Foundation pit engineering multi-source information fusion and safety situation assessment and early warning method and system

Abstract

The application discloses a foundation pit engineering multi-source information fusion and safety situation assessment and early warning method and system, relates to the early warning field of foundation pit engineering, and comprises the following steps: acquiring real-time monitoring data, physical parameters and construction condition data of the foundation pit engineering; inputting the real-time monitoring data and the physical parameters into a pre-constructed physical mechanism model to calculate physical constraint prediction values; and inputting the real-time monitoring data and the construction condition data into a time sequence neural network model after feature coding. The application realizes the organic integration and cooperation of multiple technical features such as a dynamic attention fusion mechanism, construction condition coding, physical regularization deep learning, correlation rule mining diagnosis and meta-learning generalization strategy, and constructs a precise, interpretable and self-adaptive foundation pit engineering safety situation assessment and early warning system. The system realizes breakthroughs in the prior art in terms of early warning advancement, accuracy, environmental risk diagnosis capability and cross-scene adaptability.

Description

Technical Field

[0001] This invention belongs to the field of early warning technology for foundation pit engineering, specifically, it relates to a method and system for multi-source information fusion and safety situation assessment and early warning of foundation pit engineering. Background Technology

[0002] With the acceleration of urbanization, underground space development is becoming increasingly frequent, and deep foundation pit engineering is being used more and more widely in complex urban environments. The safety and stability of foundation pit engineering are directly related to the safety of surrounding buildings, underground pipelines, and construction workers. Therefore, real-time and accurate monitoring and early warning of foundation pit engineering is an important issue in the field of geotechnical engineering.

[0003] Current technologies for monitoring and early warning of foundation pit safety are mainly divided into two categories: numerical simulation methods based on physical mechanisms and data-driven monitoring and analysis methods.

[0004] Physical mechanism-based methods (such as finite element analysis and limit equilibrium methods) typically construct models based on soil mechanics theory and geological survey reports. However, these methods have significant limitations when dealing with complex urban environments. First, physical models are often based on idealized assumptions (such as homogeneous and isotropic soil), making it difficult to accurately describe the complex nonlinear coupling relationship between "support-soil-environment" in actual engineering. Second, the acquisition of physical parameters is uncertain, and the model is difficult to update in real time. This leads to delayed or distorted calculation results when facing sudden construction conditions (such as heavy rain or vibration from heavy machinery), which can easily result in missed reports.

[0005] Data-driven methods (such as neural networks and support vector machines) utilize historical monitoring data for training and prediction. While these methods excel in handling nonlinear time-series data, they also have significant drawbacks. On one hand, general-purpose time-series prediction algorithms (such as LSTM and Transformer) are often considered "black box" models, lacking physical interpretability and making it difficult to distinguish between "normal construction disturbances" and "precursors of structural instability." On the other hand, methods relying solely on data-driven approaches are susceptible to monitoring noise, and in the absence of physical constraints, the models may predict results that violate mechanical principles, leading to a high false alarm rate.

[0006] To address the aforementioned issues, some existing technologies have attempted to combine physical models with neural networks using "dual-drive" or "Physical Information Neural Network (PINN)" schemes. However, most existing combinations are rather rigid or conventional. For example, some techniques simply use the output of the physical model as an additional input feature of the neural network, or merely add a simple physical penalty term to the loss function.

[0007] This conventional approach to combining physical and data-driven models still faces challenges in practical applications: in complex and ever-changing construction environments, the reliability of both physical and data-driven models is dynamic. Existing technologies often employ static fusion strategies, failing to dynamically adjust their weights based on the deviation (residual) between real-time monitoring data and theoretical predictions. For example, when significant discrepancies between on-site geological conditions and the survey report cause the physical model to fail, existing systems still assign it a high weight, leading to false alarms. Conversely, when abnormal noise appears in the monitoring data, over-reliance on the data-driven model can easily trigger false alarms. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a method and system for multi-source information fusion and safety situation assessment and early warning of foundation pit engineering that can overcome the above problems or at least partially solve the above problems.

