Hydrogeological environment intelligent monitoring method

Abstract

The invention discloses an intelligent monitoring method for a hydrogeological environment, and the method comprises the steps: 1, constructing a space-air-ground integrated sensing network, and dynamically adjusting the monitoring frequency through a self-adaptive sampling strategy; 2, developing cross-platform data middleware, realizing conversion from multi-protocol data to a WaterML 2.0 standard, constructing an anomaly detection model in combination with a deep residual network, and correcting regional parameter deviation; 3, designing a hybrid fusion architecture, processing time sequence data by using an improved Transform at a bottom layer to capture long-term dependence, constructing a graph neural network at a top layer to fuse spatial connectivity, and outputting a prediction result with a confidence interval through an attention mechanism and a Bayesian network; step 4, establishing an edge-cloud hierarchical computing system, deploying lightweight model real-time early warning at an edge end, integrating complex models such as 3D geological modeling at the cloud end to generate a regional report, and adopting federal learning cooperative training and embedding differential privacy protection; and 5, developing a three-dimensional geological visualization platform and an intelligent decision-making system.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology, and in particular to an intelligent monitoring method for hydrogeological environment. Background Technology

[0002] Hydrogeological environmental monitoring is a key means to ensure water resource security, prevent geological disasters and maintain ecological balance. Especially in coastal areas, problems such as seawater intrusion and soil salinization caused by groundwater over-extraction directly affect agricultural production and residents' water safety.

[0003] However, traditional monitoring methods have significant shortcomings, such as insufficient integration of multi-source data. For example, a coastal city once relied solely on ground sensor networks to monitor groundwater levels, ignoring data on coastline retreat reflected by satellite remote sensing and the distribution of salt patches in farmland revealed by drone aerial photography, resulting in an underestimation of the extent of seawater intrusion by 30%. Due to the isolated use of data, the local water resources department failed to adjust groundwater extraction limits in a timely manner, ultimately causing a reduction in yield on 500 hectares of farmland and resulting in serious economic and property losses. The data formats, spatiotemporal resolutions, and reliability of different monitoring methods vary significantly, making it difficult for traditional methods to dynamically integrate heterogeneous data, leading to one-sided monitoring results and delayed early warnings. Therefore, an intelligent monitoring method for hydrogeological environment is proposed. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an intelligent monitoring method for hydrogeological environment.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A method for intelligent monitoring of hydrogeological environment, comprising:

[0007] Step 1: Construct a multi-layered monitoring network covering the ground, air, and satellite, and deploy integrated sensor nodes to simultaneously collect multi-parameter data such as groundwater level, water quality, and soil moisture. Combine UAV hyperspectral imagery and satellite radar data to form an integrated air-ground-space perception system. Through an adaptive sampling strategy, dynamically adjust the sensor operating frequency according to tidal cycles and rainfall forecasts.

[0008] Step 2: Develop cross-platform data conversion middleware to convert data from protocols such as MQTT and Modbus into the WaterML 2.0 standard format; construct an anomaly detection model using deep residual networks, and correct regional parameter deviations such as salinity and conductivity using historical data; perform spatiotemporal alignment of remote sensing images and ground data through edge computing gateways to solve the problem of mismatch between satellite transit time and ground sampling period, providing a high-quality, standardized data foundation for subsequent fusion analysis;

[0009] Step 3: Design a hybrid fusion architecture. The bottom layer uses an improved Transformer model to process time-series data such as water level and rainfall to capture long-term dependencies. The top layer constructs a graph neural network (GNN) to fuse spatial data on geological connectivity between sensor nodes. The data source weights are dynamically allocated through an attention mechanism (e.g., prioritizing high-resolution satellite imagery or high-frequency calibration sensor data), and combined with a Bayesian neural network to quantify prediction uncertainties, outputting a fusion result with confidence intervals.

[0010] Step 4: Construct a hierarchical computing system. Lightweight MobileNetV3 models are deployed at the edge to process high-frequency data such as soil moisture in real time, triggering local early warnings. Complex models, such as 3D geological modeling, run in the cloud, integrating multi-source long-term data to generate regional reports. Federated learning enables collaborative training of monitoring stations across Qingdao's districts and counties, avoiding cross-regional data transmission and protecting sensitive geological information using differential privacy technology. The cloud architecture is compatible with Kubernetes for dynamically expanding computing resources, supporting scenarios with surged monitoring demands during sudden disasters such as typhoons.

[0011] Step 5: Integrate OpenSceneGraph and CesiumJS to develop a 3D geological visualization platform that dynamically displays the structure of underground aquifers, the process of seawater intrusion, and changes in the coastline; build an intelligent early warning module based on a rule engine (such as triggering a yellow warning when soil salinity exceeds the standard), and integrate the NSGA-III multi-objective optimization algorithm to provide Pareto optimal solutions for groundwater extraction; develop a WeChat mini-program to support on-site inspectors in uploading geological photos, and combine it with AI image recognition to assist decision-making, forming a full-process support system from data display to intelligent decision-making.

[0012] The above technical solution further includes:

[0013] Furthermore, the construction of a multi-layered monitoring network covering ground, air, and satellite includes the following steps:

[0014] A low-power IoT sensor network is deployed on the ground layer, using the LoRaWAN / NB-IoT communication protocol, covering key areas such as Hongdaowan New Area. The node spacing is optimized according to geological connectivity, such as dense deployment near seawater intrusion channels, and supports real-time data back to the edge computing gateway.

