shallow monitoring well pipes 502, intermediate monitoring well pipes 503, and deep monitoring well pipes 504 are arranged side by side along the borehole axis, with the water inlet depth of each level of monitoring well pipe corresponding to different target aquifers. Screen pipe sections 507 are connected to the bottom of the shallow monitoring well pipes 502, intermediate monitoring well pipes 503, and deep monitoring well pipes 504 or to the corresponding target aquifer location. In addition, automatic well-washing sampling components can be optionally installed in each level of monitoring well pipe. The automatic well-washing sampling component includes a sampling pipeline, a micro lift pump, a solenoid valve, a flow detection unit, and a sampling controller installed in the corresponding well pipe. The sampling controller is communicatively connected to the intelligent monitoring terminal 101. Before sampling, the sampling controller controls the micro lift pump to perform well-washing operation according to a preset drainage volume; when the drainage volume reported by the flow detection unit reaches the preset well-washing volume, the sampling controller switches the solenoid valve to the sampling channel and collects water samples in time intervals according to a preset time interval. The specific implementation of pump drive, valve control and sampling bottle switching in the automatic well washing sampling assembly can be accomplished by those skilled in the art using existing automatic sampling equipment. The mechanical transmission and electrical control connection methods are well-known technologies in the field and will not be elaborated here.

[0134] To ensure the independent acquisition of water quality parameters for each target aquifer, layered isolation filling is carried out in the annular gap between the borehole wall 501 and each level of monitoring well pipe, as well as in the gap between adjacent monitoring well pipes: bentonite water-stopping material 505 is filled at the position corresponding to the aquifer 508 in the natural stratum, and quartz sand filter material 506 is backfilled at the position corresponding to the aquifer 509 (such as shallow aquifer, middle aquifer and deep aquifer), so that each screen pipe section 507 is only connected to the corresponding target aquifer. As a typical engineering application size, the borehole diameter 501 of this nested monitoring well can be set to 700mm, and the overall well depth can reach 120m. Specifically, the screen section of the shallow monitoring well casing 502 is set to a depth of 0.30m to monitor shallow groundwater, the screen section of the intermediate monitoring well casing 503 is set to a depth of 4.07m to monitor intermediate confined water, and the screen section of the deep monitoring well casing 504 is set to a depth of 80-110m to monitor deep confined water. This casing isolation and sealing design effectively blocks vertical water flow, ensuring that independent water quality signals from each aquifer 509 are transmitted to the surface for reading by the externally connected intelligent monitoring terminal 101.

[0135] Finally, the cloud server cluster 301 transforms the system optimization strategy, which includes spatial coordinates and motion timing, into optimization configuration instructions and sends them to the edge layer and perception layer endpoints. Upon receiving these optimization configuration instructions, the intelligent monitoring terminal 101 updates the acquisition timing of its internal controller, thus forming a physical logic closed loop of perception, evaluation, and dynamic scheduling.

[0136] After completing the pollution risk assessment and control parameter calculation at the cloud layer, the system enters the cloud-edge-device end-to-end closed-loop feedback and dynamic configuration command distribution and control phase. The engineering objective of this stage is to ensure that the optimization scheme of the upper-level decision-making system can be translated into physical execution actions of remote devices, thereby completing the logical closed loop of dynamic environmental monitoring.

[0137] Based on the risk assessment conclusions output by the decision support system 303 according to the evaluation results and early warning information, the system triggers corresponding control actions according to the branch logic, which are specifically divided into the following two dimensions of processing mechanisms:

[0138] The specific execution process of sending configuration instructions to terminals and edge nodes is as follows: when the pollution risk status is determined to change, the cloud server cluster 301 converts the system optimization strategy into optimization configuration instructions and sends them to adjust the hardware operating status, and then returns to the step of collecting real-time hydrogeological parameters.

[0139] In this embodiment, after receiving the system optimization strategy generated by the decision support system 303, the cloud server cluster 301 initiates a downlink signaling encapsulation program. This system optimization strategy includes updated acquisition frequencies, sensor sleep adjustment parameters, and spatial mapping coordinate tables for temporary monitoring points to be explored, all redesigned for each coordinate grid. To adapt to the weak network environment of wireless transmission links in complex terrain, as a preferred approach, the cloud server cluster 301 internally carries an IoT message broker server based on the Eclipse Mosquitto architecture, and uses the existing MQTT protocol to serialize and package the aforementioned scheduling parameters, transforming them into optimization configuration instructions. Subsequently, these optimization configuration instructions are sent in parallel to the wireless communication module 103 on the edge computing node 201 and the sensing layer side via cellular mobile network channels. For the protocol frame header construction, heartbeat keep-alive detection, and breakpoint retransmission verification involved in the communication link, those skilled in the art can use existing IoT communication middleware for conventional configuration; the channel handshake logic is well-known in the field and will not be elaborated here.