[0009] To solve the above-mentioned technical problems, the basic concept of the technical solution adopted by this invention is: a method for multi-source information fusion and safety situation assessment and early warning in foundation pit engineering, comprising the following steps: Obtain real-time monitoring data, physical parameters, and construction condition data for the foundation pit project; The real-time monitoring data and the physical parameters are input into a pre-built physical mechanism model to calculate the predicted value of physical constraints. After feature encoding of the real-time monitoring data and the construction condition data, the data is input into a time-series neural network model to calculate the data-driven prediction value. Based on the attention mechanism, the physical constraint prediction value and the data-driven prediction value are dynamically weighted according to the residual of the physical constraint prediction value and the real-time monitoring data, and then weighted and fused to generate a security situation assessment result. When the security situation assessment result exceeds the preset security threshold, a tiered warning is triggered.

[0010] Preferably, the specific process of dynamically allocating weights based on the attention mechanism includes: Calculate the residual sequence between the predicted physical constraint value and the real-time monitoring data in real time; When the absolute value of the residual in the residual sequence is greater than the first preset threshold and the duration exceeds the preset time window, the weight of the physical constraint prediction value is reduced and the weight of the data-driven prediction value is increased through an attention mechanism. When the absolute value of the residual is less than or equal to the first preset threshold, the weight of the physical constraint prediction value is maintained or increased through an attention mechanism to correct the data-driven prediction value.

[0011] Preferably, the construction condition data includes excavation depth, number of support channels, and number of dewatering wells opened; Before inputting the construction condition data into the temporal neural network model, the method further includes: vectorizing the discrete construction condition data, and concatenating the encoded condition vector as a condition variable into the hidden layer of the temporal neural network model to adjust the activation function of the neural network.

[0012] Preferably, during the training process of the temporal neural network model, a loss function containing physical constraint terms is used for optimization; the loss function consists of a temporal prediction error term and a physical regularization term constructed based on the soil mechanics equilibrium equation, so as to force the temporal neural network model to learn features that conform to physical laws.

[0013] Preferably, the safety situation assessment result is generated based on the "support-soil-environment" collaborative diagnosis rule base, which is constructed by mining the correlation rules between support structure deformation, surrounding building settlement and pipeline displacement from historical engineering cases using the Apriori algorithm.

[0014] Preferably, it also includes a cross-scene adaptive update step: When applied to new foundation pit projects, a pre-trained general dual-drive model is used as the base model, and a small amount of monitoring data from the initial stage of the new foundation pit project is collected as a fine-tuning dataset. A meta-learning strategy is used to fine-tune the parameters of the basic model, enabling the model to quickly adapt to the specific geological and environmental characteristics of the new foundation pit project.

[0015] Preferably, before inputting the time-series neural network model, the real-time monitoring data is further preprocessed: The real-time monitoring data is denoised using a wavelet transform algorithm to separate high-frequency noise signals from low-frequency trend signals. The denoised data is spatiotemporally aligned and dimensionality reduced to extract spatiotemporal feature vectors that reflect the deformation and evolution characteristics of the foundation pit.

[0016] Preferably, the physical mechanism model is constructed using the limit equilibrium method or the elastic foundation beam theory to provide the theoretical stress state and safety factor boundary of the foundation pit support structure under the current working conditions.

[0017] A multi-source information fusion and safety situation assessment and early warning system for foundation pit engineering also includes: The multi-source data acquisition module is used to acquire real-time monitoring data, physical parameters, and construction condition data of the foundation pit project; The dual-drive fusion assessment module is used to execute the multi-source information fusion and safety situation assessment and early warning method for foundation pit engineering, and output the safety situation assessment results. The graded early warning module is used to trigger corresponding alarm commands based on the security situation assessment results.

[0018] A multi-source information fusion and safety situation assessment and early warning system for foundation pit engineering includes a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, it performs the functions of the early warning method described above.

