Saline frozen soil roadbed settlement prediction method and system

Abstract

The invention discloses a salinized frozen soil roadbed settlement prediction method and system, and relates to the technical field of roadbed settlement prediction, and the method comprises the steps: obtaining multi-source monitoring data of a target salinized frozen soil roadbed, the multi-source monitoring data comprises temperature, moisture content, conductivity, stress and ground surface settlement time sequence observation values distributed along the depth, and a ground surface remote sensing image; establishing a physical mechanism model reflecting the heat-water-salt-force coupling effect of the salted frozen soil based on the multi-source monitoring data; integrating the physical mechanism model as a constraint condition into a neural network training process, and constructing a settlement response prediction sub-model with physical consistency in combination with multi-source monitoring data; performing spatial reconstruction on the multi-source monitoring data to form a two-dimensional profile image containing a temperature field, a humidity field and a salt field, and inputting the two-dimensional profile image into a convolutional neural network to extract spatial structure features; and constructing graph structure data by utilizing spatial structure characteristics, modeling a mutual influence relationship among different salinized areas through a graph neural network, and outputting settlement tendency distribution.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of subgrade settlement prediction, in particular to a salted frozen soil subgrade settlement prediction method and system. BACKGROUND

[0002] The subgrade settlement prediction technology refers to a technology for quantitatively predicting the vertical deformation of a subgrade of a traffic infrastructure such as a road or a railway under the action of external load, environmental change and material evolution by establishing a mathematical model or a calculation method. The core goal of the subgrade settlement prediction technology is to identify potential settlement risks in advance, evaluate the stability of the structure and provide a scientific basis for maintenance decision-making. Therefore, how to use advanced technical means to improve the intelligent level and safety of subgrade settlement prediction has become one of the problems to be solved at present.

[0003] In the field of subgrade settlement prediction, the existing salted frozen soil subgrade settlement prediction technology mainly relies on empirical statistical models or simplified numerical simulation methods. There are prominent problems such as insufficient description of the influence mechanism of salt, incomplete consideration of multi-field coupling, poor model generalization ability, etc. Moreover, the traditional method treats salted soil as ordinary frozen soil, ignoring the influence of salt on key physical processes such as freezing point reduction, unfrozen water content increase, frost heaving inhibition and long-term corrosion weakening, resulting in a serious deviation of the settlement prediction results from the actual situation under a salted environment. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides a salted frozen soil subgrade settlement prediction method to solve the problem that the existing salted frozen soil subgrade settlement prediction technology mainly relies on empirical statistical models or simplified numerical simulation methods, has prominent problems such as insufficient description of the influence mechanism of salt, incomplete consideration of multi-field coupling, poor model generalization ability, etc., and the traditional method treats salted soil as ordinary frozen soil, ignoring the influence of salt on key physical processes such as freezing point reduction, unfrozen water content increase, frost heaving inhibition and long-term corrosion weakening, resulting in a serious deviation of the settlement prediction results from the actual situation under a salted environment.

[0006] To solve the above technical problems, the present application provides the following technical solutions: In a first aspect, the present application provides a salted frozen soil subgrade settlement prediction method, which comprises: obtaining multi-source monitoring data of a target salted frozen soil subgrade, the multi-source monitoring data comprising time series observation values of temperature, water content, electrical conductivity, stress and surface settlement along the depth distribution and a surface remote sensing image; establishing a physical mechanism model reflecting the heat-water-salt-force coupling of the salted frozen soil based on the multi-source monitoring data; The physical mechanism model is taken as a constraint condition and is integrated into a neural network training process, multi-source monitoring data are combined to construct a settlement response prediction sub-model with physical consistency; Multi-source monitoring data are subjected to spatial reconstruction to form two-dimensional profile images containing temperature field, humidity field and salt field, and are input into a convolutional neural network to extract spatial structure features; The spatial structure features are used to construct graph structure data, the mutual influence relationship between different salted areas is modeled through a graph neural network, and a settlement tendency distribution is output; The settlement response prediction sub-model and the settlement tendency distribution are fused to form a joint prediction framework, and after pre-training in a non-salted frozen soil area, the model is migrated to a target road section, and model adaptation is realized by using limited measured data; The joint prediction framework after adaptation is used to generate subgrade settlement prediction results, and model dynamic updating and risk warning output are completed by continuously accessing new monitoring data.

