Support structure deformation multi-source data fusion collaborative prediction system

Abstract

The application discloses a supporting structure deformation multi-source data fusion cooperative prediction system and relates to the field of civil construction engineering monitoring. The application discloses a supporting structure deformation multi-source data fusion cooperative prediction system and relates to the field of civil construction engineering monitoring. The application discloses a supporting structure deformation multi-source data fusion cooperative prediction system and relates to the field of civil construction engineering monitoring. The application discloses a supporting structure deformation multi-source data fusion cooperative prediction system and relates to the field of civil construction engineering monitoring. The application discloses a supporting structure deformation multi-source data fusion cooperative prediction system and relates to the field of civil construction engineering monitoring. The application discloses a supporting structure deformation multi-source data fusion cooperative prediction system and relates to the field of civil construction engineering monitoring. The application discloses a supporting structure deformation multi-source data fusion cooperative prediction system and relates to the field of civil construction engineering monitoring. The application discloses a supporting structure deformation multi-source data fusion cooperative prediction system and relates to the field of civil construction engineering monitoring. The application discloses a supporting structure deformation multi-source data fusion cooperative prediction system and relates to the field of civil construction engineering monitoring. The application discloses a supporting structure deformation multi-source data fusion cooperative prediction system and relates to the field of civil construction engineering monitoring.

Description

Technical Field

[0001] This invention relates to the field of civil engineering monitoring, and in particular to a multi-source data fusion and collaborative prediction system for support structure deformation. Background Technology

[0002] With the increasing development of urban underground space, the scale and depth of ultra-deep foundation pit projects are constantly increasing. During service, the foundation pit support structure is subject to the complex coupling effects of multiple physical fields, including soil creep, groundwater dynamics, and construction loads, making its safety crucial. A review and analysis of existing technologies reveals the following shortcomings in current foundation pit monitoring methods: First, existing systems mostly use static sampling at a fixed frequency, which cannot capture high-frequency transient characteristic displacement signals of the support structure on the eve of instability; Second, heterogeneous sensor data lacks a unified spatial coordinate mapping based on the 3D Building Information (BIM) model, making it difficult to identify the physical relationship between "local water level changes - sudden changes in structural stress - macroscopic displacement trends"; Third, the system mostly remains in the "monitoring-early warning" stage, lacking proactive intervention methods to directly intervene in the physical and mechanical fields when signs of instability appear. Summary of the Invention

[0003] To address the aforementioned issues, this invention provides a multi-source data fusion and collaborative prediction system for support structure deformation.

[0004] The present invention provides a multi-source data fusion and collaborative prediction system for support structure deformation, which adopts the following technical solution: A multi-source data fusion and collaborative prediction system for support structure deformation includes: A sensing terminal cluster is used to acquire multi-source monitoring data reflecting the physical state of the support structure; wherein, the sensing terminal cluster includes distributed fiber optic sensors, a three-axis MEMS accelerometer, a vision measuring instrument, and a step-frequency continuous wave ground-penetrating radar. An edge computing gateway, whose input is electrically connected to the sensing terminal cluster, is used to perform wavelet threshold hierarchical noise reduction and time axis alignment processing on the multi-source monitoring data and output a standard spatiotemporal dataset. The multi-source data fusion module, based on the unified coordinate benchmark of the 3D Building Information Model (BIM) and Geographic Information System (GIS), uses the coordinate transformation matrix M to establish the spatial mapping relationship between different dimensions of data in the standard spatiotemporal dataset, and uses a convolutional neural network (CNN) for feature extraction, outputting a coupled feature vector; the collaborative prediction module is used to input the coupled feature vector into the GA-LSTM prediction model, and outputs the deformation trend prediction value of the support structure by calling the cloud computing cluster. A feedback control loop is used to generate frequency adjustment commands and servo drive commands based on the deformation trend prediction value; The actuators include an intelligent pressure relief pump unit and a clock divider register; The frequency adjustment command is used to modify the clock divider register value of the monitoring terminal hardware in the sensing terminal cluster to adjust the sampling frequency; the servo drive command drives the intelligent pressure relief pump group to adjust the pore water pressure around the support structure through the Modbus-TCP protocol.

[0005] Preferably, the formula for wavelet threshold hierarchical noise reduction performed by the edge computing gateway is: , in, The high-frequency wavelet coefficients of the j-th level decomposition reflect the amplitude of a specific frequency component. Based on the standard deviation of background noise at the construction site The risk threshold, dynamically generated with a sampling window length N, is expressed as follows: , The standard deviation of background noise extracted during non-working periods at the support construction site is given by N, where N is the sampling window length.

