Highway pavement structure settlement deformation monitoring identification method and system based on digital twinning

Abstract

The invention provides a highway pavement structure settlement deformation monitoring and identification method and system based on digital twinning, and belongs to the field of fiber bragg grating sensing structure health monitoring. According to the method, the vibration mode characteristics of the pavement structure are obtained through the vibration pickup, the modal information is calculated by matching numerical values, and the pavement deformation is reconstructed by adopting a modal vibration mode matrix and monitoring strain data; sensor layout optimization is carried out based on an actual test road section, theoretical calculation deviation existing under the road surface strong constraint condition is corrected by means of a neural network model, and the road surface deformation type disease development trend is predicted. The method can effectively monitor pavement structure settlement evolution development under unfavorable geological conditions, early warns sudden collapse risks in advance, and provides an important basis for intelligent monitoring of road settlement. A digital model is built based on the multi-source data, and real-time diagnosis of the state of the monitored road section and early warning and treatment of settlement risks are achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to a kind of road surface structure settlement deformation monitoring method in asphalt pavement, belong to optical fiber grating sensing structure health monitoring field. BACKGROUND

[0002] With the construction and improvement of infrastructure in China, the national highway mileage has broken through 5.5 million kilometers. Road health monitoring field has become an important concern for current road safety operation. Precise monitoring of road disease evolution behavior can effectively protect personal and property safety and guide the development of road structure maintenance and repair program. For the deformation evolution of pavement structure under adverse geological conditions such as seepage soft soil subgrade and karst cave, it is necessary to build a health monitoring system to prevent sudden collapse risk.

[0003] Currently, there is still a lack of long-term settlement deformation monitoring means for key areas of pavement. Displacement monitoring sensors such as tensioned rope displacement meter, vibrating wire displacement meter and single-point settlement meter cannot meet the requirements of embedded pavement structure due to their large size. External monitoring insar technology, digital point cloud settlement reconstruction and unmanned aerial vehicle remote sensing technology cannot meet the real-time monitoring requirements of pavement deformation.

[0004] Digital twin is a simulation process that fully utilizes physical models, sensors, operation history and other data, integrates multi-disciplinary, multi-physical, multi-scale and multi-probability, and completes mapping in virtual space, so as to reflect the whole life cycle process of the corresponding entity equipment. Digital twin is a concept beyond reality, which can be regarded as a digital mapping system of one or more important and interdependent equipment systems. Digital twin technology can be applied in many fields, including product design, product manufacturing, medical analysis and engineering construction. SUMMARY

[0005] The technical problem to be solved by the present application is that the existing pavement structure subsidence disease embedded monitoring means is lacking, and the timeliness is insufficient. A highway pavement structure settlement deformation monitoring and identification method based on digital twin is proposed, which obtains pavement structure vibration mode characteristics through a vibration pickup, matches numerical calculation modal information, and reconstructs pavement deformation using modal vibration mode matrix and monitoring strain data. In view of the deviation between theoretical calculation and actual pavement, a neural network model is used for data correction and prediction of pavement deformation evolution.

[0006] In order to solve the above technical problems, the present application adopts the following technical means:

[0007] Firstly, the present application proposes a highway pavement structure settlement deformation monitoring and identification method based on digital twin, which includes the following steps:

[0008] S1, collect the vibration response signal of the road surface under the passing of the vehicle by arranging a vibration pickup on the settlement monitoring section; perform feature extraction and modeling on the vibration response signal to solve the state matrix of the road surface structure, and then extract the measured natural frequency and vibration mode of the road surface structure;

[0009] S2, select the main mode of the road surface in the finite element model based on the measured natural frequency obtained in step S1; embed an optical fiber grating in the to-be-measured road surface along the driving direction to obtain strain data, combine the vibration mode obtained in step S1, and establish a strain-displacement conversion relationship by a modal superposition method to identify the displacement settlement distribution of the monitoring section;

[0010] S3, arrange an optical fiber grating sensing network on the test section, and synchronously collect the measured settlement values of multiple representative points and the strain response data of all grating measuring points; taking the measured settlement value as the reference, the strain-displacement conversion relationship established in step S2 is used to optimize the grating measuring point arrangement scheme by a genetic algorithm, and an optimal measuring point arrangement scheme with the least number of measuring points under the premise of ensuring identification accuracy is obtained;

[0011] S4, construct a numerical simulation model containing the optimal measuring point arrangement scheme, generate a low-fidelity simulation data set, and combine the high-fidelity data set obtained in step S3 to train a deep belief network; the modal superposition method in step S2 is modified by adversarial training and loss function optimization to obtain a strain-displacement mapping model;

[0012] S5, introduce multi-source time series monitoring data on the section arranged with the optimal measuring point arrangement scheme; input the displacement sequence output by the model modified in step S4 and the multi-source time series monitoring data into a long short-term memory network to predict the future settlement of the road surface; based on the prediction result, the settlement gradient and trend are identified by a graph attention network to realize the regional accurate positioning and early warning of the road surface deformation risk.

[0013] As a preferred technical scheme of the present application: in step S1, the time domain operation modal identification method of the road surface vibration system under the excitation of the passing vehicle is as follows:

[0014] ① Road vibration signal acquisition: for the settlement monitoring section, arrange 941B vibration pickups at intervals of 20cm along the edge of each lane, the signal sampling frequency is 200HZ, and the test vehicle passing speed is 60km / h. By adjusting the parameters of the vibration pickup micro-dial switch and the magnifying lens, the acceleration, speed and displacement response data under the passing of the vehicle are obtained.

[0015] ② Vibration signal feature extraction: The time-series vibration response measured by the vibration pickup is expressed in Hankel matrix form, and its Toeplitz matrix is ​​calculated to extract the dynamic features from the original vibration signal. Environmental excitation noise interference terms such as temperature, humidity, and load changes in the actual road surface are considered. and the systematic measurement error noise of the vibration pickup itself Singular value decomposition is performed on the Hankel matrix to remove noise interference and extract the system's natural vibration frequencies.

