High-geothermal exploitation ground subsidence data monitoring method and device

Abstract

This invention provides a method and device for monitoring ground subsidence data in high geothermal extraction. The method includes the following steps: S1: Determining the monitoring area and setting up monitoring points; S2: Installing and debugging corresponding monitoring devices at the monitoring points; S3: Collecting multi-source raw data; S4: Preprocessing the multi-source raw data; S5: Constructing a collaborative fusion model of a long short-term memory network and a gradient boosting decision tree, and outputting the final fused subsidence amount; S6: Performing time series analysis, spatial distribution analysis, and anomaly detection. This invention, through a complete process design of "multi-source data acquisition - intelligent fusion modeling - dynamic analysis and early warning," achieves high-precision and dynamic subsidence monitoring, overcoming the limitations of accuracy and response lag in traditional monitoring methods, and providing key technical support for the safe and sustainable development of high geothermal resources.

Description

Technical Field

[0001] This invention relates to the field of measurement and ground subsidence data monitoring technology, and in particular to a method and device for monitoring ground subsidence data in high geothermal extraction. Background Technology

[0002] Geothermal resources, as a clean and renewable energy source, are of great significance to the transformation of the energy structure. However, the extraction of underground fluids (hot water, steam) during geothermal exploitation can cause changes in ground stress, which can easily lead to geological and environmental problems such as land subsidence. Land subsidence can not only damage surface buildings and infrastructure (such as pipelines and roads), but may also exacerbate the risk of geological disasters and restrict the sustainable development of geothermal resources.

[0003] Existing methods for monitoring ground subsidence mainly rely on a single technology: GNSS (Global Navigation Satellite System) can provide high-frequency three-dimensional coordinates, but is susceptible to weather interference; InSAR (Synthetic Aperture Radar Interferometry) can achieve large-scale monitoring, but is limited by the satellite revisit period and has low temporal resolution; leveling measurement has high accuracy but is expensive and difficult to achieve high-frequency dynamic monitoring; groundwater level monitoring can reflect the impact of mining on the strata, but cannot be directly correlated with ground subsidence.

[0004] Traditional data fusion methods (such as Kalman filtering) rely heavily on linear assumptions, making it difficult to capture nonlinear characteristics in the settlement process (such as the differentiated response of different soil layers to water level changes) and long-term dynamic trends (such as settlement fluctuations caused by seasonal changes in mining intensity). This results in limited fusion accuracy, delayed anomaly warnings, and an inability to meet the needs of refined monitoring in high geothermal mining areas. Summary of the Invention

[0005] This invention provides a method and device for monitoring ground subsidence data in high-temperature geothermal mining. First, GNSS, InSAR corner reflectors, leveling, and groundwater level monitoring points are scientifically deployed in and around the mining area to collect multi-source raw data. After preprocessing (correction, noise reduction, and standardization), a dataset containing time series, spatial location, and derived features is constructed. A long short-term memory network model is used to capture the temporal dynamic trend of subsidence. A gradient boosting decision tree model fuses spatial features with the long short-term memory network output. Then, an attention mechanism is used to dynamically balance the weights of both to obtain a high-precision fused subsidence measurement. Based on the fusion results, time-space feature analysis is performed to identify anomalies and trigger tiered early warnings, while simultaneously achieving secure storage and management of all data.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for monitoring ground subsidence data during high-temperature geothermal extraction includes the following steps:

[0008] S1: Determine the monitoring area and set up monitoring points, and record the geographical coordinates of the monitoring points and their distance from the nearest well.

[0009] S2: Install and debug the corresponding monitoring devices at the monitoring points;

[0010] S3: Collect multi-source raw data at a set frequency using the calibrated monitoring device. The raw data is transmitted to the data processing center with the monitoring point number and UTC timestamp appended.

[0011] S4: Preprocess the multi-source raw data to obtain GNSS vertical displacement data, InSAR vertical displacement data, leveling settlement and smoothed groundwater depth data respectively.

[0012] S5: Construct a collaborative fusion model of Long Short-Term Memory Network and Gradient Boosting Decision Tree. Specifically, this includes: using GNSS vertical displacement data, InSAR vertical displacement data, leveling settlement, and groundwater depth data as basic input features; defining the geographical coordinates recorded in S1 and the distance to the nearest mining well as spatial features; and defining the year-day conversion of the UTC timestamp in S3 and the mining duration obtained from the mining company's operation records as time-derived features, thus forming complete input features; capturing the time dynamic features of the basic input features through the Long Short-Term Memory Network and outputting the hidden state; integrating the hidden state with the spatial features and time-derived features and inputting it into the Gradient Boosting Decision Tree to obtain the preliminary fusion result; and then dynamically adjusting the correlation score between the hidden state and the preliminary fusion result through an attention mechanism to output the final fused settlement, including the periodic settlement and the cumulative settlement.

[0013] S6: Based on the final fused settlement amount, perform time series analysis, spatial distribution analysis, and anomaly judgment to obtain the judgment results.

[0014] In this manual, the method for monitoring ground subsidence data in high geothermal extraction also includes S7: when S6 determines that there is an abnormal subsidence, a multi-level early warning is triggered and the abnormal point data is collected in a more intensive manner. At the same time, the original data, preprocessed data, fusion results, analysis results and extraction parameters are stored in the database.

[0015] In this specification, the construction of the S5 long short-term memory network model includes: controlling the flow of information through forget gate, input gate, candidate memory unit and output gate, using the difference in level subsidence at adjacent time points as the training label, using the mean squared error loss function and Adam optimizer for training, and outputting hidden states containing time dynamic features.

[0016] In this specification, the construction of the gradient boosting decision tree model in S5 includes: integrating the hidden state output by the long short-term memory network with the geographical coordinates, the distance to the nearest mining well, the number of days in a year, and the duration of mining into an input feature vector; generating multiple regression trees by iteratively fitting the residuals of the preceding model; and outputting the preliminary fusion result.

[0017] In this specification, the interaction process of the attention mechanism in S5 includes: calculating the attention score of the initial fusion result of the hidden state of the Long Short-Term Memory Network and the Gradient Boosting Decision Tree; obtaining the weight ratio of the Long Short-Term Memory Network through the softmax function; and weighted fusion of the two to output the final fused sedimentation amount. The weight ratio is tilted towards the Long Short-Term Memory Network during the sedimentation abrupt change period and towards the Gradient Boosting Decision Tree during the sedimentation stabilization period.

[0018] In this specification, the joint training process of the collaborative fusion model in S5 includes: using the mean square error of the final fused settlement amount and the level settlement amount as the total loss function, and simultaneously updating the weight matrix of the long short-term memory network, the tree structure of the gradient boosting decision tree, and the parameters of the attention mechanism through backpropagation until the total loss function converges.

[0019] In this manual, the threshold for correlation anomaly judgment in S6 is set as follows: when the groundwater level drops by more than 5m within one month and the monthly settlement rate of the corresponding monitoring point increases by more than 100%, it is judged as a correlation anomaly.

[0020] In this manual, the S7 multi-level early warning includes: Level 1 early warning corresponds to a cumulative settlement exceeding 200mm and a monthly settlement rate exceeding 15mm, triggering an audible and visual alarm and pushing information to multiple management departments; Level 2 early warning corresponds to a cumulative settlement of 100-200mm or a monthly settlement rate of 10-15mm, pushing information to mining companies and the Natural Resources Bureau; Level 3 early warning is for cases where the first two levels are not reached but the abnormal threshold is met, and it is only displayed on the monitoring platform.

[0021] In this manual, the data acquisition frequency in S3 is set as follows: GNSS data is acquired once every 30 minutes, InSAR data is acquired once every 7 days, leveling data is acquired once every 10 days, and groundwater level data is acquired once every 6 hours. All data timestamps are accurate to milliseconds and kept synchronized.

[0022] A high-temperature geothermal extraction ground subsidence data monitoring device, employing any one of the above-described high-temperature geothermal extraction ground subsidence data monitoring methods, the high-temperature geothermal extraction ground subsidence data monitoring device comprising:

[0023] The deployment module is used to determine the monitoring area and deploy monitoring points, and record the geographical coordinates of the monitoring points and their distance from the nearest mining well;

[0024] The debugging module is used to install and debug the corresponding monitoring devices at the monitoring points;

[0025] The data acquisition module is used to collect multi-source raw data at a set frequency through the debugged monitoring device. The raw data is transmitted to the data processing center with the monitoring point number and UTC timestamp attached.

