Urban land subsidence three-dimensional monitoring method, equipment and medium

Abstract

The invention relates to the technical field of settlement monitoring, and provides an urban ground settlement three-dimensional monitoring method and device and a medium, and the method comprises the steps: firstly obtaining remote sensing monitoring data and ground actual measurement data of a target area; then remote sensing monitoring data and ground actual measurement data are processed into a comprehensive feature data set; and finally, the comprehensive feature data set is processed according to the adaptive linkage network, and the land subsidence trend is predicted. Wherein the adaptive linkage network is a neural network provided with an adaptive physical constraint module; and the physical constraint in the adaptive physical constraint module is adaptively adjusted according to the features of the comprehensive feature data set and the scene of the target area. According to the method, the data set is processed through the self-adaptive linkage network carrying the self-adaptive physical constraint module of which the physical constraint is self-adaptively adjusted along with the comprehensive characteristics and the target area scene, so that the precision of land subsidence prediction in the complex urban environment is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of settlement monitoring technology, and in particular to a three-dimensional monitoring method, equipment and medium for urban ground settlement. Background Technology

[0002] Urban land subsidence is a slow-accumulating geological hazard driven by a combination of factors, including tectonic movements, groundwater over-extraction, engineering construction disturbances, and topography. It is characterized by its high degree of concealment, wide-ranging impact, and long duration, and has become one of the core geological problems restricting sustainable urban development. With the acceleration of urbanization, human activities such as the dense layout of high-rise buildings, large-scale development of underground space, and continuous increase in groundwater extraction have further exacerbated the complexity and uncertainty of land subsidence.

[0003] Existing settlement prediction models mostly use traditional machine learning or general deep learning architectures, which cannot adapt to the dynamic changes in different geological types, settlement stages, and construction activities in cities. Summary of the Invention

[0004] This invention provides a method, equipment, and medium for three-dimensional monitoring of urban ground subsidence, which addresses the problem that existing subsidence prediction models cannot adapt to dynamic scene changes in different geological types, subsidence stages, and construction activities in cities.

[0005] A first aspect of this invention provides a three-dimensional monitoring method for urban land subsidence, comprising: Acquire remote sensing monitoring data and ground measurement data for the target area; The remote sensing monitoring data and the ground measurement data are processed into a comprehensive feature dataset; The comprehensive feature dataset is processed using an adaptive linkage network to predict the trend of ground subsidence. Among them, the adaptive linkage network is a neural network with an adaptive physical constraint module; the physical constraints in the adaptive physical constraint module are adaptively adjusted according to the features of the comprehensive feature dataset and the scene of the target region.

[0006] In one possible implementation, the adaptive linkage network includes: a scene recognition module, a feature evaluation module, an adaptive physical constraint module, and a prediction module; the adaptive linkage network processes the comprehensive feature dataset to predict the ground subsidence trend, including: The scene recognition module analyzes the comprehensive feature dataset to obtain a multi-dimensional scene feature vector of the dynamic scene in the target area. The feature evaluation module quantifies and evaluates the multi-dimensional scene feature vectors to obtain evaluation results; the evaluation results are used to represent the physical rationality, spatiotemporal correlation and factor coupling of the features. Constraint information is determined based on the evaluation results, the comprehensive feature dataset, and the adaptive physical constraint module; Based on the prediction module and constraint information, scene-adaptive feature extraction and optimization training are performed on the comprehensive feature dataset to determine the ground subsidence trend.

[0007] In one possible implementation, the scene recognition module parses the comprehensive feature dataset to obtain a multi-dimensional scene feature vector of the dynamic scene in the target area, including: The comprehensive feature dataset is optimized based on dual knowledge graphs to obtain an optimized dataset; the dual knowledge graphs are the urban geology knowledge graph and the urban construction knowledge graph. Extract scene classification features and cross-domain interaction features from the optimized dataset and generate dynamic weights; By integrating multi-dimensional scene classification features, dynamic weights, and cross-domain interaction features, a multi-dimensional scene feature vector is generated.

[0008] In one possible implementation, the comprehensive feature dataset is optimized based on a dual knowledge graph to obtain an optimized dataset, including: Based on the entity linking engine of the dual knowledge graph, entity mapping is performed on the unstructured data in the comprehensive feature dataset. At the same time, the missing values ​​in the comprehensive feature dataset are inferred based on the relationship of the dual knowledge graph, resulting in an optimized dataset.

[0009] In one possible implementation, constraint information is determined based on the evaluation results, the comprehensive feature dataset, and the adaptive physical constraint module, including: Based on the multi-dimensional scene feature vector, match the combination of constraint types that are suitable for the current scene from the preset physical constraint library; Based on the evaluation results, the weights and penalty coefficients of each constraint in the constraint type combination are adjusted using a preset dynamic formula. By combining the spatiotemporal distribution characteristics of the comprehensive feature dataset, the threshold range and spatiotemporal effective domain of each constraint are dynamically defined; Generate structured constraint information that includes constraint type, dynamic weight, penalty coefficient, threshold range, spatiotemporal effective domain, and knowledge graph tracing.

[0010] In one possible implementation, based on a multi-dimensional scene feature vector, a combination of constraint types adapted to the current scene is matched from a pre-defined physical constraint library, including: Analyze the scene parameters in the three dimensions of geological type, subsidence-dominant factors, and construction stage in the multi-dimensional scene feature vector; Based on geological type parameters, basic physical constraints corresponding to the formation mechanical behavior are matched from a preset physical constraint library. Based on the parameters of the dominant settlement factors, matching the driving process constraints corresponding to the core process driving settlement is performed from the preset physical constraint library; Based on the construction stage parameters, match the engineering disturbance constraints corresponding to the engineering activity disturbances from the preset physical constraint library; Select at least one primary constraint and at least one secondary constraint from the basic physical constraints, driving process constraints, and engineering disturbance constraints to form a combination of constraint types that are suitable for the current scenario.

[0011] In one possible implementation, scene-adaptive feature extraction and optimization training are performed on the comprehensive feature dataset based on the prediction module and constraint information to determine the ground subsidence trend, including: Constraint information is embedded in the neural network of the prediction module to construct a loss function, which includes a data fitting term and a physical constraint penalty term. The prediction module is trained based on a comprehensive feature dataset. By minimizing the loss function, the network can fit the data while satisfying physical constraints, thereby determining the ground subsidence trend.

[0012] In one possible implementation, constraint information is embedded in the neural network of the prediction module to construct a loss function, including: Construct a data fitting term to measure the weighted deviation between the network's predicted values ​​and the measured values; A physical constraint penalty term is constructed, and the physical consistency of the network output within the spatiotemporal domain is quantitatively penalized based on the dynamic weights, penalty coefficients, and threshold ranges in the constraint information. Construct a knowledge graph consistency verification item to quantify the penalty for consistency between the prediction result and the preset rules in the dual knowledge graphs; A loss function is constructed by integrating data fitting terms, physical constraint penalty terms, and knowledge graph consistency verification terms.

[0013] A second aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the three-dimensional monitoring method for urban land subsidence as described in the first aspect above.

[0014] A third aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the three-dimensional monitoring method for urban ground subsidence as described in the first aspect above.

[0015] Compared to traditional technologies, this invention provides a method, device, and medium for three-dimensional monitoring of urban land subsidence. First, remote sensing monitoring data and ground-measured data of the target area are acquired. Then, the remote sensing monitoring data and ground-measured data are processed into a comprehensive feature dataset. Finally, the comprehensive feature dataset is processed using an adaptive linkage network to predict the land subsidence trend. The adaptive linkage network is a neural network equipped with an adaptive physical constraint module. The physical constraints in the adaptive physical constraint module are adaptively adjusted according to the characteristics of the comprehensive feature dataset and the scene of the target area. This invention significantly improves the accuracy of land subsidence prediction in complex urban environments by processing the dataset using an adaptive linkage network equipped with an adaptive physical constraint module that adaptively adjusts physical constraints according to the comprehensive features and the scene of the target area. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the implementation of the three-dimensional monitoring method for urban ground subsidence provided in this embodiment of the invention. Detailed Implementation

[0017] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0018] Figure 1 This is a flowchart illustrating the implementation of the three-dimensional monitoring method for urban ground subsidence provided in this embodiment of the invention. Figure 1 As shown, the method includes: S110, acquire remote sensing monitoring data and ground measurement data of the target area; S120 processes remote sensing monitoring data and ground-measured data into a comprehensive feature dataset; S130, based on the adaptive linkage network, processes the comprehensive feature dataset to predict the ground subsidence trend; Among them, the adaptive linkage network is a neural network with an adaptive physical constraint module; the physical constraints in the adaptive physical constraint module are adaptively adjusted according to the features of the comprehensive feature dataset and the scene of the target region.