[0015] The aerial layer deploys drones for regular patrols, equipped with hyperspectral cameras and thermal infrared sensors, to scan coastal erosion, changes in vegetation cover, and the development of surface cracks; the flight frequency is adjusted in conjunction with the tidal cycle, such as twice a day during spring tides and once a day during neap tides.

[0016] The satellite layer accesses Sentinel-1 radar satellite data (C-band, resolution 5m×20m) and monitors land subsidence and seawater intrusion dynamics through interferometric synthetic aperture radar technology; satellite transit data and real-time data from ground sensors are spatiotemporally aligned through a federated learning framework.

[0017] Furthermore, the deployment of integrated sensor nodes synchronously collects multiple parameter data such as groundwater level, water quality, and soil moisture, including the following steps:

[0018] Modular integrated sensors, such as the Decagon EC-5 soil moisture sensor, Solinst water level gauge, and HachHQ40d water quality analyzer, are used to perform multi-parameter synchronous sampling via a microcontroller (such as STM32F4) at a frequency of 10 minutes / time, which can be dynamically adjusted.

[0019] Parameter coverage:

[0020] Groundwater: water level, water quality (pH, conductivity, salinity, accuracy ±2%);

[0021] Soil: Moisture (volume water content), temperature;

[0022] Environment: Rainfall (tipping bucket rain gauge), surface displacement (laser rangefinder);

[0023] Priority deployment will be given to seawater intrusion channels and groundwater over-extraction areas in Hongdaowan New Area, with a node spacing of 500m (optimized based on geological borehole data), and the depth will be adjusted according to the thickness of the aquifer, such as 10m for shallow aquifer nodes and 30m for deep aquifer nodes.

[0024] Furthermore, the adaptive sampling strategy for dynamically adjusting the sensor's operating frequency based on tidal cycles and rainfall forecasts includes the following steps:

[0025] Based on the Qingdao Port tide table, the sampling frequency of groundwater level and salinity will be increased to 5 minutes / time during spring tides (tidal range > 3m) and reduced to 30 minutes / time during neap tides (tidal range < 1m).

[0026] Access the 48-hour rainfall forecast data from the Qingdao Meteorological Bureau. When the predicted rainfall is >10mm, activate the rainstorm mode. The soil moisture and surface displacement sampling frequency is 1 minute / time. After the rainfall ends, restore the basic frequency.

[0027] The remaining power is calculated in real time through the edge computing gateway. When the power is less than 20%, it automatically switches to low power mode, retaining only water level and salinity monitoring, at a frequency of once per hour.

[0028] A tide-rainfall prediction model is built based on LSTM neural network. It takes historical tide level and rainfall data as input and outputs sampling frequency suggestions for the next 24 hours. The frequency adjustment strategy is optimized by reinforcement learning (DQN algorithm) to minimize the trade-off between data accuracy loss and equipment energy consumption.

[0029] Furthermore, the proposed hybrid fusion architecture employs an improved Transformer model at the bottom layer to process time-series data on water level and rainfall, capturing long-term dependencies; the top layer constructs a graph neural network to fuse spatial data on geological connectivity between sensor nodes, including the following steps:

[0030] Low-level time-series data processing (improved Transformer):

[0031] Temporal convolutional layers are introduced into the standard Transformer Encoder to enhance the ability to capture local temporal features; relative positional encoding is used instead of absolute positional encoding to improve the modeling accuracy of long sequences, such as monthly rainfall data; with a time step of 10 minutes, the input includes time-series data including groundwater level, rainfall, and soil moisture, with a length of 720 steps, covering 5 days of dynamic changes;

[0032] Generate a temporal feature embedding vector (Dimension=256) to capture the long-term impact of tidal cycles and rainfall events on the hydrogeological system, such as the nonlinear change in groundwater recharge after 3 consecutive days of rainfall;

[0033] Top-level spatial data processing:

[0034] Based on historical borehole data from the Qingdao Geological Survey Institute, sensor nodes are treated as graph nodes, and geological connectivity between nodes, such as fault and fracture development zones, is used as edge weights (quantified through geological survey data, e.g., "if the fault width is >1m, the edge weight is +0.3"). A graph attention network is adopted, which aggregates neighbor node information through a multi-head self-attention mechanism and outputs a spatial feature embedding vector (Dimension = 128) to reflect the correlation of hydrogeological conditions between nodes. For example, a sudden change in salinity at node A may affect the water level at node B through connectivity.

[0035] The temporal feature vector (256 dimensions) and the spatial feature vector (128 dimensions) are concatenated to form a 384-dimensional fusion feature, which is then input into a fully connected layer (with ReLU activation function) to generate the final fusion result (such as the seawater intrusion risk index).

[0036] Furthermore, the step of dynamically allocating data source weights through an attention mechanism and combining it with a Bayesian neural network to quantify prediction uncertainty, and outputting a fusion result with a confidence interval, includes the following steps:

[0037] Ground sensors are weighted based on calibration frequency (e.g., monthly calibration, +0.2 weight) and failure rate (e.g., abnormal data percentage <5% in the past 30 days, +0.3 weight); satellite data weights are adjusted based on resolution (Sentinel-1 resolution 5m, +0.4 weight) and transit frequency (coverage once every 6 days, +0.1 weight); UAV data is dynamically weighted based on flight altitude (below 100m, +0.3 weight) and cloud cover (<10%, +0.2 weight). Multi-head spatial attention is used to softmax normalize the reliability score of each data source, generating dynamic weights, such as 0.6 for ground sensors, 0.3 for satellites, and 0.1 for UAVs.

