Method for predicting hierarchical spatio-temporal correlation of embankment in permafrost region based on graph neural network

Abstract

The application discloses a permafrost area roadbed layered space-time correlation prediction method based on a graph neural network. The method comprises the following steps: providing multi-dimensional time series monitoring data; establishing horizontal direction space adjacency edges and vertical direction interlayer coupling edges to generate a weighted graph with weights; cleaning, aligning and standardizing the data to combine into a node feature tensor; performing graph convolution operation based on the weighted graph to obtain a spatial feature sequence; performing time coding operation to obtain a space-time coupled feature; obtaining an evolution law to obtain a roadbed deformation prediction result; and constructing a loss function based on the difference between the roadbed deformation prediction result and a real settlement value and the deviation degree of the roadbed deformation prediction result from a physical law to perform parameter adjustment. The application simultaneously utilizes coupling information of space-interlayer-time, significantly improves the permafrost area highway roadbed deformation prediction accuracy and advance, and is suitable for a real-time monitoring and early warning system.

Description

Technical Field

[0001] This invention relates to the field of road engineering monitoring and disaster early warning technology in cold regions, and in particular to a method for predicting the spatiotemporal correlation of roadbed stratification in permafrost regions based on graph neural networks. Background Technology

[0002] In permafrost regions, highway subgrades are subjected to intense freeze-thaw cycles and complex climatic loads over long periods, resulting in a complex deformation process involving frost heave, thaw settlement, and creep. Monitoring and providing early warning of these deformations are crucial for ensuring road safety and extending road lifespan. Existing monitoring systems largely rely on discrete observation methods such as manual leveling and single-point temperature or displacement sensors, offering relatively limited monitoring dimensions and failing to comprehensively characterize the layered coupling mechanisms of the surface layer, active layer, and permafrost layer. Especially in road sections where frost heave and thaw settlement alternate frequently, the dynamic evolution of the deep active layer and underlying permafrost layer is often overlooked, leading to blind spots in understanding the overall subgrade stability.

[0003] With the development of new monitoring technologies such as distributed optical fiber, satellite radar interferometry (InSAR), and ground resistivity tomography, acquiring multi-source, multi-scale data is no longer a bottleneck. However, significant differences in data formats and sampling frequencies exist between different monitoring methods, and the lack of a unified fusion framework leads to isolated information and fragmented analysis. Traditional statistical regression or single-point time series models can only process local data and lack the ability to express spatial correlations between monitoring points and layers, making it difficult to reveal the overall evolution of roadbed deformation in permafrost regions. Under complex geological structures and extreme climatic conditions, these models are prone to misjudgments or omissions, thereby reducing the reliability of early warning systems.

[0004] Furthermore, most existing data-driven methods fail to effectively integrate the thermo-hydraulic-mechanical coupling mechanism of permafrost, lacking physical consistency constraints on model outputs. When encountering new operating conditions or extreme weather events exceeding historical observation ranges, predictions often deviate from actual physical processes, resulting in limited extrapolation capabilities. In addition, communication conditions are limited in high-altitude and remote areas, leading to offline model updates and inferences, which lack real-time performance and adaptive capabilities, making it difficult to provide timely and accurate early warnings in the initial stages of accelerated deformation.

[0005] In summary, there is an urgent need for a novel technical solution that can integrate multi-source hierarchical monitoring data, explicitly express spatial-inter-layer-temporal coupling relationships, enhance extrapolation capabilities through physical mechanism constraints, and support online incremental learning. This solution would significantly improve the prediction accuracy and early warning lead time for roadbed deformation in permafrost regions, thus meeting the practical needs for long-term safe operation of roads in cold regions. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the present invention aims to provide a spatiotemporal correlation prediction method for roadbed stratification in permafrost regions based on graph neural networks.

[0007] To achieve the aforementioned objectives, the technical solution adopted by this invention includes:

[0008] In a first aspect, the present invention provides a training method for a spatiotemporal correlation prediction model of roadbed layering in permafrost regions based on graph neural networks, comprising:

[0009] Provides multi-dimensional time-series monitoring data from multiple monitoring points along highways in permafrost regions;

[0010] A horizontal spatial adjacency edge is established based on the actual spatial distribution relationship of the monitoring points, and a vertical interlayer coupling edge is established based on the thermal-hydraulic-mechanical coupling relationship between different depth layers of the same monitoring point. A weighted graph with weights is generated, wherein the horizontal spatial adjacency edge is negatively correlated with the distance between adjacent monitoring points, and the vertical interlayer coupling edge is negatively correlated with the magnitude of the difference in physical properties between layers.

