Early warning method for surface collapse, medium, equipment and product

Abstract

The invention discloses a surface collapse early warning method, medium, equipment and product, and relates to the technical field of surface collapse early warning, and the method comprises the steps: carrying out the data space-time alignment of multi-source monitoring data of surface collapse, and obtaining an observation vector after the space-time alignment; constructing a surface collapse causal diagram according to the geological prior data, and performing causal diagram structure learning; establishing a dynamic credibility state variable based on each data source, and calculating a multi-source dynamic weight; extracting a threatened feature from the observation vector, and calculating a consistency score in a preset time window based on the threatened feature and the dynamic weight; based on the cause and effect graph reasoning result, the dynamic weight and the consistency score, constructing a residual instability time prediction model, and outputting residual instability time prediction of the target area; and performing graded early warning according to the residual instability time threshold value and the risk threshold value, and updating the credibility state variable, the threshold value and the prediction model parameter by taking the field check result as observation feedback. According to the invention, the false alarm rate of early warning can be effectively reduced.

Description

Technical Field

[0001] This invention relates to the field of ground subsidence early warning technology, and in particular to a ground subsidence early warning method, medium, equipment and product. Background Technology

[0002] Ground subsidence, a common geological hazard, poses a serious threat to the safety of urban infrastructure, traffic operations, and the lives and property of residents. Currently, ground subsidence monitoring relies heavily on single data sources or simple multi-source data overlays, which have many limitations.

[0003] Some monitoring methods rely solely on single deformation monitoring tools such as InSAR (Intrusive Synthetic Aperture Radar) or GNSS (Global Navigation Satellite System), making it difficult to comprehensively capture the complex precursory information before ground subsidence. This can easily lead to delayed warnings or false alarms due to incomplete data. Furthermore, some multi-source monitoring schemes fail to consider the varying reliability of different data sources under different circumstances, applying equal weight to all data. When a data source experiences anomalies due to environmental interference (such as severe weather affecting InSAR observations or instantaneous fluctuations in pipeline networks affecting flow data), the accuracy of the warning results can be severely impacted.

[0004] Meanwhile, existing early warning methods lack systematic modeling of the causes of ground subsidence, failing to clearly identify the inherent correlation between "rainfall—seepage—leakage—stratum state—subsidence / cavitation," resulting in unclear early warning logic and difficulty in accurately predicting the remaining instability time. Furthermore, most schemes lack effective feedback mechanisms, failing to dynamically optimize model parameters and thresholds based on on-site verification results. Long-term use can easily lead to a decline and degradation in early warning performance, making them unsuitable for complex and changing geological and environmental conditions. Summary of the Invention

[0005] The purpose of this invention is to address the limitations of existing ground subsidence monitoring methods that rely heavily on single data sources or simple multi-source data overlay. This invention proposes an early warning method for ground subsidence, comprising the following steps: S1. Obtain geological prior data and multi-source monitoring data of ground collapse, perform spatiotemporal alignment of multi-source monitoring data, and obtain the spatiotemporally aligned observation vector; S2. Construct a cause-and-effect diagram of ground subsidence based on prior geological data, and perform cause-and-effect diagram structure learning; S3. Based on the monitoring data of each data source, establish dynamic confidence state variables. Based on the prior of the causal graph and the recent observation residuals, use Bayesian update and Kalman filter to perform online estimation to obtain the dynamic weights of the multi-source data. S4. Extract precursor features from the spatiotemporally aligned observation vectors, and calculate the consistency score within a preset time window based on the precursor features and dynamic weights. S5. Based on the causal graph reasoning results, dynamic weights and consistency scores, construct a remaining instability time prediction model and output the remaining instability time prediction for the target area. S6. Based on the predicted remaining instability time, compare it with the pre-set remaining instability time threshold and risk threshold to obtain the warning level of the target area; and use the on-site verification results as observation feedback to update the dynamic credibility state variables, remaining instability time threshold, risk threshold, and parameters of the remaining instability time prediction model of each data source.