[0026] The preprocessing module is used to preprocess the multi-source raw data to obtain GNSS vertical displacement data, InSAR vertical displacement data, leveling settlement and smoothed groundwater level depth data.

[0027] The fusion model construction module is used to build a collaborative fusion model of a long short-term memory network and a gradient boosting decision tree. Specifically, it includes: using GNSS vertical displacement data, InSAR vertical displacement data, leveling settlement, and groundwater depth data as basic input features; defining the geographical coordinates recorded in S1 and the distance to the nearest mining well as spatial features; and defining the year-day conversion of the UTC timestamp in S3 and the mining duration obtained from the mining company's operation records as time-derived features, thus forming complete input features; capturing the time dynamic features of the basic input features through the long short-term memory network and outputting the hidden state; integrating the hidden state with the spatial features and time-derived features and inputting it into the gradient boosting decision tree to obtain the preliminary fusion result; and then dynamically adjusting the correlation score between the hidden state and the preliminary fusion settlement result through an attention mechanism to output the final fusion settlement amount, including the periodic settlement amount and the cumulative settlement amount.

[0028] The analysis module is used to perform time series analysis, spatial distribution analysis, and anomaly detection based on the final fused settlement amount, and obtain the judgment results.

[0029] In summary, the present invention has at least the following beneficial effects:

[0030] 1. Improve monitoring accuracy and reliability: By synergistically integrating long short-term memory networks and gradient boosting decision trees, the temporal dynamic characteristics (such as the trend of settlement rate changes with mining time) and spatial nonlinear correlations (such as the difference in the impact of distance from the mining well on settlement) of multi-source data are fully explored, which solves the limitations of single technology and traditional fusion methods, and makes the settlement monitoring results more consistent with actual geological changes.

[0031] 2. Enhance anomaly identification and early warning capabilities: The fusion model can effectively capture sudden subsidence changes (such as accelerated subsidence caused by a sudden drop in groundwater level). Combined with a multi-level early warning mechanism, it can achieve timely detection and graded response to anomalies, providing accurate basis for the safety management of geothermal mining areas and reducing the risk of geological disasters.

[0032] 3. Supporting sustainable development decisions: Full-process data management and analysis (from monitoring point deployment to subsidence pattern analysis) can systematically reveal the correlation mechanism between geothermal extraction and land subsidence, providing scientific support for optimizing extraction plans (such as adjusting extraction volume and laying out supplementary water sources) and balancing resource development and geological environmental protection. Attached Figure Description

[0033] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 This is a flowchart illustrating the method for monitoring ground subsidence data in high geothermal extraction involved in this invention.

[0035] Figure 2 This is a schematic diagram illustrating the process of setting up monitoring areas and points involved in this invention.

[0036] Figure 3 This is a schematic diagram of the data fusion processing involved in this invention.

[0037] Figure 4 This is a schematic diagram of the settlement analysis and early warning process involved in this invention. Detailed Implementation

[0038] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0039] The following disclosure provides many different implementations or examples for carrying out different structures of the embodiments of the present invention. To simplify the disclosure of the embodiments of the present invention, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the embodiments of the present invention. Furthermore, reference numerals and / or reference letters may be repeated in different examples of the embodiments of the present invention; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various implementations and / or arrangements discussed.

[0040] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0041] like Figure 1 As shown in the figure, this embodiment provides a method for monitoring ground subsidence data in high geothermal extraction, including the following steps:

[0042] S1: Determine the monitoring area and set up monitoring points, and record the geographical coordinates of the monitoring points and their distance from the nearest well.

[0043] S2: Install and debug the corresponding monitoring devices at the monitoring points;

[0044] S3: Collect multi-source raw data at a set frequency using the calibrated monitoring device. The raw data is transmitted to the data processing center with the monitoring point number and UTC timestamp appended.

[0045] S4: Preprocess the multi-source raw data to obtain GNSS vertical displacement data, InSAR vertical displacement data, leveling settlement and smoothed groundwater depth data respectively.

[0046] S5: Construct a collaborative fusion model of Long Short-Term Memory Network and Gradient Boosting Decision Tree. Specifically, this includes: using GNSS vertical displacement data, InSAR vertical displacement data, leveling settlement, and groundwater depth data as basic input features; defining the geographical coordinates recorded in S1 and the distance to the nearest mining well as spatial features; and defining the year-day conversion of the UTC timestamp in S3 and the mining duration obtained from the mining company's operation records as time-derived features, thus forming complete input features; capturing the time dynamic features of the basic input features through the Long Short-Term Memory Network and outputting the hidden state; integrating the hidden state with the spatial features and time-derived features and inputting it into the Gradient Boosting Decision Tree to obtain the preliminary fusion result; and then dynamically adjusting the correlation score between the hidden state and the preliminary fusion result through an attention mechanism to output the final fused settlement, including the periodic settlement and the cumulative settlement.

[0047] S6: Based on the final fused settlement amount, perform time series analysis, spatial distribution analysis, and anomaly judgment to obtain the judgment results.

[0048] In some embodiments, the method for monitoring ground subsidence data in high geothermal extraction further includes S7: when S6 determines that there is an abnormal subsidence, a multi-level early warning is triggered and abnormal point data is collected in an encrypted manner, while the original data, preprocessed data, fusion results, analysis results and extraction parameters are stored in the database.

[0049] In some embodiments, the construction of the long short-term memory network model in S5 includes: controlling the flow of information through a forget gate, an input gate, candidate memory units and an output gate, using the difference in level subsidence at adjacent times as the training label, training with a mean squared error loss function and an Adam optimizer, and outputting hidden states containing time dynamic features.

[0050] In some embodiments, the construction of the gradient boosting decision tree model in S5 includes: integrating the hidden state output by the long short-term memory network with the geographical coordinates, the distance to the nearest mining well, the number of days in a year, and the duration of mining into an input feature vector; generating multiple regression trees by iteratively fitting the residuals of the preceding model; and outputting the preliminary fusion result.

[0051] In some embodiments, the interaction process of the attention mechanism in S5 includes: calculating the attention score of the hidden state of the long short-term memory network and the preliminary fusion result of the gradient boosting decision tree; obtaining the weight ratio of the long short-term memory network through the softmax function; and weightedly fusing the two to output the final fused sedimentation amount, wherein the weight ratio tilts towards the long short-term memory network during the sedimentation abrupt change period and towards the gradient boosting decision tree during the sedimentation stabilization period.

[0052] In some embodiments, the joint training process of the collaborative fusion model in S5 includes: using the mean square error of the final fused settlement amount and the level settlement amount as the total loss function, and simultaneously updating the weight matrix of the long short-term memory network, the tree structure of the gradient boosting decision tree, and the parameters of the attention mechanism through backpropagation until the total loss function converges.

[0053] In some embodiments, the threshold for correlation anomaly judgment in S6 is set as follows: when the groundwater level drops by more than 5m within one month and the monthly settlement rate of the corresponding monitoring point increases by more than 100%, it is judged as a correlation anomaly.

[0054] In some embodiments, the multi-level early warning in S7 includes: Level 1 early warning corresponds to a cumulative settlement exceeding 200mm and a monthly settlement rate exceeding 15mm, triggering an audible and visual alarm and pushing information to multi-level management departments; Level 2 early warning corresponds to a cumulative settlement of 100-200mm or a monthly settlement rate of 10-15mm, pushing information to mining enterprises and the natural resources bureau; Level 3 early warning is for cases that do not reach the first two levels but meet the abnormal threshold, and is only displayed on the monitoring platform.

[0055] In some embodiments, the data acquisition frequency in S3 is set as follows: GNSS data is acquired once every 30 minutes, InSAR data is acquired once every 7 days, leveling data is acquired once every 10 days, and groundwater level data is acquired once every 6 hours, and the timestamps of all data are accurate to milliseconds and kept synchronized.