[0140] After receiving the optimization configuration instruction, the edge computing node 201 synchronously updates the model runtime parameters in its local data storage unit 202, aligning the next exercise cycle step size of the hydrogeological model engine 203 when running the dynamic coupling model to simulate pollution migration with the frequency of the newly introduced data source acquisition. Simultaneously, the intelligent monitoring terminal 101 on the sensing layer side receives instruction fragments via the wireless communication module 103. In this embodiment, the intelligent monitoring terminal 101 internally carries a low-power microcontroller with an ARM Cortex-M4 core. This microcontroller extracts payload data based on its built-in message parsing tree and resets the interrupt reload register value of the general-purpose timer according to the new acquisition frequency.

[0141] To avoid power depletion due to frequent wake-ups and long-term communication unavailability of the intelligent monitoring terminal 101, the intelligent monitoring terminal 101 calls the analog-to-digital conversion interface to read the current battery terminal voltage feedback of the solar power supply module 102 for boundary verification before reconstructing the sleep-wake duration parameters.

[0142] As a preferred embodiment, the solar power module 102 includes a photovoltaic panel, an energy storage battery, and an intelligent charge / discharge management unit. The intelligent charge / discharge management unit adjusts the sensor power supply duration, the radio frequency activation time of the wireless communication module 103, and the sleep cycle of the main control unit based on the acquisition frequency, data reporting frequency, and local computing load. Under conditions where the illumination, energy storage capacity, and acquisition frequency meet the preset energy consumption budget, the intelligent monitoring terminal 101 can operate autonomously for extended periods without an external power source.

[0143] Under the conditions that the annual average effective sunshine duration is not less than a preset value, the energy storage battery capacity is not less than a preset capacity, and the intelligent monitoring terminal 101 operates according to the low-frequency reference acquisition frequency, the solar power supply module 102 can support the long-term continuous operation of the intelligent monitoring terminal 101. As a preferred method, the long-term continuous operation duration can be set to not less than 3 years. If the battery terminal voltage feedback is detected to be lower than the set undervoltage threshold, a delayed sleep offset is added based on the optimized configuration command. As a preferred parameter configuration standard, the undervoltage threshold is calibrated according to the type of energy storage battery built into the solar power supply module 102 and set to 0.8 to 0.95 times the rated voltage of the energy storage battery.

[0144] For example, for a 12V rated energy storage battery, the undervoltage threshold can be set to 9.6V to 11.4V. Meanwhile, to prevent excessively large additional sleep delay offsets from causing the device to remain disconnected from the network topology for extended periods, the system sets a maximum sleep duration threshold (e.g., 24 hours). When the calculated total sleep duration exceeds this threshold, it is forcibly adopted as the final sleep-wake cycle.

[0145] After confirming that the sampling clock and sleep / wake parameters of the intelligent monitoring terminal 101 have been updated, the system returns to the real-time hydrogeological parameter acquisition step, and the intelligent monitoring terminal 101 acquires groundwater level and water quality indicators and forms end-side acquisition data.

[0146] On the other hand, if it is determined that the pollution risk status has not changed, the process directly returns to the step of collecting real-time hydrogeological parameters.

[0147] In a scenario where groundwater quality and hydrodynamic field evolution are in a stable state, the system control core focuses on reducing the bandwidth consumption of redundant control signaling. In this business scenario, the decision support system 303 locks its status bit and does not dispatch new communication configuration events. The intelligent monitoring terminal 101 at the perception layer and the edge computing node 201 at the edge layer retain the timing and cycle parameters of the previous working cycle by default. The solar power module 102 maintains the predetermined charge and discharge protection threshold, and the wireless communication module 103 continues to open the radio frequency upload window according to the original time slot. If the pollution risk status remains unchanged, the system maintains the acquisition frequency, model running cycle, and communication upload cycle of the previous working cycle and returns to execute the real-time hydrogeological parameter acquisition step to continuously acquire baseline monitoring data of the target monitoring area under low power consumption.

