Subway construction AI deep foundation pit deformation early warning system

Abstract

The invention relates to the technical field of computer calculation models, and discloses a subway construction AI deep foundation pit deformation early warning system, which comprises a sensing interface unit for acquiring displacement sequence data of an enclosure structure; a parameter mapping unit extracts rigidity parameters of the supporting system structure, translates a component geometric constraint relation into a calculation graph model logic edge weight and generates a weighted adjacency matrix; the graph reasoning calculation module takes the weighted adjacency matrix as a topological operator to be input into a graph neural network model, neighborhood features are aggregated through a convolutional layer, time sequence trend features are extracted, and a calculation state vector is output; and the decision instruction generation module maps a risk score value according to the calculation state vector, and outputs an early warning instruction when the score value exceeds a judgment threshold. Through deep translation from physical rigidity to calculation logic, organic fusion of a calculation framework and an engineering mechanism is realized, the analysis precision of a system on a structure evolution state is enhanced, and the reliability of the system is improved. And false alarms caused by environmental noise are effectively avoided.

Description

AI-powered early warning system for deep foundation pit deformation in subway construction Technical Field

[0001] This invention belongs to the field of computer computing model technology, and in particular relates to an AI-based early warning system for deep foundation pit deformation in subway construction. Background Technology

[0002] Current deep foundation pit excavation construction for subway stations and tunnel sections involves complex soil stress release and structural resistance evolution. Soil displacement and retaining structure deformation exhibit strong spatiotemporal coupling characteristics. Utilizing computer systems for real-time monitoring and trend prediction is a fundamental means of ensuring construction safety. Existing technologies typically employ computer early warning systems based on neural networks or regression analysis. These systems acquire displacement time-series signals from various monitoring points and use statistical models to analyze foundation pit risks. To improve data acquisition coverage and accuracy, various monitoring devices have been developed in the industry, such as those authorized under announcement number CN213867730. U's utility model patent discloses a monitoring device for the impact of deep foundation pit excavation near subway on the vertical deformation of the subway. Through the combination of monitoring rods, observation prisms, anti-tilting concentric supports and protective sleeves, it realizes the synchronous monitoring of the vertical displacement of deep soil in the pit and the subway tunnel during the deep foundation pit excavation process. This kind of hardware-driven solution improves the dimensionality of the original physical parameters, but it has shortcomings in the data processing level. The core technology essentially treats the foundation pit monitoring system as a collection of isolated data streams. The internal early warning logic is based on the statistical inference of the discrete data distribution law, and lacks the internalized expression of the mechanical transmission mechanism and displacement coordination law of the underground engineering support system.

[0003] However, existing computing systems have shortcomings when facing the complex and highly dynamic and uncertain conditions of subway construction. The core technology of such models is based on statistical inference of the distribution patterns of historical monitoring data, treating the foundation pit monitoring system as a set of isolated data streams. The internal weight update mechanism lacks an internalized expression of the mechanical transmission mechanism and displacement coordination laws of the underground engineering support system. Under interference environments such as heavy tunneling machinery vibration, sudden changes in temporary loads, or sensor logic drift, the general computing model, lacking physical topological constraints, struggles to distinguish between displacement changes that conform to the principle of mechanical continuity and those that violate physical laws. Structural noise makes the system prone to false alarms in complex environments, or forces the system to reduce its sensitivity to structural instability signs in order to lower the false alarm rate. To address the above bottlenecks, attempts have been made to improve the system’s immunity by increasing the depth of the computational model or introducing data filtering algorithms. However, such improvement approaches do not address the underlying contradiction between the prediction logic and the physical mechanism. Increasing the number of model parameters often leads to overfitting in the early stages of construction when samples are scarce. At the same time, since the calculation process does not include the stiffness characteristics of physical entities, there is no clear engineering causal chain between the prediction results and the physical world, making it difficult to intuitively reflect the attenuation state of the local support effectiveness of the foundation pit.

[0004] Therefore, the technical problem to be solved by this invention is how to construct a topological computing architecture that internalizes physical stiffness constraints, and how to realize physical consistency reasoning of the deformation field of deep foundation pits under complex disturbances by transforming the stiffness characteristics of supporting components into physical priors of computing weights. Summary of the Invention

[0005] This invention provides an AI-powered deep foundation pit deformation early warning system for subway construction. The system includes: a perception interface unit, used to execute step 101: real-time acquisition of multi-dimensional sequence data characterizing the displacement features of the deep foundation pit retaining structure, and denoising preprocessing of the multi-dimensional sequence data; a parameter mapping unit, connected to the perception interface unit, used to execute step 102: extracting the structural stiffness parameters of the deep foundation pit retaining structure, translating the geometric and physical constraints between support components into logical edge weights in a computational graph model, calculating the numerical values ​​of the logical edge weights based on the structural stiffness parameters, and generating a weighted adjacency matrix characterizing the spatial topological association characteristics of the support system; and a graph reasoning calculation module, connected to the parameter mapping unit, used to execute step 103: using the weighted adjacency matrix as... The topology operator is input into a preset graph neural network model. The convolutional layer of the graph neural network model performs neighborhood node feature aggregation on multidimensional sequence data and extracts temporal displacement trend features in combination with a recurrent neural architecture. The output is a calculated state vector reflecting the stable evolution state of the support system structure. The decision instruction generation module is connected to the graph inference calculation module and is used to execute step 104: perform nonlinear risk probability mapping based on the calculated state vector to calculate the risk score value reflecting the overall stability of the deep foundation pit. When the risk score value continuously exceeds the preset judgment threshold, the warning instruction containing the warning level, risk location coordinates and displacement evolution rate information is output. The warning level is determined based on the relationship between the risk score value and the preset graded risk interval.