[0056] The technical concept of this invention is as follows:

[0057] This solution addresses the need for land subsidence risk prevention and control during high-temperature geothermal extraction. It constructs a comprehensive data monitoring system encompassing "monitoring-processing-fusion-analysis-early warning-storage," achieving high-precision acquisition and dynamic management of subsidence data through multi-source data collaboration and intelligent algorithm integration. This provides technical support for the safety of high-temperature geothermal extraction and the protection of the surrounding environment. Specifically, it includes the following:

[0058] S1: Determine the monitoring area and set up monitoring points

[0059] Core objective: To provide a comprehensive and rationally distributed monitoring foundation for subsequent data collection, fusion, and modeling, ensuring that the Long Short-Term Memory Network and Gradient Boosting Decision Tree model in S5 can effectively capture temporal and spatial characteristics. The monitoring area and point deployment process is as follows: Figure 2 As shown.

[0060] 1. Delineate monitoring areas

[0061] Centered on the high geothermal extraction area, and based on the geological survey report (including soil layer distribution and fault location) and extraction planning map, a three-level monitoring scope is defined:

[0062] Core area: Within 500m of the mining well, it is necessary to capture subsidence directly affected by intense mining activities;

[0063] Impact zone: 500m to 1500m, reflecting the subsidence and diffusion caused by indirect mining effects;

[0064] Background area: 1500m to 2000m, used as the settlement benchmark (assuming it is not affected by mining).

[0065] At the same time, settlement-sensitive areas should be identified, including: densely built areas (such as residential buildings and factories), linear engineering areas (such as oil pipelines and highways), and concentrated agricultural land areas (such as paddy fields and vegetable fields). Monitoring points should be increased in these areas.

[0066] 2. Deploy monitoring points

[0067] Four types of monitoring points were deployed according to the principle of "spatial stratification and functional complementarity." All points required WGS-84 coordinates (longitude L, latitude B, accurate to 0.001 seconds) to be obtained using a total station, and the distance D to the nearest production well was recorded (calculated using the Haversine formula, accurate to 0.1m). This provided the basic data for the spatial feature construction of S5.1.

[0068] GNSS monitoring points: one every 150m in the core area, one every 300m in the influence area, and one every 500m in the background area; a reinforced concrete base with a diameter of 30cm and a burial depth of 2m is used, and a forced centering device is installed on the top (to ensure that the deviation between the phase center of the GNSS antenna and the center of the base is ≤±0.5mm) for acquiring three-dimensional coordinate time series.

[0069] InSAR corner reflector monitoring points: evenly distributed between GNSS points (spacing 150m to 200m), selecting unobstructed areas (no buildings / trees taller than 3m within 50m of the surrounding area); the reflectors are made of 20cm×20cm metal plates, fixed by brackets (perpendicular to the ground, error ≤±1°) to ensure stable radar echo intensity and provide high-quality signals for S4.2 phase unwrapping.

[0070] Leveling monitoring points: These are set up to coincide with GNSS points (crosshairs are engraved on the top of the GNSS base as leveling markers) to provide a high-precision elevation benchmark (second-order leveling accuracy) and serve as label data for model training in S5.

[0071] Groundwater level monitoring points: One point is set up at 200m, 500m and 1000m around each mining well, with the drilling depth reaching the confined aquifer (determined according to the previous hydrological survey, usually 30m to 100m); stainless steel casing (100mm inner diameter) is installed, with a filter hole at the bottom (to prevent silt from entering) to provide a stable environment for S4.4 water level data collection.

[0072] S2: Installation and commissioning of monitoring devices

[0073] Core objective: To ensure that the accuracy, time synchronization, and transmission stability of the data collected by each monitoring device meet the requirements of the model in S5 for input data (such as the high sensitivity of long short-term memory networks to the continuity of time series).

[0074] 1. Install monitoring devices

[0075] GNSS monitoring unit: A GNSS receiver (supporting GPS+BeiDou dual-mode, sampling rate 1Hz) is installed on the base forced alignment device, and the antenna cable is introduced into the ground distribution box through a waterproof pipe; the receiver needs to be synchronized with UTC time (error ≤ ±1ms) to ensure the time consistency of data from different points.

[0076] InSAR corner reflector unit: The bottom of the reflector bracket is cast and connected to the concrete foundation (dimensions 1m×1m×0.5m), the bracket height is 1.5m (to avoid ground interference), the reflective surface faces south (consistent with the direction of the satellite orbit), and the surface is covered with reflective film (to enhance the echo intensity).

[0077] Groundwater level monitoring unit: A submersible water level sensor (range 0-100m, accuracy ±0.1%FS) is inserted into the casing. The sensor probe is located 1m below the stable groundwater level. The cable is fixed along the inner wall of the casing (to avoid shaking that could cause reading fluctuations). It outputs a 4-20mA analog signal.

[0078] The data transmission module is equipped with a 4G / 5G transmission terminal (supporting TCP / IP protocol) at each monitoring point, which is connected to the monitoring device via RS485 bus (communication rate 9600bps); the core area monitoring points are additionally equipped with fiber optic backup (bandwidth 100Mbps) to prevent wireless signal interruption.

[0079] 2. Debugging and monitoring device

[0080] GNSS receiver debugging: Power on continuously for 24 hours and record the number of satellites tracked (≥10), signal-to-noise ratio (≥45dB), and positioning accuracy (horizontal ±2mm, vertical ±5mm); if the number of satellites is less than 8 at a certain time, check for surrounding obstructions and adjust the antenna position.

[0081] InSAR reflector debugging: Use a drone equipped with SAR equipment (simulating satellite observation) to photograph the reflector and analyze the echo signal strength (must be ≥ 15 times the background noise); if the signal is weak, clean the stains on the reflector surface or adjust the verticality of the bracket.

[0082] Water level sensor calibration: Manually measure the water level with a measuring rope (accuracy ±1cm) and compare it with the sensor reading (error must be ≤±2cm); if the error exceeds the tolerance, recalibrate the sensor (adjust the zero point and range using the calibration software provided by the manufacturer).

[0083] Data transmission debugging: Send 100 sets of test data (including timestamps, location numbers, and analog values) to the data processing center to verify transmission delay (≤5s), packet loss rate (≤0.05%), and encryption effectiveness (using AES-256 encryption, and the decrypted data is consistent with the original).

[0084] S3: Collect monitoring data

[0085] Core objective: To acquire high-quality, time-aligned multi-source raw data to provide input for S4 preprocessing and S5 modeling, especially ensuring the continuity of the time series.

[0086] 1. Set the acquisition frequency and time synchronization.

[0087] Based on model requirements (Long Short-Term Memory networks need to capture short-term fluctuations and long-term trends) and device characteristics, the data acquisition frequency is set as follows:

[0088] GNSS data: Acquired every 30 minutes for 5 minutes each time (sampling interval 1 second), obtaining raw observation values ​​such as pseudorange and carrier phase; high-precision calculations are performed at 0:00, 6:00, 12:00 and 18:00 every day (to ensure synchronization with the precise ephemeris of the International GNSS Service Organization (IGS)).

[0089] InSAR data: collected once every 7 days (selected in rain-free and fog-free weather, during satellite transit time), covering all reflector locations, and acquiring radar complex images (including amplitude and phase information).

[0090] Leveling data: collected once every 10 days (synchronized with GNSS data collection day), using second-order leveling method (forward and backward measurements, with an error of ≤±1mm per kilometer of elevation difference), and recording raw data such as foresight and backsight distances and readings.

[0091] Groundwater level data: collected once every 6 hours (on the hour: 0:00, 6:00, 12:00, 18:00), aligned with the GNSS data timestamp, facilitating the correlation analysis between water level and settlement in S5.

[0092] 2. Data Acquisition and Transmission Process

[0093] Each device automatically collects data at a set frequency and attaches a unique identifier (monitoring point number + UTC timestamp, accurate to milliseconds); for example, GNSS data is identified as “GNSS-001-20xx0520060000.000”, ensuring accurate alignment of the time series in S5.1.

[0094] Data transmission adopts a "real-time sending + local caching" mechanism: the raw data is encrypted and sent to the data processing center through the transmission module, while being stored on a local SD card (capacity 64GB, cyclically overwritten); if the transmission fails (such as network interruption), the cached data is automatically retransmitted after recovery to avoid time series breaks during S5 modeling.

[0095] After receiving the data, the data processing center immediately returns a "receive confirmation" signal (including a verification code); if the monitoring device does not receive confirmation within 10 minutes, it will automatically resend the data (up to 3 retries) to ensure data integrity.