[0006] Furthermore, the multi-source monitoring data includes: InSAR deformation sequences, GNSS displacement, GPR radar echoes, ground acoustic / vibration sensor signals, urban pipeline flow / pressure, rainfall, and groundwater levels.

[0007] Furthermore, during the learning of causal graph structures, prior constraints are set for irreversible edges and prohibited edges.

[0008] Furthermore, precursor features include abrupt changes in deformation curvature, phase envelope drift, upward shift of acceleration spectrum peaks, micro-abrupt changes in pipe pressure / flow rate, and geosonic anomaly energy windows.

[0009] Furthermore, the multi-source dynamic weights are expressed by the formula:

[0010] in, This represents the dynamic weight of the i-th data source. and These represent the dynamic credibility state variables of the i-th and k-th data sources, respectively.

[0011] Furthermore, the consistency score is expressed by the formula:

[0012] Where S represents the consistency score. Indicates the preset time window The precursor features of the i-th data source within.

[0013] Furthermore, based on the causal graph reasoning results, dynamic weights, and consistency scores, a prediction model for the remaining instability time is constructed, specifically as follows: The monitoring data within the target area are inferred along the main control mechanism chain of the causal graph to form causal graph inference features; The monitoring data are weighted according to dynamic weights to obtain weighted monitoring features; The causal graph inference features, weighted monitoring features, and consistency scores are concatenated to form a feature vector; A model is built based on machine learning methods, and feature vectors are used to train the model to obtain a prediction model for the remaining instability time.

[0014] The present invention also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described early warning method for ground subsidence.

[0015] The present invention also proposes an electronic device, including a processor and a memory, wherein the processor and the memory are interconnected, the memory is used to store a computer program, the computer program including computer-readable instructions, and the processor is configured to invoke the computer-readable instructions to execute the above-described early warning method for ground subsidence.

[0016] The present invention also proposes a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the above-described early warning method for ground subsidence.

[0017] The beneficial effects of the technical solution provided by this invention are: This invention constructs a causal graph using prior geological data, estimates the credibility of each data source, calculates the dynamic weight of each data source, extracts precursor features from each data source, calculates anomaly consistency scores based on precursor features and dynamic weights, and establishes a prediction model for remaining instability time based on the inference results of the causal graph and the consistency scores. Early warning of ground subsidence is then provided based on the prediction results, and on-site verification results are used as observation feedback to update the dynamic credibility state variables, thresholds, and prediction model parameters. This invention fully utilizes various data sources, considering the differences in credibility of different data sources under different scenarios; anomalies in one data source will not affect the accuracy of the warning results. This invention employs a feedback optimization mechanism to dynamically optimize the credibility state variables, thresholds, and prediction model parameters, achieving adaptive convergence of thresholds and weights. This invention can significantly improve the early warning lead time while ensuring low false alarms and reducing verification and handling costs. Attached Figure Description

[0018] Figure 1 This is a flowchart of an early warning method for ground subsidence according to an embodiment of the present invention; Figure 2 A block diagram of an electronic device according to an exemplary embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0020] The flowchart of the early warning method for ground subsidence according to an embodiment of the present invention is as follows: Figure 1Specifically, it includes the following steps: S1. Obtain geological prior data and multi-source monitoring data of ground collapse, perform spatiotemporal alignment of the multi-source monitoring data, and obtain the spatiotemporally aligned observation vector.

[0021] Geological prior data for ground subsidence includes, but is not limited to, basic geological data for determining the distribution of easily collapsible rock strata and the background of geological structure, hydrogeological data for analyzing the dynamics of subsidence, karst / mining-out characteristic data for directly identifying the spatial location and scale of potential hazards, overburden / rock mass characteristic data for evaluating the material conditions and stability of subsidence, and historical subsidence records.

[0022] Multi-source monitoring data for ground subsidence includes, but is not limited to, InSAR deformation sequences, GNSS displacement, GPR (Ground Penetrating Radar) echoes, ground acoustic / vibration sensor signals, urban pipeline flow / pressure, rainfall, and groundwater levels. Time synchronization, coordinate coordination, and spatial benchmark unification of each data source are performed to form a spatiotemporally aligned observation vector.