[0006] Preferably, the system further includes a compensation calibration unit for performing step 201: obtaining the ambient temperature parameters of the support system, calculating the temperature-sensitive correction factor based on the ambient temperature parameters and the preset thermal compensation coefficient, and using the temperature-sensitive correction factor to perform numerical offset compensation on the logical edge weights in the weighted adjacency matrix, so as to filter the non-structural displacement error components caused by the ambient temperature difference.

[0007] Preferably, the graph reasoning calculation module includes a temporal topology adjustment operator, used to perform step 301: identify the increase or decrease status of support components according to the step sequence parameters of subway construction, and control the logical edge weights of the corresponding nodes in the weighted adjacency matrix to evolve smoothly according to a preset decay function, so as to eliminate the output value jump generated by the calculation model at the time of topology switching.

[0008] Preferably, the system further includes a perturbation analysis unit for performing step 401: obtaining the frequency response signal excited by the vibration of construction machinery from the sensing interface unit, extracting the inherent characteristic frequency of the support structure through a high-pass filter operator, and feeding the inherent characteristic frequency back to the parameter mapping unit for numerical calibration of the structural stiffness parameters.

[0009] Preferably, the perturbation analysis unit performs step 501: updating the dynamic adjustment coefficients in the weighted adjacency matrix according to the following formula. : Where H is the dynamic adjustment coefficient. The extracted characteristic frequency values ​​of mechanical vibration, These are the reference natural frequencies of the support structure, all in Hz.

[0010] Preferably, the decision instruction generation module includes a residual distribution statistician, used to perform step 601: calculate the spatial distribution residual between the measured displacement data and the displacement prediction value output by the graph inference calculation module, and numerically separate the zero-point drift error of the sensing component from the increase in the risk score value caused by the deformation of the support structure according to the dispersion index of the spatial distribution residual.

[0011] Preferably, the system further includes an edge processing gateway for carrying the sensing interface unit and the parameter mapping unit, and performing step 701: performing linear projection dimensionality reduction processing on the multidimensional sequence data, and sending the generated topological feature vector to the cloud server to reduce the bandwidth occupation of the field communication link.

[0012] Preferably, the sensing interface unit is connected to a Beidou monitoring terminal with positioning function to perform step 801: acquiring geographic coordinate data containing the same high-precision time synchronization label as multidimensional sequence data, so as to ensure that the graph reasoning calculation module performs the correlation feature extraction of multidimensional sequence data under a unified time reference.

[0013] Preferably, the decision instruction generation module is used to execute step 901: when the numerical jump rate of the calculated state vector is greater than the preset step size, the displacement redundancy information of adjacent monitoring points is called to perform confidence weighting processing on the risk score value in order to filter out data anomalies caused by the failure of a single sensing node.

[0014] Preferably, the system further includes a feedback iteration unit for performing step 1001: comparing the risk score with the physical measured displacement value obtained by on-site manual inspection, and correcting the scaling factor in the parameter mapping unit that maps the geometric physical constraint relationship to the logical edge weight based on the prediction deviation obtained by comparison, so that the output result of the computational graph model approximates the physical boundary conditions in the deep foundation pit evolution process.

[0015] Compared with existing technologies, the AI-based deep foundation pit deformation early warning system for subway construction of this invention has the following advantages: 1. In the AI-based deep foundation pit deformation early warning, by constructing a logical topology diagram that mirrors the physical support structure of the foundation pit, and using the equivalent stiffness coefficient of the physical support components to determine the initial connection weight of the logical edges, the prediction process of the computer system is transformed from simple data statistics-driven to physical constraint-driven. This physically consistent mapping mechanism enables the model to have an understanding of the transmission law of underground engineering mechanics in the initialization stage, ensuring that the prediction output is always within the reasonable space defined by the displacement coordination constraint, eliminating the logical distortion that general algorithms are prone to when facing non-statistical interference at the construction site, and improving the reliability of early warning judgment.

[0016] 2. By integrating local consistency arbitration logic into the topological constraint calculation unit, logical self-healing for abnormal front-end sensing data is achieved. When the sensor at a specific monitoring point experiences data deviation due to construction machinery collision or drastic environmental changes, the system does not rely on simple threshold filtering. Instead, it uses the feature vectors of neighboring nodes and preset physical stiffness weights to perform logical reconstruction on abnormal nodes that deviate from the displacement coordination constraints. This ensures that the global calculation graph can still maintain stable deformation field inference through physical redundancy information even when some sensing units fail, thus enhancing the system's operational robustness under complex interference conditions.