[0096] S4: Preprocessing monitoring data

[0097] Core objective: To clean, correct, and standardize the raw data to generate the input features (such as GNSS vertical displacement, water level depth, etc.) required for the Long Short-Term Memory Network and Gradient Boosting Decision Tree Model in S5, while removing outliers (to avoid affecting model training).

[0098] 1. GNSS data preprocessing

[0099] The GAMIT / GLOBK software is used for processing to output high-precision vertical displacement.

[0100] Data editing: Remove gross errors (pseudorange observations exceeding 3 times the mean square error) and interruptions (continuous missing data > 5 minutes), and retain continuous observation sequences; for example, if satellite signal interference causes positioning errors > 10 mm during a certain period, mark it as invalid data and remove it.

[0101] Precision calibration: Import the 15-minute sampled precise ephemeris (5cm accuracy) provided by IGS for orbit calibration; use the Saastamoinen model to correct ionospheric delay and the VMF1 model to correct tropospheric delay (input meteorological data such as station elevation, temperature, and air pressure).

[0102] Output: Calculate the vertical displacement at each time step. (Relative to the background area reference point, unit: mm), stored in CSV format (including monitoring point number, timestamp, ... ), which is one of the input features of S5.1.

[0103] 2. InSAR Data Preprocessing

[0104] Vertical settlement was extracted using SARscape software.

[0105] Image registration: Using the first phase image as a reference, perform geometric correction on subsequent images (error ≤ 0.5 pixels) to ensure that the reflector points are aligned.

[0106] Terrain removal: Import 30m resolution DEM data (obtained from the National Geographic Information Public Service Platform) to eliminate the influence of terrain undulations on the phase and obtain an interferometric phase map.

[0107] Phase optimization: Goldstein filtering (5×5 window size) is used to suppress speckle noise, followed by phase unwrapping using the minimum cost flow algorithm (to avoid phase jumps), ultimately converting the phase into vertical displacement. (Unit: mm), stored in GeoTIFF format (associated reflector coordinates), as input features for S5.1.

[0108] 3. Leveling data preprocessing

[0109] The data was processed according to the second-order leveling measurement standard and used as the baseline label for model training.

[0110] Closure error calculation: Calculate the closure error according to the "closed route" (e.g., starting from point A, passing through points B and C, and returning to point A). If the closure error > ±4... mm (L is the route length in km), remeasure; otherwise, allocate the closure error by distance weighting. For example, if the length of a closed route is L = 5 km, the allowable closure error is ±4. ±8.94mm.

[0111] Elevation calculation: Obtain the absolute elevation of each monitoring point, and then calculate the difference with the previous period, i.e., the leveling settlement. (Unit: mm), save as an Excel spreadsheet (including location, timestamp, ... ), serving as training labels for the Long Short-Term Memory network and Gradient Boosting Decision Tree in S5.

[0112] 4. Groundwater level data preprocessing

[0113] Eliminate noise and smooth fluctuations to extract effective water level features:

[0114] Outlier removal: Remove data that exceeds the sensor's range (0-100m) or abruptly changes (the difference between two adjacent values ​​is >1m, which may indicate a sensor malfunction); for example, if the water level suddenly changes from 20m to 50m at a certain moment, mark it as an anomaly and delete it.

[0115] Smoothing: The data was smoothed using a triple moving average method (window=3) to obtain... (Unit: m), save as a text file (including location, timestamp, ... ), which is a key feature in S5.1 that reflects the impact of mining.

[0116] S5: Fusion of Preprocessed Data (A Collaborative Fusion Method Based on Long Short-Term Memory Networks and Gradient Boosting Decision Trees)

[0117] The core of S5 is to fuse the multi-source preprocessed data (GNSS subsidence, InSAR subsidence, leveling subsidence, and groundwater level data) output by S4 into a high-precision ground subsidence measurement through the collaborative operation of a Long Short-Term Memory (LSTM) network and a Gradient Boosting Decision Tree (GBDT). This process requires first constructing a unified feature set, then using the LSTM network to capture temporal dynamics, the Gradient Boosting Decision Tree to mine spatial and nonlinear correlations, and finally using an attention mechanism to achieve collaborative optimization of both, ensuring that the fused result reflects both temporal trends and spatial distribution characteristics. The data fusion processing flow is as follows: Figure 3 As shown.

[0118] S5.1 Data Preprocessing and Feature Construction

[0119] To ensure that the Long Short-Term Memory network and gradient boosting decision tree model can effectively extract data patterns, the preprocessing results of S4 need to be integrated to construct an input feature set containing temporal, spatial, and derived information. This feature set is the foundation for subsequent model training and application, and directly affects the model's ability to capture sedimentation patterns.

[0120] 1. Basic Data Integration

[0121] Using the monitoring point number and collection timestamp as keywords, the four types of preprocessed data output by S4 are correlated to form time series data for a single monitoring point. Let the basic data of a monitoring point at the t-th collection time (t=1,2,...,T, where T is the total number of collections) be:

[0122] ;

[0123] in:

[0124] The vertical displacement of GNSS output by S4.1 (unit: mm) reflects the change in the Z-axis of the three-dimensional coordinates of the monitoring point at that moment. S4.2 output InSAR vertical displacement (unit: mm), the preliminary settlement calculated from phase unwrapping;

[0125] The leveling settlement (unit: mm) output by S4.3 is calculated by the difference between two adjacent elevation measurements, and has the highest accuracy, serving as a reference benchmark for model training.

[0126] The groundwater level depth (unit: m) output by S4.4 reflects the impact of geothermal extraction on groundwater bodies and is strongly correlated with land subsidence.

[0127] T: Vector transpose operation, converting a row vector into a column vector (dimension: ...). ).

[0128] 2. Spatial Feature Supplement

[0129] Spatial attributes of monitoring points are introduced to reflect the impact of geographical location on settlement (e.g., settlement may be more significant the closer the location is to the well). Spatial characteristics include:

[0130] longitude with latitude (Unit: degrees): Coordinates of the monitoring points recorded in S1 in the WGS-84 coordinate system. , It does not change over time;

[0131] Distance to the nearest well (Unit: m): Calculated using the coordinates of the production well and monitoring points in S1 (using the Haversine formula), i.e. ,in The coordinates of the most recently exploited well (obtained from a high geothermal extraction company).

[0132] 3. Construction of Time-Derived Features

[0133] Extract derived information from the time dimension to capture the periodic or long-term trends of subsidence (such as the impact of seasonal changes in mining intensity on subsidence):

[0134] Yearly accumulated days Convert the data collection timestamp to the day of the year (e.g., January 1st). =1, June 30 =181, December 31 =365), reflecting seasonal factors;

[0135] Mining duration (Unit: days): The number of days from the start of geothermal extraction in the area where the monitoring point is located to the current collection time (the start date of extraction is obtained from the company's operation records), reflecting the cumulative impact of extraction on the geological structure.

[0136] 4. Feature set integration

[0137] Arrange the above features in time series to form the input feature set for the Long Short-Term Memory Network and the Gradient Boosting Decision Tree: Each feature vector is:

[0138] Dimensions: It includes 4 types of basic data, 2 types of spatial features, and 2 types of time-derived features.

[0139] S5.2 Long Short-Term Memory Network (LSTM) Model Construction and Training

[0140] The core function of the LSTM model is to capture the dynamic temporal correlation of settlement data. For example, if the settlement rate at a monitoring point accelerates for three consecutive months, LSTM can identify this trend and predict subsequent changes, thus overcoming the shortcomings of traditional methods in capturing "time dependence." Its output will serve as a key input feature for the gradient boosting decision tree model, conveying the patterns in the time dimension.

[0141] (1) Model structure design

[0142] LSTM stores historical information through "memory units" and controls the flow of information through input gates, forget gates, and output gates, avoiding the gradient vanishing problem of traditional recurrent neural networks (RNNs). Considering the characteristics of geothermal mining subsidence data (cycles of approximately 7–30 days, exhibiting short-term fluctuations and long-term trends), a single-hidden-layer LSTM is designed with a hidden layer dimension of 64 (determined through cross-validation to balance accuracy and computational cost). The specific formula is as follows:

[0143] Forgotten Gate ( ): Determines how much information to retain from historical memory (for example, if the settling is stable for the first 10 days, the forgetting gate will retain this stable trend).