[0023] S2. Construct a cause-and-effect diagram of ground subsidence based on prior geological data, and perform cause-and-effect diagram structure learning.

[0024] Factors contributing to ground subsidence are derived from prior geological data. A causal graph is constructed with "rainfall—seepage—leakage—stratum state—settlement / cavitation" as the main controlling mechanism chain. The causal graph structure is learned using either a scoring-based or constraint-based DAG learning method, with prior constraints set for irreversible edges (e.g., "rainfall → seepage") and prohibited edges (e.g., "settlement → rainfall") to reduce spurious causality. Specifically: (1) Based on geological survey data, underground pipeline data, construction information, and historical subsidence cases, the main causative factors of ground subsidence were identified and classified into external triggering layer, water-soil process layer, engineering disturbance layer, and observation response layer according to their mechanisms, among which: External triggering factors include rainfall intensity, cumulative rainfall, river water level, and groundwater extraction volume; The water-soil process layer includes infiltration rate, seepage field, groundwater level, water pressure and leakage in the pipe network, degree of soil softening, soil strength reduction factor, and volume fraction of voids, etc. The engineering disturbance layer includes the excavation depth of the foundation pit, the location and attitude of the tunnel boring machine, the traffic load level, and the intensity of construction vibration, etc. The observation response layer includes surface subsidence, building tilt, pipeline deformation, InSAR deformation rate, GNSS displacement, pore water pressure, and radar anomaly indicators.

[0025] (2) Based on the above-mentioned multiple nodes, and combined with the water-induced collapse mechanism commonly known in the field, at least one main control causal path covering “rainfall-seepage-leakage-stratum state-settlement / cavitation” is constructed, wherein “stratum state” is used to comprehensively characterize stratum strength, pore structure, softening degree and plastic zone range, etc., and the main control path is retained as a fixed structure in the causal diagram.

[0026] (3) In addition to the main control path, other possible causal relationships are automatically mined and optimized using scoring-type and / or constraint-type directed acyclic graph (DAG) structure learning algorithms. The following prior constraints are introduced during the structure learning process: Set "Rainfall → Infiltration / Seepage", "Seepage → Formation State", "Pipeline Leakage → Formation State", "Formation State → Settlement / Voids", and "External Load / Engineering Disturbance → Formation State" as irreversible edges, meaning that only the causal direction is allowed to point from the former to the latter. Edges that clearly violate physical causality, such as “subsidence / cavitation → rainfall”, “surface deformation → rainfall”, and “surface deformation → historical stratigraphic properties”, are set as prohibited edges. Regarding "pipeline leakage" "Surface subsidence" and "groundwater level" For mutually directional relationships such as "surface subsidence", only the direction of "leakage / groundwater level → subsidence" is allowed, and the reverse direction is prohibited.

[0027] (4) Under the guidance of the above prior constraints, the candidate causal structure is scored or constrained to obtain a ground collapse causal graph covering the water-induced collapse path, engineering disturbance path and observation response path. This not only preserves the main control chain of "rainfall-seepage-leakage-stratum state-settlement / cavitation", but also automatically supplements the branch paths and reduces the introduction of pseudo-causal edges.

[0028] In structural learning, the main control mechanism chain is written as a hard prior, while other edges and nodes are mined from the data through algorithms. The factors and results of ground collapse are treated as nodes in a causal graph, and these nodes can be categorized as follows: (1) Basic conditions for ground subsidence Geological conditions include the distribution of soluble rocks (limestone), fault / fracture zones, and the thickness and structure of loose overburden; hydrogeological conditions include the existence of karst aquifers and the velocity and direction of groundwater flow; historical legacy issues include the distribution of old mining subsidence areas and historical collapse points.

[0029] (2) Causes of ground subsidence Natural triggers include heavy rainfall events and fluctuations in river flood levels; human-induced triggers include groundwater extraction / recharge, tunnel excavation, foundation pit dewatering, vehicle vibration, and pipeline leakage.