[0006] Preferably, the edge computing gateway utilizes a cubic spline interpolation formula. Time axis alignment is performed on multi-source monitoring data, and the cubic spline interpolation formula is used. satisfy: , in, To monitor time points, The polynomial coefficients are used to satisfy the second derivative continuity constraint at the nodes.

[0007] Preferably, the coordinate transformation matrix M is The homogeneous transformation matrix, whose mapping relationship is expressed as follows: , in, Let represent the local homogeneous coordinate vector collected by the sensing terminal cluster. Where, These are the local spatial coordinates of the sensor's original output. The 3x3 rotation matrix represents the cosine of the angle between the axis of the distributed fiber optic sensor and the global axis of the BIM. Let be the translation vector of the 3x1 distributed fiber optic sensor mounting point in the global coordinate system. The symbols are collectively referred to as P, where P represents the globally unified coordinates corresponding to the 3D Building Information (BIM) model, used to complete the rotation matrix and translation vector. Homogeneous matrix.

[0008] Preferably, the CNN performs cross-modal feature extraction using 3×3 convolutional kernels and outputs coupled feature vectors after processing with the LeakyReLU activation function.

[0009] Preferably, the GA-LSTM prediction model utilizes a genetic algorithm to globally optimize the number of hidden layer neurons and the learning rate of the LSTM, and the LSTM forget gate... The calculation formula is: ,in, As input features, This is the hidden state from the previous moment. It is the Sigmoid activation function. This is the weight matrix. For bias terms, This indicates that the hidden state from the previous time step is concatenated with the current input features as a vector.

[0010] Preferably, the coupled feature vector is obtained through the Pearson correlation coefficient. Calculate the correlation contribution weight coefficient The formula is: , : The Pearson correlation coefficient between the i-th monitoring dimension and the deformation target item y; k: The total number of dimensions of the current multi-source sensing terminal cluster; : The contribution weighting factor assigned to the feature extraction layer of the convolutional neural network.

[0011] Preferably, the hardware clock division coefficient K in the frequency adjustment command satisfies a mapping relationship with the deformation trend prediction value y: ,in, The hardware reference frequency division coefficient, The response sensitivity coefficient, y: Safety deformation threshold; y: Deformation trend prediction value.

[0012] Preferably, the feedback control loop uses a PID control algorithm to adjust the flow rate of the pressure relief pump in order to reduce the pore water pressure. To forcibly increase the effective stress of the soil Its mechanical mechanism satisfies ,in, For the effective stress of the soil, Let be the total soil stress, and u be the pore water pressure.

[0013] Preferably, the distributed optical fiber sensor is covered with a spiral protective tube with a ring stiffness ≥10KN / ㎡, and is fastened to the main reinforcement of the support pile steel cage by U-shaped metal clips with a fixed spacing of 0.5m.

[0014] In summary, the present invention has at least one of the following beneficial technical effects: 1. This application uses a 4×4 homogeneous transformation matrix to uniformly map strain data from distributed fiber optic sensors, tilt data from MEMS accelerometers, and coordinate data from visual measuring instruments onto a BIM+GIS 3D coordinate reference, eliminating spatial heterogeneity errors between heterogeneous distributed fiber optic sensors. This "virtual-real binding" mechanism enables CNN to accurately extract the spatiotemporal coupling characteristics between "water level-strain-displacement". At a depth of 40 meters underground, the root mean square error between the predicted displacement and the measured value is only 0.21 mm, achieving sub-millimeter level accuracy. This allows monitoring personnel to intuitively identify the causal relationship between "road surface cracks-pile bending moment-sudden water level change" in 3D space. By identifying creep trends through the GA-LSTM model, collapse signs can be detected 3-5 hours in advance, increasing the lead time by 200% compared to traditional alarm systems. 2. The system dynamically modifies the clock divider register value of the sensing terminal cluster hardware through feedback control loop, and establishes a sampling frequency adaptive adjustment mechanism driven by deformation trend prediction value. When the deformation trend prediction value is close to the safety threshold, the system can automatically compress the sampling period from the hour level to the minute level to ensure that the transient high-frequency deformation characteristics of the support structure on the eve of failure can be captured. 3. This application actively intervenes in the physical and mechanical field of the soil through a feedback control loop, and uses an intelligent pressure relief pump set to precisely adjust the pore water pressure. Based on Terzaghi's effective stress principle, the effective stress is forcibly increased under the condition that the total soil stress remains unchanged, thereby suppressing the lateral displacement of the support structure at the physical level. Experimental data show that compared with traditional static monitoring and passive alarm methods, this system can reduce the maximum deformation of the support structure by more than 30%. The total delay time from the deformation trend trigger threshold to the action of the intelligent pressure relief pump set is 1.45s, which fully meets the safety requirements of deep foundation pit construction. 4. The distributed fiber optic sensor uses a spiral protective tube with a ring stiffness ≥10KN / ㎡ and is fixed with U-shaped clips spaced 0.5m apart. When subjected to the enormous lateral dynamic pressure (approximately 27.6MPa) generated by concrete pouring, the deflection angle of the distributed fiber optic sensor's centerline can be controlled within [value missing]. This addresses the systematic error in displacement inversion caused by installation offset at its physical source. Attached Figure Description