[0144] (5-1)

[0145] Input gate ( ): Determines how much new information (such as a sudden drop in water level) should be stored at the current moment.

[0146] (5-2)

[0147] Candidate memory units ( ): Temporarily stores new information at the current moment, including a compressed representation of the input features.

[0148] (5-3)

[0149] Memory unit update ( ): Integrating historical memory with current new information to form an updated memory

[0150] ;5-4

[0151] Output gate ( ): Determines how much information to output from the memory unit as the current hidden state.

[0152] (5-5)

[0153] Hidden state (h_{LSTM,t}): The final output of the LSTM, containing the temporal characteristics of the current moment (such as recent subsidence trends).

[0154] (5-6)

[0155] Weight matrices for forget gate, input gate, candidate memory units, and output gate (all dimensions are 1). 9 are input features (64 is the dimension of the hidden layer).

[0156] : The corresponding bias vector (all dimensions are 1) );

[0157] : sigmoid activation function (output range [0,1], used to control the degree to which the door is open or closed);

[0158] Hyperbolic tangent activation function (output range [-1, 1], used to compress memory cell information);

[0159] Element-wise multiplication (multiplying elements at corresponding positions);

[0160] The hidden state at the previous moment (initial value) Dimension );

[0161] : The memory unit from the previous moment (initial value) Dimension );

[0162] The hidden state at the current moment (output, dimension) ).

[0163] (2) Model training process

[0164] The training objective is to enable the LSTM to accurately capture the temporal trend of settlement. The training data, loss function, and optimization strategy are as follows:

[0165] Training dataset partitioning: Time series cross-validation was used, selecting the top 80% of the feature set. As the training set (e.g., when the total number of data collections T=100, take the first 80 data collections), the remaining 20% ​​( ) as the validation set.

[0166] Label data: Using the level settlement output by S4.3 as the benchmark (because it has the highest accuracy), the settlement difference between adjacent time points is calculated as the training labels:

[0167] (Unit: mm, reflecting the trend of settlement rate at time t);

[0168] Loss function: The mean squared error (MSE) is used to measure the difference between the predicted value and the label. The formula is as follows:

[0169] (5-7)

[0170] Where N=0.8T is the number of training samples. Let be the LSTM prediction value at time t.

[0171] Optimizers and Iterations: Using the Adam optimizer (suitable for optimizing non-stationary data), setting the learning rate. (Control parameter update magnitude), attenuation rate , (Control the momentum term) Iterate through the training for 50 rounds. After each training round, calculate the loss using the validation set. If the loss does not decrease for 5 consecutive validation rounds, stop training early (to prevent overfitting).

[0172] (3) Model application output

[0173] validation set data Input the trained LSTM and output the hidden states at each time step. (dimension) The output contains the temporal dynamic characteristics of the monitoring points (such as "accelerated settlement in the last 3 collections" and "seasonal fluctuation patterns"), which will serve as one of the input features for the subsequent GBDT model, realizing the transfer of temporal information to the spatial model.

[0174] S5.3 Gradient Boosting Decision Tree (GBDT) Model Construction and Training

[0175] The core function of the GBDT model is to integrate spatial features, temporally derived features, and the temporal features of the LSTM output to capture nonlinear correlations (for example, the increase in settlement when the groundwater level drops by 5m may differ from that when it drops by 10m; GBDT can fit this nonlinear relationship). Its input includes the temporal features of the LSTM, ensuring a "temporal-spatial" synergy with the LSTM.

[0176] (1) Model input and structural design

[0177] The input features of GBDT need to integrate spatial, temporal derivatives, and LSTM temporal features to form a comprehensive feature vector:

[0178] (5-8)

[0179] in:

[0180] : Spatial and temporal derived features (dimensions) as defined in S5.1 );

[0181] The transpose of the hidden state of the LSTM output (dimension) After conversion to a column vector, the dimension is ;

[0182] therefore, The total dimensions are 5 + 64 = 69 ( ).

[0183] GBDT is an ensemble of M regression trees (CART trees), which improves accuracy by iteratively fitting the residuals. The output at time t is the weighted sum of the predictions from each tree.

[0184] (5-9)

[0185] in:

[0186] M: Number of trees (set to 100 in experiments to balance accuracy and computational load).

[0187] The m-th regression tree pairs features The predicted value (leaf node output);

[0188] : The weight of the m-th tree (initially 0, optimized through training).

[0189] (2) Model training process

[0190] GBDT uses an additive model for iterative optimization, where each new tree is fitted with the prediction residuals of the previous model. The specific steps are as follows:

[0191] 1. Initialize the model: The first tree is initialized with the mean of the training labels to ensure a reasonable starting point for the model.

[0192] ,in (The average leveling settlement of the training area);

[0193] The initial output is: .

[0194] 2. Iterative fitting residuals (for m=2,3,...,M):

[0195] Calculate the residuals: The residuals in the m-th round are the difference between the true value and the current model prediction, reflecting the errors that the model has not yet captured.

[0196] (5-10)

[0197] Fitting Residual Trees: Using Regression Trees Fitting residuals Leaf nodes are divided by the Gini coefficient. (k=1,2,...,K, where K is the number of leaf nodes, set to 8), each leaf node corresponds to a set of samples with similar features.

[0198] Optimize tree weights: Calculate the optimal weight for each leaf node in the m-th tree. To minimize the sum of squared residuals of samples within that node:

[0199] (5-11)

[0200] in leaf node Number of samples included.

[0201] Update model output: Introduce learning rate (To prevent overfitting), gradually accumulate the predictions of the new tree:

[0202] (5-12)

[0203] in For indicator functions ( (If it is 1, then it is 0).

[0204] 3. Training Termination: Training stops when the number of iterations reaches M=100, and the final output is... .

[0205] (3) Model application output

[0206] The feature vector of the validation set (including LSTM output) Input the trained GBDT to obtain the preliminary fused sedimentation amount. (Unit: mm). This result integrates spatial, temporal derivation, and temporal characteristics transmitted by LSTM, but its sensitivity to "temporal abrupt changes" (such as a sudden drop in water level leading to a sudden increase in settlement) still needs to be improved. Therefore, further co-optimization with LSTM output is required.

[0207] Synergistic integration of S5.4LSTM and GBDT (attention mechanism)

[0208] To address the issue of complementing the strengths of LSTM (strong time sensitivity) and GBDT (strong spatial fit), an attention mechanism is introduced to dynamically adjust their weights. During the stable settling period, GBDT's spatial fit is more reliable and is given a higher weight; during the abrupt settling period, LSTM's temporal trend prediction is more critical and is given a higher weight.

[0209] (1) Calculation of attention weights

[0210] The attention mechanism measures the correlation between the LSTM output and the GBDT output using an "attention score." A higher score indicates that the temporal characteristics of the LSTM are more important at that moment. Specific steps:

[0211] Attention score ( ): Calculate the correlation strength between the LSTM hidden state and the GBDT output through a fully connected layer:

[0212] (5-13)

[0213] in:

[0214] Attention weight vector (dimensions) (used for compressing features).

[0215] : Weight matrix of LSTM features (dimensions) );

[0216] The weight matrix (dimensions) output by GBDT );

[0217] Bias vector (dimension) ).

[0218] Attention weights ( The scores are normalized using the softmax function to obtain the relative importance weights of the LSTM (range [0,1]):

[0219] (5-14)

[0220] For example: if at a certain moment Much greater than at other times, This indicates that the temporal characteristics of LSTM play a dominant role.

[0221] (2) Calculation of final fusion settlement

[0222] By incorporating attention weights, the temporal features of the LSTM are weighted and fused with the preliminary fusion results of GBDT:

[0223] (5-15)

[0224] in The normalized result of the LSTM hidden state ( Mapped to the order of settlement).

[0225] ;

[0226] For min-max normalization, Compress to [0,1], then multiply by the maximum value of the training settling level to ensure consistency with... They are of the same magnitude.

[0227] (3) Collaborative training and optimization

[0228] To ensure coordinated parameter optimization of LSTM, GBDT, and the attention mechanism, a joint training strategy is adopted:

[0229] Total loss function: based on the final fusion result With horizontal settlement The target is MSE:

[0230] (5-16)

[0231] Backpropagation update: calculated using the chain rule For LSTM parameters ( GBDT parameters (tree structure and weights), attention parameters ( The gradient of ) is used to update all parameters simultaneously, so that the three are optimized together in the direction of reducing the total loss.