[0038]

[0039] Where, α i The dynamic weights for the i-th data source, such as ground sensors, satellites, and drones; q i The query vector represents the feature representation of the current data source; k i The key vector represents the feature representation from other data sources; d k is the dimension of the key vector, used to scale the dot product and prevent gradient vanishing; n is the total number of data sources (e.g., 3: ground, satellite, drone); weights are dynamically assigned by calculating the correlation (dot product) between data sources, with data sources with high reliability (e.g., sensors that are calibrated frequently) receiving higher weights.

[0040] Add a Bayesian linear layer after the fully connected layer in the fusion architecture, replace the weight parameters with a probability distribution (e.g., normal distribution μ≈0, σ≈0.1), and approximate the posterior distribution using Monte Carlo dropout (MC Dropout); run 100 forward propagations, calculate the standard deviation (σ) of the prediction results, and output a 95% confidence interval (e.g., seawater intrusion risk index = 0.75±0.08);

[0041] The fusion results are presented in the form of "numerical value + confidence interval", such as "the current salinity diffusion rate in the area is 0.3 m / day, with a confidence interval of 0.25 to 0.35 m / day". When the confidence interval width exceeds the threshold, a data backtracking mechanism is triggered to re-verify low-weight data sources, such as calibrating ground sensors or supplementing drone patrols.

[0042] Furthermore, the construction of the hierarchical computing system involves deploying a lightweight MobileNetV3 model at the edge to process high-frequency data such as soil moisture in real time and trigger local early warnings; and running a complex model, including 3D geological modeling, in the cloud to integrate multi-source long-term data and generate regional reports, including the following steps:

[0043] Real-time edge processing (MobileNetV3):

[0044] Employing MobileNetV3 (Small version), and utilizing depthwise separable convolution and linear bottleneck techniques, the model size is compressed to 2.5MB, and the inference speed is increased to 120 frames per second (GPU accelerated), adapting to the resource constraints of the edge gateway (NVIDIA Jetson Nano). It receives high-frequency data in real time from soil moisture sensors (every 10 minutes), surface displacement laser rangefinders (every 5 minutes), and UAV thermal infrared imagery (0.5m resolution, updated every 2 hours).

[0045] Processing logic:

[0046] Soil moisture data: The MobileNetV3 classification model is used to determine whether the threshold is exceeded, such as >35% volumetric water content in clay areas and >25% in sandy areas. If the limit is exceeded, a local yellow alert is triggered.

[0047] Surface displacement data: Combined with historical displacement trends, such as displacement rate > 2 mm / day in the past 24 hours, the model predicts the displacement in the next 2 hours. If the predicted value is > 5 mm, an orange alert is triggered.

[0048] Drone imagery: The MobileNetV3 target detection module identifies coastal erosion areas. If cracks wider than 1m are detected, they are marked and pushed to the WeChat mini-program of on-site inspectors.

[0049] The warning information is broadcast to nearby terminals via LoRaWAN, such as water conservancy station displays and patrol personnel's mobile phones, and is also stored in the edge gateway's local database (SQLite), retaining the data for the most recent 72 hours for retrospective purposes;

[0050] Cloud-based complex analysis (3D geological modeling):

[0051] Long-term data uploaded from edge computing sources (such as monthly average soil moisture and quarterly seawater intrusion paths), satellite radar data (Sentinel-1 interannual variation of surface subsidence), and geological borehole data (aquifer thickness and lithological distribution) are aggregated to construct a multi-source spatiotemporal data cube. A voxel-based modeling method is employed, and a CUDA-accelerated ray casting algorithm is used to generate the 3D structure of the underground aquifer, achieving a resolution of 1m. 3 / Voxenograms dynamically display the front edge of seawater intrusion, such as using a blue gradient to represent areas with salinity >3g / L;

[0052] Generate regional reports:

[0053] Trend Analysis: Statistics on the changes in groundwater recharge and seawater intrusion area over the past 5 years, combined with rainfall forecasts, are used to predict hydrogeological trends for the next quarter.

[0054] Risk assessment: Based on the NSGA-III multi-objective optimization algorithm, output the Pareto optimal solution for groundwater extraction and ecological water replenishment, such as "reducing the extraction volume by 20% can reduce the rate of seawater intrusion by 30%, but the cost of agricultural water will increase by 15%";

[0055] Visualization output: Interactive 3D reports are generated using CesiumJS, supporting stratigraphic slicing and timeline backtracking, and can be embedded into the Qingdao government cloud platform for relevant departments to view.

[0056] Furthermore, the collaborative training of monitoring stations through federated learning, combined with differential privacy technology to protect sensitive geological information, includes the following steps:

[0057] Each district and county monitoring station is divided into independent clients, each holding local data, such as sensor data from Jimo District and borehole data from Jiaozhou City. Each client trains a lightweight model (such as MobileNetV3) for 5 epochs based on its local data, and the model converges to a local optimum. Before uploading parameters, the client uses the formula w′. k =w k +Lap(∈ / Δf) applies to the model parameter w k Add Laplace noise to generate the perturbation parameter w′ k , where Lap(b) is the Laplace distribution of the scaling parameter b = ∈ / Δf; Δf is the sensitivity (e.g., the range of model parameter variation Δf = 0.1); ∈ is the privacy budget (e.g., ∈-differential privacy is satisfied when ∈ = 1.0);

[0058] The cloud receives all perturbation parameters w′ uploaded by clients. k The global model parameter w is calculated using the following formula. global :

[0059]

[0060] Among them, w global For global model parameters (parameters aggregated in the cloud); w′ k Here are the local model parameters for the k-th client (monitoring site); n k The data volume of the kth client (e.g., the sensor data volume of Jimo District); n is the total data volume of all clients (K is the total number of clients); by weighted averaging the local model parameters of each station, a global model is generated to avoid the cross-regional transmission of raw data and protect sensitive geological information;

[0061] Each client updates its local model with global parameters and continues the next round of training; the total privacy budget is set to 3.0, and each round of training consumes 0.1 of the budget, which supports privacy requirements even after 30 rounds of collaborative training; through member inference attack testing, it is proved that the model cannot distinguish whether the training data belongs to a specific site after adding noise.