[0019] In this embodiment of the invention, remote sensing monitoring data, with its advantages of large-scale and long-term observation, can comprehensively capture the spatial continuous distribution characteristics and temporal evolution trend of land subsidence in the target area, making up for the shortcomings of insufficient coverage in traditional single-point monitoring. Ground-based measured data, with high precision as its core, obtains accurate displacement information of key points in the target area through direct contact measurement, providing accuracy calibration and supplementing local details for the remote sensing data. Simultaneously, relevant auxiliary data such as geological, hydrological, and engineering activity data are integrated during the acquisition process to construct a multi-dimensional and multi-type raw data system, providing comprehensive and three-dimensional basic support for subsequent data processing and model training, ensuring the reliability of monitoring and prediction from the source.

[0020] In this embodiment of the invention, a systematic preprocessing procedure is used to perform noise suppression, atmospheric correction, and phase unwrapping on remote sensing monitoring data, and outlier removal, error correction, and standardization on ground-measured data, effectively improving the purity and accuracy of the original data. Then, spatiotemporal alignment technology is used to unify the spatial scale and temporal sampling frequency of data from different sources, eliminating data misalignment in the spatiotemporal dimension. Finally, feature engineering is used to extract core features such as spatial, temporal, attribute, and multi-factor coupling features. After redundancy removal and fusion processing, the scattered multi-source data is transformed into a comprehensive feature dataset that can accurately characterize the subsidence driving mechanism and evolution law, providing high-quality and highly recognizable input samples for the adaptive linkage network.

[0021] The preprocessing workflow addresses the differences in characteristics between remote sensing monitoring data and ground-measured data by employing a combination of source-specific processing and cross-source collaborative optimization. Each step precisely addresses specific technical issues through well-defined algorithms. In the remote sensing monitoring data preprocessing stage, atmospheric correction algorithms are used to remove atmospheric scattering and absorption effects to address radiation distortion caused by atmospheric interference; phase unwrapping algorithms are used to restore true deformation information to address phase ambiguity in interferometric SAR data; RPC geometric correction algorithms are used to achieve precise spatial positioning to address geometric deviations caused by terrain undulations and sensor attitude; and median filtering algorithms are used to preserve the spatial continuity of settlement signals to address random noise interference. In the ground-measured data preprocessing stage, the Grubbs criterion is used to identify and remove outliers exceeding three standard deviations to address gross errors; Kalman filtering algorithms are used to correct systematic errors such as sensor zero-point drift and smooth temporal fluctuations; and Z-score normalization algorithms are used to transform measured data of different dimensions into a unified scale to eliminate the impact of magnitude differences. In the cross-source data collaborative processing stage, a linear interpolation algorithm is used to unify the temporal sampling frequency of multi-source data, a thin plate spline interpolation algorithm is used to achieve spatial scale alignment, and finally, principal component analysis algorithm is used to remove feature redundancy, thereby improving the quality of the dataset and the efficiency of subsequent model training.

[0022] In this embodiment of the invention, an adaptive linkage network is used to achieve intelligent and accurate transformation from data to prediction results. This network breaks through the limitations of the fixed architecture of traditional models. Taking a comprehensive feature dataset as input, it deeply analyzes the dynamic scene characteristics and feature quality of the target area through a built-in scene recognition and feature evaluation mechanism. Combined with the dynamic adjustment capability of the adaptive physical constraint module, the model can specifically optimize feature extraction strategies and training logic when processing settlement data from different regions and stages. Through this series of collaborative technologies, the network can accurately capture the core evolutionary characteristics of settlement—"long-term slow accumulation + short-term abrupt change"—effectively uncovering the inherent laws driven by multiple coupled factors, and ultimately outputting ground settlement trend prediction results with both high accuracy and high reliability, achieving a profound leap from data fitting to mechanism alignment.

[0023] In this embodiment of the invention, the adaptive physical constraint module dynamically adjusts the type combination, weight coefficient, penalty intensity, and spatiotemporal effective domain of physical constraints by real-time sensing the feature quality of the comprehensive feature dataset (such as physical consistency and spatiotemporal correlation) and the scene characteristics of the target area (such as geological type, dominant subsidence factors, and engineering activity stage). This deeply adapts core physical mechanisms such as geomechanics, hydrodynamics, and engineering practice rules to specific scenarios. This adaptive adjustment mechanism ensures that the model training process does not deviate from objective physical laws, avoiding contradictions in prediction results, while also allowing the model to flexibly adapt to the complex and diverse subsidence scenarios in cities. This significantly improves the physical rationality, scene adaptability, and engineering practicality of the prediction results, providing reliable technical support for subsequent decision support.

[0024] In some embodiments, the adaptive linkage network includes: a scene recognition module, a feature evaluation module, an adaptive physical constraint module, and a prediction module; processing the comprehensive feature dataset according to the adaptive linkage network to predict the ground subsidence trend includes: parsing the comprehensive feature dataset according to the scene recognition module to obtain a multi-dimensional scene feature vector of the dynamic scene of the target area; quantitatively evaluating the multi-dimensional scene feature vector according to the feature evaluation module to obtain an evaluation result; wherein, the evaluation result is used to represent the physical rationality, spatiotemporal correlation, and factor coupling of the features; determining constraint information according to the evaluation result, the comprehensive feature dataset, and the adaptive physical constraint module; and performing scene adaptive feature extraction and optimization training on the comprehensive feature dataset according to the prediction module and the constraint information to determine the ground subsidence trend.

[0025] In this embodiment of the invention, the scene recognition module extracts key scene parameters from the dataset, such as geological type, dominant subsidence factors, engineering activity stage, hydrological dynamics, and topography. These parameters are then structured and encoded using prior domain knowledge, integrating scattered scene information into a high-dimensional vector that comprehensively characterizes the subsidence background of the target area. This vector not only covers static scene features (such as whether the geological type is soft soil or rock) but also dynamic evolution features (such as the transition from excavation to operation during construction, or the shift from a stable to a rapid drop in groundwater level). This provides accurate scene data for subsequent constraint adjustments and feature extraction, ensuring the network can adapt to the complex dynamic environment of urban subsidence.

[0026] The structured coding adopts a hybrid scheme combining rule-based coding and model-based coding, accurately adapting the coding logic to different types of parameters: For discrete static parameters such as geological type and construction stage, a "parameter category-coding mapping table" is constructed based on the domain expert rule base, and standardized conversion is achieved by using one-hot coding or ordered coding. For example, when the geological type is divided into 4 categories such as soft soil area and sandy soil layer, a 4-dimensional one-hot coding vector is used. The construction stage adopts 1-dimensional ordered coding according to the time priority. For continuous dynamic parameters such as groundwater level change rate and construction load intensity, the normalized numerical sequence is transformed into a 32-dimensional fixed-dimensional feature vector through a pre-trained 3-layer MLP embedding coding model. The model is pre-trained with more than 100,000 settlement monitoring samples to ensure that the coding fits the domain mechanism. The encoding dimensions follow the rule of "classification parameter dimension = total number of domain categories, continuous parameters fixed at 32 dimensions, cross-domain association parameters at 64 dimensions". The total dimension is the sum of the encoding dimensions of each single parameter. Under the default configuration, static parameters (4 dimensions of geological type + 3 dimensions of settlement dominant factors, etc.) are 12 dimensions, continuous static parameters (soil compressibility coefficient, etc.) are 64 dimensions, continuous dynamic parameters (groundwater level change, etc.) are 64 dimensions, and cross-domain association parameters are 128 dimensions, for a total of 268 dimensions. The dimensions can be dynamically adjusted between 192 and 384 dimensions according to the complexity of the scenario. Domain prior knowledge is incorporated through three paths: First, expert rule base weighting, with 500+ core rules labeled with influence weights, and element-wise weighting of the encoded feature vectors; second, pre-trained model initialization, where MLP weights are initialized through a dual knowledge graph relationship embedding algorithm, transforming the "parameter-settlement" association into an initial weight distribution; and third, real-time rule verification, where contradictory features are checked through the knowledge graph rule engine, triggering gradient backpropagation to fine-tune the encoding results.