[0016] ③ Vibration signal modeling: Based on the principle of linear time invariance of road vibration, the collected vibration signals are discretized in time. The road vibration signal at time t, which includes environmental factors and unknown road excitation noise, is used to predict the road vibration at the next time step (time t+1), thus establishing the main excitation signal of the road surface decoupled from noise interference. Vibration response signal The association is shown in equation (1):

[0017] (1)

[0018] In the formula, A is a discrete state vector containing displacement and velocity terms; A is the state matrix. For road surface environmental noise; C is the output vector (vibration signal); C is the output matrix; The noise is measured for the system. The noise signal follows a zero-mean Gaussian distribution.

[0019] ④ Vibration signal matrix acquisition: The Toeplitz matrix acquired by the vibration pickup is decomposed into a state matrix A and an output matrix C. The key matrices A and C for road vibration signal modeling are obtained by fitting and solving the matrix equations using the least squares method, as shown in equation (2):

[0020] (2)

[0021] In the formula, Let G be the Toeplitz matrix, G be the process noise (including road surface environmental noise and measurement noise), and i be the time step. In road vibration signal processing, 1 / 10 of the sampling signal length is used, i.e., 20.

[0022] ⑤ Road surface modal signal extraction: Solve the characteristic equation of the road surface state matrix A to obtain the natural frequencies of the road surface structure. and vibration modes See formula (3). The vibration modes are verified based on the output matrix C. For accuracy, Formula (4) is the error test standard based on the Euclidean norm.

[0023] (3)

[0024] (4)

[0025] wherein, is the time series acceleration data measured by the vibration pickup, and m is the number of the vibration pickup arranged on the road surface. The vibration modal error limit is set to be not less than 95% of the corresponding accuracy of the acceleration signal.

[0026] As a preferred technical solution of the present application: in step S2, the grating finite point deformation sensing technology based on modal superposition method is proposed as follows:

[0027] ① Fiber grating basic parameters and embedding requirements: the multi-point grating device selects a 3mm diameter thick braided mesh armored structure to ensure survival in the complex stress environment inside the road surface. The grating deformation sensing length is 10mm, the spacing between measuring points is 40cm, and the number of measuring points of a single grating is 20. The matching grating demodulator equipment identifies the wavelength range of 1500~1600nm, the demodulation rate is 100hz, the wavelength resolution is 1pm, and the signal dynamic identification range is 50db. The slotting cutter is used to cut along the driving direction and the transverse direction, the slot width is not greater than 1cm, and the grating is embedded at a depth of 3cm at the top of the base structure. The grating is positioned horizontally at both ends by rivets and straps, and then the surface is embedded with asphalt mortar to ensure that the grating strain transmission efficiency is not less than 98% level. After the surface is wrapped stably, fine-grained asphalt mixture or pouring sealant is used to backfill to the original road surface.

[0028] ② Selection of main modal of road surface: a conventional road surface structure finite element model is established, linear perturbation analysis based on dynamic response equation is carried out, and the first fifty order modal shapes are obtained by using Lanczos eigenvalue solver. The measured natural frequency of step S1 is compared with the numerical solution, and the corresponding modal order in the numerical model is selected from low to high. And taking the effective mass contribution rate in the depth direction of the road surface not less than 60% as the selection basis of the representative modal order n under the constraint condition of the road surface.

[0029] ③ Modal superposition method vibration mode matrix: the displacement modal shape under the corresponding modal order of linear perturbation analysis is extracted , wherein N is the number of monitoring road section settlement points, and n is the modal shape order. The strain modal shape under the corresponding modal order of linear perturbation analysis is extracted , wherein M is the number of grating measuring points of the monitoring road section, and n is the modal shape order.

[0030] ④ Road surface deformation identification: based on the modal superposition method, the strain-displacement conversion relationship of the road surface structure can be solved by formula (5):

[0031] (5)

[0032] In the formula, To monitor displacement at key points of interest in road settlement; To monitor the strain data of the grating measuring points on the road section.

[0033] As a preferred embodiment of the present invention: In step S3, the fiber optic grating measurement point optimization layout scheme based on the genetic algorithm is as follows:

[0034] ① Fiber Optic Grating (FGP) Sensor Network Deployment for the Test Section: A 10-meter-long dual-lane test area was selected. In the cross-sectional direction, three FGP strings with 20 measuring points (as shown in S2) were deployed at 5-meter intervals. In the longitudinal direction, three FGP strings with 20 measuring points were deployed along the centerline of the driving lanes. See the schematic diagram of the FGP deployment. Figure 2 .

[0035] ② Pavement Strain and Settlement Data Acquisition: Dongfeng EQ140 test vehicles, conforming to JTG 3450-2019 "Specifications for Field Testing of Highway Subgrade and Pavement," were parked at different locations along the test section. Settlement values ​​at six selected representative points were measured sequentially using a Beckman beam, while strain response data at 120 grating measuring points were recorded simultaneously. The measured data will be used to optimize the measuring point layout using a genetic algorithm.

[0036] ③ Measurement point distribution generation: The data from the above 120 strain measurement points are sequentially encoded as xy in the longitudinal and transverse directions, where x represents the grating string number and y represents the measurement point number. Under the condition of a total of 60 measurement points, the individual measurement point distributions are obtained through the random number generation function. Multiple initial layout methods can be formed using this method.

[0037] ④ Genetic algorithm measurement point optimization layout: Calculate the settlement value of the above 6 representative points under each individual using the modal superposition method described in S2, and calculate the population fitness index f according to formula (6):

[0038] (6)

[0039] In the formula, The settlement value is calculated for the i-th representative point in individual x. The value represents the measured settlement value at a representative point; the value of 1000 is the sensitivity adjustment coefficient.

[0040] All initial individuals are placed into the parent pool. Based on the fitness index f, two parents are selected sequentially from the parent pool using the roulette wheel method. The selection probability is calculated using the roulette wheel method according to formula (7). Using the fitness of the individual as the hybridization and mutation probability, each gene is sequentially determined to perform hybridization or mutation operations to form a new individual. Through repeated iterations, when the average relative error of 5 consecutive generations is less than 20%, the optimal individual can be considered to have been obtained. In subsequent road settlement monitoring, the above measurement point form can be adopted to reduce the number of grating string measurement points while ensuring that the road surface deformation identification accuracy is not less than 80%.

[0041] (7)

[0042] In the formula, Let be the probability of selecting the a-th individual; n is the total number of individuals.