[0232] S5.5 generates a fusion result table.

[0233] After fusion, the results need to be correlated with the original information to form structured data that can be directly used for S6 analysis.

[0234] 1. Stage-by-stage settlement: The difference between the fusion results at time t and time t-1, i.e. (Reflects the settlement rate within this period);

[0235] 2. Cumulative Settlement: The total settlement from the start of monitoring to time t, i.e. ;

[0236] 3. Related Information: This includes linking the periodic settlement volume, cumulative settlement volume, monitoring point number, data collection timestamp, and groundwater level data. The data is then linked to generate a "Ground Settlement Integration Result Table" and stored in the database.

[0237] Summary of core functions:

[0238] LSTM captures the temporal dynamics of subsidence (such as accelerated subsidence due to increased mining intensity) through memory cells, thus solving the problem of insufficient modeling of "time dependence" in traditional methods.

[0239] GBDT integrates multi-source features to explore the nonlinear effects of factors such as spatial location and groundwater level (e.g., the different responses of different soil layers to water level changes).

[0240] The attention mechanism dynamically balances the weights of both methods, relying on the spatial fitting of GBDT during the stable settling period and the temporal sensitivity of LSTM during the abrupt change period, ultimately improving the fusion accuracy to ±1.5mm (a 25% improvement over the original S5 method of ±2mm), providing a high-precision data foundation for the settlement analysis of S6.

[0241] S6: Analyze ground subsidence data

[0242] Core Objective: Based on the high-precision fused settlement data output from S5, this study analyzes settlement patterns from temporal and spatial dimensions, identifies anomalies, and provides a basis for S7 early warning. The high precision (±1.5mm) of the fused results enhances the reliability of the analysis. The settlement analysis and early warning process is as follows: Figure 4 As shown.

[0243] 1. Time series analysis

[0244] Extract the periodic settlement data of the same monitoring point from the "Ground Settlement Integration Results Table". and cumulative settlement Combined with groundwater level data Analyze the evolution characteristics of subsidence over time:

[0245] Trend Analysis: Plot a settlement-time curve (horizontal axis: time, vertical axis: ...). The monthly average settlement rate was calculated using linear fitting. (n is the number of times data is collected in the current month); for example, if a monitoring point has collected data for 3 consecutive months... The increase from 2mm to 8mm indicates accelerated settlement.

[0246] Correlation analysis: Plot the groundwater level curve (vertical axis is...). ), and calculate the correlation coefficient by comparing it with the settlement curve. (cov is the covariance, σ is the standard deviation); if (Strong negative correlation) indicates that the drop in water level is the main cause of subsidence.

[0247] 2. Spatial Distribution Analysis

[0248] Cumulative settlement based on fusion results Based on the coordinates of the monitoring points, a spatial distribution map was drawn to reveal the spatial diffusion pattern of settlement:

[0249] Settlement zoning: The Jenks method was used to zonify the settlement points. The settlement is divided into four levels: slight settlement (0-50mm), moderate settlement (50-100mm), severe settlement (100-200mm), and extremely severe settlement (>200mm), which are marked on the map with different colors (light green → yellow → orange → red).

[0250] Delineation of the impact area: Taking the well as the center, calculate the average cumulative settlement at different distances D. Fitted decay curve (a, b, and c are fitting parameters); for example, if b = 0.005, it means that for every 200m increase in distance, the settlement decreases by about 63%.

[0251] 3. Anomaly detection

[0252] Based on geological conditions (obtained from the survey report) and engineering standards, a three-level anomaly threshold is established to determine whether the settlement exceeds the safe range:

[0253] Abnormal velocity: Soft soil area >10mm, hard soil area >5mm (soft soil has high compressibility, so the threshold is more lenient);

[0254] Abnormal cumulative amount: around the building >100mm (may cause wall cracking), S_t>200mm for agricultural land (may affect irrigation);

[0255] Related anomalies: If the water level drops by more than 5 meters within one month and An increase of more than 100% (such as from 3mm to 6mm) is considered a chain of anomalies caused by mining.

[0256] For any monitoring point that meets any threshold, mark it as an "abnormal point" and record its number, abnormality type and occurrence time as the trigger condition for S7 early warning.

[0257] S7: Settlement Anomaly Early Warning and Data Storage

[0258] Core objective: To respond promptly to anomalies identified by S6, reduce disaster risks through a multi-level early warning mechanism, and save full-process data for subsequent analysis and tracing.

[0259] 1. Abnormal Warning

[0260] Early warning classification: Divided into three levels according to the severity of the anomaly:

[0261] Level 1 Warning (Extremely Severe): >200mm and If the distance is greater than 15mm, an audible and visual alarm will be triggered (red light flashing on the on-site terminal + 110dB buzzer), and an SMS message and platform push will be sent to the mining company, the Natural Resources Bureau, and the Emergency Management Bureau (delivered within 5 minutes).

[0262] Level 2 Warning (Severe): or The on-site terminal flashes a yellow light, sending an early warning to the mining company and the Natural Resources Bureau;

[0263] Level 3 Warning (Minor): If the level does not reach Level 1 or 2 but meets the abnormal threshold, the warning information will only be displayed on the monitoring platform.

[0264] Encrypted monitoring: After an early warning is triggered, the GNSS acquisition frequency of the abnormal point is increased to once every 10 minutes, and the water level is acquired once every hour, with continuous monitoring for 72 hours; if the settlement rate drops to within the threshold, the original frequency is restored; otherwise, the early warning level is upgraded.

[0265] 2. Data storage and management

[0266] Storage contents: including raw data from S3 (GNSS pseudorange, InSAR imagery, etc.) and preprocessing results from S4 ( , The fusion results of S5 (etc.) , Analysis reports (trend curves, anomaly records) and mining parameters (mining volume, water temperature, etc., obtained from the company) for S6.

[0267] Storage strategy: Uses a distributed database (MySQL + MongoDB), with storage partitioned by "year + quarter" (e.g., 2024Q1, 2024Q2), supporting 1000 data writes per second; configured with RAID5 redundancy (automatic recovery in case of hard drive failure), and automatic backup at 3 AM every day (retaining the most recent 3 months of backups).

[0268] Data services: Provides a web query interface (supporting filtering by location, time, and data type) and export function (formats include Excel, Shapefile, and CSV) to facilitate subsequent research (such as settlement mechanism analysis and mining scheme optimization).

[0269] In some embodiments, a surface fitting algorithm is introduced to construct a collaborative fusion method based on long short-term memory networks, gradient boosting decision trees, and surface fitting.

[0270] Data preprocessing and feature construction

[0271] Based on the existing feature set, spatial mesh features required for surface fitting are added to provide a foundation for the subsequent construction of continuous spatial surfaces:

[0272] 1. Spatial grid division: Divide the monitoring area into 50m × 50m grids, with each grid cell numbered (u, v) (u being the horizontal index and v the vertical index), and record the coordinates of the grid center point. (Calculated via grid boundaries).

[0273] 2. Feature set expansion: In the original feature vector Add grid index (u,v) and center point coordinates (u,v) to the grid. This forms an extended feature set:

[0274] ;

[0275] Dimensions: Three new spatial grid features have been added.

[0276] Surface Fitting Algorithm Model Construction and Training

[0277] The surface fitting algorithm is used to construct a continuous settlement spatial surface in the monitoring area, filling the data gaps in sparse areas of the monitoring points. Its output will serve as a supplementary feature to LSTM and GBDT, improving spatial continuity.

[0278] (1) Model building

[0279] A quadratic polynomial surface model is used to fit the spatial settlement distribution. The settlement surface function at time t is assumed to be:

[0280] (5-17)

[0281] Parameter definition:

[0282] : Settlement of the surface fitted at coordinates (L,B) (unit: mm);

[0283] L, B: Longitude and latitude (unit: degrees);

[0284] : Surface coefficients at time t (which change over time, reflecting the temporal dynamics of spatial distribution).

[0285] (2) Model training (least squares method)

[0286] Horizontal settlement output by S4 Based on this, the surface coefficients are solved using the least squares method:

[0287] 1. Construct the observation equation: For all monitoring points i = 1, 2, ..., N (N is the total number of monitoring points), we have:

[0288] ;

[0289] in For the fitting residuals (expected value is 0), ( ) represents the coordinates of the i-th monitoring point.