[0019] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art: 1. This multi-source information fusion and safety situation assessment and early warning method and system for foundation pit engineering breaks through the limitations of traditional static or simple weighted fusion. It introduces an attention mechanism based on residual sequences to monitor the deviation between the predicted values ​​of the physical mechanism model and the actual monitoring data in real time. When the physical model fails due to geological uncertainties or severe construction disturbances (residuals continuously exceed the threshold), the system automatically reduces the weight of the physical model and increases the weight of the data-driven model, thereby sensitively capturing abnormal trends. Conversely, when there is noise in the monitoring data or the data-driven model makes unreasonable predictions, the system automatically increases the weight of the physical model for correction. This "dynamic gating" fusion strategy enables the early warning logic to adaptively adjust according to the actual on-site conditions, effectively avoiding missed or false alarms caused by the failure of a single model.

[0020] 2. The multi-source information fusion and safety situation assessment and early warning method and system for foundation pit engineering involves vectorizing and encoding discrete or continuous construction condition data, such as excavation depth, number of supports, and number of dewatering wells, and then concatenating them as condition variables into the hidden layer of a temporal neural network (LSTM). This design enables the neural network to dynamically adjust its internal state transition rules according to different construction stages. In other words, the model "knows" whether it is currently in the support removal stage or the foundation slab pouring stage, thus making differentiated judgments on normal deformation fluctuations and abnormal instability precursors under different conditions. Compared with existing technologies that use only static background input for construction conditions, this scheme significantly improves the model's robustness in dynamically changing construction environments and reduces false alarms caused by construction disturbances.

[0021] 3. The multi-source information fusion and safety situation assessment and early warning method and system for foundation pit engineering introduces a physical regularization term based on soil mechanics equilibrium equations (such as the Mohr-Coulomb failure criterion) into the loss function of the time-series neural network model. This regularization term forces the neural network to avoid predicting results that violate basic mechanical principles (such as infinite plastic flow of soil or stress exceeding shear strength) while learning the patterns of monitoring data. Compared with a purely "black box" data-driven model, the prediction results of this scheme have better physical interpretability, allowing engineers to understand the mechanical basis of the model output, thereby enhancing their trust in the early warning system. At the same time, physical regularization also plays a role in structuring priors, suppressing overfitting and improving the model's generalization ability when training data is limited.

[0022] 4. The multi-source information fusion and safety situation assessment and early warning method and system for foundation pit engineering not only outputs a comprehensive safety situation index and early warning level, but also, through a pre-built collaborative diagnostic rule base (integrating expert experience rules and association rules mined from historical accident cases using the Apriori algorithm), quickly matches abnormal patterns when an early warning is triggered, outputting the root cause location of the risk (e.g., "sudden increase in axial force in the eastern section of the second support") and a quantitative environmental impact assessment of surrounding buildings and pipelines (e.g., "expected settlement of adjacent pipelines to increase by 3-5 mm"). This function breaks through the limitations of existing technologies that only focus on a single indicator of the foundation pit itself, realizing a leap from "individual monitoring" to "system risk assessment," providing engineers with comprehensive and quantitative decision-making basis, and is especially suitable for deep foundation pit projects in urban areas with sensitive surrounding environments and complex risk coupling.

[0023] 5. This invention relates to a multi-source information fusion and safety situation assessment and early warning method and system for foundation pit engineering. Addressing the pain point of existing data-driven early warning models that require "one-project-one-parameter tuning" and are difficult to promote across projects, this invention employs a model-independent meta-learning (MAML) strategy. A general dual-drive basic model is pre-trained on historical data from multiple different types of foundation pits, giving its initial parameters the characteristic of "easy and rapid adaptation to new tasks." When applied to new foundation pit projects, only a small amount of monitoring data from the initial stage of the project (such as the first two weeks) needs to be collected. The basic model can then be quickly fine-tuned through a few gradient updates, achieving high-precision adaptation to the specific geological and environmental characteristics of the new site. Compared to training from scratch or simple transfer learning, this meta-learning strategy significantly reduces the time cost and data requirements for deploying the model in new projects, providing key technical support for the industrialization and promotion of early warning models.