[0030] (3) Mechanism of ground subsidence Karst development includes the formation and expansion of caves and soil caves; stress changes include stress redistribution in the roof of mined-out areas and a reduction in bearing capacity; erosion / loss includes the removal of fine-grained material from the overburden by groundwater; vacuum erosion includes a sudden drop in groundwater level leading to negative pressure in the cavity.

[0031] (4) Results Surface deformation (e.g., subsidence and cracking) and collapse (e.g., sudden sinkhole).

[0032] Cause-effect graph construction process: 1. Starting from basic conditions, construct the basic chain. Connect the above nodes with arrows to form a chain of “cause → mechanism → result”. For example, the following two chains: (1) the existence of soluble rock + groundwater with corrosive properties → leads to → karst development; (2) karst development + the existence of loose overburden → constitutes → hidden soil caves or open karst caves.

[0033] 2. Add dynamic causes and the mechanism of ground subsidence to the chain to construct a logical chain. List the following logical chains: (1) Heavy rainfall / human pumping → cause → water level fluctuation; (2) Water level fluctuation + the overburden is a "sticky on top and sandy underneath" structure → generate → groundwater erosion, carrying away sand particles → leading to → soil caves migrating upward and expanding; (3) Sudden drop in water level + existence of open karst caves → generate → negative pressure in the cavity → leading to → absorption and collapse of the overburden; (4) Soil caves expand to the critical size / overburden bearing capacity reaches the limit → trigger → sudden instability and collapse of the overburden → form → surface pit; (5) Mining activities → form goaf area → stress redistribution of roof strata → roof bending, fracture, collapse → collapse zone propagates upward to the surface → surface subsidence or settlement.

[0034] 3. Integrate all the above chains into a comprehensive cause-and-effect diagram. During integration, local paths may change. For example, after a collapse forms, a positive feedback loop may emerge: "Collapse → Becomes a concentrated rainwater infiltration channel → Exacerbates erosion." Identify the most likely or most fatal cause-and-effect chains. For example: Rainfall → Seepage → Leakage → Ground condition → Settlement / Voids.

[0035] S3. Based on the monitoring data of each data source, establish dynamic confidence state variables. Based on the prior of the causal graph and the recent observation residuals, use Bayesian update and Kalman filtering for online estimation to obtain the dynamic weights of multi-source data.

[0036] (1) For each monitoring data source, a dynamic reliability state variable is defined to characterize the monitoring reliability of the data source. The dynamic reliability state variable evolves over time according to the preset state transition model and is initialized in combination with geological prior information and historical monitoring error statistics.

[0037] (2) Based on the ground collapse causal diagram obtained in step S2, the predicted observation value at the current moment is calculated using the parent node observation value and / or state estimate value of each monitoring node. The difference between the actual observation value and the predicted observation value of each monitoring node is taken as the observation residual, and a recent observation residual sequence is formed within a preset time window. The mean, variance and / or root mean square value of the residual sequence are calculated to obtain the recent observation residual index.

[0038] A conditional model is established for each node in the causal graph (the parameters can be fitted by historical data after structure learning), and the linear Gaussian fitting equation is as follows:

[0039] in, Let Pa(j) represent the observation value of the j-th node at time t, and let Pa(j) represent the set of parent nodes of the j-th node. Let represent the causal influence coefficient from node j to node p. This represents the Gaussian noise term at time t. Let be the value of the parent node at time t. If the parent node is observable, then... = If the parent node is a hidden state (such as the "strata state" comprehensive index), then = (Current estimate from the causal reasoning / state estimation module).

[0040] The predicted observation (conditional expectation) is then:

[0041] in, Let represent the predicted value of the j-th node at time t. Indicates in Under the conditions Expectations Let represent the set of observations of all parent nodes of the j-th node at time t.