[0043] As a preferred technical solution of the present invention: the implementation steps of the algorithm for improving the accuracy of road deformation perception by fusing multi-fidelity data under strong road constraints in step S4 are as follows:

[0044] ① Construction of High- and Low-Fidelity Datasets: A numerical simulation model consistent with the test road section described in S3 was established. Through multi-condition simulation calculations, strain data of optimized grating measuring points and displacement and settlement values ​​of six representative points were collected at multiple times under different driving speeds, tire pressures, axle types, and traffic positions. Simultaneously, static load strain and displacement data of corresponding points on the road surface under different tire pressures, axle types, and stationary positions were collected to construct a large-capacity low-fidelity simulation dataset. The strain and displacement data of the measured road section in step S3 were collected to construct a low-capacity high-fidelity real dataset.

[0045] ② Low-fidelity network model construction: Using strain data, calculated displacement values, and simulated displacement values ​​as input data, a deep belief network (DBN) is pre-trained layer by layer in an unsupervised manner using a simulation dataset to obtain the general hierarchical features of the modal superposition method. The training process includes the pre-training stage where a restricted Boltzmann machine (RBM) based on unsupervised learning is used to build the network layer by layer, and the fine-tuning stage where a backpropagation backpropagation (BP) neural network is established between the last two hidden layers of the DBN to adjust the weights of the entire network in a supervised learning manner, so as to realize the calculation and correction of simulation data of road surface under strong constraints under multiple working conditions. The improved modal superposition method is formula (8):

[0046] (8)

[0047] In the formula, The correction term is calculated for the modal superposition method of road surface with strong constraints based on DBN network.

[0048] ③ Network Model Correction Based on High-Fidelity Dataset: The measured real strain, calculated displacement values, and measured displacement values ​​of representative points on the road surface obtained from step S3 are used for fine-tuning of the top-level task head of the existing DBN network. This allows the DBN network to capture the characteristics of real road surface data while ensuring the general features of the underlying modal calculation. Adversarial training is performed on both the simulation and real datasets. A discriminator in the generated adversarial network is used to distinguish the different distributions of high- and low-fidelity data samples. A feature covariance alignment method is employed to make the feature distributions of high- and low-fidelity data more consistent, achieving indistinguishability between the features of the two data sources.

[0049] ④ Optimization of loss function for multi-fidelity regression networks: In the step of modifying the network model by introducing high-fidelity datasets, a domain adaptation loss function term based on Wasserstein distance is added. This is used to reduce the impact of differences in the distribution of high-fidelity and low-fidelity data. It involves data sample supervision. Weights are adjusted by using larger weights for real data and smaller weights for simulation data. An L2 norm starting point regularization term is also introduced to limit weight deviations in general feature extraction. Physical consistency is employed. The loss function term characterizes the dynamic properties of the data.

[0050] As a preferred technical solution of the present invention: In step S5, the implementation steps of the road surface deformation prediction model based on physical term constraints and time-series data of long short-term memory network are as follows:

[0051] ① Acquisition of Multi-Source Time-Series Monitoring Dataset: Based on the deployment of multi-point fiber optic gratings on the S3 road section, temperature, pressure, and subgrade humidity sensors were introduced for each structural layer of the pavement. Temperature and subgrade humidity baselines for each structural layer were obtained using wavelet threshold denoising; pressure response data for each structural layer were processed using SG smoothing filtering. Material parameters and traffic load spectra of the actual road section were acquired, and the mechanical response of the pavement structure was obtained using finite element numerical calculation.

[0052] ② Long Short-Term Memory Network Pavement Deformation Prediction Model: The hourly time-series displacement values, temperature and pressure of each structural layer, and soil moisture data corrected in step S4 are used as inputs to the memory network, and the settlement amount in the subsequent 24 hours is used as the output to build a pavement deformation prediction model.

[0053] ③ Optimization of pavement deformation prediction model based on physical constraints: In the long short-term memory network, the difference between the measured pressure data and the numerically calculated pressure at the structural layer points, and the difference between the measured temperature change and the numerically calculated temperature evolution of the structural layer are used as physical driving objective function terms to guide the update of prediction model parameters, ensuring that the prediction results not only conform to the time evolution law of monitoring data, but also meet the mechanical behavior characteristics of pavement structure, thereby improving the robustness of settlement prediction.

[0054] ④ Parameter Update of Incremental Learning Road Deformation Prediction Model Based on Forgetting Mechanism: New time-series data acquired from on-site sensing devices is used to update the prediction model weights. The chain rule is used to calculate the weight gradient in the prediction network, and the error is propagated from the output layer to the weights of each layer through backpropagation, thereby updating the memory unit state. To ensure that the prediction model does not suffer from catastrophic forgetting, a fixed-length memory is used in data training, that is, only the data within one week is retained. At the same time, the sample weights are adjusted through a dynamic weight update mechanism so that new input data has higher weights.

[0055] ⑤ Precise location of pavement deformation risk: A graph attention network is constructed based on the predicted hourly settlement values ​​of the pavement obtained from the prediction model. The settlement gradient tensor and curvature change rate between adjacent nodes are extracted as edge features, and the local differential settlement and deformation transmission trend are output. A regional cumulative settlement threshold δ1 and a settlement rate threshold δ2 are set to evaluate the structural safety. This enables early regional warning of pavement deformation risk.

[0056] Furthermore, the present invention also includes constructing a pavement structure settlement monitoring platform based on digital twins, as detailed below:

[0057] S601, Digital Twin Semantic Model Construction: The model includes road segment location information, road surface material composition, road surface structure layer thickness, sensor device installation point information, temperature, pressure, and time information parameters; based on the minimum database model, it realizes the query, rendering, and calculation application requirements of multi-source information;

[0058] S602, Digital Platform Internal Component Data Transmission Design: Field data is reported from the demodulation device to the platform access layer via a TCP encrypted channel, and the structure and unit are verified according to the data contract; a streaming asynchronous call multi-fidelity data fusion algorithm and a long short-term memory neural network are adopted;

[0059] S603, User Interaction Components: Enables basic information of monitored road sections, display of sensor installation information, display of road surface time-series settlement and deformation, and query of settlement evolution risk assessment and development prediction.

[0060] S604, Platform Risk Warning and Governance Solution: Adjust the cumulative settlement threshold, settlement rate threshold, and monitoring data threshold conditions according to user needs; based on the coupled judgment of multiple indicators and multiple regions, provide feedback on the road surface settlement disease situation and treatment plan.

