Road settlement prediction method and system based on data physical driving model

Abstract

The invention belongs to the technical field of road settlement prediction, and particularly relates to a road settlement prediction method and system based on a data physical driving model. Comprising the steps of obtaining original strain data of a to-be-monitored road; constructing a physical constraint equation about strain to settlement of the road, and introducing the physical constraint equation into a loss function of the deep learning prediction network as a physical constraint item; training a deep learning prediction network by using historical strain data and a real settlement value of a road to be monitored to obtain a trained data physical driving model; and inputting the original strain data into the trained data physical driving model, and predicting to obtain a road settlement curve of the road to be monitored. According to the method, a reasonable model framework can be established by utilizing physical knowledge, the model is ensured to accord with an actual physical rule, the model can be continuously optimized through data learning, the prediction precision is improved, and thus the road settlement can be predicted more accurately.

Description

Technical Field

[0001] This invention belongs to the field of road settlement prediction technology, and particularly relates to a road settlement prediction method and system based on a data physics-driven model. Background Technology

[0002] Currently, most existing road settlement monitoring methods heavily rely on discrete point measurements. Discrete point measurements select specific isolated points on the road to obtain relevant data; however, this method has significant limitations, as it struggles to achieve comprehensive monitoring of the continuous road space. Because a road is a continuous spatial structure, discrete point measurements alone cannot fully and accurately reflect the settlement situation of the entire road space.

[0003] While Brillouin optical time-domain analysis (BOTDA) technology can provide high-density strain data, meaning it can acquire strain information at a relatively large number of points, thus compensating to some extent for the insufficient coverage of discrete point measurements, accurately converting this strain data into settlement data still presents a significant challenge. This is because the conversion from strain to settlement involves complex physical relationships and numerous uncertainties, and currently, there is no perfectly reliable method to complete this conversion flawlessly.

[0004] Pure data-driven models are those that learn and predict based on large amounts of data. However, such models are prone to failure when data is scarce or when road structures change. When data is insufficient, the model cannot learn enough features and patterns, leading to inaccurate predictions; and when road structures change, previously learned patterns may no longer be applicable, making it difficult for the model to make effective predictions. Summary of the Invention

[0005] To overcome the shortcomings of the existing technologies, this invention provides a road settlement prediction method and system based on a data physics-driven model. This method can utilize physical knowledge to establish a reasonable model framework, ensuring that the model conforms to actual physical laws. Furthermore, it can continuously optimize the model through data learning, thereby improving the accuracy of prediction and enabling more accurate prediction of road settlement.

[0006] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: The first aspect of this invention provides a method for predicting road settlement based on a data physics-driven model.

[0007] The road settlement prediction method based on data physics-driven models includes the following steps: Obtain the raw strain data of the road to be monitored; We construct physical constraint equations for the road from strain to settlement, and introduce these physical constraint equations into the loss function of the deep learning prediction network as physical constraint terms. A deep learning prediction network is trained using historical strain data and actual settlement values ​​of the road to be monitored. This allows the deep learning prediction network to learn the mapping relationship between strain and settlement that conforms to the physical constraint equation, resulting in a well-trained data physics-driven model. The raw strain data is input into the trained data physics-driven model to predict the road settlement curve of the road to be monitored.

[0008] A second aspect of the present invention provides a road settlement prediction system based on a data physics-driven model.

[0009] A road settlement prediction system based on a data physics-driven model includes: The strain data acquisition module is configured to acquire the raw strain data of the road to be monitored. The physical constraint module is configured to: construct physical constraint equations for the road with respect to strain and settlement, and introduce these physical constraint equations into the loss function of the deep learning prediction network as physical constraint terms; The model training module is configured to: train a deep learning prediction network using historical strain data and actual settlement values ​​of the road to be monitored, so that the deep learning prediction network learns the mapping relationship between strain and settlement that conforms to the physical constraint equation, and obtains a trained data physics-driven model. The settlement prediction module is configured to input raw strain data into a trained data physics-driven model to predict the road settlement curve of the road to be monitored. A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps of the road settlement prediction method based on a data physics-driven model as described in the first aspect of the present invention.

[0010] A fourth aspect of the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the road settlement prediction method based on a data physics-driven model as described in the first aspect of the present invention.