[0290] 2. Matrix form: ,in:

[0291] ;

[0292] 3. Solving for coefficients: by minimizing the sum of squared residuals. ,have to:

[0293] (5-18)

[0294] For matrix transpose, ( It is the inverse matrix.

[0295] (3) Model application

[0296] The coordinates of the grid center point ( Substituting into equation (5-17), we obtain the surface fitting settlement for each grid:

[0297] (5-19)

[0298] This result reflects the spatially continuous distribution characteristics and serves as one of the input features for subsequent GBDT.

[0299] Optimization of Long Short-Term Memory Network Model (Introducing Surface Fitting Spatial Features)

[0300] The input features of the Long Short-Term Memory network are expanded into spatiotemporal fusion features that include surface fitting results, enhancing the time series' perception of spatial context:

[0301] 1. Input Feature Update: The input to the Long Short-Term Memory network at time t is:

[0302] (5-20)

[0303] Added surface fitting mesh settlement Dimensions: .

[0304] 2. Model parameter adjustment: The dimensions of the weight matrices for the forget gate, input gate, etc., were adjusted accordingly. The remaining formulas (5-1 to 5-7) remain unchanged, and the input features are updated as follows: This enables the Long Short-Term Memory Network to simultaneously learn the association between temporal trends and spatial context.

[0305] Gradient boosting decision tree model optimization (integrating surface fitting and long short-term memory network features)

[0306] Gradient boosting decision trees integrate the spatially continuous features of surface fitting with the temporal dynamic features of long short-term memory networks to enhance nonlinear fitting capabilities.

[0307] 1. Input feature update:

[0308] (5-21)

[0309] The following are newly added:

[0310] : Surface fitting settlement (from Equation 5-19);

[0311] Longitudinal subsidence gradient (reflecting the rate of spatial change);

[0312] : Latitudinal subsidence gradient;

[0313] Overall dimensions: .

[0314] 3. Model Training and Output: Following the gradient boosting decision tree iterative process of equations (5-9 to 5-12), the input features are updated as follows: Output preliminary fusion results (Including spatiotemporal and gradient features).

[0315] The three elements work together and integrate

[0316] By fusing the outputs of a long short-term memory network, gradient boosting decision tree, and surface fitting through a three-layer attention mechanism, the characteristics of time sensitivity, spatial nonlinearity, and continuous distribution are dynamically balanced.

[0317] 1. First-layer attention (intra-algorithm weights): Calculate the reliability weights of each algorithm's output.

[0318] Long Short-Term Memory Network Reliability Weights: ,in Variance of the prediction error of the Long Short-Term Memory network (calculated during training);

[0319] Gradient boosting of decision tree reliability weights: ,in To increase the variance of the prediction error of the decision tree by using gradients;

[0320] Surface fitting reliability weights: ,in The variance of the surface fitting residuals (derived from the residual calculation in Equation 5-18).

[0321] 2. Second-layer attention (feature importance weight): Calculate the contribution weight of each output to the fusion result.

[0322] ;

[0323] Attention score , The output of algorithm k, For the corresponding parameters.

[0324] 3. Final aggregate settlement:

[0325] (5-22)

[0326] Standardized results for Long Short-Term Memory (LSTM) network output

[0327] 4. Collaborative Training: The total loss function is expanded to a weighted MSE of the three outputs and the true value.

[0328] (5-23)

[0329] in (Based on precision weights), all model parameters are updated simultaneously through backpropagation.

[0330] Generate fusion result table

[0331] The fusion results table now includes information on the newly added grid settlement. and spatial gradient , This forms an integrated spatiotemporal-spatial data set, providing continuous grid data support for the spatial distribution analysis of S6.

[0332] Core function:

[0333] Surface fitting constructs a continuous spatial surface using a quadratic polynomial, which solves the problem of missing data in sparse areas of monitoring points (e.g., when the distance between two monitoring points is 300m, continuous data with an interval of 50m can be obtained through surface interpolation).

[0334] Interaction with Long Short-Term Memory Network: The surface fitting results are used as input to the Long Short-Term Memory Network, enabling the time series model to perceive the spatial background (such as the overall subsidence trend of a certain area).

[0335] Interaction with Gradient Boosting Decision Trees: Surface gradient features enhance the ability of gradient boosting decision trees to capture spatial rates of change (such as subsidence differences on both sides of a fault zone).

[0336] After the three methods are combined, the fusion accuracy is improved to ±1.2mm (20% improvement over the original method), and the spatial continuity is improved by 40%, providing a more reliable data foundation for the settlement zoning and influence range delineation of S6.

[0337] A high-temperature geothermal extraction ground subsidence data monitoring device, employing any one of the above-described high-temperature geothermal extraction ground subsidence data monitoring methods, the high-temperature geothermal extraction ground subsidence data monitoring device comprising:

[0338] The deployment module is used to determine the monitoring area and deploy monitoring points, and record the geographical coordinates of the monitoring points and their distance from the nearest mining well;

[0339] The debugging module is used to install and debug the corresponding monitoring devices at the monitoring points;

[0340] The data acquisition module is used to collect multi-source raw data at a set frequency through the debugged monitoring device. The raw data is transmitted to the data processing center with the monitoring point number and UTC timestamp attached.

[0341] The preprocessing module is used to preprocess the multi-source raw data to obtain GNSS vertical displacement data, InSAR vertical displacement data, leveling settlement and smoothed groundwater level depth data.

[0342] The fusion model construction module is used to build a collaborative fusion model of a long short-term memory network and a gradient boosting decision tree. Specifically, it includes: using GNSS vertical displacement data, InSAR vertical displacement data, leveling settlement, and groundwater depth data as basic input features; defining the geographical coordinates recorded in S1 and the distance to the nearest mining well as spatial features; and defining the year-day conversion of the UTC timestamp in S3 and the mining duration obtained from the mining company's operation records as time-derived features, thus forming complete input features; capturing the time dynamic features of the basic input features through the long short-term memory network and outputting the hidden state; integrating the hidden state with the spatial features and time-derived features and inputting it into the gradient boosting decision tree to obtain the preliminary fusion result; and then dynamically adjusting the correlation score between the hidden state and the preliminary fusion settlement result through an attention mechanism to output the final fusion settlement amount, including the periodic settlement amount and the cumulative settlement amount.

[0343] The analysis module is used to perform time series analysis, spatial distribution analysis, and anomaly detection based on the final fused settlement amount, and obtain the judgment results.

[0344] The embodiments described above are for illustrative purposes only and are not intended to limit the invention. Therefore, any changes in numerical values ​​or substitutions of equivalent elements should still fall within the scope of this invention.

[0345] The above detailed description will enable those skilled in the art to understand that the present invention can indeed achieve the aforementioned objectives and has complied with the provisions of the Patent Law.

[0346] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention. The above descriptions are merely preferred embodiments of the invention and are not intended to limit the invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the invention should be included within the scope of protection of the invention.

[0347] It should be noted that the above description of the process is for illustrative purposes only and does not limit the scope of this specification. Those skilled in the art can make various modifications and changes to the process under the guidance of this specification. However, these modifications and changes remain within the scope of this specification.

[0348] The basic concepts have been described above. Obviously, for those skilled in the art who have read this application, the above disclosure is merely illustrative and does not constitute a limitation of this application. Although not explicitly stated herein, those skilled in the art may make various modifications, improvements, and corrections to this application. Such modifications, improvements, and corrections are suggested in this application, and therefore, such modifications, improvements, and corrections still fall within the spirit and scope of the exemplary embodiments of this application.

[0349] Furthermore, this application uses specific terms to describe its embodiments. For example, "an embodiment," "one embodiment," and / or "some embodiments" refer to a particular feature, structure, or characteristic related to at least one embodiment of this application. Therefore, it should be emphasized and noted that "an embodiment," "one embodiment," or "an alternative embodiment" mentioned twice or more in different positions in this specification do not necessarily refer to the same embodiment. In addition, certain features, structures, or characteristics in one or more embodiments of this application can be appropriately combined.