[0061] Furthermore, the present invention proposes an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein, when executed, the computer program performs the steps of the method of the present invention.

[0062] A computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the method described in this invention.

[0063] A computer program product includes a computer program / instructions that, when executed by a processor, implement the steps of the method described in this invention.

[0064] The present invention adopts the above technical solution and has the following technical effects compared with the prior art:

[0065] (1) This invention proposes a long-term real-time settlement evolution health monitoring scheme for key areas of the road surface, which solves the problem that the previous road settlement sensing equipment could not be compatible with the road structure size and the measurement area was limited.

[0066] (2) The multi-point fiber optic grating sensor network deployment scheme proposed in this invention achieves a balance between the accuracy of road settlement identification and the cost of engineering. It combines multi-fidelity data fusion algorithm to solve the calculation difference caused by the strong constraints of real road surface, and realizes the function of fast and low-cost regional settlement monitoring of road surface.

[0067] (3) Based on the long short-term memory neural network prediction model with physical constraints, this invention can realize the prediction of settlement evolution of monitored road sections in the short term. The incremental learning method ensures the real-time correction of the prediction model and realizes the early warning function of sudden collapse risk. Attached Figure Description

[0068] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments will be briefly described below. Obviously, the drawings described below only relate to some embodiments of the present invention and are not intended to limit the present invention.

[0069] Figure 1 A flowchart of the multi-point fiber optic grating road surface deformation identification and health monitoring technology based on modal superposition method provided in this application.

[0070] Figure 2 This is an illustration of the data acquisition and road surface time-domain operating mode information acquisition of the 941b type vibration pickup in the embodiments of this application.

[0071] Figure 3 This is a diagram of the road surface layout scheme for a multi-point fiber optic grating sensor network in this application embodiment. Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention.

[0073] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0074] Example 1: As Figures 1-3As shown, this invention proposes a method for monitoring and identifying settlement and deformation of highway pavement structures based on digital twins. This technology focuses on the dynamic response characteristics of the pavement structure, extracting the pavement's own time-domain operational modal information and calculating pavement area deformation based on the modal superposition method. Multi-fidelity data fusion algorithms and long short-term memory neural networks are used as auxiliary means to correct deviations in theoretical calculations and predict settlement trends. This provides a health monitoring method for the risk of sudden pavement collapse. Specifically, it includes the following steps:

[0075] S1. A time-domain operating mode identification method for road vibration system under vehicle traffic excitation;

[0076] like Figure 2 As shown, the specific method for obtaining the road surface time-domain operating modal information of the 941b type vibration pickup is as follows: ① Road surface vibration signal acquisition: For the settlement monitoring section, 941B type vibration pickups are deployed at intervals of three meters, 20cm from the edge of each lane. The signal sampling frequency is set to 200Hz, and the test vehicle speed is 60km / h. Acceleration, velocity, and displacement response data under vehicle traffic are obtained by adjusting the pickup's micro-switch and magnifying glass parameters. ② Vibration signal feature extraction: Considering the environmental excitation noise interference factors such as temperature, humidity, load changes, etc., present in the actual road surface... and the systematic measurement error noise of the vibration pickup itself The time-series vibration response measured by the vibration pickup is represented in Hankel matrix form, and its hysteresis covariance Toeplitz matrix is ​​calculated. Noise reduction of the raw data is performed using singular spectrum analysis, followed by obtaining the Hankel matrix modal parameters based on the subspace identification ERA method. ③ Vibration signal modeling: Based on the principle of linear time invariance of road vibration, the collected vibration signal is discretized in time. The road vibration signal at time t, which includes environmental factors and unknown road excitation noise, is used to predict the road vibration at the next time step (time t+1), establishing the main road excitation signal decoupled from noise interference. Vibration response signal The association is shown in equation (1):

[0077] (1)

[0078] In the formula, A is a discrete state vector containing displacement and velocity terms; A is the state matrix. For road surface environmental noise; C is the output vector (vibration signal); C is the output matrix; The noise is measured for the system. The noise signal follows a zero-mean Gaussian distribution.

[0079] ④ Vibration signal matrix acquisition: The acquired Toeplitz matrix is ​​decomposed into a state matrix A and an output matrix C using the least squares fitting method, as shown in equation (2):

[0080] (2)

[0081] In the formula, Let G be the Toeplitz matrix, G be the process noise (including road surface environmental noise and measurement noise), and i be the time step. In road vibration signal processing, 1 / 10 of the sampling signal length is used, i.e., 20.

[0082] ⑤ Solve the characteristic equation of the pavement state matrix A to obtain the natural frequencies of the pavement structure. and vibration modes See formula (3). The vibration modes are verified based on the output matrix C. For accuracy, Formula (4) is the error test standard based on the Euclidean norm.

[0083] (3)

[0084] (4)

[0085] In the formula, This refers to the time-series acceleration data measured by the vibration pickups, where m represents the number of vibration pickups deployed on the road surface. The vibration mode error limit is set to be no less than 95% of the accuracy corresponding to the acceleration signal.

[0086] S2. Finite-point grating road surface deformation sensing technology based on modal superposition method;

[0087] The specific calculation process of the grating finite measurement point deformation sensing technology based on the modal superposition method is as follows:

[0088] ① Fiber Bragg Grating (FBG) Basic Parameters and Installation Requirements: The multi-point FBG equipment uses a 3mm diameter, thick braided mesh armor structure to ensure survival under complex stress conditions inside the road surface. The deformation sensing length of the grating area is 10mm, the spacing between measuring points is 40cm, and the number of measuring points per grating string is 20, covering a two-lane cross-section. The matching FBG demodulator has a wavelength range of 1500~1600nm, a demodulation rate of 100Hz, a wavelength resolution of 1pm, and a signal dynamic recognition range of 50dB. It uses solar power to meet the power needs of equipment in remote areas, and the equipment uses the TCP transmission protocol to achieve wireless data transmission in the field. Figure 3As shown, for the example road section, a grooving cutter is used to cut along the driving direction and cross-sectional direction, with a grooving width not exceeding 1cm. The grating is embedded 3cm inside the top of the base structure. The two ends of the grating are horizontally positioned using rivets and cable ties, and then asphalt mortar is used for surface embedding to ensure that the strain transfer efficiency of the grating is not less than 98%. After the surface is wrapped and stabilized, fine-grained asphalt mixture or potting compound is used to backfill to the original road surface.