[0011] The above one or more technical solutions have the following beneficial effects: This invention provides a road settlement prediction method and system based on a data-driven physics model. It constructs a physical constraint equation for the road's strain-to-settlement relationship and incorporates this equation into the loss function of a deep learning prediction network as a physical constraint term. This physical constraint term forces the model's internal computational logic to adhere to pre-defined physical laws during learning and prediction. It not only requires the model to output accurate settlement values ​​but also demands that the implicit strain-to-settlement transformation relationship during model learning align with the actual physical integral relationship. This effectively prevents prediction errors that "violate physical common sense" from purely data-driven models when data is insufficient. The joint loss function combining physics and data ensures that the model's prediction results possess both accuracy and physical plausibility.

[0012] The physical constraint equation of this invention is based on the mechanical assumption of the continuity of deformation and strain, connecting strain and settlement, indicating that the settlement at a certain point is not determined solely by the strain at the current point, but is jointly determined by the cumulative effect of all strains from the starting point to the current point.

[0013] Traditional models struggle to generalize due to significant differences in road structures, soil conditions, and load environments. This invention utilizes transfer learning to achieve cross-scenario and cross-road segment settlement prediction, enabling accurate prediction of future settlement trends for different road types (highways, urban roads, soft soil foundation sections, etc.). This results in a generalizable and scalable intelligent settlement prediction framework, providing support for the system's adaptability in engineering environments.

[0014] This invention constructs a settlement risk level model, further transforming the results of physical predictions (settlement curves) into engineering decision-making information (risk levels). By performing feature engineering on the settlement curves and utilizing a powerful machine learning algorithm (XGBoost) for evaluation, the system can automatically identify road sections with potential hazards that require key attention and intervention, achieving a leap from "monitoring" to "early warning".

[0015] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0016] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0017] Figure 1 This is a flowchart of the method in Example 1.

[0018] Figure 2 This is a prediction architecture diagram for Example 1.

[0019] Figure 3This is a diagram of the U-Net network structure for the data physics-driven model in Example 1.

[0020] Figure 4 This is a flowchart of the TL transfer learning process in Example 1.

[0021] Figure 5 This is a diagram of a distributed optical fiber connection in Example 1.

[0022] Figure 6 This is a schematic diagram of distributed optical fiber deployment in Example 1.

[0023] The attached diagram lists the components represented by each number as follows: 1. BOTDA main unit; 2. Sensor fiber optic cable; 3. Tail-end fiber optic splice box; 4. Probe laser emitter; 5. Pump laser emitter; 6. Probe laser; 7. Pump laser. Detailed Implementation

[0024] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0025] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0026] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0027] Example 1 This embodiment discloses a road settlement prediction method based on a data physics-driven model.

[0028] like Figure 1 As shown, the road settlement prediction method based on a data physics-driven model includes the following steps: Obtain the raw strain data of the road to be monitored; We construct physical constraint equations for the road from strain to settlement, and introduce these physical constraint equations into the loss function of the deep learning prediction network as physical constraint terms. A deep learning prediction network is trained using historical strain data and actual settlement values ​​of the road to be monitored. This allows the deep learning prediction network to learn the mapping relationship between strain and settlement that conforms to the physical constraint equation, resulting in a well-trained data physics-driven model. The raw strain data is input into the trained data physics-driven model to predict the road settlement curve of the road to be monitored.

[0029] like Figure 1 and Figure 2As shown in the figure, this invention proposes a distributed optical fiber road settlement prediction method based on a data physics driven model (DPD) and transfer learning, including several processing steps such as BOTDA distributed optical fiber monitoring, data preprocessing, data physics driven model (DPD), transfer learning, settlement inversion, and risk prediction. The technical solution of this embodiment will be explained in detail below.

[0030] Step 1: Deployment and data acquisition of BOTDA distributed fiber optic monitoring equipment.

[0031] This step aims to construct a high-density, continuous road structure strain monitoring network to provide fundamental data support for subsequent settlement prediction. A schematic diagram of the distributed fiber optic structure is shown below. Figure 5 and Figure 6 As shown.

[0032] Figure 5 In the middle, the BOTDA host 1 and the tail fiber splice box 3 are connected through the sensing fiber optic cable 2.

[0033] Figure 6 In the process, the probe laser emitter 4 and the pump laser emitter 5 emit probe laser 6 and pump laser 7 respectively. The two laser beams reach the sensing fiber 2 and are then looped at the end.