[0007] As a preferred scheme of the salted frozen soil subgrade settlement prediction method, the physical mechanism model reflecting the heat-water-salt-force coupling of the salted frozen soil is established based on the multi-source monitoring data, and the specific steps are: According to the calibration relationship between the measured conductivity data and the salt mass ratio, the conductivity observation value along the depth distribution is converted into the salt mass ratio, wherein the salt mass ratio represents the mass fraction of soluble salt in unit mass of soil; Combined with the experimental response relationship of temperature and unfrozen water content, an unfrozen water function adjusting the inflection point position with the change of salt mass ratio is constructed; The latent heat release term of phase change is introduced into the heat conduction control equation , the expression is: ; Wherein, is the latent heat of ice melting, is the unfrozen water content, is the time; A freeze-out supply term driven by the temperature gradient is added to the water migration equation, and the weakening effect of salt on capillary potential is considered, and a modified water flux expression is constructed, the expression is: ; Wherein, is the water flux, is the self-diffusion coefficient, is the moisture content gradient, is the thermal diffusivity, is the temperature gradient, is the salt inhibition coefficient, is the salt mass ratio; A stiffness degradation factor that decreases with the number of freeze-thaw cycles is introduced into the stress-strain relationship to construct a compression strain increment expression, which is: ; wherein, is the compression strain increment, is the compression index, is the effective stress increment, is the current effective stress, is the stiffness degradation factor, is the number of freeze-thaw cycles; The above modified thermal, water and force control equations are combined to form a physical mechanism model.

[0008] As a preferred scheme of the salted frozen soil subgrade settlement prediction method, the physical mechanism model is incorporated into the neural network training process as a constraint condition, and a settlement response prediction sub-model with physical consistency is constructed by combining multi-source monitoring data, specifically: A feedforward neural network is built, with spatial coordinates and time as inputs, and temperature, moisture content and vertical displacement at the corresponding position as outputs; The residual errors generated at each space-time point after the neural network output is substituted into the physical mechanism model are calculated, including thermal equilibrium residual error, water conservation residual error and stress balance residual error; The residual errors are weighted and summed to form a physical loss term, which is expressed as: ; wherein, is the physical loss term, is the thermal equilibrium residual error, is the water conservation residual error, is the stress balance residual error, 、 、 is the weight coefficient; The mean square error between the neural network output value and the measured temperature, moisture content and settlement data is calculated at the same time to form a data fitting loss term; The physical loss term and the data fitting loss term are combined into a compound loss function, and the compound loss function is minimized through an optimization algorithm to complete the training of the settlement response prediction sub-model; wherein, the expression of the compound loss function is: ; wherein, is the compound loss function, is the data fitting loss term, is the physical constraint strength factor.

[0009] As a preferred scheme of the salted frozen soil roadbed settlement prediction method, the multi-source monitoring data is spatially reconstructed to form a two-dimensional profile image containing a temperature field, a humidity field and a salt field, and is input into a convolutional neural network to extract spatial structure features, specifically: The temperature, moisture content and salt mass ratio data of the multiple monitoring layers arranged along the depth are arranged in the vertical direction to generate a two-dimensional grid matrix with spatial coordinates; The two-dimensional grid matrix is subjected to interpolation processing to improve the spatial resolution and obtain continuous distribution of the temperature field, humidity field and salt field; The three types of field quantities are superimposed as a three-dimensional image tensor with the first dimension corresponding to the depth direction, the second dimension corresponding to the time direction, and the third dimension corresponding to the physical field type; The image tensor is input into a convolutional neural network with an encoder-decoder structure to extract local gradient, frozen front position and salt enrichment zone morphological features step by step through multiple convolution kernels; A low-dimensional feature map is output, wherein each spatial unit corresponds to the state representation of a certain section in the roadbed profile.

[0010] As a preferred scheme of the salted frozen soil roadbed settlement prediction method, the spatial structure features are used to construct a graph structure data, and a graph neural network is used to model the mutual influence relationship between different salted areas, and a settlement tendency distribution is output, specifically: Each spatial unit in the feature map is regarded as a graph node to construct an undirected graph, wherein the node set is composed of all feature units, and the edge set is composed of node pairs that are adjacent or have thermal correlation; The edge weight is set according to the Euclidean distance between nodes and the physical property difference, which is used to reflect the difficulty of heat transfer and water migration, and the expression is: ; Wherein, is the edge weight between nodes and nodes , is the Euclidean distance between two nodes, is the absolute value of the difference between the salt mass ratios of two nodes, , is the attenuation parameter; The feature vector corresponding to the node is taken as the initial state, and a graph attention network is used for message passing, and the expression is: ; Wherein, is the updated state vector of node , is the neighbor set of node , is the attention coefficient, is a learnable weight matrix, is a state vector of a neighbor node is an activation function.

[0011] After multiple rounds of iterative updates, the final node state is mapped to a settlement tendency score, and the settlement tendency distribution graph of the entire profile is generated by splicing.