[0042] Predicted observations are used for online self-checking / reliability calibration of data sources: Mechanism constraints expressed by causal graphs (e.g., the chain of "rainfall → seepage → leakage → formation state → settlement / cavitation") are used to calculate the reasonable range / expected value of a certain monitoring quantity under the current operating conditions. If a data source deviates from the expectation for a long time (the residual is consistently large), it is judged to be noisy, drifting, distorted, or affected by external interference, thereby reducing its fusion weight.

[0043] (3) Using the dynamic credibility state variable as the state variable and the recent observation residual index as the observation, state equations and observation equations are constructed. Under the constraint of the credibility prior distribution given by the causal graph, Bayesian update and Kalman filtering are used for online recursive estimation to obtain the posterior estimate and uncertainty of the dynamic credibility state variable of each monitoring data source at the current time. Among them, the credibility prior distribution reflects the reliability of the causal inference results.

[0044] To avoid nonlinearity, practical filtering can be performed in log-probability space. The state variables are represented as:

[0045] in, This represents the state quantity of the i-th data source at time t. This represents the credibility status of the i-th data source at time t.

[0046] The state equation (random walk) is expressed as:

[0047] in, This represents the state quantity of the i-th data source at time t-1. This represents the process noise at time t. This indicates that the mean is 0 and the variance is . The normal distribution Let represent the state process covariance of the i-th data source.

[0048] Observations: Take the recent observation residual index.

[0049] Linear observation equation:

[0050] in, This represents the observations from the i-th data source at time t. This represents the observation matrix of the i-th data source, which can typically be taken as... =1, This represents the state variable of the i-th data source. This represents the observation noise of the i-th data source. This indicates that the mean is 0 and the variance is . Gaussian white noise distribution Let represent the observation noise covariance of the i-th data source.

[0051] Under the above linear Gaussian model, the Kalman filtering process is as follows: Prior prediction: based on Derived from the state equation ; Posterior update (Bayesian update): the likelihood given by the observation equation Updated .

[0052] Kalman filter recursive formula: predict:

[0053]

[0054] renew:

[0055]

[0056]

[0057] Complete according to the following formula. Mapping to credibility:

[0058] in, Represents conditional probability. This represents the set of all observations from time 1 to time t-1. Represents the set of all observations from time 1 to time t. Represents the state at time t ( The prior estimate of ), where F represents the state transition matrix. This represents the posterior state estimate at time t-1. This represents the covariance of the prediction error at times t-1 and t. Let Q represent the covariance estimate at time t-1, and let Q represent the state process noise covariance. Let H represent the Kalman gain at time t, H represent the observation matrix, and R represent the observation noise covariance. This represents the posterior estimate of the state at time t. This represents the covariance estimate at time t.

[0059] (4) Based on the posterior estimate of the dynamic reliability state variable of each monitoring data source, the dynamic weight of the multi-source monitoring data is obtained through normalization or nonlinear mapping. When the observation residual index of a certain data source is continuously large within the preset time window, its dynamic reliability automatically decreases and the fusion weight decreases. When its observation residual index remains near the expected noise level for a long time, the dynamic reliability gradually increases and the fusion weight increases accordingly.

[0060] The dynamic weights of multi-source data are expressed by the formula:

[0061] in, This represents the dynamic weight of the i-th data source. and These represent the dynamic credibility state variables of the i-th and k-th data sources, respectively.

[0062] S4. Extract precursor features from the spatiotemporally aligned observation vectors, and calculate the consistency score within a preset time window based on the precursor features and multi-source dynamic weights.

[0063] Precursor features are extracted from the observation vectors of various data sources after spatiotemporal alignment. For example, the abrupt change coefficients of curvature K and acceleration A are extracted from InSAR deformation sequences; the upward fluctuation characteristics of high-frequency displacement are identified from GNSS displacement data; the rise in bandwidth energy ratio and spectral peak drift characteristics of 1–3 kHz are extracted from ground acoustic signals; micro-abrupt changes (<3%) and minimum flow offset characteristics at night are captured from pipeline pressure / flow data; and the amplitude difference index between profiles at the same location is calculated from GPR radar echoes to form a set of multi-source precursor features of ground collapse.