[0034] Specifically, it includes: (1) Design of fiber optic cable deployment: Single-mode fiber optic cables are deployed along key parts of the road structure (subgrade, pavement, slope, etc.) to achieve strain monitoring with meter-level resolution.

[0035] (2) BOTDA parameter settings: including pulse width, sampling frequency, step size and signal-to-noise ratio optimization strategy.

[0036] (3) Data connection scheme: The fiber optic acquisition unit is connected to the edge computing node through the industrial gateway and uses gigabit fiber optic or 5G network for transmission.

[0037] (4) Network communication scheme: Ensure that the collected data can be transmitted to the cloud or central server in real time, and support long-distance links and multi-point synchronous collection.

[0038] This step establishes the system's foundational perception layer, which is the core data source for subsequent settlement identification and prediction.

[0039] Step 2: Construct a data physics-driven model (DPD) to form a road settlement identification system that integrates physical constraints.

[0040] like Figure 3 As shown, this step utilizes the raw strain data obtained from BOTDA, combined with a road mechanics model and a deep learning network, to construct a settlement identification model that combines physical consistency with data learning capabilities. Figure 3In the middle, on the left: the input is BOTDA strain data; in the middle: the U-Net convolutional network structure, labeled as "physical constraint layer" or "physical loss term"; on the right: the output is the predicted settlement curve + physical consistency output.

[0041] Mainly includes: (1) Strain physics preprocessing and feature extraction.

[0042] By employing wavelet denoising, temperature compensation, filtering, and signal smoothing methods, effective strain features reflecting structural deformation behavior are extracted.

[0043] (2) Construct physical constraint equations.

[0044] This invention introduces a strain-settlement integral relationship:

[0045] in: w(x): Represents the vertical settlement of the road at location x (unit: usually millimeters mm). This is the target value that the model ultimately aims to predict.

[0046] ε(ξ): Represents the axial microstrain at a point ξ along the road (unit: microstrain με). This is a raw physical quantity directly measured by the BOTDA (Board-Operated Fiber Optic) system.

[0047] ∫0 x … dξ: represents the spatial integration from the starting point (0) of the road to the current position (x). This means that the settlement at a point is not determined solely by the strain at that point, but by the cumulative effect of all strains from the starting point to that point.

[0048] ξ: Integral variable, representing the spatial location along the road.

[0049] This formula is based on the mechanical assumptions of small deformation and strain continuity. It connects the sensor data (strain ε) with the engineering target (settlement w) as constraints, guiding the neural network's learning process to conform to physical laws. This physical equation is embedded in the neural network's loss function as a constraint term.

[0050] (3) Construct a deep learning prediction network.

[0051] like Figure 3 As shown, a U-Net convolutional neural network structure is used, with strain distribution as input, and the output includes: Predicted settlement curve ; Constraint output that satisfies the physical mechanism .

[0052] (4) Establish the data physics-driven loss function

[0053] This formula defines the optimization objective of the training data physics-driven model (DPD), which is a combination of physical constraints and data fitting.

[0054] in: L: Represents the total loss value. The goal of model training is to minimize this value by adjusting the parameters.

[0055] w: Actual settlement value. This is typically a label derived from historical monitoring data (such as leveling measurements) or high-precision simulation data.

[0056] : The settlement value predicted by the model (w-hat). That is, the settlement prediction result directly output by the neural network based on the strain data.

[0057] || w - ||²: Data fitting term. Calculates predicted settlement. The sum of squares (mean square error) of the differences between the model's predictions and the actual settlement w. This term ensures that the model's predictions closely match the actual observed data.

[0058] P(w): Represents a physical operation or transformation P applied to the actual settlement w. In context, this operation is usually the strain-settlement integral relationship described above. It can be understood as: P(w) is "an intermediate physical quantity derived from the actual settlement w through physical laws".

[0059] : Intermediate physical quantities predicted by the model. is another output of the neural network, corresponding to the result of the physical operation P.

[0060] || P(w) - ||²: Physical constraint term. The physical quantity predicted by the computational model. The sum of squares of the differences between the physical quantity P(w) and the physical quantity calculated from real data according to physical laws. This forced model must adhere to the preset physical laws (i.e., Formula 1) in its internal calculation logic during learning and prediction.

[0061] It not only requires the model to output accurate settlement values ​​(first requirement), but also requires that the implicit "strain-settlement" conversion relationship during the model learning process be consistent with the actual physical integral relationship (second requirement). This effectively prevents "physically inaccurate" prediction errors from purely data-driven models when data is insufficient.