[0027] The integration of static scene features and dynamic evolution features adopts a three-level scheme of "layered splicing + scene-aware attention weighting + cross-domain fusion normalization" to ensure a balance between feature integrity and adaptability. First, layered splicing is performed, dividing the features into "static foundation layer - dynamic evolution layer - cross-domain association layer". The static foundation layer contains the encoding results of discrete and continuous static parameters such as geological type and soil properties. The dynamic evolution layer is the encoding vector of continuous dynamic parameters such as groundwater level change and construction load intensity. The cross-domain association layer is the cross-dimensional interactive features such as "geology-construction" and "hydrology-settlement". The features are spliced ​​in the order of "static foundation layer → dynamic evolution layer → cross-domain association layer" to form an initial high-dimensional vector. Next, a scene-aware attention mechanism is introduced. Based on dual knowledge graphs of urban geology and construction, the correlation strength between static core feature identifiers and dynamic feature evolution trends is queried. For example, the correlation strength between "soft soil area" and "continuous decline in groundwater level" is higher than that between "rocky area" and "continuous decline in groundwater level." Based on this, attention weights are dynamically allocated, giving higher weights to features more closely related to the current scene, thus strengthening the influence of core features. Finally, cross-domain fusion and normalization are performed. Layer normalization eliminates the numerical scale differences between different feature layers, and a gating mechanism is introduced to dynamically adjust the integration ratio of cross-domain related features, avoiding excessive interference of cross-domain features with basic features. Ultimately, a multi-dimensional scene feature vector that is both comprehensive and targeted is formed.

[0028] Real-time extraction of dynamic evolution features relies on a technical system of high-frequency sampling, temporal modeling, and incremental updates to ensure accurate capture of dynamic changes in the scene. Data update frequencies are differentiated based on feature type: high-frequency features such as groundwater level and settlement rate are sampled and updated hourly, with data collected and transmitted in real-time by deployed sensors; medium- and low-frequency features such as construction phases and load intensity are updated daily, synchronized with construction logs and on-site inspection data; and slow-changing features such as geological environment background are updated weekly, adjusted based on trend analysis results from long-term monitoring data. For time-series data processing, a dynamic sliding window algorithm is used to extract time-series trends, with the window size adapted to the feature update frequency (e.g., a 24-hour window for hourly features and a 7-day window for daily features). An exponential smoothing algorithm smooths short-term fluctuations, highlighting long-term evolution patterns; and a cumulative sum control chart algorithm detects abrupt changes in the time series, automatically marking the time and magnitude of key changes such as construction phase transitions and sudden drops in groundwater level. Incremental feature updates rely on a distributed real-time computing framework to perform incremental processing on new data, eliminating the need to repeatedly calculate historical data. The feature extraction model parameters are fine-tuned through online learning algorithms such as incremental SVM. At the same time, a feature validity verification mechanism is established to compare the consistency between new features and dual knowledge graph rules in real time. If feature drift occurs, the feature reconstruction process is automatically triggered to ensure the timeliness and reliability of features.

[0029] The feature evaluation module assesses the multi-dimensional scene feature vectors output by the scene recognition module from three core dimensions: physical rationality, spatiotemporal correlation, and factor coupling. Physical rationality focuses on whether the features conform to basic laws such as geomechanics and hydrodynamics (e.g., whether the characteristics of groundwater level decline and settlement increase are synchronized). Spatiotemporal correlation focuses on whether the features can accurately capture the spatial distribution differences of settlement (e.g., whether the characteristics of settlement-sensitive areas are significant) and the temporal evolution patterns (e.g., whether long-term cumulative trends and short-term abrupt changes are clear). Factor coupling assesses the rationality of the interaction features of multiple driving factors (e.g., whether the coupling feature of "soft soil + groundwater over-extraction" conforms to the actual impact mechanism). The evaluation results are output in the form of quantitative scores, intuitively reflecting the effectiveness and reliability of the features and providing a core reference for the dynamic adjustment of constraint information.

[0030] Specifically, core physical rules (such as "decreased pore water pressure → increased effective stress → soil compression → settlement" and "increased excavation depth → non-linear settlement growth") are extracted from the knowledge graph to form a rule set. Key features in the multi-dimensional scene feature vector (such as "water level fluctuation-settlement rate correlation feature" and "construction load-settlement correlation feature") are matched with the rule set to calculate the matching degree (such as the percentage of consistency between feature trends and rule trends). The physical rationality score (0-1 points) is quantitatively output: 1 point for a complete match, 0.6-0.9 points for a partial match (such as consistent trends but amplitude deviation within the allowable range), and 0 points for a complete conflict.

[0031] Specifically, spatial autocorrelation analysis (such as Moran's I index) is used to assess the spatial clustering of features. If features in high-settlement areas (such as "high mining intensity") also show significant clustering and match the "spatial distribution of settlement-sensitive areas" in the geological knowledge graph, the spatial correlation score is high (0.7-1 points); if the spatial distribution of features is scattered and has no correlation with actual settlement zones, the score is low (0-0.3 points). Temporal continuity of features is assessed through temporal difference analysis (such as sliding window trend consistency test). If the time series of features (such as "monthly water level change features") is synchronized with the temporal trend of settlement rate (such as "settlement acceleration corresponding to water level decline") and conforms to the "engineering stage-settlement response lag" rule in the construction knowledge graph, the temporal correlation score is high (0.7-1 points); if the temporal fluctuations are disordered and disconnected from the settlement trend, the score is low (0-0.3 points). The total spatiotemporal correlation score (0-1 points) is output by combining spatial and temporal scores.

[0032] Specifically, cross-domain association rules for "geology-hydrology-construction" (such as "soft soil (geology) + sudden drop in groundwater level (hydrology) + deep foundation pit (construction) → significantly enhanced settlement") are extracted from the dual knowledge graph to construct a coupling pattern library; the matching degree between multi-factor coupling features (such as the three-dimensional features of "geological type-water level variation-construction intensity") in the feature vector and the coupling pattern library is calculated, and the correlation strength between factors is quantified by mutual information entropy (the higher the mutual information value, the stronger the explanatory power of factor coupling on settlement); combining the matching degree and mutual information value, the factor coupling score (0-1 point) is output: strong matching and high mutual information score 0.8-1 points, weak matching or low mutual information score 0-0.5 points.

[0033] Based on geomechanical principles, engineering experience, and hydrodynamics, the fundamental constraints for urban settlement scenarios are categorized into three core types and sub-types, comprehensively covering the core physical mechanisms driving settlement. First, there are the fundamental constraints based on strata mechanics, focusing on the mechanical properties and deformation patterns of the strata themselves. These include Terzaghi one-dimensional consolidation constraints applicable to compressible strata such as soft soil and cohesive soil; rock mass elasticity constraints applicable to rock strata; layered mechanical weighted constraints applicable to mixed strata; and pore water pressure dissipation constraints describing the adjustment of effective soil stress caused by changes in groundwater level. Second, there are constraints related to the settlement-driving process, designed for the core driving factors that trigger settlement. These include dynamic groundwater constraints such as groundwater over-extraction constraints, sudden drop in water level constraints, and aquifer desiccation constraints; engineering load constraints such as building load constraints, construction load constraints, and operational load constraints; crustal deformation rate constraints applicable to tectonically active areas; and natural consolidation constraints describing the natural settlement process of soil. Thirdly, there are engineering disturbance adaptation constraints, which adapt to the dynamic disturbance characteristics of urban engineering activities. These include Peck settlement trough constraints that describe the spatial distribution of settlement around the excavation pit, advance speed-surface settlement constraints and ground loss rate constraints that control the real-time settlement of shield tunneling, underground engineering support constraints that characterize the relationship between the stiffness of the support structure and the settlement control effect, and settlement convergence rate constraints that ensure settlement stability during the operation phase (such as annual settlement ≤ 5 mm).