[0015] Figure 1 This is a block diagram of the overall architecture of the present invention; Figure 2 This is a flowchart illustrating the processing of multi-source monitoring data by the edge computing gateway of the present invention. Figure 3 This is a diagram of the internal logic architecture of the GA-LSTM prediction model of the present invention; Figure 4This is a schematic diagram illustrating the principle of spatial mapping of multi-source monitoring data based on a 4×4 homogeneous transformation matrix and a three-dimensional building information (BIM) model according to the present invention. Figure 5 This is a schematic diagram of the closed-loop logic of the feedback control loop of the present invention, which performs frequency regulation and active intervention of the intelligent pressure relief pump group.

[0016] Figure label: 1. Sensing terminal cluster; 11. Distributed fiber optic sensors; 12. Triaxial MEMS accelerometer; 13. Visual measurement instrument; 14. Stepping frequency continuous wave ground-penetrating radar; 2. Edge computing gateway; 3. Multi-source data fusion module; 4. Collaborative prediction module; 5. Feedback control loop; 6. Actuator; 61. Intelligent pressure relief pump set; 62. Clock divider register; 7. Support structure; 8. 3D building information BIM model; 10. U-shaped metal buckle. Detailed Implementation

[0017] The following is in conjunction with the appendix Figure 1 - Figure 5 The present invention will be described in further detail below.

[0018] This invention discloses a multi-source data fusion and collaborative prediction system for support structure deformation.

[0019] Reference Figure 1 A multi-source data fusion and collaborative prediction system for support structure deformation is disclosed. A sensing terminal cluster 1 is used to acquire multi-source monitoring data reflecting the physical state of the support structure 7. The multi-source monitoring data includes spatial geometric displacement sequences such as the three-dimensional coordinate sequence of the cap beam, structural internal force and strain sequences such as the axial force of the support and the bending moment sequence of the pile, and underground environmental parameter sequences such as the dynamic pressure of the groundwater level and the pore water pressure u. The sensing terminal cluster 1 includes a distributed fiber optic sensor 11, a triaxial MEMS accelerometer 12, a vision measuring instrument 13, and a step-frequency continuous wave ground-penetrating radar 14. In specific engineering applications, the support structure 7 can be a diaphragm wall or pile bank for deep foundation pits. The sensing terminal cluster 1 specifically includes the following hardware deployed at the construction site: The distributed fiber optic sensor 11 is installed in the inclinometer tube inside the support pile.

[0020] A triaxial MEMS accelerometer 12 is used to capture micro-vibration signals of the support structure 7 in the frequency band of 0.1Hz to 500Hz.

[0021] Visual measuring instrument 13: It adopts the RTS-112 high-precision measuring robot, together with 32 highly reflective optical targets deployed on the crown beam, to achieve sub-millimeter level coordinate verification.

[0022] Stepped-frequency continuous wave ground-penetrating radar 14: The antenna center frequency is set to 500MHz-1.5GHz, used for non-contact detection of soil cavities on the back side of the support structure 7.

[0023] The distributed fiber optic sensor 11 uses a dedicated high-temperature and high-pressure resistant strain gauge cable. The sensor is sheathed in a flexible HDPE spiral protective tube with a ring stiffness ≥10KN / ㎡ to prevent signal attenuation due to radial pressure. Furthermore, the distributed fiber optic sensor 11 is secured to the main reinforcement of the support pile's steel cage using U-shaped metal clips 10 with a fixed spacing of 0.5m. This allows it to withstand concrete with a compressive strength of 27.6MPa, ensuring the centerline deflection angle of the distributed fiber optic sensor 11 is controlled within 0.2 degrees.

[0024] The fixing spacing L of the U-shaped metal clip 10 must meet the deflection constraint condition. Under the action of the concrete fluid dynamic pressure q, the initial deflection angle of the distributed fiber optic sensor 11... The calculation formula is: Wherein, EI is the bending stiffness of the composite structure of the distributed optical fiber sensor 11. Setting L=0.5m can ensure that 0≤0.2 when the injection speed v=4.5ft / hr, thereby minimizing the initial geometric error of displacement inversion.