[0062] The present invention has the following beneficial effects:

[0063] This invention develops a cross-platform data conversion middleware to convert MQTT / Modbus protocols to the WaterML 2.0 standard format, and corrects regional parameter biases through a deep residual network, eliminating differences in data format and measurement range. A hybrid fusion architecture is designed, with an improved Transformer model at the bottom layer capturing long-term dependencies in water level / rainfall time-series data, and a graph neural network at the top layer fusing geological connectivity spatial data. An attention mechanism dynamically allocates data source weights, and a Bayesian neural network is used to quantify prediction uncertainties, outputting a fusion result with confidence intervals, effectively solving the problems of heterogeneous data reliability assessment and conflict resolution. Finally, through a hierarchical computing system and federated learning combined with differential privacy technology, the system integrates multi-source long-term data while ensuring data security, effectively solving the problem of insufficient multi-source data fusion. Attached Figure Description

[0064] Figure 1 This is a flowchart illustrating the steps of an intelligent monitoring method for hydrogeological environment proposed in this invention. Detailed Implementation

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

[0066] Please see Figure 1 As shown, this invention is an intelligent monitoring method for hydrogeological environment, comprising:

[0067] A method for intelligent monitoring of hydrogeological environment, comprising:

[0068] Step 1: Construct a multi-layered monitoring network covering the ground, air, and satellite, and deploy integrated sensor nodes to simultaneously collect multi-parameter data such as groundwater level, water quality, and soil moisture. Combine UAV hyperspectral imagery and satellite radar data to form an integrated air-ground-space perception system. Through an adaptive sampling strategy, dynamically adjust the sensor operating frequency according to tidal cycles and rainfall forecasts.

[0069] Step 2: Develop cross-platform data conversion middleware to convert data from protocols such as MQTT and Modbus into the WaterML 2.0 standard format; construct an anomaly detection model using deep residual networks, and correct regional parameter deviations such as salinity and conductivity using historical data; perform spatiotemporal alignment of remote sensing images and ground data through edge computing gateways to solve the problem of mismatch between satellite transit time and ground sampling period, providing a high-quality, standardized data foundation for subsequent fusion analysis;

[0070] Step 3: Design a hybrid fusion architecture. The bottom layer uses an improved Transformer model to process time-series data such as water level and rainfall to capture long-term dependencies. The top layer constructs a graph neural network (GNN) to fuse spatial data on geological connectivity between sensor nodes. The data source weights are dynamically allocated through an attention mechanism (e.g., prioritizing high-resolution satellite imagery or high-frequency calibration sensor data), and combined with a Bayesian neural network to quantify prediction uncertainties, outputting a fusion result with confidence intervals.

[0071] Step 4: Construct a hierarchical computing system. Lightweight MobileNetV3 models are deployed at the edge to process high-frequency data such as soil moisture in real time, triggering local early warnings. Complex models, such as 3D geological modeling, run in the cloud, integrating multi-source long-term data to generate regional reports. Federated learning enables collaborative training of monitoring stations across Qingdao's districts and counties, avoiding cross-regional data transmission and protecting sensitive geological information using differential privacy technology. The cloud architecture is compatible with Kubernetes for dynamically expanding computing resources, supporting scenarios with surged monitoring demands during sudden disasters such as typhoons.

[0072] Step 5: Integrate OpenSceneGraph and CesiumJS to develop a 3D geological visualization platform that dynamically displays the structure of underground aquifers, the process of seawater intrusion, and changes in the coastline; build an intelligent early warning module based on a rule engine (such as triggering a yellow warning when soil salinity exceeds the standard), and integrate the NSGA-III multi-objective optimization algorithm to provide Pareto optimal solutions for groundwater extraction; develop a WeChat mini-program to support on-site inspectors in uploading geological photos, and combine it with AI image recognition to assist decision-making, forming a full-process support system from data display to intelligent decision-making.

[0073] In one embodiment, constructing a multi-layered monitoring network covering ground, air, and satellite includes the following steps:

[0074] A low-power IoT sensor network is deployed on the ground layer, using the LoRaWAN / NB-IoT communication protocol, covering key areas such as Hongdaowan New Area. The node spacing is optimized according to geological connectivity, such as dense deployment near seawater intrusion channels, and supports real-time data back to the edge computing gateway.

[0075] The aerial layer deploys drones for regular patrols, equipped with hyperspectral cameras and thermal infrared sensors, to scan coastal erosion, changes in vegetation cover, and the development of surface cracks; the flight frequency is adjusted in conjunction with the tidal cycle, such as twice a day during spring tides and once a day during neap tides.

[0076] The satellite layer accesses Sentinel-1 radar satellite data (C-band, resolution 5m×20m) and monitors land subsidence and seawater intrusion dynamics through interferometric synthetic aperture radar technology; satellite transit data and real-time data from ground sensors are spatiotemporally aligned through a federated learning framework.