[0089] ② Selection of Pavement Principal Modes: A finite element model of the pavement structure with dimensions of 8m x 8m and a subgrade depth of 2m was established. Symmetrical constraints were applied to all model boundaries. The pavement structure layer mesh consisted of 5cm C3D8R elements, with at least four mesh layers in the thickness direction for each layer. The subgrade mesh used 10cm elements with fifteen layers of unidirectional offset in the thickness direction. Linear perturbation analysis based on the dynamic response equation was conducted, and the first fifty mode shapes of the structure were obtained using the Lanczos eigenvalue solver. The measured natural frequencies from step S1 were compared with the numerical solution results, and the corresponding mode orders in the numerical model were selected from low to high. The effective mass contribution rate in the pavement depth direction was at least 60%, which was used as the basis for selecting the representative mode order n under pavement constraints.

[0090] ③ Modal superposition method mode shape matrix: Extract the displacement mode shape at the corresponding modal order in the linear perturbation analysis. Where N represents the number of settlement focus points in the monitored road section, and n represents the modal order. The strain mode shapes at the corresponding modal order are extracted for linear perturbation analysis. , where M is the number of grating measurement points in the monitored road section, and n is the modal order.

[0091] ④ Road surface deformation identification: Based on the modal superposition method, the strain-displacement transformation relationship of the road surface structure can be solved by formula (5):

[0092] (5)

[0093] In the formula, To monitor displacement at key points of interest in road settlement; To monitor the strain data of the grating measuring points on the road section.

[0094] S3. Optimization of the layout of finite-point grating road deformation sensors based on genetic algorithm;

[0095] like Figure 3 As shown, the specific scheme for optimizing the layout of finite-point grating road deformation sensors based on genetic algorithms is as follows:

[0096] ① Fiber optic grating sensor network deployment for the test section: Select a 10-meter-long dual-lane test range, and deploy 20 grating strings with measuring points as shown in S2 at 5-meter intervals in the cross-sectional direction, for a total of three strings; deploy 20 grating strings with measuring points along the centerline of the driving lane in the longitudinal direction, for a total of three strings.

[0097] ② Pavement Strain and Settlement Data Acquisition: Dongfeng EQ140 test vehicles, conforming to JTG 3450-2019 "Specifications for Field Testing of Highway Subgrade and Pavement," were parked at different locations along the test section. Settlement values ​​at six selected representative points were measured sequentially using a Beckman beam, while strain response data at 120 grating measuring points were recorded simultaneously. The measured data will be used to optimize the measuring point layout using a genetic algorithm.

[0098] ③ Measurement point distribution generation: The data from the above 120 strain measurement points are sequentially encoded as xy in the longitudinal and transverse directions, where x represents the grating string number and y represents the measurement point number. Under the condition of a total of 60 measurement points, the individual measurement point distributions are obtained through the random number generation function. Multiple initial layout methods can be formed using this method.

[0099] ④ Genetic algorithm measurement point optimization layout: Calculate the settlement value of the above 6 representative points under each individual using the modal superposition method described in S2, and calculate the population fitness index f according to formula (6):

[0100] (6)

[0101] In the formula, The settlement value is calculated for the i-th representative point in individual x. The value represents the measured settlement value at a representative point; the value of 1000 is the sensitivity adjustment coefficient.

[0102] All initial individuals are placed into the parent pool. Based on the fitness index f, two parents are selected sequentially from the parent pool using the roulette wheel method. The selection probability is calculated using the roulette wheel method according to formula (7). Using the fitness of the individual as the hybridization and mutation probability, each gene is sequentially determined to perform hybridization or mutation operations to form a new individual. Through repeated iterations, when the average relative error of 5 consecutive generations is less than 20%, the optimal individual can be considered to have been obtained. In subsequent road settlement monitoring, the above measurement point form can be adopted to reduce the number of grating string measurement points while ensuring that the road surface deformation identification accuracy is not less than 80%.

[0103] (7)

[0104] In the formula, Let be the probability of selecting the a-th individual; n is the total number of individuals.

[0105] S4. Algorithm for improving road deformation perception accuracy based on multi-fidelity data fusion;

[0106] The specific steps of the algorithm for improving the accuracy of road deformation perception by fusing multi-fidelity data under strong road constraints are as follows:

[0107] ① Construction of High- and Low-Fidelity Datasets: A road surface simulation model consistent with the test section described in S3 was established, and multi-condition simulation calculations were carried out. The driving speed was set at intervals of 10 km / h, ranging from 40 to 100 km / h; tire pressures were selected at 0.6, 0.7, and 0.8 MPa; axle types were adopted from the three representative axle types 2, 5, and 7 in the JTG D50 standard; and the driving position was selected from the center of the symmetrical lane, with offsets of 15 and 30 cm, totaling 189 dynamic load conditions. In addition, 21 static load conditions were applied at different road surface locations using the same tire pressures and axle types. Settlement values ​​at six representative points on the road surface were collected after optimization using a genetic algorithm to construct a large-capacity low-fidelity simulation dataset. The measured road section strain and displacement data from step S3 were used to construct a low-capacity high-fidelity real dataset.

[0108] ② Low-fidelity network model construction: Using strain data, calculated displacement values, and simulated displacement values ​​as input data, a deep belief network (DBN) is pre-trained layer by layer using a simulation dataset to obtain the general hierarchical features of the modal superposition method. The training process includes a pre-training stage where a Restricted Boltzmann Machine (RBM) based on unsupervised learning is used to build the network layer by layer, and a fine-tuning stage where a backpropagation backpropagation (BP) neural network is established between the last two hidden layers of the DBN to adjust the weights of the entire network in a supervised learning manner. The specific calculation process is described below:

[0109] 1) Data initialization processing

[0110] The input data is standardized according to formula (8).

[0111] (8)

[0112] In the formula, The variance is set using a Gaussian model in the first layer. .

[0113] 2) Layer-by-layer unsupervised pre-training of RBMs

[0114] Define the energy function (9) to learn the data distribution form.

[0115] (9)

[0116] In the formula, Input strain, calculate displacement, and simulate displacement vector; These are binary hidden layer neurons; This is the weight matrix; For input layer bias; This is for hidden layer bias.

[0117] The activation probability function of hidden layer neurons is given by formula (10).