[0034] Ultimately, the three-dimensional scores correspond to physical rationality, spatiotemporal correlation, and factor coupling (all from 0 to 1 points), which intuitively reflect the quality of the features in each dimension. The comprehensive quality score is obtained through weighted calculation (e.g., physical rationality weight 0.4, spatiotemporal correlation weight 0.3, factor coupling weight 0.3) and is used to judge the effectiveness of the feature vector as a whole.

[0035] In this embodiment of the invention, the process of determining constraint information is the core link in achieving "mechanism and scenario adaptation" in the adaptive linkage network, requiring the synergistic effect of comprehensive evaluation results, comprehensive feature dataset, and adaptive physical constraint module. First, feature quality is judged based on the evaluation results: if the physical rationality of the features is insufficient, the guiding role of physical constraints needs to be strengthened; if the spatiotemporal correlation is weak, temporal or spatial constraints should be supplemented. Second, the core features of the target area extracted from the comprehensive feature dataset (such as geological parameters, engineering strength, and hydrological status) are combined to match the appropriate basic constraint type. Finally, through the adaptive physical constraint module, the above information is transformed into structured constraint information, including dynamically adjusted constraint type combinations (primary constraints + secondary constraints), weight coefficients, penalty intensity, and spatiotemporal effective domain, ensuring that the constraints not only conform to objective physical mechanisms but also accurately adapt to the current scenario and feature quality.

[0036] The feature quality assessment adopts a "three-dimensional quantitative scoring + comprehensive weighting" system, which evaluates the three core dimensions of physical rationality, spatiotemporal correlation, and factor coupling. Each dimension has a maximum score of 1.0. The comprehensive score is the weighted sum of physical rationality (weight 0.4), spatiotemporal correlation (weight 0.3), and factor coupling (weight 0.3). The scoring results are divided into four levels: excellent (≥0.8), good (0.6~0.79), qualified (0.4~0.59), and unqualified (<0.4). The physical rationality assessment compares features with core physical rules in the dual knowledge graph, calculates rule matching degree, and checks for contradictory features. A complete match with no contradictions scores 0.8-1.0 points, a partial match with a low proportion of contradictory features scores 0.6-0.79 points, and a low matching degree or a high proportion of contradictory features scores below 0.6 points. The spatiotemporal correlation assessment combines spatial autocorrelation analysis and temporal trend consistency checks. Spatial clustering that aligns with subsidence hotspots and temporal changes that are synchronized with subsidence trends scores 0.8-1.0 points, a moderate matching degree scores 0.6-0.79 points, and a scattered spatial distribution or disordered temporal fluctuations scores below 0.6 points. The factor coupling assessment calculates the matching degree between multi-factor coupling features and the coupling pattern library, and quantifies the factor correlation strength using mutual information entropy. A strong match with high correlation scores 0.8-1.0 points, a moderate match with moderate correlation scores 0.6-0.79 points, and a weak match or low correlation scores below 0.6 points. Constraint adjustments are based on three-dimensional scoring and a comprehensive score as critical conditions: When the physical rationality score is below 0.6, physical constraints are strengthened, their weight and penalty intensity are increased, and the constraint threshold range is narrowed; when the spatial autocorrelation index in spatiotemporal correlation is below 0.5, spatial topological constraints are added, spatial constraint weights are increased, and spatial grid accuracy is improved; when the temporal trend consistency test result is not significant, temporal smoothing constraints are added, and the time sliding window is extended; when the factor coupling score is below 0.6, cross-domain coupling constraints are added, different effective domains are divided according to coupling strength, and penalty coefficients are adjusted; when the comprehensive score is below 0.6, the entire constraint system is strengthened simultaneously, the data optimization process is restarted, outliers are removed, and missing data is supplemented to ensure that the reliability and accuracy of the adjusted prediction results meet the requirements.

[0037] In this embodiment of the invention, the core task of the prediction module is to achieve scene-adaptive feature extraction and model optimization training based on constraint information, ultimately outputting an accurate ground subsidence trend. During the feature extraction stage, the module adjusts its strategy according to the constraint information: spatially, it focuses on the subsidence-sensitive area defined by the constraints, strengthening the capture of core area features; temporally, it follows the temporal patterns within the constraints, highlighting long-term accumulation and short-term abrupt changes; and in terms of multi-factor coupling, it verifies the constraints, eliminating invalid coupling features that violate the mechanism and retaining effective correlation features. During the optimization training stage, the module embeds the constraint information into the loss function, using backpropagation to ensure the model strictly adheres to physical constraints while fitting the data, avoiding mechanistic deviations caused by purely data-driven approaches. Through this series of adaptive operations, the model can deeply explore the subsidence driving mechanism and accurately predict future subsidence rates, cumulative subsidence amounts, and temporal evolution trends.

[0038] In this embodiment of the invention, the four modules of the adaptive linkage network form a closed-loop linkage mechanism of "scene perception - feature evaluation - constraint adjustment - feature extraction and training," ensuring the intelligence and adaptability of the entire processing process. The scene recognition module provides the network with "environmental perception capability," the feature evaluation module ensures "controllable input quality," the adaptive physical constraint module achieves "dynamic adaptation of mechanism," and the prediction module completes "accurate conversion output." Each module performs its own function while working closely together. This closed-loop mechanism allows the network to adjust its processing strategy in real time according to the dynamic scene changes and feature quality fluctuations in the target area, effectively solving the limitations of the "fixed architecture" of traditional models. Ultimately, it achieves ground subsidence trend prediction with high accuracy, high physical rationality, and strong scene adaptability, providing reliable technical support for urban geological disaster prevention and control.

[0039] In some embodiments, the scene recognition module parses the comprehensive feature dataset to obtain a multi-dimensional scene feature vector of the dynamic scene in the target area, including: optimizing the comprehensive feature dataset based on dual knowledge graphs to obtain an optimized dataset; wherein the dual knowledge graphs are an urban geological knowledge graph and an urban construction knowledge graph; extracting scene classification features and cross-domain interaction features from the optimized dataset and generating dynamic weights; and fusing the multi-dimensional scene classification features, dynamic weights, and cross-domain interaction features to generate a multi-dimensional scene feature vector.

[0040] In this embodiment of the invention, data optimization based on dual knowledge graphs is the foundation for the scene recognition module to generate high-quality feature vectors. The core lies in using domain knowledge to address the heterogeneity, missing information, and anomalies of the original data. The urban geological knowledge graph covers geological entities (such as soil layers and aquifers), entity attributes (such as compressibility coefficient and permeability coefficient), and entity relationships (such as "the area within 500m of the fault zone is a settlement-sensitive zone"). The urban construction knowledge graph includes construction entities (such as deep foundation pits and shield tunneling projects), construction attributes (such as excavation depth and advance speed), and "construction-geology" interaction rules (such as "shield tunneling in soft soil areas - settlement amplification effect"). During the optimization process, unstructured data (such as geological borehole text records and construction logs) in the comprehensive feature dataset are first mapped to standard entities in the knowledge graph through an entity linking engine, achieving unified data naming and format. Then, missing data is completed based on relational reasoning in the knowledge graph (such as completing the foundation pit support type based on "soft soil + excavation depth 10m"). Finally, abnormal data is corrected through joint verification of dual knowledge graphs (such as determining whether a sudden drop in groundwater level is a valid engineering disturbance based on construction records). Ultimately, an optimized dataset with improved structure, completeness, and reliability is obtained.