[0024] 6. The multi-source information fusion and safety situation assessment and early warning method and system for this foundation pit project employs wavelet transform to denoise the multi-source heterogeneous monitoring data during the data input stage, separating the low-frequency trend signal reflecting the actual structural deformation from the high-frequency environmental noise. Principal component analysis (PCA) is used to reduce the dimensionality of high-dimensional monitoring features, extracting deep state feature vectors with a cumulative contribution rate exceeding 95%. These preprocessing operations effectively solve problems such as inconsistent sensor sampling frequencies, high data noise, and dimensionality explosion, enabling subsequent physical models and data-driven models to learn from high-quality, low-redundancy features, thus improving the information utilization efficiency and computational efficiency of the entire early warning system.

[0025] 7. The multi-source information fusion and safety situation assessment early warning method and system for this foundation pit project, due to its high accuracy and low false alarm rate, can effectively avoid unnecessary work stoppages, emergency rescues, and resource waste caused by false alarms. Simultaneously, through advanced and precise early warning (more than 24 hours in advance) and risk root cause diagnosis, timely and targeted measures such as grouting reinforcement, adjusting the excavation sequence, and increasing support can be taken to prevent major safety accidents such as foundation pit collapse and excessive settlement of surrounding buildings. This significantly reduces direct costs such as compensation and effectively shortens the overall project duration.

[0026] The specific embodiments of the present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0027] In the attached diagram: Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments will be clearly and completely described below with reference to the accompanying drawings. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0029] This embodiment provides an early warning method for multi-source information fusion and safety situation assessment in foundation pit engineering under complex urban environments. The core of this method lies in constructing an intelligent analysis framework driven by both mechanism and data, and introducing an attention mechanism to achieve dynamic adaptive fusion of the two driving modes. The specific process of this method includes the following steps: Step S1, Data Acquisition and Preprocessing: It can collect or receive monitoring data from various sensors at the foundation pit site in real time, physical parameters from the survey report, and working condition data from the construction records.

[0030] Step S2, Calculation of physical constraint prediction values: Input some real-time monitoring data (such as current support structure deformation and earth pressure) and physical parameters into a pre-built physical mechanism model (such as a numerical model based on elastic foundation beam theory) to calculate the predicted physical constraint values ​​(such as theoretical deformation values ​​of the support structure and safety factor) under the current working conditions.

[0031] Step S3: Data-driven prediction calculation: After feature encoding of real-time monitoring data and construction condition data, the data is input into a time-series neural network model (such as an LSTM network) trained with physical regularization to calculate data-driven prediction values ​​(such as prediction of the deformation trend of the support structure in the next 24 hours).

[0032] Step S4: Adaptive weighted fusion based on attention mechanism: The two predicted values ​​obtained in steps S2 and S3, along with real-time monitoring data, are input into an attention mechanism module. This module dynamically allocates the weights of the two predicted values ​​in the final fusion by calculating the residual between the physical predicted value and the measured value in real time, generating a highly reliable security situation assessment result.

[0033] Step S5, Tiered Early Warning and Collaborative Diagnosis: The comprehensive safety situation assessment result generated in step S4 is compared with the preset multi-level safety thresholds. When the threshold is exceeded, the corresponding level of warning is triggered. At the same time, a pre-built "support-soil-environment" collaborative diagnostic rule library is invoked to quickly locate the root cause of the risk (e.g., which support has abnormal axial force, which section of soil has excessive displacement) and quantitatively assess the impact on the surrounding environment (such as nearby buildings and pipelines) based on the current anomaly pattern.

[0034] Step S6, Cross-scene adaptive update: When this method is applied to a new foundation pit project, an adaptive update step of the model is performed. A general dual-drive basic model pre-trained based on multiple historical project data is loaded. Then, a small amount of monitoring data from the early stage of the new project (such as the first two weeks) is collected as a fine-tuning dataset. A meta-learning strategy is used to quickly fine-tune the basic model to adapt it to the specific geological and environmental characteristics of the new project.