[0077] In one embodiment, the deployment of integrated sensor nodes to synchronously collect multiple parameter data such as groundwater level, water quality, and soil moisture includes the following steps:

[0078] Modular integrated sensors, such as the Decagon EC-5 soil moisture sensor, Solinst water level gauge, and HachHQ40d water quality analyzer, are used to perform multi-parameter synchronous sampling via a microcontroller (such as STM32F4) at a frequency of 10 minutes / time, which can be dynamically adjusted.

[0079] Parameter coverage:

[0080] Groundwater: water level, water quality (pH, conductivity, salinity, accuracy ±2%);

[0081] Soil: Moisture (volume water content), temperature;

[0082] Environment: Rainfall (tipping bucket rain gauge), surface displacement (laser rangefinder);

[0083] Priority deployment will be given to seawater intrusion channels and groundwater over-extraction areas in Hongdaowan New Area, with a node spacing of 500m (optimized based on geological borehole data), and the depth will be adjusted according to the thickness of the aquifer, such as 10m for shallow aquifer nodes and 30m for deep aquifer nodes.

[0084] In one embodiment, the step of dynamically adjusting the sensor operating frequency based on tidal cycles and rainfall forecasts using an adaptive sampling strategy includes the following steps:

[0085] Based on the Qingdao Port tide table, the sampling frequency of groundwater level and salinity will be increased to 5 minutes / time during spring tides (tidal range > 3m) and reduced to 30 minutes / time during neap tides (tidal range < 1m).

[0086] Access the 48-hour rainfall forecast data from the Qingdao Meteorological Bureau. When the predicted rainfall is >10mm, activate the rainstorm mode. The soil moisture and surface displacement sampling frequency is 1 minute / time. After the rainfall ends, restore the basic frequency.

[0087] The remaining power is calculated in real time through the edge computing gateway. When the power is less than 20%, it automatically switches to low power mode, retaining only water level and salinity monitoring, at a frequency of once per hour.

[0088] A tide-rainfall prediction model is built based on LSTM neural network. It takes historical tide level and rainfall data as input and outputs sampling frequency suggestions for the next 24 hours. The frequency adjustment strategy is optimized by reinforcement learning (DQN algorithm) to minimize the trade-off between data accuracy loss and equipment energy consumption.

[0089] In one embodiment, the design employs a hybrid fusion architecture, with the bottom layer using an improved Transformer model to process time-series data on water level and rainfall, capturing long-term dependencies; the top layer constructs a graph neural network to fuse spatial data on geological connectivity between sensor nodes, including the following steps:

[0090] Low-level time-series data processing (improved Transformer):

[0091] Temporal convolutional layers are introduced into the standard Transformer Encoder to enhance the ability to capture local temporal features; relative positional encoding is used instead of absolute positional encoding to improve the modeling accuracy of long sequences, such as monthly rainfall data; with a time step of 10 minutes, the input includes time-series data including groundwater level, rainfall, and soil moisture, with a length of 720 steps, covering 5 days of dynamic changes;

[0092] Generate a temporal feature embedding vector (Dimension=256) to capture the long-term impact of tidal cycles and rainfall events on the hydrogeological system, such as the nonlinear change in groundwater recharge after 3 consecutive days of rainfall;

[0093] Top-level spatial data processing:

[0094] Based on historical borehole data from the Qingdao Geological Survey Institute, sensor nodes are treated as graph nodes, and geological connectivity between nodes, such as fault and fracture development zones, is used as edge weights (quantified through geological survey data, e.g., "if the fault width is >1m, the edge weight is +0.3"). A graph attention network is adopted, which aggregates neighbor node information through a multi-head self-attention mechanism and outputs a spatial feature embedding vector (Dimension = 128) to reflect the correlation of hydrogeological conditions between nodes. For example, a sudden change in salinity at node A may affect the water level at node B through connectivity.

[0095] The temporal feature vector (256 dimensions) and the spatial feature vector (128 dimensions) are concatenated to form a 384-dimensional fusion feature, which is then input into a fully connected layer (with ReLU activation function) to generate the final fusion result (such as the seawater intrusion risk index).

[0096] In one embodiment, the step of dynamically allocating data source weights through an attention mechanism and combining this with a Bayesian neural network to quantify prediction uncertainty, and outputting a fusion result with confidence intervals, includes the following steps:

[0097] Ground sensors are weighted based on calibration frequency (e.g., monthly calibration, +0.2 weight) and failure rate (e.g., abnormal data percentage <5% in the past 30 days, +0.3 weight); satellite data weights are adjusted based on resolution (Sentinel-1 resolution 5m, +0.4 weight) and transit frequency (coverage once every 6 days, +0.1 weight); UAV data is dynamically weighted based on flight altitude (below 100m, +0.3 weight) and cloud cover (<10%, +0.2 weight). Multi-head spatial attention is used to softmax normalize the reliability score of each data source, generating dynamic weights, such as 0.6 for ground sensors, 0.3 for satellites, and 0.1 for UAVs.

[0098]

[0099] Where, α i The dynamic weights for the i-th data source, such as ground sensors, satellites, and drones; q i The query vector represents the feature representation of the current data source; k i The key vector represents the feature representation from other data sources; d k is the dimension of the key vector, used to scale the dot product and prevent gradient vanishing; n is the total number of data sources (e.g., 3: ground, satellite, drone); weights are dynamically assigned by calculating the correlation (dot product) between data sources, with data sources with high reliability (e.g., sensors that are calibrated frequently) receiving higher weights.