[0041] In this embodiment of the invention, feature extraction and dynamic weight generation are the core steps in transforming optimized data into scene representations, achieving the technical goal of "classification features defining scenes, interaction features revealing correlations, and dynamic weights reflecting key points". When extracting scene classification features, a rule base and sample library built based on dual knowledge graphs are used to extract key dimensions such as geological risk, construction impact, hydrological response, intensity of human activities, and settlement stage from the optimized dataset. Classification labels are output through a multi-task classifier (e.g., geological risk is "extremely unstable" and construction impact is "strong impact"). Cross-domain interaction feature extraction is carried out by using relational path mining of knowledge graphs to capture cross-dimensional associations such as "geology-hydrology" and "construction-settlement" (e.g., "soft soil + groundwater over-extraction + deep foundation pit - settlement aggravation"), and strong confidence paths are encoded into quantitative features. Dynamic weight generation adopts the knowledge graph attention mechanism, which assigns weights according to the correlation between each feature and settlement (e.g., "deep foundation pit" has a correlation of 0.32 with settlement in the construction knowledge graph, and "soft soil layer" has a correlation of 0.28 in the geological knowledge graph). This allows the weights to be dynamically adjusted as the scene changes, avoiding the limitations of fixed weights.

[0042] In this embodiment of the invention, the fusion process of multi-dimensional features is key to achieving accurate scene representation. A scene feature vector with comprehensiveness, dynamism, and interpretability is generated by integrating multiple types of features. During fusion, the scene classification labels are first converted into numerical features through one-hot encoding, and then concatenated dimensionally with dynamic weight vectors, cross-domain interaction features, and temporal features (such as season and construction period percentage). To ensure the traceability of the vector, each feature element is accompanied by a dual knowledge graph traceability identifier, clearly indicating the feature's origin (e.g., a feature comes from "geological map node + construction map rule"). Simultaneously, the vector has a dynamic update mechanism. When new information is added to the dual knowledge graph (e.g., a new underground utility tunnel construction entity) or when there are sudden changes in the target area scene (e.g., a sudden increase in settlement rate or the commencement of a large-scale project), the classifier and weight calculation logic are fine-tuned through incremental learning. The vector is updated while preserving historical temporal correlation, ensuring that the vector always accurately matches the dynamic scene of the target area, providing a reliable scene basis for subsequent feature evaluation and constraint adjustment.

[0043] In some embodiments, data optimization of the comprehensive feature dataset based on dual knowledge graphs is performed to obtain an optimized dataset, including: performing entity mapping on unstructured data in the comprehensive feature dataset according to the entity linking engine of the dual knowledge graph, and simultaneously inferring missing values ​​in the comprehensive feature dataset based on the relationship of the dual knowledge graph to obtain the optimized dataset.

[0044] In this embodiment of the invention, the entity mapping of unstructured data based on a dual knowledge graph-based entity linking engine is a core technical means to solve the heterogeneity of comprehensive feature datasets. Unstructured data in comprehensive feature datasets includes textual descriptions of geological boreholes (e.g., "the upper part is gray silty soil, the lower part contains sand"), engineering records in construction logs (e.g., "the subway tunnel has advanced to K1+200, encountering a soft plastic soil layer"), etc. This type of data is often difficult to use directly due to non-standardized expressions and inconsistent terminology. The entity linking engine achieves normalized mapping of heterogeneous expressions by comparing unstructured text with standard entities in the urban geological knowledge graph (including entities and attributes such as "silty soil - attribute - high compressibility" and "soft plastic soil layer - classification - Quaternary alluvial layer") and the urban construction knowledge graph (including entities and relationships such as "subway tunnel - construction method - shield tunneling" and "advancement mileage - corresponding geological section"). For example, "gray silty soil" is mapped to the "silty clay layer" entity in the geological knowledge graph, and "tunnel advancement to K1+200" is associated with the "shield tunneling - mileage section" entity in the construction knowledge graph. Through this process, unstructured data is transformed into a structured form of "entity-attribute-relationship," unifying the data naming system and format standards, and laying a consistent foundation for subsequent feature extraction.

[0045] In some embodiments, using relational reasoning based on dual knowledge graphs to complete missing values ​​is key to improving the completeness of the comprehensive feature dataset. Missing values ​​in the comprehensive feature dataset may stem from incomplete monitoring coverage (e.g., no groundwater level monitoring points were set up in a certain area) or omissions in records (e.g., the construction log did not indicate the type of foundation pit support), directly affecting data usability. At this point, the relational rule base of the dual knowledge graphs becomes the basis for reasoning: For missing geological values, the spatial association rules of the urban geological knowledge graph (such as "within the same aquifer, the correlation of water level variation of adjacent monitoring points is >0.8") are used to infer and complete the missing values, combined with existing data (such as "based on the average water level variation of 2.3m of 3 neighboring monitoring points, the target point is completed as 2.2-2.4m"); for missing construction values, the attribute association rules of the urban construction knowledge graph (such as "excavation depth >10m + soft soil area → support type = pile + internal support") are used to infer and complete the missing support type based on known parameters (such as "excavation depth 12m + geology is soft soil"); for missing cross-domain values ​​(such as "a certain construction area did not record the settlement influence radius due to missing data"), the interaction rules of the dual knowledge graphs (such as "soft soil area + shield diameter 6m → settlement influence radius = 3 times the diameter") are used for joint reasoning. This knowledge-rule-based reasoning ensures the physical rationality and engineering adaptability of missing value completion, ultimately resulting in an optimized dataset that combines completeness and reliability.

[0046] In some embodiments, constraint information is determined based on the evaluation results, the comprehensive feature dataset, and the adaptive physical constraint module, including: matching a combination of constraint types adapted to the current scene from a preset physical constraint library based on the multi-dimensional scene feature vector; adjusting the weights and penalty coefficients of each constraint in the constraint type combination according to the evaluation results using a preset dynamic formula; dynamically defining the threshold range and spatiotemporal effective domain of each constraint by combining the spatiotemporal distribution characteristics of the comprehensive feature dataset; and generating structured constraint information containing constraint type, dynamic weight, penalty coefficient, threshold range, spatiotemporal effective domain, and knowledge graph tracing.

[0047] In this embodiment of the invention, analyzing the core scene parameters in the multi-dimensional scene feature vector is a prerequisite for achieving accurate constraint type matching. Its core objective is to extract key information that decisively influences physical constraints from high-dimensional scene features. The multi-dimensional scene feature vector includes multiple dimensions such as geology, hydrology, construction, and settlement status. Geological type, dominant settlement factors, and construction stage are the core parameters determining the settlement mechanical mechanism and its impact process. Geological type determines the basic mechanical properties of the strata; dominant settlement factors clarify the core driving force of settlement; and construction stage defines the method and intensity of engineering disturbance. The analysis process uses feature filtering and parameter extraction algorithms to separate the specific parameter values ​​of these three dimensions from the vector. For example, the geological type is "soft soil area," the dominant settlement factor is "groundwater over-extraction," and the construction stage is "excavation period," providing a clear and explicit target basis for subsequent multi-dimensional constraint matching.

[0048] The pre-defined physical constraint library stores fundamental mechanical constraints corresponding to different geological types. These constraints originate from the core theories of geomechanics and directly reflect the deformation patterns of strata under external forces or environmental influences. If the analyzed geological type is "soft soil," its mechanical properties are characterized by high compressibility and low bearing capacity. The corresponding fundamental physical constraints are matched with Terzaghi consolidation theory constraints, which can accurately describe the slow compression deformation process of soft soil under effective stress changes. If the geological type is "rocky area," the strata are mainly characterized by elastic deformation and extremely low compressibility. In this case, rock mass elastic mechanical constraints are matched, which characterize its deformation pattern through the linear relationship between stress and strain. If it is "mixed strata," layered mechanical constraints are matched, weighted according to the thickness proportion of different strata to ensure that the fundamental constraints are consistent with the actual mechanical behavior of the strata.

[0049] The dominant factor in settlement directly determines the core logic of settlement occurrence, and the pre-set physical constraint library stores specific constraints for various driving processes. If the dominant factor is "groundwater over-extraction," the core driving process is the increase in effective soil stress due to the drop in groundwater level. The corresponding driving process constraint is the "groundwater level drop - soil compression" coupling constraint, which controls settlement development by quantifying the nonlinear correlation between the two. If the dominant factor is "engineering load," the core driving force is the redistribution of stratum stress caused by building or construction loads. In this case, the "load intensity - settlement" power function constraint is matched, which conforms to the settlement evolution law under load. If the dominant factor is "natural consolidation," the "time - degree of consolidation" logarithmic constraint is matched, which conforms to the time effect characteristics of natural drainage consolidation of soil.