[0064] The formula for calculating the consistency score is as follows:

[0065] Where S represents the consistency score. Indicates the preset time window The precursor features of the i-th data source within.

[0066] S5. Based on the causal graph reasoning results, dynamic weights and consistency scores, construct a remaining instability time prediction model and output the remaining instability time prediction for the target area.

[0067] The process of constructing the remaining instability time prediction model is as follows: (1) Based on the ground collapse causal diagram obtained in step S2, reasoning is performed on the causal diagram main control mechanism chain of each monitoring unit in the target area in terms of disaster path activation state, stratum state index, seepage state index, engineering disturbance index and collapse condition probability, and a causal diagram reasoning feature vector is formed. (2) Combining the dynamic weights of each monitoring data source obtained in step S3, the multi-source monitoring characteristics such as surface deformation, pipeline pressure and flow, pore water pressure, groundwater level, and geophysical anomalies are weighted and fused to obtain weighted monitoring characteristics that reflect the degree of comprehensive response. (3) Using the multi-source anomaly consistency score calculated in step S4, the synchronicity and consistency of the anomalies in each data source within the preset time window are characterized. The cause-effect graph inference features, weighted monitoring features and anomaly consistency scores are spliced ​​together in a predetermined order to form a feature vector. (4) During the offline training phase, based on historical ground collapse cases, physical model tests and / or numerical simulation results, the instability time of each monitoring unit is determined. The feature vectors of each time before instability are used as inputs, and the corresponding real remaining instability time is used as outputs. At least one of the following algorithms is used: gradient boosting tree, random forest regression, deep neural network or survival analysis model to train the remaining instability time prediction model. (5) In the online prediction stage, the feature vector at the current moment is input into the remaining instability time prediction model, and the predicted value of the remaining instability time of the target area under the current working condition is output.

[0068] In this way, the remaining instability time prediction model uses causal graph reasoning results, dynamic weights, and multi-source consistency information to characterize the evolution speed of the ground subsidence instability process. Compared with the threshold discrimination method based on a single monitoring quantity, it can more accurately predict the remaining time before instability occurs, thereby extending the early warning time and improving the reliability of the early warning.

[0069] S6. Based on the predicted remaining instability time, compare it with the pre-set remaining instability time threshold and risk threshold to obtain the warning level for the target area; and use the on-site verification results as observation feedback to update the dynamic reliability state variables, remaining instability time threshold, risk threshold, and parameters of the remaining instability time prediction model for each data source. This achieves adaptive updating and stable convergence of the dynamic reliability state variables, remaining instability time threshold, risk threshold, and parameters of the remaining instability time prediction model, ensuring that the false alarm rate and missed alarm rate remain within the preset target range after long-term system operation.

[0070] The minimum standard for determining the warning level is: Information alert level corresponds to a consistency score S ≥ 0.6 and a duration ≥ 6 hours. The following are the different warning levels: Blue warning level corresponds to a Remaining Unstable Life (RUL) value of RUL ∈ [120h, 240h) or a consistency score S ≥ 0.7 and a duration ≥ 12 hours; Yellow warning level corresponds to RUL ∈ [48h, 120h); Orange / Red warning level corresponds to a Remaining Unstable Life RUL < 48h or a consistency score S ≥ 0.85 and a duration ≥ 24 hours.

[0071] Response and coordination strategies: Blue alerts trigger speed limits and zoned flow tests; Yellow alerts trigger ground-penetrating radar rescanning and fixed-point ground acoustic arrays; Orange or red alerts trigger directional grouting, emergency drainage, and fencing.

[0072] The results of on-site verification (such as whether or not voids were found during ground-penetrating radar rescanning) serve as the basis for determining hits / false alarms and are used to update the credibility of each data source. If the false alarm rate of a certain data source is >30% for two consecutive weeks, its weight is automatically reduced and the anomaly trigger threshold of that data source is increased. The Cox proportional hazards model (a typical "survival analysis / remaining lifetime analysis" model) is incrementally retrained every 30 days, and drift detection is performed (full retraining is performed when PSI (Population Stability Index) >0.2). In a trial run of three consecutive months, compared with the baseline scheme of "no credibility calibration", the average early warning lead time of this method increased from 36 hours to 92 hours, the false alarm rate decreased from 14.7% to 6.1%, the verification cost of key road sections (monthly ground-penetrating radar mileage) decreased by about 28%, and the first-time pass rate increased to >85%.