[0017] 3. By leveraging the deep coupling between the temporal evolution soft connection operator and the construction step parameters, the computational impact risk caused by the frequent erection and dismantling of support components in subway construction is eliminated. Discrete physical construction events are translated into a continuous evolution process of topological activity coefficients, allowing the connection weights to smoothly transition with the increase or decrease of the actual bearing capacity of the components. This effectively avoids prediction curve jumps and false alarm pulses caused by hard switching of the model topology, and achieves accurate approximation of the dynamic evolution physical boundary conditions by the computational model, ensuring the continuity of early warning commands in high-risk stages such as support replacement. Attached Figure Description

[0018] Figure 1 is a schematic diagram of the system architecture and logic flow of the AI-based deep foundation pit deformation early warning system for subway construction according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0020] It should be noted that all directional and positional terms used in this invention, such as: up, down, left, right, front, back, vertical, horizontal, inner, outer, top, bottom, transverse, longitudinal, center, etc., are only used to explain the relative positional relationship and connection between components in a specific state (as shown in the accompanying drawings). They are only for the convenience of describing this invention and do not require that this invention be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention. In addition, the descriptions of "first," "second," etc., in this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated.

[0021] In the description of this invention, unless otherwise explicitly specified and limited, the terms installation, connection, and linking should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections; they can refer to direct connections or indirect connections through an intermediate medium; they can refer to the internal connection of two components. For those skilled in the art, the specific meaning of the above terms in this invention can be understood according to the specific circumstances.

[0022] In the description of this specification, references to the terms "an embodiment," "some embodiments," "illustrative embodiments," "examples," "specific examples," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example, and the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0023] This invention provides an AI-powered early warning system for deep foundation pit deformation in subway construction, comprising a sensing interface unit, a parameter mapping unit, a graph inference calculation module, and a decision command generation module. The sensing interface unit collects and preprocesses displacement signals of the foundation pit retaining structure; the parameter mapping unit converts the stiffness parameters of the physical support system into logical weights of the computational graph model; the graph inference calculation module extracts spatiotemporal displacement features under topological constraints; and the decision command generation module outputs early warning information based on the computational state vector. The various units interact with each other through a data interface. The sensing interface unit is used to acquire multi-dimensional sequence data characterizing the displacement features of the deep foundation pit retaining structure in real time. This system aims to address the deformation caused by vibrations from tunneling machinery at subway construction sites. To reduce noise interference, the sensing interface unit uses a median filtering method based on a sliding time window to denoise the multidimensional sequence data. During this process, the system sets the sampling frequency to 10Hz and selects 50 consecutive sampling points to form a sliding window. The median value within this window is used to replace sampling values ​​whose dispersion exceeds a preset deviation threshold, outputting displacement components that reflect the true deformation trend of the retaining wall. The parameter mapping unit, connected to the sensing interface unit, extracts the structural stiffness parameters of the deep foundation pit retaining structure. These parameters include the bending stiffness EI of the retaining wall and the axial stiffness EA of the supporting members. The parameter mapping unit converts the geometric and physical constraints between the supporting members into logical edge weights in the computational graph model. The system executes a dimensional balance procedure, dividing the bending stiffness of the retaining wall (in kN·m²) and the axial stiffness of the supporting member (in kN) by their corresponding initial design reference values, converting them into dimensionless mechanical ratios within the range of 0 to 1. The calculated logical edge weight is equal to the product of the scaling factor 0.15 and this dimensionless mechanical ratio. Through this mapping procedure, physical stiffness parameters of different magnitudes and dimensions are characterized within a unified weight space. The calculation formula is as follows: ,in, The logical edge weight between node i and node j in the computation graph model is calculated; k is a preset scaling factor; P is the corresponding structural stiffness parameter, i.e. bending stiffness EI or axial stiffness EA; the parameter mapping unit generates a weighted adjacency matrix A that characterizes the spatial topological association characteristics of the support system according to the connection relationship of each support node, thereby introducing the mechanical constraints of the support system into the computation architecture.

[0024] To determine the value of the scaling factor k for the parameter mapping unit, an initial calibration was performed using a three-dimensional numerical model of the deep foundation pit established based on the finite element analysis method, and the theoretical displacement of the retaining wall under a unit load was obtained. Adjusting the scaling factor k makes the predicted displacement output by the graph inference calculation module... and The sum of squared residuals between them reaches its minimum, with an initial value of 0.15. The dynamic adjustment coefficient H calculated by the perturbation analysis unit reflects the disturbance law of the vibration of heavy machinery on the support system stiffness. In the real-time weight update logic, the system directly multiplies the initial logical edge weights generated by the parameter mapping unit by this dynamic adjustment coefficient H to obtain the corrected logical edge weights. If the extracted mechanical vibration characteristic frequency increases from the reference 5Hz to 7.5Hz, the dynamic adjustment coefficient H will increase from 1.0 to 1.5. The system will then increase the value of the logical edge weights by 50% to compensate for the deviation in the calculation of the equivalent stiffness of the structure caused by the high-frequency excitation of construction machinery. The reference natural frequency value is... Under static and stable conditions in the foundation pit, random vibration signals were collected using an accelerometer, and power spectrum estimation was performed to determine the extracted characteristic frequency values ​​of mechanical vibration. This module is used to characterize frequency drift caused by loosening of support nodes or loss of prestress during the construction phase in real time. The graph inference calculation module is connected to the parameter mapping unit, and the weighted adjacency matrix A is input as a topological operator into the preset graph neural network model. The convolutional layers of the graph neural network model perform neighborhood node feature aggregation on multidimensional sequence data. The state evolution of each computation node is affected by the displacement deviation of adjacent nodes and the logical edge weights. Given the physical constraints, the graph reasoning and computation module combines the recurrent neural architecture to extract temporal displacement trend features and outputs a computational state vector S that reflects the stable evolution of the support system architecture.