[0025] This solves the geometric transformation error caused by the injection displacement of the distributed optical fiber sensor 11 in the prior art, and ensures the reliability of the initial input of the displacement inversion algorithm; The fixed spacing of 0.5m is calculated based on the lateral dynamic pressure wave during 27.6MPa concrete pouring. This spacing ensures that the pile deflection W under the fluid impact force F of the distributed fiber optic sensor 11 meets the following conditions. This ensures that the initial geometric deflection angle generated by the distributed fiber optic sensor 11 is less than 0.2 degrees after the concrete is poured, thus avoiding the destruction of displacement inversion accuracy by systematic errors in the initial coordinate system.

[0026] Edge computing gateway 2, whose input is electrically connected to sensing terminal cluster 1, is used to perform wavelet threshold hierarchical noise reduction and time axis alignment processing on multi-source monitoring data, and output a standard spatiotemporal dataset. The hardware architecture of edge computing gateway 2 is based on the RK3588 processor and integrates an NPU acceleration unit with 6 TOPS computing power. For wavelet threshold hierarchical noise reduction at construction sites, such as high-frequency random vibration noise above 50Hz caused by excavators, the system selects the db4 wavelet basis function for 5-level multi-scale decomposition. The calculation formula is as follows: , in, The high-frequency wavelet coefficients of the j-th level decomposition reflect the amplitude of a specific frequency component, and λ is the standard deviation of the background noise at the construction site. The risk threshold, dynamically generated with the sampling window length N, is used to determine the filtering boundary of noise in the wavelet coefficients. Its expression is: The system calculates the standard deviation of background noise in real time during non-operational periods. , The standard deviation of background noise extracted during non-working periods at the support construction site is given by N, which is the sampling window length, typically 1024.

[0027] For example, when When the value is 0.05 and N=512 sampling points, the risk threshold is calculated. The components of wavelet coefficients less than 0.176 are forced to zero, and the reconstructed signal is used as a pure input.

[0028] Background noise Extracted during non-operational periods, it can represent the inherent white noise of the distributed fiber optic sensor 11 and the ground micro-vibrations. With its global adaptive adjustment, the system can accurately lock the centimeter-level displacement of the support piles even under the high-frequency vibration of construction machinery operations, typically >50Hz, thus solving the false alarm problem caused by traditional fixed thresholds.

[0029] Edge computing gateway 2 utilizes cubic spline interpolation formula The system aligns the time axis of multi-source monitoring data. For the visual monitoring cycle of 10 minutes and the water level gauge cycle of 1 hour, which have inconsistent sampling frequencies, the system uses cubic spline interpolation at the edge to fill in missing points. The cubic spline interpolation formula is as follows: satisfy: , in, To monitor time points, To satisfy the polynomial coefficients of the second derivative continuity constraint at the nodes, at the discrete monitoring points The coefficients are determined by solving a system of linear equations. , , , This ensures strict synchronization of multi-source data at millisecond-level steps, thereby solving the problem of time sparsity of heterogeneous data before it is input into the CNN.

[0030] The CNN layer configuration is as follows: Input layer: A feature map with dimensions of 24×100, representing data from 24 monitoring items over the past 100 steps.

[0031] Convolutional layer: Uses 3 3×3 convolutional kernels with a stride of 1.

[0032] Activation function: LeakyReLU, negative slope =0.01, to prevent neuronal dead zones.

[0033] This structure can automatically extract spatial correlations between multi-source data, such as identifying the coupling characteristics between groundwater drawdown and pile bending moment changes.

[0034] The multi-source data fusion module 3, based on the unified coordinate benchmark of the 3D Building Information Model (BIM) and the Geographic Information System (GIS), uses the coordinate transformation matrix M to establish a standard spatiotemporal dataset and uses a 4×4 homogeneous transformation matrix to establish a nonlinear spatial projection mapping relationship between heterogeneous features.