[0062] The combined loss of physics and data ensures that the model's predictions are both accurate and physically plausible.

[0063] This step enables high-precision inversion and construction of road settlement, forming a complete data-physical fusion identification system.

[0064] Step 3: Multi-scenario road settlement prediction based on transfer learning (TL).

[0065] Due to significant differences in road structures, soil conditions, and load environments, traditional models are difficult to generalize.

[0066] This step utilizes transfer learning to achieve settlement prediction across scenarios and road sections. The TL transfer learning flowchart is shown below. Figure 4 As shown.

[0067] Includes the following: (1) Train the DPD network on the source domain road model: Obtain pre-trained weights that include basic road settlement patterns.

[0068] (2) Migrate to the target road structure and freeze some network layers: The low-level feature extraction layer is fixed, and only the high-level semantic layer is fine-tuned.

[0069] (3) Fine-tuning and optimizing the model with a small amount of target data: Address the issue of scarce target road data and improve generalization capabilities.

[0070] (4) Establish a cross-scenario settlement prediction model: It enables accurate prediction of future settlement trends for different road types (highways, urban roads, soft soil foundation sections, etc.).

[0071] This step establishes a generalizable and scalable intelligent settlement prediction framework, providing support for the system's adaptability in engineering environments.

[0072] Step 4: Settlement inversion, risk level assessment and system utility verification.

[0073] This step utilizes the settlement information obtained in the previous steps to form a complete settlement risk early warning platform on actual roads.

[0074] Includes the following: (1) Settlement inversion.

[0075] Based on the predicted strain and physical relationship:

[0076] The settlement curve along the road was obtained.

[0077] (2) Construct a road settlement risk index system.

[0078] Parameters such as strain gradient, settlement rate, cumulative settlement, and diurnal rate of change acceleration factor are extracted using statistical features. Among them: Strain / settlement gradient: reflects the degree of deformation severity; Settlement rate: the amount of settlement per unit time, reflecting the development trend; Cumulative settlement: Total settlement amount; Daily rate of change acceleration factor: reflects whether settlement is accelerating.

[0079] This is used to construct a settlement-sensitive feature vector F.

[0080] (3) An entropy weight method + XGBoost was used to construct a settlement risk level model:

[0081] in: R: Represents the output settlement risk level. It is usually a classification result (such as "low risk", "medium risk", "high risk") or a risk probability value.

[0082] F: Represents the settlement-sensitive feature vector. It is a multidimensional dataset containing multiple statistical and engineering features extracted from raw or predicted strain / settlement data.

[0083] f(·): Represents the risk prediction model function, specifically the XGBoost algorithm model. It is a trained, complex machine learning model capable of learning the complex nonlinear mapping relationship between feature vector F and risk level R.

[0084] This formula further transforms the physical prediction results (settlement curve) into engineering decision-making information (risk level). By performing feature engineering on the settlement curve and using a powerful machine learning algorithm (XGBoost) for evaluation, the system can automatically identify road sections with potential hazards that require key attention and intervention, achieving a leap from "monitoring" to "early warning".

[0085] Output three levels of settlement risk: low, medium, and high.

[0086] To systematically evaluate the mechanical response and deformation characteristics of road structures under different distress conditions, the experimental model accurately simulates the mechanical properties of the roadbed. Multiple layers of earth pressure cells and distributed fiber optic sensors are embedded within the structure to monitor stress distribution and strain field changes in real time under load. To simulate typical roadbed distress in actual engineering projects, distresses such as roadbed cracks, roadbed loosening, roadbed settlement, and pipeline leaks are artificially pre-programmed into some specimens.

[0087] The following classifies risks into three different levels: Low risk: Sensor monitoring data shows that throughout the fatigue loading process, the structural stress distribution is uniform, the strain response remains within the elastic range and the value is extremely small, and no residual deformation or damage accumulation was detected. This indicates that the road structure is intact, has high load-bearing capacity, and performs stably under long-term standard loads, with an extremely low risk of harmful deformation or structural damage.

[0088] Medium risk: Monitoring data clearly reveals localized stress concentrations and strain anomalies, with pre-defined cracks or loose areas showing a tendency to expand under load. The structure has undergone plastic deformation exceeding its elastic range; although not immediately destroyed, observable performance degradation has occurred, indicating that the damage is in an active and developing stage, with a clear risk of further deterioration.