[0025] The decision command generation module is connected to the graph reasoning calculation module. It performs nonlinear risk probability mapping based on the calculated state vector S. The decision command generation module calculates a risk score R reflecting the overall stability of the deep foundation pit and outputs an early warning command when the risk score R continuously exceeds a preset judgment threshold of 0.85. The early warning command includes the early warning level, risk location coordinates, and displacement evolution rate information. The algorithm path for the nonlinear risk probability mapping in the decision command generation module is as follows: A first-layer fully connected network containing 256 neurons converts the calculated state vector into an intermediate-state feature vector. A linear rectified function with a threshold set to 0.05 is used to suppress minor fluctuations caused by background noise. A second-layer fully connected network further compresses the features and projects them onto a probability domain between 0 and 1 using a normalized exponential function to determine the risk score. The system sets the early warning trigger conditions. To ensure that the risk score remains above 0.85 for five consecutive sampling periods, the judgment threshold is determined based on the 95% confidence interval of historical displacement distribution data when the deep foundation pit is excavated to the base. The first-layer fully connected network operator performs a spatial transformation on the calculated state vector S, and the linear rectified function ReLU is used to process the implicit features to suppress the interference of low-amplitude data fluctuations. The second-layer fully connected network operator, combined with the normalized exponential function Softmax, projects the calculation result into the interval [0,1] to determine the risk score R. The judgment threshold of 0.85 is determined by superimposing the normalized result of the reciprocal of the foundation pit design safety factor with the 95th percentile of the statistical distribution of the construction displacement field. When the risk score R continuously exceeds this threshold, the system calls the displacement data of the neighboring monitoring points in the weighted adjacency matrix A to perform spatial correlation verification, and outputs the risk location coordinates after confidence weighting to the early warning command.

[0026] The compensation and calibration unit is used to obtain the ambient temperature parameter T of the support system. To address the thermal expansion displacement error of the steel support caused by temperature difference, the compensation and calibration unit calculates the temperature sensitivity correction factor β based on the ambient temperature parameter T and the preset thermal compensation coefficient α. The temperature sensitivity correction factor β is then used to adjust the logical edge weights in the weighted adjacency matrix A. Numerical offset compensation is performed, and the specific compensation logic satisfies the following formula: ,in, The logical edge weights are compensated; ΔT is the difference between the current ambient temperature and the reference temperature; this method filters temperature interference data by adjusting the logical weights; the graph reasoning calculation module includes a temporal topology adjustment operator; the addition or subtraction status of support components is identified based on the step sequence parameters of subway construction, and the logical edge weights of the corresponding nodes in the weighted adjacency matrix A are controlled. The system undergoes a smooth evolution process according to a preset attenuation function. This process transforms discrete construction events such as support replacement into continuous weight change curves, eliminating output value jumps in the computational model at topology switching moments. The perturbation analysis unit acquires the frequency response signal excited by construction machinery vibration from the sensing interface unit. The perturbation analysis unit extracts the inherent characteristic frequencies of the support structure through a high-pass filter operator. and inherent characteristic frequencies The result is fed back to the parameter mapping unit to update the dynamic adjustment coefficient H in the weighted adjacency matrix A. The calculation formula is as follows: Where H is the dynamic adjustment coefficient; The extracted characteristic frequency values ​​of mechanical vibration; The reference natural frequency value of the support structure is Hz. The system corrects the structural stiffness parameter value according to H to realize online calibration of support node loosening or prestress loss. The decision command generation module includes a residual distribution statistical analyzer. It calculates the spatial distribution residual δ between the measured displacement data and the displacement prediction value output by the graph inference calculation module. The residual distribution statistical analyzer separates the zero-point drift error of the sensing component from the increase of the risk score value R caused by the deformation of the support structure according to the dispersion index of the spatial distribution residual δ. When the residual shows local area aggregation characteristics, it is judged as structural deformation risk.

[0027] The edge processing gateway carries the sensing interface unit and the parameter mapping unit. It performs linear projection dimensionality reduction on multidimensional sequence data and sends the generated topological feature vector to the cloud server. This method reduces the bandwidth consumption of the on-site communication link and ensures the real-time data transmission. The sensing interface unit connects to the Beidou monitoring terminal with positioning function to obtain geographic coordinate data containing the same time tag as multidimensional sequence data. This ensures that the graph inference calculation module performs the correlation feature extraction of multidimensional sequence data under a unified time reference. When the numerical jump rate of the calculated state vector S is greater than the preset step size, the decision instruction generation module calls the displacement redundancy information of adjacent monitoring points to perform confidence weighting processing on the risk score value R. This processing utilizes the redundancy relationship of the physical topology to filter data anomalies caused by the failure of a single sensing node. The feedback iteration unit compares the risk score value R with the physical measured displacement value obtained by on-site manual inspection. Based on the prediction deviation obtained from the comparison, the feedback iteration unit corrects the scaling factor k in the parameter mapping unit so that the output result of the computational graph model approaches the physical boundary conditions in the deep foundation pit evolution process.