[0100] Add a Bayesian linear layer after the fully connected layer in the fusion architecture, replace the weight parameters with a probability distribution (e.g., normal distribution μ≈0, σ≈0.1), and approximate the posterior distribution using Monte Carlo dropout (MC Dropout); run 100 forward propagations, calculate the standard deviation (σ) of the prediction results, and output a 95% confidence interval (e.g., seawater intrusion risk index = 0.75±0.08);

[0101] The fusion results are presented in the form of "numerical value + confidence interval", such as "the current salinity diffusion rate in the area is 0.3 m / day, with a confidence interval of 0.25 to 0.35 m / day". When the confidence interval width exceeds the threshold, a data backtracking mechanism is triggered to re-verify low-weight data sources, such as calibrating ground sensors or supplementing drone patrols.

[0102] In one embodiment, the construction of a hierarchical computing system, with a lightweight MobileNetV3 model deployed at the edge to process high-frequency data such as soil moisture in real time and trigger local early warnings; and a complex model including 3D geological modeling running in the cloud to integrate multi-source long-term data and generate regional reports, includes the following steps:

[0103] Real-time edge processing (MobileNetV3):

[0104] Employing MobileNetV3 (Small version), and utilizing depthwise separable convolution and linear bottleneck techniques, the model size is compressed to 2.5MB, and the inference speed is increased to 120 frames per second (GPU accelerated), adapting to the resource constraints of the edge gateway (NVIDIA Jetson Nano). It receives high-frequency data in real time from soil moisture sensors (every 10 minutes), surface displacement laser rangefinders (every 5 minutes), and UAV thermal infrared imagery (0.5m resolution, updated every 2 hours).

[0105] Processing logic:

[0106] Soil moisture data: The MobileNetV3 classification model is used to determine whether the threshold is exceeded, such as >35% volumetric water content in clay areas and >25% in sandy areas. If the limit is exceeded, a local yellow alert is triggered.

[0107] Surface displacement data: Combined with historical displacement trends, such as displacement rate > 2 mm / day in the past 24 hours, the model predicts the displacement in the next 2 hours. If the predicted value is > 5 mm, an orange alert is triggered.

[0108] Drone imagery: The MobileNetV3 target detection module identifies coastal erosion areas. If cracks wider than 1m are detected, they are marked and pushed to the WeChat mini-program of on-site inspectors.

[0109] The warning information is broadcast to nearby terminals via LoRaWAN, such as water conservancy station displays and patrol personnel's mobile phones, and is also stored in the edge gateway's local database (SQLite), retaining the data for the most recent 72 hours for retrospective purposes;

[0110] Cloud-based complex analysis (3D geological modeling):

[0111] Long-term data uploaded from edge computing sources (such as monthly average soil moisture and quarterly seawater intrusion paths), satellite radar data (Sentinel-1 interannual variation of surface subsidence), and geological borehole data (aquifer thickness and lithological distribution) are aggregated to construct a multi-source spatiotemporal data cube. A voxel-based modeling method is employed, and a CUDA-accelerated ray casting algorithm is used to generate the 3D structure of the underground aquifer, achieving a resolution of 1m. 3 / Voxenograms dynamically display the front edge of seawater intrusion, such as using a blue gradient to represent areas with salinity >3g / L;

[0112] Generate regional reports:

[0113] Trend Analysis: Statistics on the changes in groundwater recharge and seawater intrusion area over the past 5 years, combined with rainfall forecasts, are used to predict hydrogeological trends for the next quarter.

[0114] Risk assessment: Based on the NSGA-III multi-objective optimization algorithm, output the Pareto optimal solution for groundwater extraction and ecological water replenishment, such as "reducing the extraction volume by 20% can reduce the rate of seawater intrusion by 30%, but the cost of agricultural water will increase by 15%";

[0115] Visualization output: Interactive 3D reports are generated using CesiumJS, supporting stratigraphic slicing and timeline backtracking, and can be embedded into the Qingdao government cloud platform for relevant departments to view.

[0116] In one embodiment, the collaborative training of monitoring stations through federated learning, combined with differential privacy technology to protect sensitive geological information, includes the following steps:

[0117] Each district and county monitoring station is divided into independent clients, each holding local data, such as sensor data from Jimo District and borehole data from Jiaozhou City. Each client trains a lightweight model (such as MobileNetV3) for 5 epochs based on its local data, and the model converges to a local optimum. Before uploading parameters, the client uses the formula w′. k =w k +Lap(∈ / Δf) applies to the model parameter w k Add Laplace noise to generate the perturbation parameter w′ k , where Lap(b) is the Laplace distribution of the scaling parameter b = ∈ / Δf; Δf is the sensitivity (e.g., the range of model parameter variation Δf = 0.1); ∈ is the privacy budget (e.g., ∈-differential privacy is satisfied when ∈ = 1.0);

[0118] The cloud receives all perturbation parameters w′ uploaded by clients. k The global model parameter w is calculated using the following formula. global :

[0119]

[0120] Among them, w global For global model parameters (parameters aggregated in the cloud); w′ k Here are the local model parameters for the k-th client (monitoring site); n k The data volume of the kth client (e.g., the sensor data volume of Jimo District); n is the total data volume of all clients (K is the total number of clients); by weighted averaging the local model parameters of each station, a global model is generated to avoid the cross-regional transmission of raw data and protect sensitive geological information;

[0121] Each client updates its local model with global parameters and continues the next round of training; the total privacy budget is set to 3.0, and each round of training consumes 0.1 of the budget, which supports privacy requirements even after 30 rounds of collaborative training; through member inference attack testing, it is proved that the model cannot distinguish whether the training data belongs to a specific site after adding noise.