[0035] The core steps described above will be explained in detail below: 1. Detailed implementation of step S1: Acquisition and preprocessing of multi-source heterogeneous data This step forms the basis of the early warning method. In actual engineering deployments, data sources mainly include three categories: Real-time monitoring data: acquired through an automated monitoring system, including but not limited to: horizontal displacement of the top of the support structure, deep horizontal displacement (inclinometer), axial force of the support, anchor cable stress, surface settlement, groundwater level, pore water pressure, tilt and settlement of adjacent buildings, etc. The data sampling frequency can range from 5 minutes to 1 hour.

[0036] Physical parameters: derived from engineering geological survey reports and design documents, including: soil unit weight, cohesion, internal friction angle, compression modulus, elastic modulus and moment of inertia of support structures (such as bored piles and diaphragm walls), and stiffness of supports.

[0037] Construction status data: extracted from construction logs or BIM 4D models, including: current excavation depth (meters), number of installed supports, support numbers being dismantled, number and flow rate of dewatering wells, and surface overload (such as surcharges and heavy machinery loads).

[0038] Preprocessing steps: Spatiotemporal alignment: Since the sampling frequency and clock of different sensors may differ, this embodiment uses a timestamp-based sliding window method (window size for example 1 hour, sliding step size 5 minutes) to perform spatiotemporal synchronization of all data, ensuring that all features are aligned at the same point in time.

[0039] Noise reduction: Wavelet transform algorithm is used. Specifically, the 'db4' wavelet basis function is selected, and each monitoring data sequence is decomposed into 5 levels to separate the high-frequency noise components (usually corresponding to random environmental vibrations) from the low-frequency trend components that reflect the true deformation of the structure, and then the signal is reconstructed.

[0040] Missing value imputation: For missing data caused by sensor failure or transmission interruption, cubic spline interpolation is used to fill in the missing data.

[0041] Feature dimensionality reduction and extraction: For high-dimensional monitoring data (such as deep displacement data from multiple inclinometer boreholes), principal component analysis (PCA) is used to extract the top few principal components with a cumulative contribution rate of over 95%, which serve as deep feature vectors characterizing the overall deformation state of the system.

[0042] 2. Detailed implementation of step S2: Calculation of physical constraint prediction values In this embodiment, the physical mechanism model is constructed using the "elastic foundation beam method" (such as the method recommended in the "Technical Specification for Foundation Pit Support" JGJ120-2012). This model is used to calculate the theoretical horizontal displacement and internal forces of the support structure under the current excavation depth and support conditions.

[0043] Model input: Extract the current excavation depth, number of support channels, and physical parameters (m value of soil layer, structural stiffness, etc.) from the data after preprocessing in step S1.

[0044] Calculation process: Based on the theory of elastic foundation beams, the retaining piles / walls are considered as vertical elastic foundation beams, the active lateral earth pressure is taken as the load, the passive lateral soil is simulated by a series of springs (horizontal resistance coefficients), and the supports are simplified as elastic supports. By solving the differential equations, the theoretical horizontal deformation values ​​at each point of the support structure are obtained. and the theoretical value of the supporting axial force .

[0045] Output: The calculated theoretical deformation value is used as the predicted physical constraint value. Essentially, this represents the "most reasonable" structural response under the current idealized model. This value serves as a key input for subsequent attention mechanisms.

[0046] 3. Detailed implementation of step S3: Calculation of data-driven predicted values The data-driven model employs a three-layer Long Short-Term Memory (LSTM) network with targeted physical regularization design.

[0047] Network structure: The input layer dimension is the feature dimension after dimensionality reduction in step S1 (e.g., 10); the hidden layer contains 64 LSTM units; the output layer is a fully connected layer that outputs predicted values ​​(e.g., the horizontal displacement value of the top of the support structure in the next 24 hours).