[0089] High Risk: Distributed fiber optic and earth pressure cell monitoring detected drastic and unstable strain abrupt changes and abnormal stress redistribution. Under fatigue loading, deformation continues to increase irreversibly, and pre-set settlement or leakage defects develop rapidly, leading to a significant loss of structural bearing capacity and a tendency towards instability. This level indicates that the road structure is in a critical failure state and is highly susceptible to sudden and serious engineering problems.

[0090] (4) Conduct system verification on actual road sections.

[0091] The system was validated by comparing the measured and predicted results: 1) Accuracy of settlement inversion; 2) Accuracy of trend prediction; 3) Anomaly detection capability; 4) Project deployment availability.

[0092] Application results on multiple actual road sections show that the present invention can effectively predict road subsidence trends and achieve real-time risk warning.

[0093] Finally, a specific application scenario of the method in this embodiment is provided: (I) Construction of Data Physics Driven Model (DPD) in Fiber Settlement Inversion.

[0094] To address the challenge of directly and accurately inverting uneven subgrade settlement using distributed strain measured by BOTDA, this invention draws upon the data-physical-driven (DPD) approach based on the Kirchhoff–Helmholtz equations in the acoustic field. A DPD-based settlement inversion model suitable for road structures is constructed. The model uses the soil mechanics governing equations as physical constraints and employs a deep neural network to capture the mapping relationship between the BOTDA strain field and the settlement field, achieving high-precision inversion.

[0095] 1. Physical governing equations for roadbed-structure settlement Under the assumption of small deformation, the soil satisfies the equilibrium equations:

[0096] And it has a constitutive relation:

[0097] in: For stress tensor, For strain tensor, For the material elasticity matrix, This is a physical activity item.

[0098] BOTDA measures minute strain along the fiber axial direction:

[0099] Where n is the unit vector of the fiber direction.

[0100] Settlement variable can be expressed as vertical displacement w(x), and from geometric relationships we have:

[0101] 2. Data-Physical Fusion Inversion Network (DPD-SRNet).

[0102] By integrating BOTDA to collect strain εf(x), a deep convolutional network CNN is constructed to learn the mapping from strain to settlement w(x). At the same time, the above physical equations are embedded in the loss function so that the network training process follows the laws of real mechanics.

[0103] The loss function is constructed as follows:

[0104] in: Data items:

[0105] Physical constraint terms (residuals of governing equations):

[0106] Fiber continuity constraints:

[0107] By incorporating the real physical equations into the loss function, the model can maintain stable inversion performance even with insufficient data and high noise.

[0108] (II) Construction of a settlement prediction model based on transfer learning (TL).

[0109] Road settlement exhibits significant temporal evolution characteristics and site-specific variations. This invention combines Temporal Prediction Network (TCN / LSTM) with transfer learning to form a DPD–TL combined model, enabling model transfer and rapid deployment across different road segments.

[0110] 1. Time Series Prediction Network Settlement sequence obtained by inversion (t), construct a temporal convolutional network (TCN) to predict subsidence over a future period:

[0111] 2. Transfer Learning Framework Similar to the DPDT method described in the document, which uses a "frozen feature extraction layer + fine-tuned task layer," this invention transfers mature large-sample road segment models to new target road segments: Freeze low-level network parameters: Preserve common characteristics of foundation mechanics and settlement evolution.

[0112] Only minor adjustments are made to the top prediction layer to adapt to differences in soil quality, load environment, etc., of new road sections.

[0113] The transferred model requires only 5–10% of the samples to achieve predictive capabilities close to the original model, significantly reducing engineering experimentation costs. This idea draws inspiration from the transfer strategy of the DPDT model in the document.

[0114] (III) Road field test verification.

[0115] In this embodiment, a newly built test section of a road in the starting area of ​​Jinan City is selected as the implementation object. Sensing optical fibers are laid longitudinally along the road on the right side of the test road section, close to the ground surface.

[0116] The optical fiber is first fixed to the ground surface with adhesive tape, and then manually covered with a thin layer of cement slurry for sealing to enhance adhesion and ensure that environmental disturbances and structural strain can be effectively transmitted to the sensing fiber. During the deployment process, the fiber optic cables are buried at a consistent depth and are tightly fitted to avoid local loosening or warping, in order to ensure the accuracy and continuity of strain measurements.