[0028] Example 1: In a deep foundation pit construction scenario for a subway station adjacent to an existing main traffic artery and containing a layer of silty soft soil, heavy machinery such as trenching machines deployed on-site generate high-frequency vibrations, and frequent starts and stops of earthmoving vehicles cause temporary load pulses. The sensing interface unit continuously acquires the displacement sequence of the retaining structure at a sampling frequency of 10Hz, and uses the sliding time window mid-range filter operator described in the aforementioned specific implementation method to remove spike noise from the original signal with a window width of 50 sampling points. At the same time, the parameter mapping unit retrieves the bending stiffness EI of the retaining wall and the axial stiffness EA of the supporting components, and translates these physical and mechanical parameters into logical edge weights in the computational graph model according to the preset scaling factor k. Generate a weighted adjacency matrix A that reflects the spatial topology of the support system, due to the logical edge weights. This constitutes a physical constraint based on structural stiffness, enabling the graph reasoning calculation module to distinguish between the true deformation trend that conforms to the displacement coordination principle and random vibration noise that deviates from the physical topological laws when performing feature aggregation. By directly internalizing the mechanical mechanism of the support system into the edge weights of the calculation graph, it achieves physical-level suppression of environmental noise interference, ensuring that the risk score value R calculated by the decision instruction generation module is within the judgment threshold of 0.85, and avoiding false warnings under complex dynamic load interference.

[0029] During the transition from the construction sequence to the support replacement stage, the system faces discontinuous changes in the physical structural stiffness field caused by the removal of support components. The graph reasoning calculation module uses a temporal topology adjustment operator to identify the current construction state and adjust the logical edge weights corresponding to the support nodes that are about to be removed. The system performs a smooth evolution according to a preset attenuation function. By converting discrete support removal events into a continuous weighted evolution process, it eliminates the predicted numerical jumps caused by topology switching in the computational model. During this process, the perturbation analysis unit extracts the inherent characteristic frequencies of the support structure excited by on-site mechanical disturbances. The dynamic adjustment coefficient H is calculated according to the formula, and the calculation method is as follows: Where H is the dynamic adjustment coefficient; The extracted characteristic frequency values ​​of mechanical vibration; The reference natural frequency value of the support structure is Hz. The system uses the dynamic adjustment coefficient H to correct the scaling factor k in real time, thereby compensating for the weight calculation deviation caused by the release of support stress. This topology adjustment based on construction logic and the stiffness inversion based on vibration feedback work together to enable the system to capture the real residual stress redistribution phenomenon generated by the support structure during the support replacement process.

[0030] When a deep foundation pit is under long-term monitoring and there is a hidden risk of failure due to loosening of support nodes, the spatial distribution residual δ between the measured displacement data of the monitoring points and the displacement prediction value output by the graph inference calculation module is calculated using a residual distribution statistical analyzer. A topological prediction model based on physical stiffness is established. Any local failure that does not conform to the mechanical transmission logic will cause the spatial distribution residual δ to generate asymmetric aggregation in that area. By analyzing the spatial entropy index of the spatial distribution residual δ in the topological graph, deformation anomalies with physical correlation are separated from the isolated zero-point drift error of the sensing components. This transforms the identification problem from a single numerical judgment to an analysis of the residual. Logical reasoning based on spatial distribution patterns solves the problem of traditional monitoring methods failing to identify hidden structural damage. The final decision command generation module outputs early warning information containing risk location coordinates based on the calculated state vector S, achieving accurate positioning of hidden risks in the foundation pit support system. This embodiment deeply aligns the physical support logic of subway deep foundation pits with the topological evolution mechanism of the computing system, enabling the originally discrete monitoring data to be translated into a mechanism within a physical constraint framework. It proves that by mapping physical entity features to logical weights of the computing model, the operational robustness and diagnostic accuracy of the early warning system in a dynamically evolving construction environment can be improved.

[0031] Example 2: To verify the analytical accuracy and physical consistency of the AI-based deep foundation pit deformation early warning system for subway construction under strong disturbance conditions, the experiment selected a deep foundation pit support evolution dataset generated based on finite element analysis (FEA) as the data source. This dataset is constructed based on the Navier-Stokes equations and the Mohr-Coulomb constitutive model, simulating the stress response of a standard station foundation pit with a depth of 22.5m in a silty soft soil layer. The experimental platform includes a processor with a computing frequency of no less than 3.5GHz and tensor acceleration capability, and a memory space of no less than 64GB to support high-frequency iterative calculations of the graph neural network model. During the data setting phase, the sampling frequency of the sensing interface unit was determined to be 10Hz. This setting was chosen to balance the real-time performance of displacement feature capture with the data processing load of the edge gateway. When the spectral bandwidth of the monitored signal widens to over 5Hz due to mechanical construction, to avoid sampling aliasing and ensure the denoising effectiveness of the median filter operator, the sampling frequency tends towards the upper limit of the range, i.e., 10Hz is used as the fixed sampling frequency for the test group. In the parameter mapping unit, the scaling factor k follows the structural dynamics rules, mapping the bending stiffness EI of the retaining wall to the logical edge weights of the computational graph model through a preset equivalent stiffness mapping function. The scaling factor k was selected as 0.15. This value was chosen based on the balance between physical constraint strength and neural network learning rate gradient stability, and was applied to a typical support system with modal frequencies ranging from 2.5 Hz to 8.0 Hz.