[0012] As a preferred scheme of the salted frozen soil roadbed settlement prediction method, the settlement response prediction sub-model is fused with the settlement tendency distribution to form a joint prediction framework, and after pre-training in a non-salted frozen soil area, it is migrated to the target road section, and the model is adapted using limited measured data, specifically: The time series settlement value output by the settlement response prediction sub-model is aligned with the settlement tendency distribution graph in the spatial dimension to form a dual-channel prediction result; A fusion module is designed to linearly combine the two prediction results through a learnable weight; Wherein, the fusion process expression is: ; Wherein, is the fused settlement prediction value, is the time series settlement value output by the settlement response prediction sub-model, is the spatial average value of the settlement tendency distribution graph, , is a learnable fusion weight; A large amount of measured data in a non-salted frozen soil area is used to train the fusion module and its sub-model as a whole to obtain a basic prediction model with general freeze-thaw response capability; The basic prediction model is migrated to the target salted frozen soil road section, the bottom layer parameters of the convolutional neural network and the graph neural network are fixed, and only the fusion weight and the output layer bias are fine-tuned; The measured settlement data in the limited period of the target road section is used to complete parameter adjustment to obtain a customized prediction model.

[0013] As a preferred scheme of the salted frozen soil roadbed settlement prediction method, the settlement prediction result of the roadbed is generated based on the adapted joint prediction framework, and the model is dynamically updated and risk warning output is completed by continuously accessing new monitoring data, specifically: The monitoring data at the current time is taken as input to generate a daily settlement amount prediction sequence for the future period through the customized prediction model; The sliding slope of the prediction sequence is calculated as a settlement rate indicator; ​Set three early warning thresholds, when the settlement rate exceeds the first threshold, output yellow warning, exceeds the second threshold, output orange warning, exceeds the third threshold, output red warning; Push the early warning signal to the operation and maintenance terminal together with the prediction curve; Every fixed period, automatically inject the latest measured data, recalculate the model prediction residual, if the residual continuously exceeds the set tolerance, trigger the model retraining process, ensure the long-term stability of the prediction accuracy.

[0014] In the second aspect, the application provides a salted frozen soil roadbed settlement prediction system, comprising: The data acquisition module, the physical modeling module, the neural network training module, the spatial reconstruction module, the graph structure modeling module, the model fusion and adaptation module, and the dynamic prediction output module; The data acquisition module is used to acquire temperature, moisture content, electrical conductivity, stress and surface settlement time series data in the depth direction of the roadbed, and synchronously acquire surface remote sensing images; The physical modeling module is used to establish a physical mechanism model reflecting the coupling of heat, water, salt and force, including a salt-corrected unfrozen water function, a phase change latent heat term, a corrected water flux and a stiffness degradation compression relationship; The neural network training module is used to take the control equation residual of the physical mechanism model as a constraint term, combine the measured data fitting error to construct a composite loss function, and train to generate a settlement response prediction submodel with physical consistency; The spatial reconstruction module is used to reconstruct the monitoring data into a two-dimensional profile image containing temperature, humidity and salt fields, and input a convolutional neural network to extract freezing front, gradient zone and enrichment morphological features; The graph structure modeling module is used to construct a graph structure based on the extracted features, model the interaction between regions through a graph attention network, and output a spatially continuous settlement tendency distribution; The model fusion and adaptation module is used to fuse the settlement response submodel and the settlement tendency distribution to form a joint prediction framework, migrate to the target road section after pre-training in the non-saline area, and complete model adaptation using limited data; The dynamic prediction output module is used to generate a future settlement prediction sequence, calculate the settlement rate and trigger a hierarchical warning, periodically inject new data to evaluate the residual, and start retraining when the limit is exceeded, to realize dynamic updating and risk output of the model.

[0015] In the third aspect, the application provides a computer device comprising a memory and a processor, the memory storing a computer program, wherein the computer program is executed by the processor to implement any step of the salted frozen soil roadbed settlement prediction method according to the first aspect of the application.

[0016] In a fourth aspect, the present application provides a computer readable storage medium having stored thereon a computer program, wherein the computer program, when executed by a processor, implements any step of the salt frozen soil subgrade settlement prediction method according to the first aspect of the present application.

[0017] The present application has the beneficial effects that: by fusing the physical mechanism of salt frozen soil and the intelligent learning model, a settlement prediction method with physical consistency and spatial perception ability is constructed, the precise modeling of the settlement behavior under the conditions of salt inhibition of frost heaving, heterogeneous distribution and small sample scenes is realized, the technical path of combining physical information neural network and graph structure modeling is adopted, the explainability, generalization ability and engineering applicability of the prediction model are improved, and the problem of insufficient prediction accuracy of the traditional method under the conditions of complex environment adaptability and data scarcity is solved. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0019] Figure 1 The flowchart of the salt frozen soil subgrade settlement prediction method in embodiment 1.

[0020] Figure 2 The schematic diagram of the salt frozen soil subgrade settlement prediction system in embodiment 1. DETAILED DESCRIPTION

[0021] In order to make the above-mentioned purposes, features and advantages of the present application more apparent and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the drawings of the specification.

[0022] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, but the present application can also be implemented in other ways different from those described herein, and those skilled in the art can make similar generalizations without departing from the connotation of the present application, therefore the present application is not limited to the specific embodiments disclosed below.

[0023] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an independent or alternative embodiment that excludes other embodiments.