[0048] Construction condition feature encoding: Discrete construction condition data (e.g., support number 1, 2, 3) are encoded using one-hot encoding, while continuous condition data (e.g., excavation depth) are normalized using Min-Max. Then, these encoded condition vectors are "concatenated" as condition variables to the state vector of the LSTM hidden layer at each time step. ,in It is the hidden state of LSTM. It is a condition-based coding system. This enables the network to dynamically adjust its internal state transition rules according to different construction stages.

[0049] Physical regularization in the loss function: To force LSTM to learn features that conform to the laws of mechanics, its loss function L consists of two parts:

[0050] in, It is the standard time series prediction error (such as mean square error, MSE), which calculates the difference between the predicted value and the actual value. This is a physical regularization term, which in this embodiment is constructed based on the "Mohr-Coulomb failure criterion" in soil mechanics. Specifically, whenever the LSTM predicts a soil displacement or stress state, the system calculates in parallel the ratio of shear stress to normal stress corresponding to that state. If this ratio is close to or exceeds the soil shear strength (defined by physical parameters), a large penalty term is applied. This forces the network to avoid predicting results that violate mechanical common sense (such as infinite plastic flow of soil). The hyperparameter λ is set to 0.1 through cross-validation.

[0051] Output: The LSTM model outputs a sequence of predicted values ​​for multiple future time steps, and the first value is taken as the data-driven prediction value for the current time step. .

[0052] 4. Detailed implementation of step S4: Dynamic weighted fusion based on attention mechanism This step is key to distinguishing it from existing static fusion methods, achieving adaptive collaboration driven by both mechanism and data.

[0053] Attention mechanism input: The attention module receives three inputs: physical constraint prediction value. Data-driven forecasts and the actual monitoring value at the current moment. .

[0054] Residual calculation: Real-time calculation of the residual sequence R(t) between the physical model predictions and the measured values:

[0055] The residual reflects the degree of matching between the current physical model and the actual engineering state.

[0056] Dynamic weight allocation: The core of the attention mechanism is a gating function, typically implemented as a single-hidden-layer neural network. Its input is the historical trend of the residual sequence R within a window [tT, t] over a past period.

[0057] Scenario 1 (Physical Model Failure): When the absolute value of multiple consecutive points in the residual sequence R(t) (e.g., lasting for more than 6 time steps, i.e., 30 minutes) exceeds a first preset threshold (e.g., 15%), the attention mechanism determines that the physical model may have failed due to unclear geological conditions or severe construction disturbances. In this case, the model outputs a low weight (e.g., 0.2) to... Output high weights (e.g., 0.8) to This allows the early warning logic to automatically shift from "mechanism-driven" to "data-driven," sensitively capturing abnormal trends in the data.

[0058] Scenario 2 (Unreliable Data-Driven Model): When the residual sequence R(t) is small (e.g., less than 5%), but and When significant deviations occur without physical evidence (e.g., a sudden drop in predicted deformation), the attention mechanism determines that the monitoring data may contain noise or the model may be overfitting. In this case, the model is... Assign high weights (e.g., 0.9) to Assign low weights (e.g., 0.1) and use physical models to correct unreasonable predictions.

[0059] Scenario 3 (Good Coordination): In general, the model assigns balanced weights to the two (e.g., 0.5 and 0.5).

[0060] Fusion Output: Final Security Situation Assessment Results for:

[0061] Here, the attention weight α is a dynamically calculated continuous value between 0 and 1. This fused [weight / weight]... Compared to a single It is smoother, more reliable, and incorporates information on future trends.

[0062] 5. Detailed implementation of step S5: hierarchical early warning and collaborative diagnosis Tiered early warning: Preset three levels of early warning thresholds, for example: Yellow Alert (Attention Level): It exceeds the design warning value by 80%.

[0063] Orange alert (alert level): The value exceeds the design warning value, and the fused forecast trend shows that it will exceed the control value within the next 24 hours.