[0148] As a verification example, for a 10 km² groundwater monitoring area in an industrial park, the following parameter configuration was adopted: candidate grid size is... In the initial site selection fitness evaluation formula, , , In the dynamic dispersion coefficient evaluation formula, , , In the multidimensional pollution risk assessment formula, , In the adaptive acquisition frequency adjustment formula, , Next / day, preset minimum sampling frequency Next / day, preset maximum sampling frequency Next / Day; Risk Warning Threshold The maximum tolerance threshold for missing time is 4 hours; the standard deviation threshold is 3 standard deviations. A 6-month simulation experiment was conducted under these parameters. The results show that: compared to the traditional fixed-frequency monitoring method, the pollution warning time of this invention is 4.2 hours earlier; compared to the edge architecture that only forwards data, the uplink data volume of this invention is reduced by 82%; because the intelligent monitoring terminal 101 in low-risk areas can extend its sleep time, the remaining power of the solar power module 102 is always above 40%; in simulation scenarios where the migration direction of the pollution plume changes, the system can readjust the collection frequency on high-risk paths based on the cloud-integrated risk index and output suggestions for deploying new temporary monitoring points.

[0149] As another application example, in a groundwater contaminated site at a depth of 120m, a single multi-layer nested monitoring well can be used to simultaneously monitor shallow unconfined water, intermediate confined water, and deep confined water. Shallow monitoring well pipe 502, intermediate monitoring well pipe 503, and deep monitoring well pipe 504 are connected to different target aquifers. Interlayer isolation and independent sampling are achieved through bentonite sealing material 505 and quartz sand filter media 506, thereby obtaining water quality difference data for aquifers at different depths and providing a basis for determining the vertical migration range of pollutants.

[0150] As an optional implementation, the present invention also provides a computer-readable storage medium storing a computer program. When the computer program is executed by the processor of the cloud server cluster 301, the edge computing node 201, or the intelligent monitoring terminal 101, it performs at least one of the following steps: initial monitoring network construction, real-time hydrogeological parameter acquisition, dynamic hydrodynamic model coupling, pollution migration simulation, risk status determination, system optimization strategy generation, and optimization configuration instruction issuance.

Claims

1. A groundwater intelligent monitoring system based on cloud-edge-device collaboration and dynamic optimization, characterized in that: include: The sensing layer end side includes an intelligent monitoring terminal (101), which acquires water quality parameters of at least one target aquifer (509) through multi-layer nested monitoring wells and forms end-side collected data. The edge layer includes edge computing nodes (201) deployed at locations corresponding to the monitoring clusters. Each edge computing node (201) contains a basic three-dimensional hydrogeological model and is configured to: receive end-side data collected from each intelligent monitoring terminal (101) within its monitoring cluster; dynamically correct the transport parameters of the basic three-dimensional hydrogeological model using the flow velocity or flow direction parameters and pollutant concentration gradients in the end-side data to generate a dynamic coupling model; run the dynamic coupling model to predict the diffusion trend of groundwater pollution and output assessment results and early warning information. The cloud layer is used to receive the assessment results and early warning information, and generate system optimization strategies based on changes in pollution risk status. The system optimization strategies are then converted into optimization configuration instructions and sent to the edge layer and the sensing layer to adjust the monitoring network density and data acquisition frequency.

2. The intelligent groundwater monitoring system based on cloud-edge-device collaboration and dynamic optimization according to claim 1, characterized in that, The hardware structure of the sensing layer end is as follows: The sensing layer also includes a solar power module (102) and a wireless communication module (103). The multi-layer nested monitoring well includes a borehole wall (501), inside which shallow monitoring well casing (502), medium monitoring well casing (503) and deep monitoring well casing (504) are arranged side by side along the borehole axis; The shallow monitoring well pipe (502), the middle monitoring well pipe (503), and the deep monitoring well pipe (504) are respectively connected to screen pipe sections (507) at the positions corresponding to different target aquifers (509); In the annular gap between the borehole wall (501) and each level of monitoring well pipe, and in the gap between adjacent monitoring well pipes, bentonite water-stopping material (505) is filled at the position corresponding to the aquitard (508) in the natural stratum, and quartz sand filter material (506) is backfilled at the position corresponding to the aquifer (509), so that each screen pipe section (507) is only connected to the corresponding target aquifer (509).