[0024] Embodiment, refer to Figure 1 and Figure 2This embodiment of the invention provides a method for predicting settlement of roadbeds in saline-frozen soil, comprising the following steps: S1. Acquire multi-source monitoring data of the target saline-frozen soil subgrade. The multi-source monitoring data includes time-series observations of temperature, moisture content, electrical conductivity, stress and surface settlement along the depth, as well as surface remote sensing images. Furthermore, multi-source monitoring data is acquired collaboratively with a multi-parameter profile sensor array deployed at typical cross-sections of the target roadbed and a surface remote sensing device. Among them, the profile sensor array is arranged with multiple monitoring layers at equal or variable intervals along the depth direction. Each layer integrates a temperature sensor, a dielectric moisture content probe, a four-electrode conductivity probe and a soil pressure cell to synchronously collect temperature-humidity-salt-force coupled data at each depth point. Surface subsidence observation is achieved through continuous tracking using hydrostatic level, total station or synthetic aperture radar remote sensing technology; The surface remote sensing images were acquired by UAVs equipped with thermal infrared cameras and hyperspectral imagers, and were used to invert the spatial pattern of surface temperature field distribution and salinization index. After all monitoring data are timestamped and registered with coordinates, a multi-source heterogeneous dataset with spatiotemporal consistency is constructed, providing a data foundation for subsequent physical modeling and intelligent analysis. It should be noted that the multi-parameter profile sensor array is deployed to cover typical roadbed defect sections, and the sensor types are environmentally adaptable and can work stably for a long time under high salt, low temperature, and freeze-thaw cycle conditions. Timestamp alignment uses a unified time synchronization system to ensure that the data from each source are synchronized within the second level, and coordinate registration is achieved by correcting the spatial position with geographic reference points to achieve millimeter-level matching. This ensures strict correspondence of multi-source data in the spatiotemporal dimension and avoids distortion in subsequent modeling due to misalignment.

[0025] S2. Establish a physical mechanism model based on multi-source monitoring data to reflect the coupling effect of heat-water-salt-force in saline-frozen soil; Furthermore, based on the calibration relationship between measured electrical conductivity data and salt mass ratio, the electrical conductivity observations distributed along the depth are converted into salt mass ratio, where the salt mass ratio represents the mass fraction of soluble salts per unit mass of soil. Based on the experimental response relationship between temperature and unfrozen water content, an unfrozen water function is constructed that adjusts the inflection point position according to the change in salt mass ratio. Introducing a latent heat release term from phase change into the heat conduction governing equation The expression is: ; in, The latent heat of melting ice, This refers to the unfrozen water content. For time; A modified water flux expression is constructed by adding a freezing-point depression term driven by temperature gradient to the water migration equation, considering the weakening effect of salt on capillary potential, and the expression is: ; wherein, is water flux, is self-diffusion coefficient, is water content gradient, is thermal diffusion coefficient, is temperature gradient, is salt inhibition coefficient, is salt mass ratio; A stiffness degradation factor that decreases with the number of freeze-thaw cycles is introduced into the stress-strain relationship to construct a compression strain increment expression, and the expression is: ; wherein, is compression strain increment, is compression index, is effective stress increment, is current effective stress, is stiffness degradation factor, is number of freeze-thaw cycles; The above modified thermal, water and force control equations are coupled to form a physical mechanism model; It should be noted that the unfrozen water function in the physical mechanism model is calibrated based on indoor salted soil freeze-thaw experiment data, the inflection point temperature of which is nonlinearly left shifted with the salt mass ratio, truly reflecting the inhibitory effect of salt on freezing point, and the stiffness degradation factor is determined by the accelerated freeze-thaw cycle test to ensure the mechanical response of the model under long-term service conditions has engineering credibility, and all control equations satisfy the basic principles of mass conservation, energy conservation and momentum balance, forming a coupled system with clear physical meaning.

[0026] S3, the physical mechanism model is integrated into the neural network training process as a constraint condition, and a settlement response prediction sub-model with physical consistency is constructed by combining multi-source monitoring data; Further, a feedforward neural network is built, with spatial coordinates and time as inputs, and temperature, water content and vertical displacement at corresponding positions as outputs; The residuals generated by the neural network output after being substituted into the physical mechanism model at each space-time point include thermal equilibrium residual, water conservation residual and stress balance residual; The residuals are weighted and summed to form a physical loss term, and the expression is: ; wherein, is physical loss term, is a heat balance residual, is a water conservation residual, is a stress balance residual, , , is a weight coefficient; The mean square error between the neural network output value and the measured temperature, moisture content and settlement data is calculated at the same time to form a data fitting loss term; The physical loss term and the data fitting loss term are combined into a composite loss function, and the composite loss function is minimized through an optimization algorithm to complete the training of the settlement response prediction sub-model; wherein the expression of the composite loss function is: ; wherein, is a composite loss function, is a data fitting loss term, is a physical constraint strength factor; It should be noted that the output variables of the feedforward neural network strictly correspond to the state variables in the physical mechanism model, the residual calculation is performed in the continuous space-time domain, rather than only at discrete monitoring points, thereby ensuring the globality of the physical constraints, the weight coefficient is determined by the dimensionless normalization method, avoiding the optimization bias caused by the different physical quantity levels, and the construction of the composite loss function enables the model to maintain a reasonable prediction direction relying on the physical law when a large number of measured labels are lacking, thereby improving the generalization ability in the small sample scenario.