[0064] Red Alert (Danger Level): The design control value is rapidly approached or reached, and the α weight indicates that data-driven factors dominate (indicating the presence of rapidly unknown risks).

[0065] Collaborative Diagnostic Rule Base: This base is implemented using a "rule engine" and contains two types of rules: Expert experience rules: For example, pre-written "IF [Support axial force increases sharply (increase of more than 10% in half an hour)] AND [corresponding deep horizontal displacement rate > 2 mm / day] THEN Risk is identified as plastic deformation of the support system in this area".

[0066] Data mining rules: Automatically mined from accident case data of multiple historical projects using the Apriori algorithm. For example, the algorithm may find that the support and confidence of the rule "{accelerated settlement of adjacent pipelines, increased width of surface cracks} → {overall instability of the foundation pit}" both exceed the preset thresholds (e.g., support 0.1, confidence 0.8).

[0067] Diagnosis and Assessment: When an early warning is triggered, the system inputs the abnormal patterns of all currently monitored indicators (such as excessive axial force of the support, excessive displacement rate, etc.) as a "fact set" into the rule engine. The engine quickly matches the rule base, outputs the most likely root cause of the risk (such as "sudden increase in axial force of the eastern section of the second support"), the scope of impact (such as "affecting water supply pipelines within a 10m radius"), and quantifies and predicts the specific settlement increment within the scope of impact based on a simplified numerical model or empirical formula (such as "expected pipeline settlement increase of 3-5mm").

[0068] 6. Detailed implementation of step S6: Cross-scene adaptive update To achieve the model's generalization ability, this embodiment adopts a strategy that combines transfer learning and meta-learning.

[0069] Pre-trained basic model: On a cloud server, using complete historical foundation pit engineering data from 5-8 different geological conditions (such as soft soil, gravel, weathered rock) and different support forms (pile bracing, wall anchor, internal bracing), a general dual-drive early warning basic model is trained according to steps S2-S4.

[0070] Scenario fine-tuning: When the model is deployed to a new foundation pit project, the system automatically performs fine-tuning: a. Collect small sample: Collect a small amount of monitoring data and operating condition data for the first 14 days after the start of the new project.

[0071] b. Parameter freezing and unfreezing: Freeze 2 / 3 of the parameters of the bottom layer of the LSTM network in the pre-trained model (these parameters learn general temporal patterns), and unfreeze and retrain only the top 1 / 3 of the parameters and the parameters of the attention mechanism.

[0072] c. Meta-learning optimization: A Model-independent Meta-learning (MAML) strategy is employed. MAML optimizes the initial parameters of the model during the pre-training phase, enabling it to adapt to new tasks with only a few gradient descent steps. On new tasks, we only need to perform 1-5 gradient updates on a small number of samples to obtain a highly accurate adapted model.

[0073] This invention organically integrates and coordinates multiple technical features, including dynamic attention fusion mechanism, construction condition encoding, physical regularization deep learning, association rule mining and diagnosis, and meta-learning generalization strategy, to construct a precise, interpretable, and adaptive safety situation assessment and early warning system for foundation pit engineering. This system achieves breakthroughs over existing technologies in terms of early warning anticipation, accuracy, environmental risk diagnosis capabilities, and cross-scenario adaptability.

[0074] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-described technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for multi-source information fusion and safety situation assessment and early warning in foundation pit engineering, characterized in that, Includes the following steps: Obtain real-time monitoring data, physical parameters, and construction condition data for the foundation pit project; The real-time monitoring data and the physical parameters are input into a pre-built physical mechanism model to calculate the predicted value of physical constraints. After feature encoding of the real-time monitoring data and the construction condition data, the data is input into a time-series neural network model to calculate the data-driven prediction value. Based on the attention mechanism, the physical constraint prediction value and the data-driven prediction value are dynamically weighted according to the residual of the physical constraint prediction value and the real-time monitoring data, and then weighted and fused to generate a security situation assessment result. When the security situation assessment result exceeds the preset security threshold, a tiered warning is triggered.