[0032] In a subway foundation pit monitoring scenario involving multiple work faces, the sensing interface unit connects to Beidou monitoring terminals deployed at key control points of the retaining structure to acquire multidimensional sequence data containing high-precision time synchronization tags. The system is configured with edge processing gateways carrying parameter mapping units. Each edge processing gateway uses local computing power to perform linear projection dimensionality reduction processing on the original displacement signal. By calculating the eigenvalues ​​of the covariance matrix of each sampling channel and extracting principal components with a cumulative contribution rate of not less than 95%, the high-dimensional signal is converted into a low-dimensional topological feature vector and sent to the cloud server. High-precision time synchronization tags are used to ensure that the graph inference calculation module meets the clock synchronization requirements when performing neighborhood feature aggregation, eliminating data phase deviation caused by network transmission delay. This reduces the bandwidth usage of the on-site communication link by more than 65%. During the experiment, environmental noise from the construction site was introduced at the original input to simulate high-frequency random micro-disturbances generated by the trenching machine operation. The signal-to-noise ratio was set to... Gaussian white noise of 20dB is superimposed on the theoretical displacement sequence generated by FEA to generate multidimensional sequence data containing noise components. At this time, the original displacement curve exhibits highly discrete characteristics, with an instantaneous jump amplitude of 4.25mm, which masks the actual evolution trend of the retaining wall. The sample group of this invention uses the median filter operator of the sensing interface unit to perform denoising processing on the signal and uses a sliding time window with a width of 50 sampling points to extract the quasi-static component of the displacement signal. The standard deviation of the residual of the processed displacement signal is reduced from the original 1.15mm to 0.22mm, providing high-quality input features for the graph inference calculation module. At the same time, a comparison sample group 1 is set up, which uses a conventional recurrent neural network model and lacks the physical topological constraints represented by the weighted adjacency matrix A, and a comparison sample group 2 uses a traditional statistical regression model. Parallel tests are conducted under the same noise injection environment to verify the enhancement effect of the translation of physical mechanism into computational logic on prediction accuracy.

[0033] During the derivation phase of key intermediate data, the parameter mapping unit generates a weighted adjacency matrix A based on the real-time stress state of the support components, with logical edge weights. As the structural stiffness parameters evolve, when a supporting member at a monitoring node is removed due to simulated bracing, the weight of the corresponding edge controlled by the time-series topology adjustment operator decays exponentially from an initial 1.00 to below 0.05. This evolution process takes 15.5 minutes. By translating discrete construction events into continuous topology activity coefficients, the numerical abrupt changes generated by the computational model at the topology switching moment are eliminated. It was observed that the second derivative of the computational state vector S output by the sample group of this invention remains smooth in the bracing interval, while the prediction of the first sample group at the same node is different. The displacement exhibited a transient jump pulse with an amplitude of 8.40 mm, indicating that this scheme, through the dynamic introduction of physical stiffness weights, resolves the logical conflict between discrete construction events and continuous computation flow, thereby achieving the analysis of the structural evolution state. To verify the gradient law and performance inflection point of the technical effect, multiple sets of comparisons were set up with different noise intensity levels and different scaling factor k gradients. The experimental results show that when the signal-to-noise ratio deteriorates from 30 dB to 10 dB, the mean absolute error of displacement prediction, i.e., MAE, of the sample group of this invention only increases from 0.45 mm. The length reached 0.88 mm, demonstrating interference suppression capability, while the MAE of the comparison sample group 2 surged from 2.15 mm to 12.40 mm. In the optimization test of the scaling factor k, when k was in the range of 0.10 to 0.35, the output gain of the risk score value R was linearly related to the physical deformation rate, proving that this range is the working window of physical constraints. Once k exceeded 0.40, the nonlinear fitting ability of the model saturated due to the over-strengthening of local physical constraints, and the growth trend of the accuracy of risk identification slowed down. If k further increased to above 0.80, it caused the model to amplify the local small errors of the sensor, resulting in a decrease in the stability coefficient of the early warning system by about 25%. This experiment, through quantitative comparison and gradient verification, confirmed that the present invention, by mapping the mechanical constraints of the deep foundation pit to the topological weights of the computer system, can decouple the complex construction environment interference from the real structural failure symptoms at the physical level. Its output risk score value R captures the support loosening risk points set in the simulation test, and the early warning trigger time is 4.2 hours earlier than that of the conventional statistical model.

[0034] Example 3: In a deep foundation pit engineering application scenario of a subway transfer station located in a high-density urban building area with an excavation depth exceeding 30m, a combined structure consisting of diaphragm walls and internal supports is adopted. Due to the asymmetry between groundwater level fluctuations and settlement constraints of surrounding high-rise buildings, the sensing signals exhibit extremely strong nonlinear coupling in the spatial dimension. During system operation, the initialization procedure of the sensing interface unit is executed to determine the initial displacement vector of each monitoring node as the zero reference, and the structural stiffness parameters of the support system are input into the parameter mapping unit, where the bending stiffness of the retaining wall is... The initial value is set to 100% of the design value. The system starts the inference procedure of the graph neural network model, and performs neighborhood feature aggregation on multidimensional sequence data through convolutional layers. In each convolutional operation, the latent feature vector of node i is obtained through the features of its neighboring node j and the corresponding logical edge weights. The weighted summation is used to obtain the logical edge weights. The parameter mapping unit determines the result according to the formula: ,in, For the calculation of the graph weights; P is the measured equivalent stiffness parameter of the support member; The design stiffness benchmark value of the support components is used. This step transforms the mechanical constraints of the support system into algebraic constraints of the computational graph model, so that the state evolution of each computation node is limited by the mechanical feedback of adjacent physical components. The decision instruction generation module maps the calculated state vector S to obtain the risk score value R, and compares the risk score value R with the judgment threshold. The judgment threshold is calibrated based on the reciprocal of the static safety margin of the foundation pit. Under this working condition, the system reads the design safety index and combines it with the displacement standard deviation of the previous construction period to calculate the judgment threshold as 0.85. By aligning the physical mechanism with the computational logic, the logical divergence contradiction generated by the pure data-driven model when analyzing geological abrupt changes is resolved.