[0050] The system matches engineering disturbance constraints to construction stage parameters, aiming to adapt the constraints to the dynamic disturbance characteristics of engineering activities and accurately control settlement risks at specific stages of construction. The disturbance methods, intensity, and timeliness vary significantly across different construction stages. A pre-defined physical constraint library categorizes corresponding engineering disturbance constraints according to the entire construction lifecycle. For example, in the "excavation phase," the core disturbance is the lateral unloading and bottom heave caused by soil excavation. The corresponding engineering disturbance constraint is an empirical formula constraint of "excavation depth - settlement trench width" (such as the Peck formula constraint), accurately describing the spatial distribution of settlement around the excavation pit. In the "tunnel boring machine (TBM) advancement phase," the disturbance mainly comes from the cutting, thrust, and synchronous grouting of the TBM. Therefore, a constraint of "advancement speed - surface settlement rate" is matched to control real-time settlement during construction. In the "operation phase," the disturbance tends to stabilize, and the load is constant. Therefore, a constraint of "settlement convergence rate" is matched to ensure that settlement during the operation phase stabilizes and meets specifications.

[0051] The selection of primary constraints is based on the characteristic parameters that play a decisive role in the scenario, prioritizing the constraint type with the highest weight in terms of settlement impact. For example, in the scenario of "soft soil area + groundwater over-extraction + foundation pit excavation period," the consolidation characteristics of soft soil are the underlying mechanical foundation, and the corresponding Terzaghi consolidation constraint is the primary constraint. The secondary constraints are selected from the constraints corresponding to the other two parameters, namely the "groundwater level drop - soil compression" coupling constraint (driving process constraint) and the Peck settlement trough constraint (engineering disturbance constraint), respectively supplementing the impact of the core driving process and engineering disturbance. Each constraint combination contains at least one primary constraint and one secondary constraint, with the primary constraint weight accounting for no less than 60%, and the secondary constraint weight adjusted according to the scenario adaptability to ensure that the constraint combination highlights the core contradictions while not omitting key secondary impacts.

[0052] In some embodiments, based on a multi-dimensional scene feature vector, a combination of constraint types adapted to the current scene is matched from a preset physical constraint library. This includes: parsing scene parameters in the multi-dimensional scene feature vector from three dimensions: geological type, settlement-dominant factor, and construction stage; matching basic physical constraints corresponding to the geological mechanical behavior from the preset physical constraint library based on the geological type parameters; matching driving process constraints corresponding to the core process driving settlement from the preset physical constraint library based on the settlement-dominant factor parameters; matching engineering disturbance constraints corresponding to engineering activity disturbances from the preset physical constraint library based on the construction stage parameters; and selecting at least one primary constraint and at least one secondary constraint from the basic physical constraints, driving process constraints, and engineering disturbance constraints to form a combination of constraint types adapted to the current scene.

[0053] In this embodiment of the invention, the multi-dimensional scene feature vector encompasses complex information such as geology, hydrology, and construction. However, geological type, dominant settlement factor, and construction stage are the "three key elements" affecting the settlement mechanism. Geological type determines the mechanical background of the strata, dominant settlement factor clarifies the core driving force of settlement, and construction stage defines the dynamic characteristics of engineering disturbance. The analysis process extracts specific parameters of these three dimensions from the vector through a feature selection algorithm, such as "geological type = alluvial plain soft soil", "dominant settlement factor = excessive groundwater extraction", and "construction stage = shield tunneling period". This transforms the abstract scene features into concrete constraint matching criteria, laying the foundation for subsequent multi-dimensional constraint screening.

[0054] In this embodiment of the invention, a preset physical constraint library stores basic mechanical constraints corresponding to various geological types. These constraints originate from classical geomechanics theories: if the geological type is "soft soil" (high compressibility, low permeability), then a Terzaghi one-dimensional consolidation constraint is matched to accurately describe the pore water discharge and volume compression process of soft soil under effective stress changes; if it is "hard rock strata" (low compressibility, high elasticity), then a rock mass elastic deformation constraint is matched to characterize its stress-strain linear relationship through Hooke's law; if it is "sand and gravel layer" (high permeability, granular skeleton support), then a "pore water pressure dissipation-skeleton compression" coupled constraint is matched to take into account its hydraulic and mechanical properties. This matching ensures that the basic constraints are consistent with the "nature" of the strata, avoiding constraint settings that deviate from the geological essence.

[0055] In this embodiment of the invention, the dominant factor of settlement directly determines the evolution path of settlement. A pre-set physical constraint library designs specific constraints for different dominant factors: if the dominant factor is "groundwater over-extraction," the core driver is the increase in effective stress due to the drop in water level, and the corresponding driving process constraint is the nonlinear correlation constraint of "groundwater level fluctuation - soil compression" (such as the quantitative relationship based on Biot's consolidation theory); if it is "engineering load," the core driver is the transfer of additional stress, and a three-dimensional constraint of "load strength - influence depth - settlement" is matched, which conforms to the stress diffusion law of the Businesk solution; if it is "tectonic movement," a constraint of "crustal deformation rate - surface settlement response" is matched, reflecting the long-term influence of geological structure on settlement. Through this matching, the constraints can accurately pinpoint the "source driving force" of settlement.

[0056] In this embodiment of the invention, the disturbance modes and intensities differ significantly at different construction stages. The preset physical constraint library is divided into constraints according to the entire construction cycle: if the construction stage is the "deep foundation pit excavation period", the core of the disturbance is the unloading of the soil in the pit and the redistribution of stress around the pit. The corresponding engineering disturbance constraint is the Peck settlement trough constraint, which describes the spatial distribution of settlement around the foundation pit through the relationship of "excavation depth - settlement trough width coefficient - maximum settlement". If it is the "shield tunneling period", the disturbance comes from shield cutting, thrust fluctuation and insufficient grouting. Then, the "propulsion speed - torque - surface settlement rate" real-time constraint is matched to control the instantaneous settlement during the construction process. If it is the "engineering operation period", the disturbance tends to be stable. Then, the "settlement convergence rate" constraint is matched (such as annual settlement ≤ 5mm) to ensure that the settlement during the operation stage meets the safety standards.

[0057] In this embodiment of the invention, the primary constraint is selected based on the constraint type that contributes the most to the settlement of the current scenario. For example, in the scenario of "soft soil area + groundwater over-extraction + foundation pit excavation period", the consolidation characteristics of soft soil are the underlying mechanism of settlement. Therefore, the Terzaghi consolidation constraint is the primary constraint (weight ≥ 60%). The secondary constraints are selected from the driving process constraint ("groundwater level - compression" constraint) and the engineering disturbance constraint, which respectively supplement the influence of the core driving force and the construction disturbance (each accounting for 20%-30% of the weight). Each combination contains at least one primary constraint and one secondary constraint. The synergy between primary and secondary constraints ensures that the constraints can control the core contradictions of the scenario and cover secondary influencing factors, ultimately forming a combination of constraint types that is highly adapted to the current scenario, providing a framework support for the subsequent dynamic adjustment of constraint parameters.

[0058] For example, the algorithm logic of each unit in the adaptive physics constraint module is as follows: Based on a comprehensive feature dataset and multi-dimensional scene feature vectors, core scene parameters (geological type, dominant subsidence factor, and construction stage) are extracted using a "feature-scene" mapping algorithm. Specifically, a two-layer parsing logic is employed: the first layer transforms unstructured scene descriptions (e.g., "subsidence caused by groundwater over-extraction in soft soil areas") into numerical features through word vector embedding; the second layer classifies these numerical features using a random forest classifier, outputting structured scene parameters (e.g., geological type = soft soil, dominant factor = groundwater over-extraction, construction stage = non-construction period). The classifier training samples are derived from a "scene-parameter" associated case library within a dual knowledge graph.