2. The method for multi-source information fusion and safety situation assessment and early warning of foundation pit engineering according to claim 1, characterized in that, The specific process of dynamically allocating weights based on the attention mechanism includes: Calculate the residual sequence between the predicted physical constraint value and the real-time monitoring data in real time; When the absolute value of the residual in the residual sequence is greater than a first preset threshold and the duration exceeds a preset time window, the weight of the physical constraint prediction value is reduced and the weight of the data-driven prediction value is increased through an attention mechanism. When the absolute value of the residual is less than or equal to the first preset threshold, the weight of the physical constraint prediction value is maintained or increased through an attention mechanism to correct the data-driven prediction value.

3. The method for multi-source information fusion and safety situation assessment and early warning of foundation pit engineering according to claim 1, characterized in that, The construction condition data includes excavation depth, number of support channels, and number of dewatering wells opened. Before inputting the construction condition data into the time-series neural network model, the method further includes: vectorizing the discrete construction condition data, and concatenating the encoded condition vector as a condition variable into the hidden layer of the time-series neural network model to adjust the activation function of the neural network.

4. The method for multi-source information fusion and safety situation assessment and early warning of foundation pit engineering according to claim 1, characterized in that, During the training process of the time-series neural network model, a loss function containing physical constraints is used for optimization. The loss function consists of a time-series prediction error term and a physical regularization term constructed based on the soil mechanics equilibrium equation, so as to force the time-series neural network model to learn features that conform to physical laws.

5. The method for multi-source information fusion and safety situation assessment and early warning of foundation pit engineering according to claim 1, characterized in that, The safety situation assessment results are generated based on the "support-soil-environment" collaborative diagnosis rule base, which is constructed by mining the correlation rules between support structure deformation, surrounding building settlement and pipeline displacement from historical engineering cases using the Apriori algorithm.

6. The method for multi-source information fusion and safety situation assessment and early warning of foundation pit engineering according to claim 1, characterized in that, It also includes cross-scene adaptive update steps: When applied to new foundation pit projects, a pre-trained general dual-drive model is used as the base model, and a small amount of monitoring data from the initial stage of the new foundation pit project is collected as a fine-tuning dataset. A meta-learning strategy is used to fine-tune the parameters of the basic model, enabling the model to quickly adapt to the specific geological and environmental characteristics of the new foundation pit project.

7. The method for multi-source information fusion and safety situation assessment and early warning of foundation pit engineering according to claim 1, characterized in that, Before inputting the time-series neural network model, the real-time monitoring data is preprocessed: The real-time monitoring data is denoised using a wavelet transform algorithm to separate high-frequency noise signals from low-frequency trend signals. The denoised data is spatiotemporally aligned and dimensionality reduced to extract spatiotemporal feature vectors that reflect the deformation and evolution characteristics of the foundation pit.

8. The method for multi-source information fusion and safety situation assessment and early warning of foundation pit engineering according to claim 1, characterized in that, The physical mechanism model is constructed using the limit equilibrium method or the elastic foundation beam theory to provide the theoretical stress state and safety factor boundary of the foundation pit support structure under the current working conditions.

9. A multi-source information fusion and safety situation assessment and early warning system for foundation pit engineering, characterized in that, Also includes: The multi-source data acquisition module is used to acquire real-time monitoring data, physical parameters, and construction condition data of the foundation pit project; The dual-drive fusion assessment module is used to execute the multi-source information fusion and safety situation assessment and early warning method for foundation pit engineering as described in any one of claims 1 to 8, and output the safety situation assessment result. The graded early warning module is used to trigger corresponding alarm commands based on the security situation assessment results.

10. A multi-source information fusion and safety situation assessment and early warning system for foundation pit engineering, characterized in that, It includes a computer-readable storage medium on which a computer program is stored, the computer program being executed by a processor to perform the function of the early warning method as described in any one of claims 1 to 8.

Images