[0027] S4, spatial reconstruction is performed on the multi-source monitoring data to form two-dimensional profile images containing temperature field, humidity field and salt field, and the images are input into a convolutional neural network to extract spatial structure features; Further, the temperature, moisture content and salt mass ratio data of the multiple monitoring layers arranged along the depth are arranged in the vertical direction to generate a two-dimensional grid matrix with spatial coordinates; The two-dimensional grid matrix is subjected to interpolation processing to improve the spatial resolution and obtain continuously distributed temperature field, humidity field and salt field; The three types of field quantities are superimposed as a three-dimensional image tensor, wherein the first dimension corresponds to the depth direction, the second dimension corresponds to the time direction, and the third dimension corresponds to the physical field type; The image tensor is input into a convolutional neural network with an encoder-decoder structure, and local gradient, frozen front position and salt enrichment zone morphological features are extracted step by step through multiple convolution kernels; A low-dimensional feature map is output, wherein each spatial unit corresponds to the state representation of a certain section in the subgrade profile; It should be noted that the spatial resolution of the two-dimensional grid matrix is not less than half of the vertical spacing of the sensor, the interpolation method uses piecewise spline interpolation considering the discontinuity of frozen soil phase change, avoids false smoothing, the channel arrangement order of the three-dimensional image tensor is fixed, ensures that the convolution kernel learns stable cross-field correlation features, the encoder-decoder structure contains a skip connection, effectively preserves shallow detail information, and enables key interface features such as freezing fronts to be accurately expressed in low-dimensional feature maps.

[0028] S5, constructing a graph structure data by using a spatial structure feature, modeling mutual influence relationships between different salted regions by a graph neural network, and outputting a settlement tendency distribution; Further, each spatial unit in the feature map is regarded as a graph node, and an undirected graph is constructed, wherein a node set is composed of all feature units, and an edge set is composed of node pairs that are adjacent or have thermal correlation; The edge weight is set according to the Euclidean distance between nodes and the difference in physical properties, and is used to reflect the difficulty of heat transfer and water migration, and the expression is: ; wherein, is the edge weight between nodes and node , is the Euclidean distance between two nodes, is the absolute value of the difference in salt mass ratio between two nodes, , is a decay parameter; The feature vector corresponding to the node is taken as the initial state, and a graph attention network is used for message passing, and the expression is: ; wherein, is the updated state vector of node , is the neighbor set of node , is an attention coefficient, is a learnable weight matrix, is the state vector of neighbor node , is an activation function.

[0029] After multiple rounds of iterative updates, the final node state is mapped to a settlement tendency score, and the settlement tendency distribution map of the entire profile is generated by splicing; It should be noted that the connection of the edges in the graph structure is not limited to geometric adjacency. For nodes that have potential heat conduction paths but are not physically adjacent, remote edges are introduced according to the thermal conductivity and permeability coefficient of the soil layer to reflect non-local interaction. The attention coefficient is calculated through a multi-head mechanism to enhance the model's ability to capture complex influence relationships between heterogeneous salted subareas and avoid modeling bias caused by fixed weights in traditional graph networks.

[0030] S6, the settlement response prediction sub-model is fused with the settlement tendency distribution to form a joint prediction framework, and after pre-training in the non-saline frozen soil area, the joint prediction framework is migrated to the target road section to realize model adaptation using limited measured data; Further, the time series settlement value output by the settlement response prediction sub-model is aligned with the settlement tendency distribution in the spatial dimension to form a dual-channel prediction result. A fusion module is designed to linearly combine the two prediction results through learnable weights. The fusion process expression is: ; Wherein, is the fusion settlement prediction value, is the time series settlement value output by the settlement response prediction sub-model, is the spatial average value of the settlement tendency distribution, , is the learnable fusion weight. The fusion module and its sub-model are trained as a whole using a large amount of measured data in the non-saline frozen soil area to obtain a basic prediction model with general freeze-thaw response capability. The basic prediction model is migrated to the target saline frozen soil road section, the bottom layer parameters of the convolutional neural network and the graph neural network are fixed, and only the fusion weight and the output layer bias are fine-tuned. The measured settlement data in the limited period of the target road section is used to complete parameter adjustment to obtain a customized prediction model. It should be noted that the learnable weight of the fusion module has converged to a reasonable range during the pre-training phase, and the fine-tuning after migration only needs a small number of iterations to complete the adaptation, avoiding overfitting. The bottom layer parameter freezing strategy effectively preserves the general freeze-thaw response features learned from the non-saline area, while the top adjustable parameters focus on capturing salt-specific environment characteristics, achieving a balance between knowledge transfer and individual adaptation, and improving the deployment efficiency and reliability of the model in new road sections.