[0117] The road section is approximately 200 meters long, and its structure includes an asphalt concrete surface layer, a cement-stabilized crushed stone base layer, and a plain fill subgrade.

[0118] This system achieves high-precision, long-distance, and continuous monitoring of road settlement by deploying distributed fiber optic sensing equipment. The optical fibers are buried at a consistent depth along the road centerline at the bottom of the roadbed. They are initially secured with tape and then encapsulated with a thin layer of cement slurry to ensure coordinated deformation between the fiber optics and the roadbed. The total length of the fiber optic cables is 200 meters, with a spatial sampling interval of 0.1 meters, and a total of 2000 measuring points are set up.

[0119] 1. Verification of inversion accuracy.

[0120] The settlement was inverted using DPD–SRNet based on the measured strain of BOTDA. By using physical equation constraints, the accumulation of errors caused by the "black box" of deep networks was avoided. Even with low sampling density and noise, the accuracy remained stable. Compared with the results of total station and leveling surveys, the settlement inversion error was ≤1.2mm, which is about 35-60% higher than that of traditional polynomial fitting and finite element inversion.

[0121] 2. Settlement prediction for the next 7 days.

[0122] The TCN model based on transfer learning predicts future settlement development with a prediction error (7 days) of ≤6%. The model can be transferred from the source road segment to the new road segment and only requires a small amount of calibration data. At the same time, the prediction performance fluctuates little with the settlement stage and is suitable for early warning.

[0123] 3. Identification of settlement anomalies.

[0124] Using deep networks, it can automatically identify abnormal patterns such as settlement abrupt changes, piping, and local buckling of soft soil, with an anomaly identification rate of 94%. It can quickly distinguish between "normal settlement trend" and "structural failure-level settlement" and provide risk warnings with a lead time of 30–48 hours.

[0125] After baseline establishment and temperature compensation are completed, the multi-time-period strain sequences are input into the proposed DPD (Data Physics Driven Model) and TL (Transfer Learning) joint modeling framework. The obtained settlement inversion results are compared with the measured settlement points, and the average error is controlled within ≤1.2 mm, which is significantly better than the traditional strain integration method (the error is often 3–8 mm).

[0126] Furthermore, the model was used to predict the settlement development trend over the next 7 days, with the trend error controlled within ≤6%, indicating that the present invention has the ability to effectively predict the long-term settlement evolution behavior in complex geotechnical-road systems.

[0127] Simultaneously, a distributed fiber optic intelligent sensing and early warning platform was established. By integrating various advanced technologies, it enables comprehensive real-time monitoring of the road environment. Through the combination of high-precision sensors and intelligent algorithms, it achieves accurate identification and rapid response to potential risks. Its modular design allows for flexible adjustments to meet different scenario requirements, while supporting multi-terminal data synchronization to ensure efficient and consistent information transmission.

[0128] In addition, the platform has powerful data storage and analysis capabilities, which can provide reliable data support for subsequent traffic optimization.

[0129] Example 2 This embodiment discloses a road settlement prediction system based on a data physics-driven model.

[0130] A road settlement prediction system based on a data physics-driven model includes: The strain data acquisition module is configured to acquire the raw strain data of the road to be monitored. The physical constraint module is configured to: construct physical constraint equations for the road with respect to strain and settlement, and introduce these physical constraint equations into the loss function of the deep learning prediction network as physical constraint terms; The model training module is configured to: train a deep learning prediction network using historical strain data and actual settlement values ​​of the road to be monitored, so that the deep learning prediction network learns the mapping relationship between strain and settlement that conforms to the physical constraint equation, and obtains a trained data physics-driven model. The settlement prediction module is configured to input raw strain data into a trained data physics-driven model to predict the road settlement curve of the road to be monitored. Example 3 The purpose of this embodiment is to provide a computer-readable storage medium.

[0131] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps in the road settlement prediction method based on a data physics-driven model as described in Embodiment 1 of this disclosure.

[0132] Example 4 The purpose of this embodiment is to provide an electronic device.

[0133] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the road settlement prediction method based on a data physics-driven model as described in Embodiment 1 of this disclosure.