[0059] The system matches constraint type combinations from a pre-defined physical constraint library (containing 32 basic constraint categories, each with a mechanism formula, parameter range, and applicable scenario label). A "multi-dimensional similarity weighted matching algorithm" is employed: semantic similarity (based on entity distance from a knowledge graph) is calculated between the applicable scenario label (e.g., "soft soil + groundwater over-extraction") of each constraint and the current scenario parameters. This is combined with the constraint's impact weight on settlement (constraint contribution statistics from historical data) to obtain a weighted matching score. The top 3 constraints with the highest scores are then selected for the candidate set.

[0060] The constraint weights and penalty coefficients are adjusted based on the feature evaluation results (physical rationality P, spatiotemporal correlation T, factor coupling C). The weight adjustment uses a dynamic formula: The main constraint weight W_main = 0.6 + 0.3 × (1 - P) (when P < 0.6, the constraint is strengthened by increasing the weight). The secondary constraint weight W_sub = 0.2 + 0.1 × (1 - T) (when T < 0.5, increase the weight of the time-related secondary constraint) The penalty coefficient adjustment adopts a piecewise function: if C < 0.5, the cross-factor constraint penalty coefficient λ = 2.0; if 0.5 ≤ C < 0.8, λ = 1.5; if C ≥ 0.8, λ = 1.0.

[0061] By combining the spatiotemporal distribution characteristics of the comprehensive feature dataset with the rules of dual knowledge graphs, the effective domain is accurately delineated through "spatial topology analysis + time window segmentation". Spatially, based on the "fault zone-sensitive area" topological relationship of the geological knowledge graph, a spatial mask is generated using buffer analysis (e.g., 500m on both sides of the fault is a strong effective area). Temporally, according to the "engineering stage-response period" rule of the construction knowledge graph (e.g., 3 months before and after the foundation pit excavation period is the effective window), a time series segmentation algorithm is used to delineate time slices, and finally outputs the effective domain identifiers of a 10m×10m grid + quarterly time slices.

[0062] By integrating the above results, JSON-formatted constraint information containing "Constraint ID - Primary and Secondary Type - Dynamic Weight - Penalty Coefficient - Spatiotemporal Effective Domain Boundary - Knowledge Graph Source ID" is generated to ensure that each parameter is traceable (e.g., the weight source is marked as "Feature Evaluation Result P=0.4 + Knowledge Graph Rule R-072").

[0063] In some embodiments, the method of determining the ground subsidence trend by performing scene-adaptive feature extraction and optimization training on a comprehensive feature dataset based on the prediction module and constraint information includes: embedding constraint information into the neural network of the prediction module to construct a loss function, wherein the loss function includes a data fitting term and a physical constraint penalty term; training the prediction module based on the comprehensive feature dataset, and by minimizing the loss function, enabling the network to satisfy physical constraints while fitting the data, thereby determining the ground subsidence trend.

[0064] In this embodiment of the invention, the "constraint type combination" in the constraint information determines the feature dimensions that the network needs to focus on. For example, if the constraint combination includes Terzaghi consolidation constraints (basic physical constraints) and Peck settlement trough constraints (engineering disturbance constraints), the network's feature extraction layer will automatically strengthen the weights of "soil compressibility features" and "construction distance features." The "spatiotemporal effective domain" is applied to the feature map through a binary mask, so that the network only strengthens feature extraction within the spatial grid and time window where the constraints are effective, avoiding interference from invalid areas. The "dynamic weights" adjust the importance of each feature channel through an attention mechanism, giving higher attention to features associated with the main constraints (such as the "void ratio feature" in soft soil areas). This embedding method makes the network no longer a general feature extractor, but a "scene-adaptive" learner that can dynamically adjust the focus points according to the scene. The fused weights are input into the scene-aware attention network, and the network dynamically adjusts the attention coefficients through scene feature vectors (such as "soft soil area + high mining"). For example, in a soft soil scene, the attention coefficient for the "void ratio" feature is increased by 20%; in a construction scene, the attention coefficient for the "load strength" feature is increased by 30%.

[0065] In this embodiment of the invention, the data fitting term employs a weighted mean square error design, assigning higher weights to high-precision data (such as measured GNSS values) and basic weights to large-scale data (such as InSAR imagery), quantifying the deviation between predicted and measured values ​​to ensure the model's ability to capture data patterns. The physical constraint penalty term, on the other hand, designs specific penalty functions for various constraint types within the constraint information. For example, for consolidation constraints, the penalty function quantifies the deviation between predicted settlement and Terzaghi theoretical calculations; for engineering disturbance constraints, the penalty function focuses on the degree of deviation between the predicted results and Peck's formula. These two terms are dynamically fused; when the feature evaluation results show insufficient physical rationality, the penalty term weight is increased (e.g., from 0.3 to 0.5), forcing the model to conform to physical rules.

[0066] In this embodiment of the invention, during the initial training phase, the model first learns the statistical regularities in the comprehensive feature dataset through data fitting terms, initially capturing the spatiotemporal distribution characteristics of settlement. As the iteration progresses, the physical constraint penalty term gradually takes effect. When the prediction result violates a certain type of constraint (such as the settlement rate in soft soil exceeding the allowable range of consolidation theory), the penalty term value increases, and the network parameters (such as convolution kernel weights and fully connected layer coefficients) are adjusted through backpropagation, forcing feature extraction to focus more on key features related to the constraints (such as groundwater level change rate and soil compression coefficient). In the middle of the training phase, a learning rate decay mechanism is introduced, which, together with the "threshold range" in the constraint information, dynamically adjusts the optimization step size to avoid exceeding the physical threshold due to overfitting the data. Finally, through multiple iterations, the loss function converges to the minimum value, at which point the model both fits the measured data to the maximum extent and strictly satisfies various physical constraints.

[0067] In this embodiment of the invention, during the feature extraction stage, the model automatically adjusts the feature weights based on the scenario parameters (such as geological type and construction stage) in the constraint information. For example, in the scenario of "soft soil area + shield tunneling," the model focuses on extracting the coupled features of "shield tunneling speed - ground loss rate - settlement." In the prediction output stage, the model generates multi-dimensional trend results, including the spatial distribution of cumulative settlement over the next 1-5 years, the quarterly settlement rate time series curve, and the evolution path of settlement hotspot areas. Simultaneously, the prediction results are further validated through "knowledge graph tracing" in the constraint information to ensure consistency with geomechanical rules and engineering practice experience (e.g., the predicted settlement in soft soil areas conforms to the theoretical relationship of "compression coefficient - load - time"). The final output ground settlement trend accurately reflects data patterns and deeply aligns with physical mechanisms, providing a reliable decision-making basis for urban settlement prevention and control.

[0068] In some embodiments, constraint information is embedded in the neural network of the prediction module to construct a loss function, including: constructing a data fitting term to measure the weighted deviation between the network's predicted value and the measured value; constructing a physical constraint penalty term to quantify the physical consistency of the network output within the spatiotemporal domain based on the dynamic weights, penalty coefficients, and threshold ranges in the constraint information; constructing a knowledge graph consistency verification term to quantify the consistency between the prediction result and the preset rules in the dual knowledge graphs; and fusing the data fitting term, the physical constraint penalty term, and the knowledge graph consistency verification term to construct the loss function.

[0069] In this embodiment of the invention, the construction of the data fitting term is the foundation for the loss function to fit the patterns of measured data. The core is to achieve differentiated adaptation to data sources of different precision through a weighted bias design. The data fitting term takes network predicted values ​​and measured values ​​(such as GNSS 3D displacement and InSAR temporal deformation) from the comprehensive feature dataset as input, and uses a weighted mean square error formula to quantify the bias: higher weights (e.g., 1.2) are assigned to high-precision data (e.g., GNSS millimeter-level observations), while basic weights (e.g., 1.0) are assigned to data with large coverage but low local precision (e.g., InSAR imagery). The weight values ​​are dynamically calibrated with reference to the "data type-confidence" mapping rules in the dual knowledge graph. This weighted design ensures that the model prioritizes fitting high-confidence data to guarantee core accuracy, while also considering the overall trend fitting of large-scale data, enabling the data fitting term to accurately reflect the degree of deviation between predicted values ​​and measured patterns.