[0035] The spatial mapping relationship between them is determined, and a convolutional neural network (CNN) is used for feature extraction to output coupled feature vectors. The coordinate transformation matrix M is used to transform the local coordinates of sensing terminal cluster 1. Mapped to the global unified coordinate P of the 3D Building Information (BIM) model 8. Its calculation satisfies the homogeneous transformation relationship: , in, Let represent the local homogeneous coordinate vector collected by sensing terminal cluster 1. Where, These are the local spatial coordinates of the sensor's original output. The 3x3 rotation matrix represents the cosine of the angle between the axis of the distributed fiber optic sensor 11 and the global axis of the BIM, which is obtained by solving the global coordinates of the four control corner points of the foundation pit. This achieves "virtual-real binding" and enables the system to identify the spatial mechanical causality between "increased pile displacement" and "settlement of surrounding road surface". This represents the translation vector of the mounting point of the 3x1 distributed fiber optic sensor 11 in the global coordinate system. The symbol is collectively referred to as P, where P represents the globally unified coordinates corresponding to the 3D Building Information (BIM) model 8. :express The zero vector transpose is [0,0,0], which is used to fill the rotation matrix and translation vector into a 4×4 homogeneous matrix; 1: represents the homogeneous coordinate components, used to assist matrix operations and ensure that the translation terms of spatial transformations take effect; T: Mathematical transpose symbol, indicating that a row vector is converted into a column vector, or a column vector is converted into a row vector.

[0036] Quantifying the technical effect of the coordinate transformation matrix M: Through a 4×4 homogeneous transformation, the pixel coordinates of the vision measuring instrument 13, the strain coordinates of the distributed fiber optic sensor 11, and the tilt coordinates of the MEMS accelerometer are unified into the 3D Building Information (BIM) model 8. This enables the prediction model to identify the abrupt stress change at its mechanically relevant point B when the displacement of point A of the pile increases. This spatial coupling characteristic is the underlying support that enables GA-LSTM to achieve sub-millimeter-level prediction.

[0037] In implementation, the system establishes a digital twin of the support structure 7 using a 3D Building Information Model (BIM) model 8, mapping the coordinates of discrete distributed fiber optic sensors 11 to a unified space; the CNN performs cross-modal feature extraction using 3×3 convolutional kernels, and outputs coupled feature vectors after processing with the LeakyReLU activation function. The coupled feature vectors are then analyzed using the Pearson correlation coefficient. Calculate the correlation contribution weight coefficient The formula is: , : The Pearson correlation coefficient between the i-th monitoring dimension, such as water pressure, displacement, internal force, and deformation target item y; k: The total number of dimensions of the current multi-source sensing terminal cluster; : The contribution weighting factor assigned to the feature extraction layer of a convolutional neural network (CNN).

[0038] The system calculates the Pearson correlation coefficient between each monitoring item and the deformation target item. .For example: Deep lateral displacement correlation =0.85; Support axial force correlation =0.72; Groundwater level correlation =0.40.

[0039] Weighting coefficient calculation: .

[0040] This allows for the dynamic increase of the weights of key mechanical factors, thereby enhancing the sensitivity of the prediction model to extreme risks such as collapse.

[0041] Collaborative prediction module 4 is used to input the coupled feature vector into the GA-LSTM prediction model. The GA-LSTM algorithm is used to optimize the hyperparameters of the LSTM. Population size: 50 individuals.

[0042] Optimization space: Number of neurons in hidden layers Learning rate ,Note: These are standard engineering reference values ​​and can be adjusted according to the actual algorithm.

[0043] By calling the cloud computing cluster to output the deformation trend prediction value of support structure 7, the cloud cluster uses servers with H800 computing power to perform large-scale tensor operations. The GA-LSTM prediction model uses a genetic algorithm to globally optimize the number of hidden layer neurons and the learning rate of LSTM. The GA-LSTM forget gate... The calculation formula is: ,in, As input features, This is the hidden state from the previous moment. It is the Sigmoid activation function. This is the weight matrix. For bias terms, This means that the hidden state of the previous time step is concatenated with the current input features as a vector; because the deformation of the support structure 7 has long-term creep, the forget gate of GA-LSTM is responsible for identifying and retaining the historical long-term displacement trend, while discarding the signal fluctuations caused by temporary vehicle loads.

[0044] The collaborative prediction module 4 first initializes the number of neurons in the hidden layer based on the historical creep characteristic intervals of the support structure 7. and learning rate The optimal population is selected; then the root mean square error (RMSE) between the model output value and the measured deformation value under each set of hyperparameters is calculated iteratively, and the fitness function F is constructed with the reciprocal of the RMSE. The genetic algorithm is used to lock the hyperparameter combination that maximizes F, so as to improve the accuracy of capturing the long-term displacement trend of the support structure.