[0122] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent monitoring of hydrogeological environment, characterized in that, include: Step 1: Construct a multi-layered monitoring network covering the ground, air, and satellite; deploy integrated sensor nodes to synchronously collect multi-parameter data on groundwater level, water quality, and soil moisture; combine UAV hyperspectral imagery and satellite radar data to form an integrated air-ground-space perception system; and dynamically adjust the sensor operating frequency according to tidal cycles and rainfall forecasts through an adaptive sampling strategy. Step 2: Develop a cross-platform data conversion middleware to convert MQTT and Modbus protocol data into the WaterML 2.0 standard format; build an anomaly detection model using deep residual networks and correct regional parameter deviations of salinity-conductivity using historical data; and perform spatiotemporal alignment through an edge computing gateway. Step 3: Design a hybrid and integrated architecture, with an improved Transformer model at the bottom layer to process water level and rainfall time series data and capture long-term dependencies; The top-level graph neural network is used to fuse spatial data on geological connectivity between sensor nodes; the weights of the data sources are dynamically allocated through an attention mechanism, and the uncertainty of prediction is quantified by combining a Bayesian neural network, outputting the fusion result with confidence intervals; Step 4: Construct a hierarchical computing system, deploy a lightweight MobileNetV3 model at the edge to process high-frequency data such as soil moisture in real time, and trigger local early warnings; The system runs complex models, including 3D geological modeling, in the cloud, integrating long-term data from multiple sources to generate regional reports; it also conducts collaborative training among various monitoring stations through federated learning, and combines differential privacy technology to protect sensitive geological information. Step 5: Integrate OpenSceneGraph and CesiumJS to develop a 3D geological visualization platform that dynamically displays the structure of underground aquifers, the process of seawater intrusion, and changes in the coastline; build an intelligent early warning module based on a rule engine and integrate the NSGA-III multi-objective optimization algorithm to provide Pareto optimal solutions for groundwater extraction; develop a WeChat mini-program to support on-site inspectors in uploading geological photos, combined with AI image recognition to assist decision-making, forming a full-process support system from data display to intelligent decision-making.

2. The intelligent monitoring method for hydrogeological environment according to claim 1, characterized in that, The construction of a multi-layered monitoring network covering ground, air, and satellite includes the following steps: A low-power IoT sensor network is deployed on the ground layer, using the LoRaWAN / NB-IoT communication protocol, covering key areas including the Hongdaowan New Area. The node spacing is optimized according to geological connectivity, and it supports real-time data back to the edge computing gateway. The aerial layer deploys drones for regular patrols, equipped with hyperspectral cameras and thermal infrared sensors, to scan coastal erosion, changes in vegetation cover, and the development of surface cracks. The flight frequency is adjusted in conjunction with the tidal cycle. The satellite layer accesses Sentinel-1 radar satellite data and monitors land subsidence and seawater intrusion dynamics through interferometric synthetic aperture radar technology; satellite transit data and real-time data from ground sensors are spatiotemporally aligned through a federated learning framework.

3. The intelligent monitoring method for hydrogeological environment according to claim 1, characterized in that, The deployment of integrated sensor nodes synchronously collects multiple parameter data such as groundwater level, water quality, and soil moisture, including the following steps: A modular integrated sensor is used, and multi-parameter synchronous sampling is performed through a microcontroller, covering the following parameters: Groundwater: water level, water quality; Soil: humidity, temperature; Environment: rainfall, surface displacement; Priority deployment will be given to areas with seawater intrusion channels and groundwater over-extraction in the Hongdaowan New Area, with a node spacing of 500m and the depth adjusted according to the thickness of the aquifer.

4. The intelligent monitoring method for hydrogeological environment according to claim 1, characterized in that, The adaptive sampling strategy dynamically adjusts the sensor's operating frequency based on tidal cycles and rainfall forecasts, including the following steps: Based on the Qingdao Port tide table, the sampling frequency of groundwater level and salinity will be increased to 5 minutes / time during spring tides and reduced to 30 minutes / time during neap tides; It accesses 48-hour rainfall forecast data from the Qingdao Meteorological Bureau. When the predicted rainfall exceeds 10mm, it activates the rainstorm mode and restores the base frequency after the rainfall ends. It also calculates the remaining power in real time through the edge computing gateway and automatically switches to low-power mode when the power is less than 20%. A tide-rainfall prediction model is built based on LSTM neural network. It takes historical tide level and rainfall data as input and outputs a sampling frequency suggestion for the next 24 hours. The frequency adjustment strategy is optimized through reinforcement learning to minimize the trade-off between data accuracy loss and equipment energy consumption.

5. The intelligent monitoring method for hydrogeological environment according to claim 1, characterized in that, The proposed hybrid fusion architecture employs an improved Transformer model at the bottom layer to process time-series data on water level and rainfall, capturing long-term dependencies. The top layer utilizes a graph neural network to fuse spatial data on geological connectivity between sensor nodes, including the following steps: Low-level time series data processing: Temporal convolutional layers are introduced into the standard Transformer Encoder to enhance the ability to capture local temporal features; relative positional encoding is used instead of absolute positional encoding to improve the modeling accuracy of long sequences; with a time step of 10 minutes, the input includes temporal data such as groundwater level, rainfall, and soil moisture; and temporal feature embedding vectors are generated to capture the long-term impact of tidal cycles and rainfall events on hydrogeology. Top-level spatial data processing: Based on historical borehole data, sensor nodes are treated as graph nodes, and the geological connectivity between nodes is used as edge weights. A graph attention network is adopted to aggregate neighbor node information through a multi-head self-attention mechanism and output spatial feature embedding vectors to reflect the correlation of hydrogeological conditions between nodes. The temporal feature vectors and spatial feature vectors are concatenated into a 384-dimensional fusion feature, which is then input into a fully connected layer to generate the final fusion result.