[0031] S7, based on the adapted joint prediction framework, a roadbed settlement prediction result is generated, and model dynamic updating and risk warning output are completed through continuous access to new monitoring data. Furthermore, using the current monitoring data as input, a customized prediction model is used to generate a daily settlement prediction sequence for future periods; Calculate the sliding slope of the predicted sequence as an indicator of settlement rate; Three warning thresholds are set: a yellow warning is issued when the settlement rate exceeds the first threshold, an orange warning is issued when it exceeds the second threshold, and a red warning is issued when it exceeds the third threshold. The warning signal, along with the prediction curve, is pushed to the operation and maintenance management terminal. The latest measured data is automatically injected at fixed intervals to recalculate the model prediction residuals. If the residuals continue to exceed the set tolerance, the model retraining process is triggered to ensure long-term stability of prediction accuracy. It should be noted that the length of the sliding slope calculation window is matched with the roadbed freeze-thaw cycle to avoid short-term noise interference with early warning judgment. The three-level early warning threshold is calibrated based on historical settlement events and engineering safety standards, and has clear engineering basis. The model retraining process includes data quality screening and residual attribution analysis to prevent abnormal data from misleading parameter updates and to ensure the long-term stability of the system and the credibility of the prediction results.

[0032] This embodiment also provides a settlement prediction system for saline-frozen soil subgrade, including: The module includes a data acquisition module, a physical modeling module, a neural network training module, a spatial reconstruction module, a graph structure modeling module, a model fusion and adaptation module, and a dynamic prediction output module. The data acquisition module is used to acquire temperature, moisture content, electrical conductivity, stress and surface settlement time series data in the depth direction of the roadbed, and simultaneously acquire surface remote sensing images; The physical modeling module is used to establish a physical mechanism model that reflects the coupling effect of heat, water, salt, and force. It includes the salt-corrected unfrozen water function, the latent heat of phase change term, the corrected water flux, and the stiffness degradation compression relationship. The neural network training module is used to use the residuals of the control equations of the physical mechanism model as constraints, combine the fitting error of the measured data to construct a composite loss function, and train to generate a settlement response prediction sub-model with physical consistency. The spatial reconstruction module is used to reconstruct the monitoring data into a two-dimensional profile image containing temperature, humidity, and salinity fields, and input it into a convolutional neural network to extract freezing fronts, gradient regions, and enriched morphological features. The graph structure modeling module is used to construct graph structures from extracted features, model the interactions between regions through a graph attention network, and output a spatially continuous distribution of settlement tendency. The model fusion and adaptation module is used to fuse the settlement response sub-model and the settlement tendency distribution to form a joint prediction framework. After pre-training in non-saline areas, it is transferred to the target road section and the model adaptation is completed using limited data. The dynamic prediction output module is used to generate future settlement prediction sequences, calculate settlement rates and trigger graded early warnings, while periodically injecting new data to evaluate residuals and initiating retraining when limits are exceeded, thereby realizing dynamic model updates and risk output.

[0033] This embodiment also provides a computer device applicable to the settlement prediction method for saline-frozen soil subgrade, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the settlement prediction method for saline-frozen soil subgrade as proposed in the above embodiment.

[0034] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0035] This embodiment also provides a storage medium on which a computer program is stored. When executed by a processor, the program implements the method for predicting settlement of saline-frozen soil subgrade as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0036] In summary, this invention constructs a settlement prediction method with physical consistency and spatial perception capabilities by integrating the physical mechanisms of saline permafrost with an intelligent learning model. It achieves accurate modeling of settlement behavior under conditions of salt-induced frost heave, heterogeneous distribution, and small sample size. The technical approach of combining physical information neural networks with graph structure modeling improves the interpretability, generalization ability, and engineering applicability of the prediction model, and solves the problem of insufficient prediction accuracy of traditional methods under complex environmental adaptability and data scarcity conditions.