[0035] In the critical stage before the foundation slab is poured during the construction sequence, the stress state of the support system approaches the design load limit, causing the effective axial stiffness EA of the support components to weaken nonlinearly due to material plastic damage. The system uses a perturbation analysis unit to capture the inherent characteristic frequencies of the support structure in real time. When the monitored frequency is relative to the reference natural frequency When the attenuation rate exceeds 15%, a dynamic adjustment mechanism is triggered, and the scaling factor k in the parameter mapping unit is progressively corrected. According to experimental group observation data, when the axial stiffness EA drops below 80% of the original design value, the logical edge weight... The value of the risk score R shows a non-linear decay trend. At this time, the risk score R increases from 0.62 to 0.88, triggering the system to output an early warning command containing the coordinates of the risk location. During this process, the spatial entropy index of the spatial distribution residual δ is analyzed to identify physical disturbances. When the spatial entropy value decreases for three consecutive sampling periods and the residual shows a clustered aggregation feature in the logic subgraph, it is determined that there is a real physical failure risk in the support system, rather than the zero-point drift error of the non-sensing component. Finally, the displacement evolution rate of the early warning command output is consistent with the measured inclinometer data verified by subsequent manual review, and its prediction error range is maintained within 1.5 mm.

[0036] Example 4: Before deploying an early warning system in the deep foundation pit of a newly started subway station, the initial state of the computational model is established through on-site parameter calibration. Representative support nodes in the retaining structure are selected to perform load tests with known stress increments to obtain the displacement response characteristics of the support components under specific loads. The parameter mapping unit uses the on-site measured displacement values ​​and the theoretical stiffness field generated by finite element simulation to perform least squares fitting to calculate the equivalent stiffness mapping scaling factor k suitable for the current engineering environment. Specifically, when the strain measurement value of the retaining wall sample is at 50μm... Up to 150μ In the case of range fluctuations, the graph inference calculation module reduces the model prediction error to below 0.5mm through iterative calculations, and the determined scaling factor k is stored in the configuration memory as a global preset variable of the parameter mapping unit.

[0037] When the system encounters situations with differences in sensor batches or inconsistent background noise at measurement points, the sensing interface unit executes a pre-calibration procedure. Before loading the support structure, it collects background signals from no fewer than 3600 sampling points, calculates the static drift component and white noise variance of each monitoring channel, and establishes a noise compensation matrix B for the current physical measurement point distribution. The sensing interface unit then injects the noise compensation matrix B into a median filter operator based on a sliding time window, ensuring that the denoised displacement components... The calculation formula is as follows: ,in, This is the denoised displacement output vector. For the median operation operator, The input data is the original multidimensional sequence, and B is the environmental noise compensation matrix. The system error of the hardware system is corrected during the deployment phase, and the feature vector input to the graph inference calculation module is aligned with the mechanical state of the physical support structure.

[0038] Example 5: In a deep foundation pit monitoring scenario for a subway station located in a permeable stratum near a river where some monitoring cables have been damaged by construction compaction, the sensing interface unit detects an interruption in the displacement signal of a specific monitoring node. The system executes a computational graph reconstruction procedure based on physical redundancy, extracts the implicit features of the N adjacent active nodes using the topological connectivity of the damaged node in the weighted adjacency matrix A, and combines the corresponding logical edge weights. The state evolution of missing nodes is filled using spatial operators, where the latent features of the missing nodes are derived from their neighborhood features and logical edge weights. The sum of the products determines the physical stiffness continuity constraint of the support system, which is translated into the fault tolerance mechanism of the computational model. This enables the graph reasoning computation module to output a complete computational state vector S in the case of local failure of the sensing unit, and the reasoning logic of the risk score value R remains continuous under the hardware physical damage condition.

[0039] When the system encounters a situation where the physical stiffness field has local pre-set deviations due to limitations in the accuracy of geological exploration, the feedback iteration unit executes an online parameter convergence procedure. After the first layer of support in the foundation pit is erected, 48 hours of measured displacement data are collected. The spatial distribution residual δ between each physical measuring point and the predicted value output by the graph inference calculation module is calculated. The spatial entropy index of the spatial distribution residual δ in the topology graph is then calculated. To identify mechanistic biases, spatial entropy index The threshold for judgment is selected as 1.5 times the standard deviation of the residuals during the baseline operation phase. When the value deviates from this range, the system performs gradient descent correction on the scaling factor k in the parameter mapping unit. By introducing a residual feedback mechanism, the initial stiffness parameter is corrected. The mapping accuracy between the risk score value R and the actual stratum response is improved by no less than 18.5%, and the displacement prediction deviation of the warning command in the subsequent deep excavation process converges to within 0.8 mm.