[0118] (10)

[0119] The activation function chosen is sigmoid, which is used for calculating hidden layer probabilities.

[0120] The comparative divergence formula (11) is used to calculate the approximate gradient of the data in order to estimate the difference between the data and the model.

[0121] (11)

[0122] In the formula, Expectations for the training data; Generate sample expectations for the model; The learning rate; Weight decay; The momentum is used to accelerate convergence.

[0123] After the above pre-training is completed, the hidden layer activation values Equation (12) will be used to input the next layer.

[0124] (12)

[0125] 3) Backpropagation in hidden layers

[0126] Minimizing the loss function is the objective of the fine-tuning stage, and the mean squared error of the data is used as the basic supervised loss. Equation (13).

[0127] (13)

[0128] To prevent the model from overfitting, a weight decay term is set. Equation (14) enhances the generalization ability of the model under various working conditions and ensures that the basic characteristics of the road dynamic response are preserved.

[0129] (14)

[0130] In the formula, Let Frobenius norm be the weight matrix of the l-th layer; is the Frobenius norm of the output layer weight matrix.

[0131] L2-SP regularization loss term Equation (15) is used to prevent the model from forgetting the useful features of the data obtained during the unsupervised pre-training stage.

[0132] (15)

[0133] The above three loss functions together constitute the total loss function of the backpropagation fine-tuning phase. Equation (16).

[0134] (16)

[0135] In the formula, , The weight decay hyperparameter is set to 10. -4 10 -3 .

[0136] Through the loss function The partial derivatives of the network output are calculated to obtain the output layer residuals. Equation (17-1) is followed by the chain rule to pass the residual terms to the previous structural layer equation (17-2).

[0137] (17-1)

[0138] (17-2)

[0139] In the formula, It is the derivative of the activation function term of the previous layer.

[0140] After the hidden layer error term is calculated, the gradient descent method can be used to update the weights and biases.

[0141] 4) The improved modal superposition method is given by formula (18):

[0142] (18)

[0143] In the formula, The correction term is calculated for the modal superposition method of road surface with strong constraints based on DBN network.

[0144] ③ Network model correction based on high-fidelity dataset: The measured real strain, calculated displacement value, and measured displacement value of representative points of the road surface from step S3 are used for fine-tuning of the top-level task head of the established DBN network. This step is consistent with the backpropagation of DBN. Here, measured data is used instead of simulation data, and only the weights of the output layer are updated, so that the DBN network can capture the characteristics of real road surface data while ensuring the general features of the modal calculation. Adversarial training is performed on the simulation dataset and the real dataset. The discriminator formula (19) of the generative adversarial network is used to distinguish the difference distribution of high-fidelity data samples. The feature covariance alignment formula (20) is used to make the feature distribution of high-fidelity data more consistent and realize that the features of the two data sources are indistinguishable.

[0145] (19)

[0146] In the formula, Expected based on real data; The expected value of the generator noise data; This represents the discriminator's prediction of the true data x. Simulation data derived from generator noise samples.

[0147] (20)

[0148] In the formula, To minimize the covariance loss; , These are the feature covariance matrices of the real data and the simulated data, respectively.

[0149] ④ Optimization of loss function for multi-fidelity regression networks: In the step of modifying the network model by introducing high-fidelity datasets, a domain adaptation loss function term based on Wasserstein distance is added. This is used to reduce the impact of differences in the distribution of high-fidelity and low-fidelity data. It involves data sample supervision. Weights are adjusted by using larger weights for real data and smaller weights for simulation data. An L2 norm starting point regularization term is also introduced to limit weight deviations in general feature extraction. Physical consistency is employed. The loss function term characterizes the dynamic properties of the data.

[0150] S5. A road surface deformation prediction model based on long short-term memory network time-series data constrained by physical terms.

[0151] The implementation steps of the road surface deformation prediction model based on physical term constraints using long short-term memory network time-series data are as follows:

[0152] ① Acquisition of multi-source time-series monitoring dataset: Based on the deployment of fiber optic gratings at multiple measurement points on the S3 road section, temperature and pressure sensors were installed on the top and bottom surfaces of the asphalt intermediate and lower layers, the semi-rigid base course, and the top surface of the subgrade. Humidity and pressure sensors were also installed inside the subgrade. The pressure response data were processed using the SG smoothing filter method. The actual pavement structure material parameters and traffic load spectrum were used for finite element numerical calculations to obtain corresponding mechanical response simulation comparison results.

[0153] ② Long Short-Term Memory Network Pavement Deformation Prediction Model: The hourly time-series displacement values, temperature and pressure data of each structural layer, and subgrade moisture data (corrected in step S4) within one week are used as input to the memory network, and the subsequent 24-hour settlement is used as the output to build a pavement deformation prediction model. The following is a description of the composition of the Long Short-Term Memory Network:

[0154] 1) Gate of Oblivion

[0155] (twenty one)

[0156] In the formula, This is the hidden state from the previous moment; Input data for the current time step, including temperature, pressure, and soil moisture; and These are the weight matrix and the bias term, respectively.

[0157] 2) Input Gate

[0158] (twenty two)

[0159] 3) Cell status update

[0160] (23-1)

[0161] (23-2)

[0162] In the formula, The current time step candidate cell state; This represents the current time step cell state.

[0163] 4) Output gate

[0164] (24-1)

[0165] (24-2)

[0166] In the formula, Hide the current state.

[0167] 5) Network prediction output formula

[0168] (25)

[0169] ③ Optimization of pavement deformation prediction model based on physical constraints: In the long short-term memory network, the difference between the measured pressure data and the numerically calculated pressure at the structural layer points, and the difference between the measured temperature change and the numerically calculated temperature evolution of the structural layer are used as physical driving objective function terms to guide the update of prediction model parameters, ensuring that the prediction results not only conform to the time evolution law of monitoring data, but also meet the mechanical behavior characteristics of pavement structure, thereby improving the robustness of settlement prediction.

[0170] ④ Parameter Update of Incremental Learning Road Deformation Prediction Model Based on Forgetting Mechanism: New time-series data acquired from on-site sensing devices is used to update the prediction model weights. The chain rule is used to calculate the weight gradient in the prediction network, and the error is propagated from the output layer to the weights of each layer through backpropagation, thereby updating the memory unit state. To ensure that the prediction model does not suffer from catastrophic forgetting, a fixed-length memory is used in data training, that is, only the data within one week is retained. At the same time, the sample weights are adjusted through a dynamic weight update mechanism so that new input data has higher weights.