[0073] The above values ​​are typical parameters and effect ranges used in the example. In actual deployment, the parameters can be adjusted according to the urban geology and pipeline network conditions.

[0074] To verify the effectiveness of the method of the present invention, a densely populated area with frequent ground subsidence was selected as the test area. The experiment was conducted according to the following scheme.

[0075] 1. Divide the road into several grid units of 50m×50m, and deploy or connect the following monitoring data: InSAR surface deformation monitoring; GNSS and precision leveling are used to monitor the displacement of monitoring points; Monitoring of pressure and flow in water supply and drainage networks; Rainfall and groundwater level monitoring; GPR mobile monitoring results and its cavity anomaly indication; Information on historical collapse events and emergency response.

[0076] The trial period lasted for 6 consecutive months, with the first 3 months used for model training and parameter tuning, and the last 3 months used for online trial operation and comparative evaluation.

[0077] 2. Comparison of Schemes Three early warning schemes are set up: Option A: An early warning method based on a single deformation threshold, which uses only the InSAR surface deformation rate and triggers an early warning when the deformation rate exceeds a static threshold; Option B: A fixed-weighted method based on multi-source monitoring data, using multi-source data such as InSAR, GNSS, pipeline pressure, flow, and rainfall, but the weights of each data source are fixed and do not change over time, and there is no causal graph or remaining instability time prediction module. Scheme C: The method of this invention employs causal graph modeling, dynamic credibility calibration, multi-source anomaly consistency scoring, and Cox remaining instability time prediction, and introduces on-site feedback closed loop and threshold / model parameter adaptive update mechanism.

[0078] 3. Evaluation Indicators The comparison indicators include: Early warning hit rate (the proportion of actual collapse / cavitation events that are warned in advance); Underreporting rate (the proportion of real events that were not reported under warning); False alarm rate (the percentage of warnings issued but no abnormalities found on-site); Average warning lead time (the average time from the issuance of an effective warning to on-site confirmation of instability / void). Cost of reviewing key road sections (expressed as monthly GPR mileage). The success rate of treatment (the percentage of cases where recurrence or continuous remedial treatment does not occur in the same location within a certain observation period after the first treatment).

[0079] 4. Experimental Results Table 1. Comparison of the effects of three early warning schemes (trial operation for 3 months)

[0080] Table 1 shows a comparison of the effects of the three early warning schemes. Compared with the multi-source baseline scheme B without confidence calibration, the method of this invention (scheme C), under the same trial operation conditions, increases the average early warning lead time from 36 hours to 92 hours, reduces the false alarm rate from 14.7% to 6.1%, reduces the monthly GPR mobile mileage on key road sections by approximately 28%, and increases the first-time pass rate from 71.2% to 86.4%. Compared with scheme A, which uses only a single deformation threshold, this invention shows significant improvements in hit rate, lead time, and handling efficiency.

[0081] Table 2 Examples of Reliability and Weight Changes of Monitoring Data Sources for Typical Road Sections

[0082] Table 2 shows examples of changes in the reliability and weight of various monitoring data sources for typical road sections. After dynamic reliability calibration, the weight of highly reliable data sources such as InSAR increased, while the weight of pipeline pressure data sources with a false alarm rate exceeding 30% for two consecutive weeks decreased significantly, and the abnormal trigger threshold was automatically raised. The method of this invention provides a clear trend of shortening remaining instability time earlier than expected before a collapse occurs, which helps to organize early response.

[0083] During the trial operation, the system incrementally retrained the Cox model every 30 days and performed a population stability index (PSI) test. When the PSI was greater than 0.2, a full retraining was triggered. During the three-month trial operation, the overall early warning performance of the system showed a gradual improvement and trend towards stability, indicating that the dynamic reliability and threshold / model parameter adaptive update mechanism of this invention can effectively suppress long-term performance degradation.