[0037] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting the settlement of a salt-frozen soil embankment, characterized in that: The application relates to a method for predicting the settlement of a target saline frozen soil subgrade, and belongs to the technical field of subgrade settlement prediction. The method comprises the following steps: acquiring multi-source monitoring data of the target saline frozen soil subgrade, wherein the multi-source monitoring data comprises time-series observation values of temperature, water content, electrical conductivity, stress and ground settlement along the depth distribution and a ground remote sensing image; establishing a physical mechanism model reflecting the thermal-water-salt-force coupling of the saline frozen soil based on the multi-source monitoring data; integrating the physical mechanism model into a neural network training process as a constraint condition, combining the multi-source monitoring data to construct a settlement response prediction submodel with physical consistency; spatially reconstructing the multi-source monitoring data to form a two-dimensional profile image containing a temperature field, a humidity field and a salt content field, and inputting the two-dimensional profile image into a convolutional neural network to extract spatial structure features; constructing graph structure data using the spatial structure features, modeling the mutual influence relationship between different saline regions through a graph neural network, and outputting a settlement tendency distribution; fusing the settlement response prediction submodel and the settlement tendency distribution to form a joint prediction framework, and migrating the joint prediction framework to the target subgrade after pre-training in a non-saline frozen soil region, and realizing model adaptation by using limited measured data; 2. The method of predicting the settlement of a salt-frozen soil subgrade according to claim 1, characterized in that: generating a subgrade settlement prediction result based on the adapted joint prediction framework, and completing model dynamic updating and risk warning output by continuously inputting new monitoring data. The method for establishing the physical mechanism model reflecting the thermal-water-salt-force coupling of the saline frozen soil based on the multi-source monitoring data comprises the following steps: converting the electrical conductivity observation values along the depth distribution into salt mass ratios according to the calibration relationship between the measured electrical conductivity data and the salt mass ratios, wherein the salt mass ratio represents the mass fraction of soluble salt in unit mass of soil; Introducing the latent heat release term into the heat conduction control equation , the expression is: ； wherein, L is the latent heat of fusion of ice, W is the unfrozen water content, t is time; combining the experimental response relationship between temperature and unfrozen water content to construct an unfrozen water function that adjusts the inflection point position with the change of the salt mass ratio; ； wherein, is the water flux, is the self-diffusion coefficient, is the water content gradient, is the thermal diffusivity, is the temperature gradient, is the salinity suppression coefficient, is the salinity mass ratio; adding a freeze-out supply term driven by the temperature gradient to the water migration equation, and considering the weakening effect of salt on the capillary potential energy to construct a modified water flux expression, which is: ； wherein, is the compressive strain increment, is the compression index, is the effective stress increment, is the current effective stress, is the stiffness degradation factor, is the number of freeze-thaw cycles; introducing a stiffness degradation factor that decreases with the number of freeze-thaw cycles into the stress-strain relationship to construct a compression strain increment expression, which is:

3. The method of predicting the settlement of a salt-frozen soil subgrade according to claim 2, characterized in that: combining the above-mentioned modified thermal, water and force control equations to form the physical mechanism model. The method for integrating the physical mechanism model into the neural network training process as a constraint condition, combining the multi-source monitoring data to construct the settlement response prediction submodel with physical consistency comprises the following steps: building a feedforward neural network with spatial coordinates and time as inputs, and outputting the temperature, water content and vertical displacement of the corresponding position; calculating the residual error generated by the neural network output after being substituted into the physical mechanism model at each space-time point, including thermal equilibrium residual error, water conservation residual error and stress balance residual error; ; wherein, is a physical loss term, is a thermal balance residual, is a moisture conservation residual, is a stress balance residual, , , is a weight coefficient; weighting and summing the residual errors to form a physical loss term, which is: simultaneously calculating the mean square error between the neural network output value and the measured temperature, water content and settlement data to form a data fitting loss term; merging the physical loss term and the data fitting loss term into a compound loss function, and minimizing the compound loss function through an optimization algorithm to complete the training of the settlement response prediction submodel; wherein the expression of the compound loss function is: ； wherein, is a composite loss function, is a data fitting loss term, is a physical constraint strength factor.

4. The method of predicting the settlement of a salt-frozen soil subgrade according to claim 3, characterized in that: The multi-source monitoring data is spatially reconstructed to form a two-dimensional profile image containing temperature field, humidity field and salt field, and is input into a convolutional neural network to extract spatial structure features, specifically: The temperature, moisture content and salt mass ratio data of multiple monitoring layers arranged along the depth are arranged in the vertical direction to generate a two-dimensional grid matrix with spatial coordinates; The two-dimensional grid matrix is interpolated to improve the spatial resolution and obtain continuous distribution of temperature field, humidity field and salt field; The three types of field quantities are stacked as a three-dimensional image tensor with the first dimension corresponding to the depth direction, the second dimension corresponding to the time direction, and the third dimension corresponding to the physical field type; The image tensor is input into a convolutional neural network with an encoder-decoder structure, and local gradient, frozen front position and salt enrichment zone morphology features are extracted step by step through multiple convolution kernels; A low-dimensional feature map is output, wherein each spatial unit corresponds to the state representation of a certain section in the subgrade profile.