3. The intelligent groundwater monitoring system based on cloud-edge-device collaboration and dynamic optimization according to claim 1, characterized in that, The edge layer includes an edge computing node (201), a local data storage unit (202), and a hydrogeological model engine (203). The local data storage unit (202) is used to store the basic three-dimensional hydrogeological model; The edge computing node (201) is used to complete, clean and spatially map the edge-acquired data, and input the processed edge-acquired data into the basic three-dimensional hydrogeological model to correct the migration parameters of the basic three-dimensional hydrogeological model grid by grid to obtain the dynamic coupling model. The hydrogeological model engine (203) is used to run the dynamic coupling model on the edge side, and to complete the prediction of groundwater pollution diffusion trend and the calculation of edge side risk index on the edge side, and generate and output the assessment results and early warning information.

4. The intelligent groundwater monitoring system based on cloud-edge-device collaboration and dynamic optimization according to claim 1, characterized in that, The cloud layer includes a cloud server cluster (301), a long-term historical database (302), and a decision support system (303). The decision support system (303) is used to receive the assessment results and early warning information, and combine the historical geological data, annual water quality monitoring records and observation feature matrix or statistical summary data uploaded by the edge layer in the long-term historical database (302) to generate the pollution risk status of the current period; The decision support system (303) is also used to compare the pollution risk status of the current period with the pollution risk status stored in the previous monitoring period, and generate the system optimization strategy when it is determined that the pollution risk status has changed. The cloud server cluster (301) is used to convert the system optimization strategy into the optimization configuration instruction and issue it.

5. A groundwater intelligent monitoring method based on cloud-edge-device collaboration and dynamic optimization, characterized in that, The groundwater intelligent monitoring system based on cloud-edge-device collaboration and dynamic optimization as described in any one of claims 1-4 includes the following steps: The cloud server cluster (301) generates an initial network deployment plan, and the decision support system (303) determines the deployment location and issues a data collection command based on the initial network deployment plan; According to the data acquisition instructions, the intelligent monitoring terminal (101) acquires water quality parameters of at least one target aquifer (509) through a multi-layer nested monitoring well and generates end-side acquisition data at the end side. The edge computing node (201) receives the end-side collected data and uses the flow velocity or flow direction parameters and pollutant concentration gradient in the end-side collected data to dynamically correct the migration parameters of the basic three-dimensional hydrogeological model and generate a dynamic coupling model. The hydrogeological model engine (203) runs the dynamic coupling model on the edge side to predict the diffusion trend of groundwater pollution, calculates the edge side risk index on the edge side, and outputs the assessment results and early warning information. The decision support system (303) generates a pollution risk status based on the assessment results and early warning information, and executes scheduling tasks in parallel when it is determined that the pollution risk status has changed, and summarizes and generates a system optimization strategy; The cloud server cluster (301) converts the system optimization strategy into optimization configuration instructions and sends them to the edge layer and the sensing layer to adjust the monitoring network density and data acquisition frequency, and then returns to execute the step of obtaining water quality parameters.

6. The intelligent groundwater monitoring method based on cloud-edge-device collaboration and dynamic optimization according to claim 5, characterized in that, The steps of the cloud server cluster (301) generating an initial network deployment plan, and the decision support system (303) determining the deployment location and issuing the data collection command based on the initial network deployment plan include: Historical geological data, pollution source distribution coordinates, and groundwater sensitive protection target distribution information of the target monitoring area are extracted, and the continuous geographic space is discretized into candidate grids; The fitness value of each candidate grid as a monitoring node is calculated one by one using the initial deployment fitness evaluation formula. Based on the fitness values ​​of the points arranged in descending order, candidate grids are selected to generate the initial network layout scheme under the condition of satisfying the minimum point spacing constraint; The decision support system (303) determines the deployment location and monitoring density of the intelligent monitoring terminal (101) based on the initial network deployment scheme, and issues the data acquisition command.

7. The intelligent groundwater monitoring method based on cloud-edge-device collaboration and dynamic optimization according to claim 5, characterized in that, The edge computing node (201) uses the flow velocity or flow direction parameters and pollutant concentration gradients from the end-side collected data to dynamically correct the migration parameters of the basic three-dimensional hydrogeological model, and the steps to generate a dynamically coupled model include: The basic three-dimensional hydrogeological model is retrieved from the local data storage unit (202); Extract the absolute value of groundwater flow velocity from the data collected at the end point or the estimated flow velocity determined based on the groundwater level difference and spatial distance between adjacent monitoring points, and extract the absolute value of the spatial gradient of pollutant concentration; The migration parameters in the model grid of the basic three-dimensional hydrogeological model are corrected point by point using the dynamic dispersion coefficient evaluation formula, and the normalized dynamic dispersion coefficient is calculated. The normalized dynamic dispersion coefficient is denormalized, and the denormalized dispersion coefficient is written into the corresponding grid node of the basic three-dimensional hydrogeological model, so that the static geological skeleton is integrated with the real-time updated hydrodynamic parameters to obtain the dynamic coupling model.