[0171] ⑤ Precise location of pavement deformation risk: Based on the predicted hourly settlement values ​​of the pavement obtained from the prediction model, a graph attention network is constructed. The settlement gradient tensor (26) and curvature change rate (27) between adjacent nodes are extracted as edge features, and the local regional differential settlement and deformation transmission trend are output. The regional cumulative settlement threshold δ1 and settlement rate threshold δ2 are set to evaluate the structural safety. This achieves regional early warning of pavement deformation risk.

[0172] (26)

[0173] (27)

[0174] Furthermore, based on the above steps, this invention proposes a road structure settlement monitoring platform based on digital twins, the specific implementation method of which is as follows:

[0175] ① Digital Twin Semantic Model Construction: The road surface settlement semantic model includes road segment location information, road material composition, road structure layer thickness, sensor installation point information, temperature, pressure, and time information parameters. Based on a minimal database model, it fulfills the requirements for querying, rendering, and tracking multi-source road information. An executable data contract is built to support 3D data-driven processing. Road surface settlement visualization modeling uses Tiles meshes for 3D rendering.

[0176] ② Data transmission from the monitoring platform: Monitoring data from the fiber optic grating sensor network is reported to the platform access layer via a wireless demodulation device through a TCP encrypted channel. Structure and unit verification are performed according to the data contract to enable data transmission to the semantic model. Uploaded data is stored in the time-series and spatial libraries. The platform processing layer employs a streaming asynchronous multi-fidelity data fusion algorithm and a long short-term memory neural network to acquire road surface settlement and future development trend data. The processed settlement and trend data will then be used for 3D rendering and display of the semantic model.

[0177] ③ Platform visualization and user interaction components: The pavement structure settlement monitoring platform includes basic information of the monitored road section, sensor installation display information, pavement time-series settlement deformation display, settlement evolution risk assessment and development prediction. Users can achieve real-time reading of on-site monitoring data, query of settlement evolution time-series values ​​at various points on the pavement, visualization of three-dimensional pavement structure settlement from multiple perspectives and profiles, and pavement settlement evolution risk assessment and early warning reports.

[0178] ④ Settlement Risk Early Warning and Remediation Solution: The digital twin platform can adjust the cumulative settlement threshold, settlement rate threshold, and monitoring data threshold conditions according to user needs. Based on the coupled judgment of multiple indicators and multiple regions, it provides feedback on pavement settlement defects and treatment solutions, realizing real-time early warning of highway pavement structure settlement and deformation based on digital twin.

[0179] Example 2: This example proposes an electronic system, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method steps of the present invention.

[0180] Example 3: This example proposes a computer-readable storage medium storing a computer program / instructions thereon, which, when executed by a processor, implements the steps of the method of the present invention.

[0181] Example 4: This example proposes a computer program product, including a computer program / instructions, which, when executed by a processor, implements the steps of the method described in this invention.

[0182] It should be noted that the processing flows of Embodiments 2 to 4 correspond to the specific steps of the method provided in the embodiments of the present invention, and have the corresponding functional modules and beneficial effects of the method. Technical details not described in detail in this embodiment can be found in the method provided in the embodiments of the present invention.

[0183] The program code used to implement the methods of this application may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0184] In the context of this application, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0185] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solutions based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. A method for monitoring and identifying settlement and deformation of highway pavement structures based on digital twins, characterized in that, Including the following steps: S1. By deploying vibration detectors in the settlement monitoring section, the road surface vibration response signal under vehicle traffic is collected; the vibration response signal is feature extracted and modeled, the state matrix of the road surface structure is solved, and then the measured natural frequency and vibration mode of the road surface structure are extracted. S2. Based on the measured natural frequency obtained in step S1, the main road surface mode is selected in the finite element model; fiber optic gratings are buried along the driving direction on the road surface to be tested to obtain strain data; combined with the vibration mode obtained in step S1, the strain-displacement conversion relationship is established by the mode superposition method to identify the displacement and settlement distribution of the monitored road section. S3. Deploy a fiber optic grating sensor network on the test section and simultaneously collect the measured settlement values ​​of multiple representative points and the strain response data of all grating measurement points. Based on the measured settlement values, use the strain-displacement conversion relationship established in step S2 to optimize the grating measurement point layout scheme through a genetic algorithm to obtain an optimal measurement point layout scheme with the fewest measurement points while ensuring recognition accuracy. S4. Construct a numerical simulation model containing the optimal measurement point layout scheme, generate a low-fidelity simulation dataset, and combine it with the high-fidelity dataset obtained from the actual measurement in step S3 to train a deep belief network; through adversarial training and loss function optimization, modify the modal superposition method in step S2 to obtain a strain-displacement mapping model. S5. In the road section where the optimal measuring point layout scheme is set up, multi-source time-series monitoring data is introduced; the displacement sequence output by the model after step S4 and the multi-source time-series monitoring data are input together into the long short-term memory network to predict the future settlement of the road surface. Based on the prediction results, a graph attention network is used to identify settlement gradients and trends, enabling precise location and early warning of road deformation risks in specific areas.

2. The method according to claim 1, characterized in that, Step S1 specifically includes: S101, Road Vibration Signal Acquisition: Acquire acceleration, velocity, and displacement response data under vehicle traffic by deploying vibration pickups; S102. Vibration signal feature extraction: The time-series vibration response measured by the vibration pickup is expressed in Hankel matrix form, and its Toeplitz matrix is ​​calculated to extract dynamic features. S103. Vibration signal modeling: Based on the principle of linear time invariance of road vibration, the vibration signal is discretized in time, and a correlation model between the main excitation signal and the vibration response signal of the road is established, where the vibration signal at time t is used to predict the vibration at time t+1; as shown in equation (1): (1) In the formula, A is a discrete state vector containing displacement and velocity terms; A is the state matrix. For road surface environmental noise; C is the output vector, i.e., the vibration signal; C is the output matrix. Measure the noise of the system; S104. Vibration signal matrix acquisition: The Toeplitz matrix is ​​decomposed into a state matrix A and an output matrix C, and matrices A and C are solved by least squares fitting, as shown in equation (2): (2) In the formula, Let G be the Toeplitz matrix, G be the process noise, which includes road surface environmental noise and measurement noise, and i be the time step. S105. Road surface modal signal extraction: The characteristic equation of the road surface state matrix A is solved using formula (3) to obtain the natural frequencies of the road surface structure. and vibration modes The vibration modes are verified based on the output matrix C. Accuracy, Formula (4) is the error test standard based on the Euclidean norm: (3) (4) In the formula, The data represents the time-series acceleration measured by the vibration pickups, where m represents the number of vibration pickups deployed on the road surface.