[0045] The system adopts a hierarchical architecture of "edge prediction and cloud training". Edge computing gateway 2 is based on the RK3588 chip, and its built-in NPU performs lightweight INT8 quantized GA-LSTM inference with an inference latency of ≤50ms. It communicates with the MCU (Microcontroller Unit) of the sensing terminal cluster 1 through its built-in clock management module CCU (Clock Control Unit). The cloud computing cluster, such as the H800 computing power group, retrains the weights every 24 hours based on the cumulative data of the whole day and synchronizes them to the edge side through the differential upgrade protocol. The feedback control loop 5 rewrites the values ​​of clock divider register 62 at address 0x01A2 and servo pump flow at address 0x01B0 through the Modbus-TCP protocol to achieve nanosecond-level instruction issuance.

[0046] The actuator 6 includes an intelligent pressure relief pump unit 61 and a clock divider register 62.

[0047] Feedback control loop 5 is used to generate frequency adjustment instructions and servo drive instructions based on the deformation trend prediction value. The frequency adjustment instructions modify the clock divider register 62 value in the monitoring terminal hardware of the sensing terminal cluster 1 to dynamically adjust the sampling frequency. The servo drive instructions drive the intelligent pressure relief pump group 61 to adjust the pore water pressure around the support structure 7. Feedback control loop 5 uses a PID control algorithm to adjust the pressure relief pump flow rate to reduce the pore water pressure. To forcibly increase the effective stress of the soil It follows Terzaghi's effective stress principle and satisfies ,in, For the effective stress of the soil, denoted as the total soil stress, determined by the soil's self-weight and ground load; u is the pore water pressure, influenced by the groundwater level and permeability coefficient. The system performs active intervention based on Terzaghi's effective stress principle.

[0048] When the predicted deformation y exceeds the first-level warning value, the feedback control loop 5 sends a servo drive command to the frequency converter at address 0x01B0 via Modbus-TCP. The intelligent pressure relief pump unit 61 then starts working, precisely adjusting the drainage flow rate according to the pressure drop gradient set by the PID controller. Lowering the groundwater level reduces pore water pressure (u), thereby increasing the effective stress and enhancing the interlocking force between soil particles while maintaining the total soil stress. This improves the overall stability of the soil behind the pile, significantly increasing the soil's shear strength and suppressing further lateral displacement of the support pile, in accordance with the Mohr-Coulomb criterion.

[0049] The frequency adjustment instruction is used to modify the value of the clock divider register 62 in the monitoring terminal hardware of the sensing terminal cluster 1 to adjust the sampling frequency; the hardware clock division coefficient K in the frequency adjustment instruction and the deformation trend prediction value y satisfy a mapping relationship: , in, The hardware reference frequency division coefficient, The response sensitivity coefficient, The safe deformation threshold is defined as the value of the predicted deformation trend y approaching the safe threshold. For example, when 0.002h = 40mm for a first-level foundation pit, h represents the depth of the foundation pit, and the feedback control loop 5 calculates the hardware clock division coefficient K.

[0050] in, Taking 3600 indicates sampling once per hour, and the response sensitivity... Take 0.8. When the deformation trend prediction value increases from the safe value of 30mm to 35mm, K≈65.9 is calculated. At this time, the "frequency adjustment instruction" generated by the feedback control loop 5 is specifically a register rewrite instruction for the hardware clock tree in the MCU. Through the Modbus-TCP protocol, the calculated frequency division coefficient K is written into the 16-bit clock frequency division register 62 with the address 0x01A2 of the sensing terminal cluster, thereby changing the frequency division ratio of the crystal oscillator output signal, realizing physical-level adjustment of the sampling frequency, and automatically switching the sampling frequency to about once per minute. At the same time, the system generates a servo drive instruction to drive the intelligent pressure relief pump group 61 to perform drainage operations at a step pressure Δu=5kpa.

[0051] PID parameter settings: proportional coefficient Integral time Differential time .

[0052] Modbus registers: Frequency control instructions are written to register 0x01A2, and intelligent pressure relief pump group 61 drive instructions are written to register 0x01B0.

[0053] Hardware collaboration: The edge gateway is based on the RK3588 chip, and its NPU inference latency is controlled within 50ms, ensuring real-time response from deformation capture to pressure relief action.

[0054] Feedback control loop 5 modifies the value of register 0x01A2 via the Modbus-TCP protocol to adjust the hardware clock division ratio. When the predicted value y approaches the safety threshold... At that time, the 'frequency adjustment command' generated by the system will trigger a hardware-level sampling rate transition, thereby achieving physical-level adjustment of the monitoring frequency by changing the frequency division ratio.