8. The intelligent groundwater monitoring method based on cloud-edge-device collaboration and dynamic optimization according to claim 5, characterized in that, The steps of the hydrogeological model engine (203) running the dynamic coupling model on the edge side to predict the diffusion trend of groundwater pollution, calculating the edge side risk index, and outputting assessment results and early warning information include: The groundwater seepage velocity vector, dynamic dispersion coefficient, source-sink term, and boundary conditions in the dynamic coupling model are used as physical constraint parameters input into the pollution migration calculation process to obtain the predicted pollutant concentration distribution field. A multidimensional pollution risk assessment formula is used to calculate the edge risk index of the target grid based on the predicted concentration of the target pollutant in the target grid and the migration rate of the pollution plume leading edge. The edge-side risk index is compared with the risk warning threshold. When the edge-side risk index is greater than or equal to the risk warning threshold, the data matrix containing the risk index of each grid is defined as the evaluation result, and the area coordinates and the time of exceeding the risk warning threshold are encapsulated as warning information.

9. The intelligent groundwater monitoring method based on cloud-edge-device collaboration and dynamic optimization according to claim 5, characterized in that, The decision support system (303) generates a pollution risk status based on the assessment results and early warning information, and executes scheduling tasks in parallel when it determines that the pollution risk status has changed. The steps of summarizing and generating system optimization strategies include: The assessment results and early warning information are analyzed to generate the pollution risk status for the current period, and the pollution risk status for the current period is compared with the pollution risk status stored in the previous monitoring period. When the risk level of a target grid increases or decreases, or when the number of target grids exceeding the risk warning threshold changes, the pollution risk status is determined to have changed. When the pollution risk status is determined to change, scheduling tasks are triggered in parallel to increase the monitoring network density and collection frequency in high-risk areas, reduce the data collection frequency in low-risk areas, and generate suggestions for the deployment of new temporary monitoring points. Among them, the scheduling task of increasing the monitoring network density and collection frequency in high-risk areas includes: using an adaptive collection frequency adjustment formula, calculating the update collection frequency of the intelligent monitoring terminal (101) corresponding to the target grid based on the risk difference between the cloud fusion risk index and the risk warning threshold of the target grid node at the current moment; The scheduling task for generating deployment suggestions for new temporary monitoring points includes: using the blind zone deployment benefit evaluation formula, calculating the deployment benefit value of candidate points based on the Kriging interpolation variance of candidate points and the predicted exceedance of pollutant concentration relative to the national standard limit concentration, and using spatial coordinate points that meet the deployment benefit threshold, minimum spacing between points, drilling conditions, and communication coverage conditions as deployment suggestions for new temporary monitoring points. The system optimization strategy is generated by summarizing the updated collection frequency, low-risk area dormancy adjustment parameters, and new temporary monitoring point deployment suggestions.

10. The intelligent groundwater monitoring method based on cloud-edge-device collaboration and dynamic optimization according to claim 5, characterized in that, The steps of the intelligent monitoring terminal (101) acquiring water quality parameters of at least one target aquifer (509) through a multi-layer nested monitoring well according to the data acquisition instruction, and forming end-side acquisition data, include: According to the data acquisition command, the corresponding sensor channel is woken up, and the same environmental indicator is continuously sampled to extract water quality parameters. After removing the maximum and minimum values ​​in the continuous sampling sequence, the arithmetic mean of the remaining sampled values ​​is calculated. The arithmetic mean, together with the standard timestamp, the device's unique identifier, and the battery power status data of the solar power module (102), are structured and serialized to form a standardized serialized text. The serialized text is encapsulated into a message to form the terminal-side collected data; When the wireless communication module (103) does not receive a successful upload response signal within a preset timeout period, it writes the end-side collected data into a non-volatile storage unit and marks it as pending transmission so that it can be merged and retransmitted when the network recovers in the next communication cycle.

Images