[0084] In one exemplary embodiment, a computer-readable storage medium is included, which stores a computer program that, when executed by a processor, implements the aforementioned early warning method for ground subsidence.

[0085] Please see Figure 2 In one exemplary embodiment, the device further includes an electronic device including at least one processor, at least one memory, and at least one communication bus.

[0086] The memory contains a computer program, which includes computer-readable instructions. The processor calls the computer-readable instructions stored in the memory through the communication bus to execute the aforementioned early warning method for ground subsidence.

[0087] In one exemplary embodiment, a computer program product is proposed, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described early warning method for ground subsidence.

[0088] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for early warning of ground subsidence, characterized in that, Includes the following steps: S1. Obtain geological prior data and multi-source monitoring data of ground collapse, perform spatiotemporal alignment of multi-source monitoring data, and obtain the spatiotemporally aligned observation vector; S2. Construct a cause-and-effect diagram of ground subsidence based on prior geological data, and perform cause-and-effect diagram structure learning; S3. Based on the monitoring data of each data source, establish dynamic confidence state variables. Based on the prior of the causal graph and the recent observation residuals, use Bayesian update and Kalman filter to perform online estimation to obtain the dynamic weights of the multi-source data. S4. Extract precursor features from the spatiotemporally aligned observation vectors, and calculate the consistency score within a preset time window based on the precursor features and dynamic weights. S5. Based on the causal graph reasoning results, dynamic weights and consistency scores, construct a remaining instability time prediction model and output the remaining instability time prediction for the target area. S6. Based on the predicted remaining instability time, compare it with the pre-set remaining instability time threshold and risk threshold to obtain the warning level of the target area; The results of on-site verification are used as observation feedback to update the dynamic credibility state variables, remaining instability time threshold, risk threshold, and parameters of the remaining instability time prediction model for each data source.

2. The method for early warning of ground subsidence according to claim 1, characterized in that, Multi-source monitoring data include: InSAR deformation sequences, GNSS displacement, GPR radar echoes, ground acoustic / vibration sensor signals, urban pipeline flow / pressure, rainfall, and groundwater levels.

3. The method for early warning of ground subsidence according to claim 1, characterized in that, When learning the cause-effect graph structure, set prior constraints for irreversible edges and prohibited edges.

4. The method for early warning of ground subsidence according to claim 1, characterized in that, Precursor features include abrupt changes in deformation curvature, phase envelope drift, upward shift of acceleration spectrum peaks, micro-abrupt changes in pipe pressure / flow rate, and geosonic anomaly energy windows.

5. The method for early warning of ground subsidence according to claim 1, characterized in that, The multi-source dynamic weight is expressed by the formula: in, This represents the dynamic weight of the i-th data source. and These represent the dynamic credibility state variables of the i-th and k-th data sources, respectively.

6. The method for early warning of ground subsidence according to claim 5, characterized in that, The consistency score is expressed by the formula: Where S represents the consistency score. Indicates the preset time window The precursor features of the i-th data source within.

7. The method for early warning of ground subsidence according to claim 1, characterized in that, Based on the causal graph inference results, dynamic weights, and consistency scores, a prediction model for the remaining instability time is constructed, specifically as follows: The monitoring data within the target area are inferred along the main control mechanism chain of the causal graph to form causal graph inference features; The monitoring data are weighted according to dynamic weights to obtain weighted monitoring features; The causal graph inference features, weighted monitoring features, and consistency scores are concatenated to form a feature vector; A model is built based on machine learning methods, and feature vectors are used to train the model to obtain a prediction model for the remaining instability time.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it implements the method as described in any one of claims 1-7.

9. An electronic device, characterized in that, The device includes a processor and a memory interconnected thereto, wherein the memory is used to store a computer program, the computer program including computer-readable instructions, and the processor is configured to invoke the computer-readable instructions to perform the method as described in any one of claims 1-7.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1-7.

Images