[0055] The intelligent pressure relief pump unit 61 uses a variable frequency centrifugal pump driven by a three-phase AC motor. The system is equipped with hardware featuring an industrial-grade real-time transmission protocol and is powered by a three-phase 380V variable frequency power supply. The intelligent pressure relief pump unit 61 is fixedly installed on the rigid support frame of the dewatering well behind the support piles, and its discharge outlet is connected to the sedimentation tank at the top of the foundation pit via a pressure-bearing hose. The sensing terminal cluster 1 is physically connected to the gateway's RS485 interface via shielded twisted-pair cable. The edge computing gateway 2 achieves low-latency interconnection with the cloud computing cluster through a 5G communication module, with a latency of less than 20ms. All deformation trend prediction values ​​are obtained based on real-time sensor feedback data calculation, and the system output is a dynamic closed-loop calculation based on real-time sensor feedback.

[0056] The hardware clock division factor K in the frequency adjustment instruction is determined by the following formula and, after being rounded, written into the clock divider register 62: , in, The hardware reference frequency division coefficient; This is the response sensitivity coefficient; .

[0057] The implementation principle of a multi-source data fusion and collaborative prediction system for support structure deformation in this embodiment of the invention is as follows: First, the distributed fiber optic sensors 11 in the sensing terminal cluster 1 of this system are fitted with spiral protective tubes, and then fastened to the reinforcing cage of the support piles using U-shaped metal clips 10 with a fixed spacing of 0.5m, ensuring that the centerline deflection angle is controlled under the impact of concrete pouring. Within this range, the reliability of displacement inversion is guaranteed from the physical source; Subsequently, in response to the high-frequency random vibration interference generated by excavators and other machinery at the construction site and the time axis sparsity caused by inconsistent sampling frequencies of heterogeneous monitoring equipment, the edge computing gateway 2 performs wavelet threshold layered noise reduction and cubic spline interpolation processing on the raw data to output a strictly synchronized standard spatiotemporal dataset. In order to solve the technical defects of heterogeneous sensor data lacking unified spatial coordinate mapping and difficulty in identifying the physical relationship between "water level-stress-displacement", the multi-source data fusion module 3 uses a 4×4 homogeneous coordinate transformation matrix M to uniformly map the data to the global coordinate reference of the three-dimensional building information BIM model 8 and GIS, and uses a convolutional neural network CNN to extract spatial coupling feature vectors. Furthermore, the collaborative prediction module 4 inputs the coupled features into the GA-LSTM prediction model, uses a genetic algorithm to optimize the hyperparameters and combines the forget gate mechanism to identify long-term creep trends, and achieves sub-millimeter level deformation trend prediction output, effectively solving the problem of insufficient early warning lead time in traditional early warning systems. Finally, addressing the bottleneck of the existing monitoring system's fixed sampling frequency and lack of active physical and mechanical intervention, the feedback control loop 5 dynamically modifies the value of the clock divider register 62 in the actuator 6 based on the predicted value to achieve adaptive adjustment of the sampling frequency. Furthermore, based on Terzaghi's effective stress principle, it uses the Modbus-TCP protocol to drive the intelligent pressure relief pump group 61 to adjust the pore water pressure. By forcibly increasing the effective stress of the soil, it actively suppresses the deformation of the support structure 7, thereby significantly reducing the risk of foundation pit instability.

[0058] As a safety redundancy guarantee, when the system detects an unexpected sudden drop in outlet pressure (below 0.01MPa) or an abnormal command verification, the system automatically switches to the safety default mode, locks the register state through a hardware reset command, and performs shutdown protection to prevent the risk of misadjustment or secondary instability caused by physical pipeline failure or communication abnormality.

[0059] The above are merely optional embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-source data fusion and collaborative prediction system for support structure deformation, characterized in that, include: The sensing terminal cluster (1) is used to acquire multi-source monitoring data reflecting the physical state of the support structure (7); wherein, the sensing terminal cluster (1) includes a distributed fiber optic sensor (11), a three-axis MEMS accelerometer (12), a vision measuring instrument (13), and a step-frequency continuous wave ground-penetrating radar (14). The edge computing gateway (2) is electrically connected to the sensing terminal cluster (1) for performing wavelet threshold hierarchical noise reduction and time axis alignment processing on the multi-source monitoring data and outputting a standard spatiotemporal dataset. The multi-source data fusion module (3) is based on the unified coordinate benchmark of the three-dimensional building information BIM model (8) and the geographic information system GIS. It uses the coordinate transformation matrix M to establish the spatial mapping relationship between different dimensions of data in the standard spatiotemporal dataset, and uses the convolutional neural network CNN to extract features and output coupled feature vectors. The collaborative prediction module (4) is used to input the coupled feature vector into the GA-LSTM prediction model and output the deformation trend prediction value of the support structure (7) by calling the cloud computing cluster. Feedback control loop (5) is used to generate frequency adjustment commands and servo drive commands based on the deformation trend prediction value; The actuator (6) includes an intelligent pressure relief pump assembly (61) and a clock divider register (62). The frequency adjustment instruction is used to modify the value of the clock divider register (62) of the monitoring terminal hardware in the sensing terminal cluster (1) to adjust the sampling frequency; the servo drive instruction drives the intelligent pressure relief pump group (61) to adjust the pore water pressure around the support structure (7) through the Modbus-TCP protocol.