[0070] In some embodiments, the construction of the physical constraint penalty term is crucial for forcing the network output to conform to physical mechanisms. Its core is to transform the dynamic parameters in the constraint information into a quantifiable penalty mechanism. This penalty term, for each type of constraint in the constraint type combination (e.g., consolidation constraints, engineering disturbance constraints), designs a sub-penalty function based on its dynamic weight, penalty coefficient, threshold range, and spatiotemporal effective domain: the dynamic weight determines the contribution ratio of the constraint to the overall penalty (e.g., primary constraint weight 0.6, secondary constraint weight 0.3); the penalty coefficient is adjusted according to the feature evaluation results (e.g., the coefficient increases to 1.8 when the physical rationality score is low), strengthening the penalty for violations of constraints; the threshold range sets an upper limit for allowable deviations (e.g., settlement rate threshold ±5mm / year), triggering penalty if the range is exceeded; the spatiotemporal effective domain limits the penalty's scope of action through a binary mask (e.g., effective only in the excavation area and during the construction period). The sub-penalty functions are then aggregated to form the physical constraint penalty term, quantifying the deviation of the network output from physical laws within the spatiotemporal effective domain, ensuring that the prediction does not violate core mechanisms such as geomechanics and engineering practice.

[0071] In some embodiments, the construction and multi-part fusion of knowledge graph consistency verification items are the core of improving the interpretability of the loss function. Comprehensive and reliable prediction results are achieved through knowledge rule verification and multi-objective balancing. The knowledge graph consistency verification item is based on preset rules in the urban geological knowledge graph and the urban construction knowledge graph (e.g., "annual settlement > 50mm in soft soil areas is abnormal" and "the settlement influence radius of shield tunneling is ≥ 3 times the diameter"). It calculates the deviation between the prediction result and the rules: core rules (e.g., "groundwater extraction is positively correlated with settlement") are assigned high weights (e.g., 0.3), while general rules are assigned low weights (e.g., 0.1). A penalty is triggered if the deviation exceeds the rule's allowable error range. The final loss function uses dynamic coefficient fusion of three items: a default data fitting item (0.5), a physical constraint penalty item (0.3), and a knowledge graph verification item (0.2). If feature evaluation shows a prominent deviation (e.g., poor physical consistency), the corresponding item's coefficient is increased (e.g., the physical constraint penalty item is increased to 0.4). This fusion mechanism enables the loss function to both fit the data patterns and follow physical mechanisms and knowledge rules, driving the network to output prediction results that are both accurate and reliable.

[0072] This invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the three-dimensional monitoring method for urban land subsidence as described in the above embodiment.

[0073] This invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the steps of the three-dimensional monitoring method for urban ground subsidence as described in the above embodiment.

[0074] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. Unless otherwise specified or in conflict with logic, the terminology and / or descriptions between different embodiments are consistent and can be referenced interchangeably. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0075] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A three-dimensional monitoring method for urban land subsidence, characterized in that, include: Acquire remote sensing monitoring data and ground measurement data for the target area; The remote sensing monitoring data and ground measurement data are processed into a comprehensive feature dataset; The comprehensive feature dataset is processed using an adaptive linkage network to predict the land subsidence trend; The adaptive linkage network is a neural network equipped with an adaptive physical constraint module; the physical constraints in the adaptive physical constraint module are adaptively adjusted according to the features of the comprehensive feature dataset and the scene of the target region.

2. The three-dimensional monitoring method for urban land subsidence according to claim 1, characterized in that, The adaptive linkage network includes: a scene recognition module, a feature evaluation module, an adaptive physical constraint module, and a prediction module; it processes the comprehensive feature dataset based on the adaptive linkage network to predict the ground subsidence trend, including: The scene recognition module parses the comprehensive feature dataset to obtain a multi-dimensional scene feature vector of the dynamic scene in the target area. The feature evaluation module performs a quantitative evaluation of the multi-dimensional scene feature vectors to obtain an evaluation result; wherein the evaluation result is used to represent the physical rationality, spatiotemporal correlation and factor coupling of the features; The constraint information is determined based on the evaluation results, the comprehensive feature dataset, and the adaptive physical constraint module. Based on the prediction module and the constraint information, scene-adaptive feature extraction and optimization training are performed on the comprehensive feature dataset to determine the ground subsidence trend.

3. The three-dimensional monitoring method for urban land subsidence according to claim 2, characterized in that, The scene recognition module parses the comprehensive feature dataset to obtain a multi-dimensional scene feature vector of the dynamic scene in the target area, including: The comprehensive feature dataset is optimized based on dual knowledge graphs to obtain an optimized dataset; wherein the dual knowledge graphs are an urban geology knowledge graph and an urban construction knowledge graph. Extract scene classification features and cross-domain interaction features from the optimized dataset and generate dynamic weights; By integrating the multi-dimensional scene classification features, dynamic weights, and cross-domain interaction features, the multi-dimensional scene feature vector is generated.

4. The three-dimensional monitoring method for urban land subsidence according to claim 3, characterized in that, The comprehensive feature dataset is optimized based on dual knowledge graphs to obtain an optimized dataset, including: Based on the entity linking engine of the dual knowledge graph, entity mapping is performed on the unstructured data in the comprehensive feature dataset. At the same time, the missing values ​​in the comprehensive feature dataset are inferred based on the relationship of the dual knowledge graph, resulting in an optimized dataset.

5. The three-dimensional monitoring method for urban land subsidence according to claim 4, characterized in that, Based on the evaluation results, the comprehensive feature dataset, and the adaptive physical constraint module, constraint information is determined, including: Based on the multi-dimensional scene feature vector, a combination of constraint types adapted to the current scene is matched from a preset physical constraint library; Based on the evaluation results, the weights and penalty coefficients of each constraint in the constraint type combination are adjusted using a preset dynamic formula. Based on the spatiotemporal distribution characteristics of the comprehensive feature dataset, the threshold range and spatiotemporal effective domain of each constraint are dynamically defined; Generate structured constraint information that includes constraint type, dynamic weight, penalty coefficient, threshold range, spatiotemporal effective domain, and knowledge graph tracing.

6. The three-dimensional monitoring method for urban land subsidence according to claim 5, characterized in that, Based on the multi-dimensional scene feature vector, a combination of constraint types adapted to the current scene is matched from a preset physical constraint library, including: The scene parameters in the three dimensions of geological type, subsidence dominant factor and construction stage in the multi-dimensional scene feature vector are analyzed. Based on the geological type parameters, basic physical constraints corresponding to the formation mechanical behavior are matched from the preset physical constraint library; Based on the settlement dominant factor parameters, matching driving process constraints corresponding to the core process driving settlement is performed from the preset physical constraint library; Based on the construction stage parameters, match the engineering disturbance constraints corresponding to the engineering activity disturbances from the preset physical constraint library; Select at least one primary constraint and at least one secondary constraint from the basic physical constraints, driving process constraints, and engineering disturbance constraints to form a combination of constraint types that are suitable for the current scenario.

7. The three-dimensional monitoring method for urban land subsidence according to claim 2, characterized in that, Based on the prediction module and the constraint information, scene-adaptive feature extraction and optimization training are performed on the comprehensive feature dataset to determine the ground subsidence trend, including: The constraint information is embedded in the neural network of the prediction module to construct a loss function, wherein the loss function includes a data fitting term and a physical constraint penalty term; The loss function is minimized, and the prediction module is trained based on the comprehensive feature dataset to determine the ground subsidence trend.

8. The three-dimensional monitoring method for urban land subsidence according to claim 7, characterized in that, The constraint information is embedded in the neural network of the prediction module to construct a loss function, including: Construct a data fitting term to measure the weighted deviation between the network's predicted values ​​and the measured values; A physical constraint penalty term is constructed, and the physical consistency of the network output within the spatiotemporal domain is quantitatively penalized based on the dynamic weights, penalty coefficients, and threshold ranges in the constraint information. Construct a knowledge graph consistency verification item to quantify the penalty for consistency between the prediction result and the preset rules in the dual knowledge graphs; The loss function is constructed by integrating the data fitting term, the physical constraint penalty term, and the knowledge graph consistency verification term.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the three-dimensional monitoring method for urban ground subsidence as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the three-dimensional monitoring method for urban land subsidence as described in any one of claims 1 to 8.

Images