[0040] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit of this application and the scope of protection of this invention, and all of these forms are within the protection scope of this application.

Claims

1. A deep foundation pit deformation early warning system for subway construction using AI, characterized in that, The system includes: a perception interface unit, used to execute step 101: real-time acquisition of multi-dimensional sequence data characterizing the displacement features of the deep foundation pit retaining structure, and denoising preprocessing of the multi-dimensional sequence data; a parameter mapping unit, connected to the perception interface unit, used to execute step 102: extracting the structural stiffness parameters of the deep foundation pit retaining structure, translating the geometric and physical constraint relationships between the support components into logical edge weights in the computational graph model, calculating the numerical values ​​of the logical edge weights based on the structural stiffness parameters, and generating a weighted adjacency matrix characterizing the spatial topological association characteristics of the support system; and a graph reasoning calculation module, connected to the parameter mapping unit, used to execute step 103: inputting the weighted adjacency matrix as a topological operator into a preset graph model. The network model aggregates neighborhood node features on multidimensional sequence data through the convolutional layer of the graph neural network model, and extracts temporal displacement trend features by combining the recurrent neural architecture, outputting a calculated state vector reflecting the stable evolution state of the support system structure; the decision instruction generation module is connected to the graph reasoning calculation module and is used to execute step 104: perform nonlinear risk probability mapping based on the calculated state vector, calculate the risk score value reflecting the overall stability of the deep foundation pit, and when the risk score value continuously exceeds the preset judgment threshold, output a warning instruction containing the warning level, risk location coordinates and displacement evolution rate information. The warning level is determined based on the relationship between the risk score value and the preset graded risk interval.

2. The AI-based deep foundation pit deformation early warning system for subway construction according to claim 1, characterized in that, The system also includes a compensation calibration unit for performing step 201: obtaining the ambient temperature parameters of the support system, calculating the temperature-sensitive correction factor based on the ambient temperature parameters and the preset thermal compensation coefficient, using the temperature-sensitive correction factor to perform numerical offset compensation on the logical edge weights in the weighted adjacency matrix, and filtering the non-structural displacement error components caused by the ambient temperature difference.

3. The AI-based deep foundation pit deformation early warning system for subway construction according to claim 1, characterized in that, The graph reasoning calculation module includes a temporal topology adjustment operator, which is used to perform step 301: identify the increase or decrease status of support components according to the step sequence parameters of subway construction, and control the logical edge weights of the corresponding nodes in the weighted adjacency matrix to evolve smoothly according to a preset decay function.

4. The AI-based deep foundation pit deformation early warning system for subway construction according to claim 1, characterized in that, The system also includes a perturbation analysis unit, which performs step 401: obtaining the frequency response signal excited by the vibration of construction machinery from the sensing interface unit, extracting the inherent characteristic frequency of the support structure through a high-pass filter operator, and feeding the inherent characteristic frequency back to the parameter mapping unit for numerical calibration of the structural stiffness parameters.

5. The AI-based deep foundation pit deformation early warning system for subway construction according to claim 4, characterized in that, Perturbation analysis unit execution step 501: Update the dynamic adjustment coefficients in the weighted adjacency matrix according to the following formula. ： Where H is the dynamic adjustment coefficient. The extracted characteristic frequency values ​​of mechanical vibration, These are the reference natural frequencies of the support structure, all in Hz.

6. The AI-based deep foundation pit deformation early warning system for subway construction according to claim 1, characterized in that, The decision instruction generation module includes a residual distribution statistician, which is used to perform step 601: calculate the spatial distribution residual between the measured displacement data and the displacement prediction value output by the graph inference calculation module, and numerically separate the zero-point drift error of the sensing component from the increase in the risk score value caused by the deformation of the support structure according to the dispersion index of the spatial distribution residual.

7. The AI-based deep foundation pit deformation early warning system for subway construction according to claim 1, characterized in that, The system also includes an edge processing gateway, which carries the sensing interface unit and the parameter mapping unit, and performs step 701: performing linear projection dimensionality reduction processing on the multidimensional sequence data and sending the generated topological feature vector to the cloud server.

8. The AI-based deep foundation pit deformation early warning system for subway construction according to claim 1, characterized in that, The sensing interface unit is connected to a Beidou monitoring terminal with positioning function to perform step 801: acquiring geographic coordinate data containing the same high-precision timing tag as multidimensional sequence data.

9. The AI-based deep foundation pit deformation early warning system for subway construction according to claim 1, characterized in that, The decision instruction generation module is used to execute step 901: when the numerical jump rate of the calculated state vector is greater than the preset step size, the displacement redundancy information of adjacent monitoring points is called to perform confidence weighting processing on the risk score value, and the data anomalies caused by the failure of a single sensing node are filtered out.

10. The AI-based deep foundation pit deformation early warning system for subway construction according to claim 1, characterized in that, The system also includes a feedback iteration unit for performing step 1001: comparing the risk score with the physical measured displacement value obtained by on-site manual inspection, and correcting the scaling factor in the parameter mapping unit that maps the geometric physical constraint relationship to the logical edge weight based on the prediction deviation obtained from the comparison.

Images