2. The multi-source data fusion and collaborative prediction system for support structure deformation according to claim 1, characterized in that: The formula for wavelet threshold hierarchical noise reduction performed by the edge computing gateway (2) is as follows: ， in, The high-frequency wavelet coefficients of the j-th level decomposition reflect the amplitude of a specific frequency component. Based on the standard deviation of background noise at the construction site The risk threshold, dynamically generated with a sampling window length N, is expressed as follows: , The standard deviation of background noise extracted during non-working periods at the support construction site is given by N, where N is the sampling window length.

3. The multi-source data fusion and collaborative prediction system for support structure deformation according to claim 1, characterized in that: The edge computing gateway (2) utilizes the cubic spline interpolation formula. Time axis alignment is performed on multi-source monitoring data, and the cubic spline interpolation formula is used. satisfy: , in, To monitor time points, The polynomial coefficients are used to satisfy the second derivative continuity constraint at the nodes.

4. The multi-source data fusion and collaborative prediction system for support structure deformation according to claim 1, characterized in that: The coordinate transformation matrix M is a 4×4 homogeneous transformation matrix, and its mapping relationship is expressed as follows: ， in, Let represent the local homogeneous coordinate vector collected by the sensing terminal cluster (1). Wherein, These are the local spatial coordinates of the sensor's original output. The 3x3 rotation matrix represents the cosine of the angle between the axis of the distributed fiber optic sensor (11) and the global axis of the BIM. Let the translation vector of the mounting point of the 3x1 distributed fiber optic sensor (11) in the global coordinate system be given. The symbol is collectively referred to as P, which is the global unified coordinate corresponding to the three-dimensional building information BIM model (8), and is used to fill the rotation matrix and translation vector into a 4×4 homogeneous matrix.

5. The multi-source data fusion and collaborative prediction system for support structure deformation according to claim 1, characterized in that: The CNN performs cross-modal feature extraction using 3×3 convolutional kernels and outputs coupled feature vectors after processing with the LeakyReLU activation function.

6. The multi-source data fusion and collaborative prediction system for support structure deformation according to claim 1, characterized in that: The GA-LSTM prediction model utilizes a genetic algorithm to globally optimize the number of hidden layer neurons and the learning rate of the LSTM. The LSTM forget gate... The calculation formula is: , in, For input features, This is the hidden state from the previous moment. It is the Sigmoid activation function. This is the weight matrix. For bias terms, This indicates that the hidden state from the previous time step is concatenated with the current input features as a vector.

7. The multi-source data fusion and collaborative prediction system for support structure deformation according to claim 1, characterized in that: The coupling feature vector is obtained through the Pearson correlation coefficient. Calculate the correlation contribution weight coefficient The formula is: , : The Pearson correlation coefficient between the i-th monitoring dimension (such as water pressure, displacement, internal force) and the deformation target term y; k: The total number of dimensions of the current multi-source sensing terminal cluster (1); : The contribution weighting factor assigned to the feature extraction layer of a convolutional neural network (CNN).

8. The multi-source data fusion and collaborative prediction system for support structure deformation according to claim 1, characterized in that: In the frequency adjustment command, the hardware clock division coefficient K and the deformation trend prediction value y satisfy a mapping relationship: ,in, The hardware reference frequency division coefficient, The response sensitivity coefficient, y: Safety deformation threshold; y: Deformation trend prediction value.

9. The multi-source data fusion and collaborative prediction system for support structure deformation according to claim 1, characterized in that: The feedback control loop (5) uses a PID control algorithm to adjust the flow rate of the pressure relief pump in order to reduce the pore water pressure. To forcibly increase the effective stress of the soil Its mechanical mechanism satisfies ,in, For the effective stress of the soil, Let be the total soil stress, and u be the pore water pressure.

10. The multi-source data fusion and collaborative prediction system for support structure deformation according to claim 1, characterized in that: The distributed optical fiber sensor (11) is covered with a spiral protective tube and is fastened to the main reinforcement of the support pile steel cage by a U-shaped metal buckle (10).

Images