6. The intelligent monitoring method for hydrogeological environment according to claim 1, characterized in that, The process of dynamically allocating data source weights through an attention mechanism and combining this with a Bayesian neural network to quantify prediction uncertainty, and outputting a fusion result with a confidence interval, includes the following steps: Ground sensors quantify reliability based on calibration frequency and failure rate; satellite data has its weights adjusted based on resolution and transit frequency; and UAV data is dynamically assigned values ​​based on flight altitude and cloud cover. A multi-head spatial attention approach is used to perform softmax normalization on the reliability score of each data source, generating dynamic weights. The main formula is as follows: Where, α i q represents the dynamic weight of the i-th data source; i The query vector represents the feature representation of the current data source; k i The key vector represents the feature representation from other data sources; d k is the dimension of the key vector; n is the total number of data sources; weights are dynamically allocated by calculating the correlation between data sources, with data sources with higher reliability receiving higher weights; A Bayesian linear layer is added after the fully connected layer in the fusion architecture, and the weight parameters are replaced with probability distributions. The posterior distribution is inferred by Monte Carlo dropout. The forward propagation is run 100 times, the standard deviation σ of the prediction results is calculated, and the 95% confidence interval is output. The fusion result is presented in the form of "numerical value + confidence interval". When the confidence interval width exceeds the threshold, the data backtracking mechanism is triggered to re-validate the low-weight data source.

7. The intelligent monitoring method for hydrogeological environment according to claim 1, characterized in that, The aforementioned hierarchical computing system involves deploying a lightweight MobileNetV3 model at the edge to process high-frequency data such as soil moisture in real time and trigger local early warnings; and running a complex model, including 3D geological modeling, in the cloud to integrate multi-source long-term data and generate regional reports, including the following steps: Real-time edge processing: Using MobileNetV3, through depthwise separable convolution and linear bottleneck techniques, the model size is compressed to 2.5MB, and the inference speed is increased to 120 frames per second, adapting to the resource constraints of edge gateways; it receives high-frequency data from soil moisture sensors, surface displacement laser rangefinders, and UAV thermal infrared images in real time; Processing logic: Soil moisture data: The MobileNetV3 classification model is used to determine whether the threshold is exceeded. If the threshold is exceeded, a local yellow alert is triggered. Surface displacement data: Combining historical displacement trends, the model predicts the displacement in the next 2 hours. If the predicted value is greater than 5 mm, an orange alert is triggered. Drone imagery: Coastal erosion areas are identified using the MobileNetV3 target detection module, marked, and pushed to the WeChat mini-program of on-site patrol personnel; The warning information is broadcast to nearby terminals via LoRaWAN and stored in the local database of the edge gateway, retaining the data for the most recent 72 hours for retrospective purposes; Complex analytics in the cloud: Long-term data, satellite radar data, and geological borehole data uploaded from the edge are aggregated to construct a multi-source spatiotemporal data cube; a voxel-based modeling method is adopted, and a CUDA-accelerated ray casting algorithm is used to generate the three-dimensional structure of underground aquifers with a resolution of 1m. 3 / Voxens, dynamically displaying the leading edge of seawater intrusion; Generate regional reports: Trend Analysis: Statistics on the changes in groundwater recharge and seawater intrusion area over the past 5 years, combined with rainfall forecasts, are used to predict hydrogeological trends for the next quarter. Risk assessment: Based on the NSGA-III multi-objective optimization algorithm, output the Pareto optimal solution for groundwater extraction and ecological water replenishment; Visualization output: Interactive 3D reports are generated using CesiumJS, supporting stratigraphic slicing and timeline backtracking, and can be embedded into the cloud platform for relevant departments to view.

8. The intelligent monitoring method for hydrogeological environment according to claim 1, characterized in that, The method of collaborative training of monitoring stations through federated learning, combined with differential privacy technology to protect sensitive geological information, includes the following steps: Each district and county monitoring station is divided into independent clients, each client holding local data; each client trains a lightweight model for 5 epochs based on its local data, and the model converges to a local optimum; before uploading parameters, the client uses the formula w′ k =w k +Lap(∈ / Δf) applies to the model parameter w k Add Laplace noise to generate the perturbation parameter w′ k Where Lap(b) is the Laplace distribution of the scaling parameter b = ∈ / Δf; Δf is the sensitivity; ∈ is the privacy budget; The cloud receives all perturbation parameters w′ uploaded by clients. k The global model parameter w is calculated using the following formula. global : Among them, w global For global model parameters; w′ k For the local model parameters of the k-th client; n k Let n be the data volume of the kth client; n is the total data volume of all clients; a global model is generated by weighted averaging of the local model parameters of each site to avoid cross-regional transmission of raw data and protect sensitive geological information. Each client updates its local model with global parameters and continues the next round of training; the total privacy budget is set to 3.0, and each round of training consumes 0.1 of the budget, which supports privacy requirements even after 30 rounds of collaborative training; through member inference attack testing, it is proved that the model cannot distinguish whether the training data belongs to a specific site after adding noise.

Images