[0350] Furthermore, those skilled in the art will understand that aspects of this application can be described and illustrated through several patentable types or situations, including any new and useful combination of processes, machines, products, or substances, or any new and useful improvements thereof. Therefore, aspects of this application can be implemented entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or a combination of hardware and software. All of the above hardware or software can be referred to as a “unit,” “module,” or “system.” Furthermore, aspects of this application can take the form of a computer program product embodied in one or more computer-readable media, wherein computer-readable program code is contained therein.

[0351] The computer program code required for the operation of each part of this application can be written in any one or more programming languages, including object-oriented programming languages ​​such as Java, Scala, Smalltalk, Eiffel, JADE, Emerald, C++, C, VB.NET, and Python; general programming languages ​​such as C; Visual Basic, Fortran2103, Perl, COBOL2102, PHP, and ABAP; dynamic programming languages ​​such as Python, Ruby, and Groovy; or other programming languages. This program code can run entirely on the user's computer, or as a standalone software package on the user's computer, or partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the latter case, the remote computer can be connected to the user's computer via any network, such as a local area network (LAN) or wide area network (WAN), or connected to an external computer (e.g., via the Internet), or in a cloud computing environment, or used as a service such as Software as a Service (SaaS).

[0352] Furthermore, unless expressly stated in the claims, the order of processing elements and sequences, the use of numbers and letters, or other names described in this application are not intended to limit the order of the processes and methods of this application. Although some currently considered useful embodiments of the invention have been discussed in the foregoing disclosure by way of various examples, it should be understood that such details are for illustrative purposes only, and the appended claims are not limited to the disclosed embodiments; rather, the claims are intended to cover all modifications and equivalent combinations that conform to the substance and scope of the embodiments of this application. For example, although the implementation of the various components described above can be embodied in a hardware device, it can also be implemented as a purely software solution, such as an installation on an existing server or mobile device.

Claims

1. A method for monitoring ground subsidence data in high-temperature geothermal extraction, characterized in that, Includes the following steps: S1: Determine the monitoring area and set up monitoring points, and record the geographical coordinates of the monitoring points and their distance from the nearest well. S2: Install and debug the corresponding monitoring devices at the monitoring points; S3: Collect multi-source raw data at a set frequency using the calibrated monitoring device. The raw data is transmitted to the data processing center with the monitoring point number and UTC timestamp appended. S4: Preprocess the multi-source raw data to obtain GNSS vertical displacement data, InSAR vertical displacement data, leveling settlement and smoothed groundwater depth data respectively. S5: Construct a collaborative fusion model of Long Short-Term Memory Network and Gradient Boosting Decision Tree. Specifically, this includes: using GNSS vertical displacement data, InSAR vertical displacement data, leveling settlement, and groundwater depth data as basic input features; defining the geographical coordinates recorded in S1 and the distance to the nearest mining well as spatial features; and defining the year-day conversion of the UTC timestamp in S3 and the mining duration obtained from the mining company's operation records as time-derived features, thus forming complete input features; capturing the time dynamic features of the basic input features through the Long Short-Term Memory Network and outputting the hidden state; integrating the hidden state with the spatial features and time-derived features and inputting it into the Gradient Boosting Decision Tree to obtain the preliminary fusion result; and then dynamically adjusting the correlation score between the hidden state and the preliminary fusion result through an attention mechanism to output the final fused settlement, including the periodic settlement and the cumulative settlement. S6: Based on the final fused settlement amount, perform time series analysis, spatial distribution analysis, and anomaly detection to obtain the judgment results; The construction of the S5 medium-long short-term memory network model includes: controlling the flow of information through forget gate, input gate, candidate memory unit and output gate, using the difference in level subsidence at adjacent time times as the training label, using mean squared error loss function and Adam optimizer for training, and outputting hidden states containing time dynamic features. The construction of the gradient boosting decision tree model in S5 includes: integrating the hidden state output by the long short-term memory network with the geographical coordinates, the distance to the nearest mining well, the number of days in a year, and the duration of mining into an input feature vector; generating multiple regression trees by iteratively fitting the residuals of the preceding model; and outputting the preliminary fusion result.

2. The method for monitoring ground subsidence data in high-temperature geothermal extraction according to claim 1, characterized in that, It also includes S7: When S6 determines that there is an abnormal settlement, it triggers a multi-level early warning and encrypts the collection of abnormal point data, while storing the original data, preprocessed data, fusion results, analysis results and mining parameters in the database.

3. The method for monitoring ground subsidence data in high-temperature geothermal extraction according to claim 1, characterized in that, The interaction process of the attention mechanism in S5 includes: calculating the attention score of the initial fusion result of the hidden state of the Long Short-Term Memory Network and the Gradient Boosting Decision Tree; obtaining the weight ratio of the Long Short-Term Memory Network through the softmax function; and weighted fusion of the two to output the final fused sedimentation amount. The weight ratio is tilted towards the Long Short-Term Memory Network during the sedimentation abrupt change period and towards the Gradient Boosting Decision Tree during the sedimentation stabilization period.

4. The method for monitoring ground subsidence data in high-temperature geothermal extraction according to claim 1, characterized in that, The joint training process of the collaborative fusion model in S5 includes: using the mean square error of the final fused settlement and level settlement as the total loss function, and simultaneously updating the weight matrix of the long short-term memory network, the tree structure of the gradient boosting decision tree, and the parameters of the attention mechanism through backpropagation until the total loss function converges.

5. The method for monitoring ground subsidence data in high-temperature geothermal extraction according to claim 1, characterized in that, The threshold for correlation anomaly detection in S6 is set as follows: when the groundwater level drops by more than 5m within one month and the monthly settlement rate of the corresponding monitoring point increases by more than 100%, it is determined to be a correlation anomaly.

6. The method for monitoring ground subsidence data in high-temperature geothermal extraction according to claim 2, characterized in that, The S7 multi-level early warning system includes: Level 1 warning corresponds to a cumulative settlement exceeding 200mm and a monthly settlement rate exceeding 15mm, triggering an audible and visual alarm and pushing information to multiple levels of management departments; Level 2 warning corresponds to a cumulative settlement of 100-200mm or a monthly settlement rate of 10-15mm, pushing information to mining enterprises and the Natural Resources Bureau; Level 3 warning is for those that do not reach the first two levels but meet the abnormal threshold, and is only displayed on the monitoring platform.

7. The method for monitoring ground subsidence data in high-temperature geothermal extraction according to claim 1, characterized in that, The data acquisition frequency in S3 is set as follows: GNSS data is acquired once every 30 minutes, InSAR data is acquired once every 7 days, leveling data is acquired once every 10 days, and groundwater level data is acquired once every 6 hours. All data timestamps are accurate to milliseconds and kept synchronized.

8. A device for monitoring ground subsidence data in high-temperature geothermal extraction, characterized in that, The high geothermal extraction ground subsidence data monitoring method according to any one of claims 1 to 7, wherein the high geothermal extraction ground subsidence data monitoring device comprises: The deployment module is used to determine the monitoring area and deploy monitoring points, and record the geographical coordinates of the monitoring points and their distance from the nearest mining well; The debugging module is used to install and debug the corresponding monitoring devices at the monitoring points; The data acquisition module is used to collect multi-source raw data at a set frequency through the debugged monitoring device. The raw data is transmitted to the data processing center with the monitoring point number and UTC timestamp attached. The preprocessing module is used to preprocess the multi-source raw data to obtain GNSS vertical displacement data, InSAR vertical displacement data, leveling settlement and smoothed groundwater level depth data. The fusion model construction module is used to build a collaborative fusion model of a long short-term memory network and a gradient boosting decision tree. Specifically, it includes: using GNSS vertical displacement data, InSAR vertical displacement data, leveling settlement, and groundwater depth data as basic input features; defining the geographical coordinates recorded in S1 and the distance to the nearest mining well as spatial features; and defining the year-day conversion of the UTC timestamp in S3 and the mining duration obtained from the mining company's operation records as time-derived features, thus forming complete input features; capturing the time dynamic features of the basic input features through the long short-term memory network and outputting the hidden state; integrating the hidden state with the spatial features and time-derived features and inputting it into the gradient boosting decision tree to obtain the preliminary fusion result; and then dynamically adjusting the correlation score between the hidden state and the preliminary fused settlement result through an attention mechanism to output the final fused settlement amount, including the periodic settlement amount and the cumulative settlement amount. The analysis module is used to perform time series analysis, spatial distribution analysis, and anomaly detection based on the final fused settlement amount, and obtain the judgment results.

Images