5. The method of predicting the settlement of a salt-frozen soil subgrade according to claim 4, characterized in that: The spatial structure features are used to construct graph structure data, and the mutual influence relationship between different salted areas is modeled through a graph neural network to output the settlement tendency distribution, specifically: Each spatial unit in the feature map is regarded as a graph node to construct an undirected graph, wherein the node set is composed of all feature units, and the edge set is composed of node pairs that are adjacent or have thermal correlation; The edge weight is set according to the Euclidean distance between nodes and the difference in physical properties, which reflects the difficulty of heat transfer and water migration, and the expression is: ; wherein, is the node is the edge weight between the nodes is the Euclidean distance between the two nodes, is the absolute value of the difference between the salt mass ratios of the two nodes, , is the decay parameter;​ The feature vector corresponding to the node is used as the initial state, and the graph attention network is used for message passing, and the expression is: ； in, For nodes The updated state vector, For nodes The neighborhood group, Attention coefficient For learnable weight matrix, Neighboring nodes The state vector, For activation functions; After multiple rounds of iteration and update, the final node state is mapped to the settlement tendency score to generate a settlement tendency distribution map for the entire profile.

6. The method of predicting the settlement of a salt-frozen soil subgrade according to claim 5, characterized in that: The settlement response prediction sub-model and the settlement tendency distribution are fused to form a joint prediction framework, and after pre-training in the non-saline frozen soil area, it is migrated to the target road section to realize model adaptation using limited measured data, specifically: The time series settlement value output by the settlement response prediction sub-model is aligned with the settlement tendency distribution map in the spatial dimension to form a two-channel prediction result; A fusion module is designed to linearly combine the two prediction results through learnable weights; The fusion process expression is: ； wherein, is the fused settlement prediction value, is the time series settlement value output by the settlement response predictor model, is the spatial average of the settlement tendency distribution map, , is the learnable fusion weight; A large amount of measured data in the non-saline frozen soil area is used to train the fusion module and its sub-model as a whole to obtain a basic prediction model with general freeze-thaw response capability; The basic prediction model is migrated to the target saline frozen soil road section, and the bottom layer parameters of the convolutional neural network and the graph neural network are fixed, and only the fusion weight and the output layer bias are fine-tuned; The measured settlement data in the target road section within a limited period is used to complete parameter adjustment to obtain a customized prediction model.

7. The method of predicting the settlement of a salt-frozen soil subgrade according to claim 6, characterized in that: The joint prediction framework based on the adapted model generates a subgrade settlement prediction result, and continuously inputs new monitoring data to complete model dynamic updating and risk warning output, specifically: The monitoring data at the current time is input to generate a daily settlement prediction sequence for the future period through the customized prediction model; The sliding slope of the prediction sequence is calculated as a settlement rate indicator; A three-level early warning threshold is set, and when the settlement rate exceeds the first threshold, a yellow early warning is output, when the settlement rate exceeds the second threshold, an orange early warning is output, and when the settlement rate exceeds the third threshold, a red early warning is output; The early warning signal is pushed to the operation and maintenance terminal together with the prediction curve; Every fixed period, the latest measured data is automatically injected, the model prediction residual is recalculated, if the residual continuously exceeds the set tolerance, the model retraining process is triggered to ensure long-term stability of the prediction accuracy.

8. A system for predicting the settlement of a permafrost subgrade based on the method for predicting the settlement of a permafrost subgrade according to any one of claims 1 to 7, characterized in that: It comprises: a data acquisition module, a physical modeling module, a neural network training module, a spatial reconstruction module, a graph structure modeling module, a model fusion and adaptation module, and a dynamic prediction output module; The data acquisition module is configured to acquire temperature, moisture content, electrical conductivity, stress, and surface settlement time series data in the depth direction of the roadbed, and synchronously acquire surface remote sensing images. The physical modeling module is configured to establish a physical mechanism model reflecting the coupling of heat, water, salt, and force, including a salt-corrected unfrozen water function, a phase change latent heat term, a corrected water flux, and a stiffness degradation compression relationship. The neural network training module is configured to use the residual of the control equation of the physical mechanism model as a constraint term, combine the measured data fitting error to construct a compound loss function, and train a settlement response prediction sub-model with physical consistency. The spatial reconstruction module is configured to reconstruct the monitoring data into a two-dimensional profile image containing temperature, humidity, and salt fields, and input the convolutional neural network to extract freezing front, gradient zone, and enrichment morphology features. The graph structure modeling module is configured to construct a graph structure based on the extracted features, model the interaction between regions through a graph attention network, and output a spatially continuous settlement tendency distribution. The model fusion and adaptation module is configured to fuse the settlement response sub-model and the settlement tendency distribution to form a joint prediction framework, pre-train in a non-saline area, and then migrate to the target road section to complete model adaptation using limited data. The dynamic prediction output module is configured to generate a future settlement prediction sequence, calculate the settlement rate and trigger a hierarchical early warning, periodically inject new data to evaluate the residual, and start retraining when the residual exceeds the limit to achieve dynamic model updating and risk output.

9. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that: The processor executes the computer program to realize the steps of the salted frozen soil roadbed settlement prediction method of any one of claims 1-7.

10. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by the processor to realize the steps of the salted frozen soil roadbed settlement prediction method of any one of claims 1-7.

Images