3. The method according to claim 1, characterized in that, Step S2 specifically includes: S201. Bury the fiber optic grating along the driving direction on the road surface to be tested; S202. Selection of main modes of road surface: Establish a finite element model of the road surface structure, perform linear perturbation analysis, obtain the mode shapes of the structure, compare with the measured natural frequencies in step S1, and select representative mode orders; S203. Construct the mode shape matrix using the modal superposition method: Extract the displacement mode shape and strain mode shape under the selected modal order; S204. Road surface deformation identification: Based on the modal superposition method, the displacement of the monitoring point is solved through the strain-displacement transformation relationship. The specific calculation is as follows: (5) In the formula, To monitor displacement at key points of interest in road settlement; It is a displacement mode shape. For strain mode vibration, To monitor the strain data of the grating measuring points in the road section, N is the number of settlement-related points in the monitored road section, n is the mode order, and M is the number of grating measuring points in the monitored road section.

4. The method according to claim 1, characterized in that, Step S3 is as follows: S301. Deploy a fiber optic grating sensor network; S302. Collect road surface strain and settlement data: On the test road section, measure the settlement value at representative points and simultaneously record the strain response data of the grating measuring points; S303, Measurement point distribution generation: Encode the strain measurement point data and generate individual measurement point distributions using random numbers; S304, Genetic Algorithm for Optimized Measurement Point Layout: The modal superposition method is used to calculate the settlement value of the representative points under each individual. The fitness index is calculated based on the error between the calculated value and the measured value. Through iterative optimization using a genetic algorithm, the optimal measurement point layout scheme is obtained when the average relative error of multiple consecutive generations is less than the threshold. Specifically, the modal superposition method described in S2 is used to calculate the settlement value of the above 6 representative points under each individual. The population fitness index f is calculated according to formula (6): (6) In the formula, The settlement value is calculated for the i-th representative point in individual x. The value represents the measured settlement at a representative point; the value is 10. 3 This is the sensitivity adjustment coefficient; All initial individuals are placed into the parent pool. Based on the fitness index f, two parents are selected from the parent pool in turn by the roulette wheel method. The selection probability is calculated by the roulette wheel method according to formula (7). The fitness of the individual is used as the hybridization and mutation probability to determine whether to perform hybridization or mutation operation on each gene in turn to form a new individual. When the average relative error of 5 consecutive generations is less than 20%, the optimal individual can be considered to be obtained through repeated iterations. (7) In the formula, Let be the probability of selecting the a-th individual; n is the total number of individuals.

5. The method according to claim 1, characterized in that, Step S4 is as follows: S401. Construct high- and low-fidelity datasets: Obtain a low-fidelity simulation dataset through numerical simulation, and obtain a high-fidelity real dataset through actual measurement in step S3. S402. Build a low-fidelity network model: Use a simulation dataset to pre-train and fine-tune the deep belief network to correct the modal superposition method; Specifically: (8) In the formula, The correction term is calculated for the modal superposition method of road surface with strong constraints based on DBN network; S403. Improve the network model based on high-fidelity dataset: Fine-tune the network using real data and make the feature distribution of high-fidelity and low-fidelity data consistent through adversarial training; S404. By introducing the domain adaptation loss function, adjusting the weights of the supervision terms, adding a regularization term and a physical consistency loss function, a more accurate strain-displacement mapping model that integrates physical mechanisms and measured data is obtained.

6. The method according to claim 1, characterized in that, Step S5 is as follows: S501. Acquire multi-source time-series monitoring data: Collect temperature, pressure, and soil moisture data for each structural layer, and perform noise reduction and smoothing; acquire material parameters and traffic load spectrum of the actual road section, and calculate the mechanical response through finite element method; S502. Establish a long short-term memory network prediction model: Use the time-series displacement, temperature, pressure, and humidity data from the model output after step S4 as input to the long short-term memory network to predict future settlement. S503. Optimize the prediction model: Based on physical constraints, use the difference between measured and calculated values ​​as the objective function term to guide the update of prediction model parameters. S504. Update prediction model parameters: Through incremental learning mechanism, use new data to update model weights, and through backpropagation, pass the error from the output layer to the weights of each layer, thereby updating the memory cell state. S505, Precise location of road deformation risk: Based on the predicted settlement value obtained by the prediction model, a graph attention network is constructed, the settlement gradient tensor and curvature change rate between adjacent nodes are extracted as edge features, the local regional differential settlement and deformation transmission trend are output, the settlement gradient and curvature are evaluated, and a threshold is set for early warning.

7. The method according to claim 1, characterized in that, This also includes building a pavement structure settlement monitoring platform based on digital twins, as detailed below: S601, Digital Twin Semantic Model Construction: The model includes road segment location information, road surface material composition, road surface structure layer thickness, sensor device installation point information, temperature, pressure, and time information parameters; based on the minimum database model, it realizes the query, rendering, and calculation application requirements of multi-source information; S602, Digital Platform Internal Component Data Transmission Design: Field data is reported by demodulation equipment to the platform access layer along a TCP encrypted channel, and the structure and unit are verified according to the data contract; A streaming asynchronous call multi-fidelity data fusion algorithm and a long short-term memory neural network are employed. S603, User Interaction Components: Enables basic information of monitored road sections, display of sensor installation information, display of road surface time-series settlement and deformation, and query of settlement evolution risk assessment and development prediction. S604, Platform Risk Warning and Governance Solution: Adjust the cumulative settlement threshold, settlement rate threshold, and monitoring data threshold conditions according to user needs; based on the coupled judgment of multiple indicators and multiple regions, provide feedback on the road surface settlement disease situation and treatment plan.

8. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method described in any one of claims 1-7.

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

10. An electronic system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed, it implements the steps of the method as described in any one of claims 1 to 7.

Images