[0134] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0135] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0136] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A road settlement prediction method based on a data physics-driven model, characterized in that, Includes the following steps: Obtain the raw strain data of the road to be monitored; We construct physical constraint equations for the road from strain to settlement, and introduce these physical constraint equations into the loss function of the deep learning prediction network as physical constraint terms. A deep learning prediction network is trained using historical strain data and actual settlement values ​​of the road to be monitored. This allows the deep learning prediction network to learn the mapping relationship between strain and settlement that conforms to the physical constraint equation, resulting in a well-trained data physics-driven model. The raw strain data is input into the trained data physics-driven model to predict the road settlement curve of the road to be monitored.

2. The road settlement prediction method based on a data physics-driven model as described in claim 1, characterized in that, Raw strain data of the road to be monitored is acquired using distributed optical fibers.

3. The road settlement prediction method based on a data physics-driven model as described in claim 1, characterized in that, The physical constraint equations specifically include: Introducing the strain-settlement integral relationship: ； Where w(x) represents the vertical settlement of the road at location x; ε(ξ) represents the axial microstrain at a point ξ along the road; ∫0 x … dξ represents the spatial integration from the starting point 0 of the road to position x; ξ is the integration variable, representing the spatial position along the road; or, The physical constraint equations are based on the mechanical assumption of deformation and strain continuity, connecting strain and settlement. They indicate that the settlement at a point is not determined solely by the strain at that point, but by the cumulative effect of all strains from the starting point to the current point.

4. The road settlement prediction method based on a data physics-driven model as described in claim 1, characterized in that, The deep learning prediction network employs a U-Net convolutional neural network structure, taking strain data as input and outputting predicted settlement curves. and the output that satisfies the physical mechanism constraints .

5. The road settlement prediction method based on a data physics-driven model as described in claim 3, characterized in that, The loss function of the deep learning prediction network is: ； Where L represents the total loss value; w represents the actual settlement value; P(w) represents the settlement value predicted by the model; P(w) represents a physical operation or transformation P applied to the actual settlement w, which is an intermediate physical quantity derived from the actual settlement w through physical laws; P is the strain-settlement integral relationship. This represents the intermediate physical quantity predicted by the model, corresponding to the result of the physical operation P; || w - ||² represents the data fitting term; || P(w) - p||² represents the physical constraint term.

6. The road settlement prediction method based on a data physics-driven model as described in claim 1, characterized in that, It also includes multi-scenario road settlement prediction based on transfer learning, specifically including: Determine the source and target domains; The data physics-driven model is trained using source domain data to obtain pre-trained weights that include basic road settlement patterns; Transfer the data physics-driven model trained in the source domain to the target domain; In the target domain, some network layers are frozen, the low-level feature extraction layer is fixed, and the high-level semantic layer is fine-tuned using only the road data in the target domain to improve the generalization ability. By fine-tuning the high-level semantic layer in the target domain, a cross-scenario settlement prediction model is obtained, enabling the prediction of settlement for different types of roads.

7. The road settlement prediction method based on a data physics-driven model as described in claim 1, characterized in that, This also includes classifying the settlement risk level of the roads to be monitored, specifically including: A road settlement risk index system is constructed. Based on the road settlement curve of the road to be monitored, parameters such as strain gradient, settlement rate, cumulative settlement, and daily rate of change acceleration factor are extracted to construct a settlement sensitive feature vector F. A settlement risk level model is constructed by inputting the settlement-sensitive feature vector F into the settlement risk level model to obtain the risk classification results; The settlement risk level model is the XGBoost algorithm model.

8. A road settlement prediction system based on a data physics-driven model, characterized in that, include: The strain data acquisition module is configured to acquire the raw strain data of the road to be monitored. The physical constraint module is configured to: construct physical constraint equations for the road with respect to strain and settlement, and introduce these physical constraint equations into the loss function of the deep learning prediction network as physical constraint terms; The model training module is configured to: train a deep learning prediction network using historical strain data and actual settlement values ​​of the road to be monitored, so that the deep learning prediction network learns the mapping relationship between strain and settlement that conforms to the physical constraint equation, and obtains a trained data physics-driven model. The settlement prediction module is configured to input raw strain data into a trained data physics-driven model to predict the road settlement curve of the road to be monitored.

9. A computer-readable storage medium having a program stored thereon, characterized in that, When executed by a processor, the program implements the steps in the road settlement prediction method based on a data physics-driven model as described in any one of claims 1-7.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the road settlement prediction method based on a data physics-driven model as described in any one of claims 1-7.

Images