[0011] The multidimensional monitoring data within the preset time window is cleaned, aligned, and standardized, and then combined into a node feature tensor.

[0012] Based on the weighted graph, a graph convolution operation is performed on the node feature tensor to obtain a spatial feature sequence;

[0013] Perform time encoding on the spatial feature sequence to obtain spatiotemporal coupled features;

[0014] Decoding is performed based on the aforementioned spatiotemporal coupling characteristics to obtain the evolution patterns along highways in permafrost regions, and the roadbed deformation prediction results are obtained by analyzing the evolution patterns.

[0015] Based on the difference between the predicted roadbed deformation and the actual settlement value, as well as the degree of deviation between the predicted roadbed deformation and physical laws, a loss function is constructed and the parameters are adjusted.

[0016] Secondly, the present invention also provides a method for predicting roadbed settlement in permafrost regions based on graph neural networks, comprising:

[0017] Provides a spatiotemporal correlation prediction model for roadbed layering in permafrost regions obtained by the above training method;

[0018] Acquire multi-dimensional time-series monitoring data from multiple monitoring points along highways in permafrost regions;

[0019] Based on the actual spatial distribution of monitoring points, horizontal spatial adjacency edges are established, and vertical interlayer coupling edges are established based on the thermal-hydraulic-mechanical coupling relationship between different depth layers of the same monitoring point, generating a weighted graph with weights.

[0020] The multidimensional monitoring data within the preset time window is cleaned, aligned, and standardized, and then combined into a node feature tensor.

[0021] The node feature tensor is input into the spatiotemporal correlation prediction model of roadbed layering in the permafrost region to obtain the roadbed deformation prediction result.

[0022] Based on the above technical solution, compared with the prior art, the beneficial effects of the present invention include at least the following:

[0023] The technical solution provided by this invention first deploys multiple types of layered sensors along the highway subgrade to acquire multi-dimensional time-series data such as temperature, water content, strain, and settlement of the surface layer, active layer, and permafrost layer. Then, based on the spatial distance relationship between monitoring points and the coupling relationship between different depth layers at the same point, a weighted graph with horizontal spatial adjacency edges and vertical inter-layer coupling edges is constructed. The multi-dimensional observations from the most recent time points are concatenated into node feature tensors and input into a spatiotemporal graph neural network model. The model first extracts the correlation features of neighboring nodes using graph convolution in the spatial dimension, and then captures the dynamic evolution law in the temporal dimension. Simultaneously, a physical consistency constraint based on the thermo-hydraulic-mechanical coupling mechanism is introduced during the training phase to ensure that the prediction results are consistent with the permafrost deformation mechanism. During online operation, the model can predict multi-step layered settlement or expansion and contraction, and automatically generate graded early warning information based on the relationship between the predicted values ​​and set thresholds. By utilizing spatial-inter-layer-temporal coupling information, it significantly improves the prediction accuracy and lead time of highway subgrade deformation in permafrost areas, making it suitable for real-time monitoring and early warning systems.

[0024] The above description is merely an overview of the technical solution of the present invention. In order to enable those skilled in the art to better understand the technical means of this application and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described below in conjunction with detailed drawings. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall process of the prediction method provided in a typical embodiment of the present invention. Detailed Implementation

[0026] In view of the shortcomings of the prior art, the inventors of this invention, through long-term research and extensive practice, have proposed the technical solution of this invention. The following will further explain and illustrate this technical solution, its implementation process, and its principles.

[0027] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0028] Moreover, relational terms such as “first” and “second” are used merely to distinguish one component or method step from another that has the same name, and do not necessarily require or imply any such actual relationship or order between these components or method steps.

[0029] The purpose of this invention is to overcome the shortcomings of existing methods for monitoring roadbeds in permafrost regions, such as insufficient spatial dimension, lack of layered coupling mechanisms, and limited predictive extrapolation capabilities. This invention proposes a layered spatiotemporal correlation prediction method based on graph neural networks. This method fully integrates multi-source layered sensor data, explicitly establishes a dual topology of spatial adjacency and interlayer coupling, and achieves high-precision multi-step prediction of deformation in the roadbed surface layer, active layer, and permafrost layer. Furthermore, physical consistency constraints ensure that the prediction results conform to the thermo-hydraulic-mechanical coupling mechanism of permafrost, thereby improving the accuracy and lead time of early warnings.

[0030] Based on the above objectives and technical ideas, this invention first provides a training method for a spatiotemporal correlation prediction model of roadbed layering in permafrost regions based on graph neural networks, which includes the following steps:

[0031] Provides multi-dimensional time-series monitoring data from multiple monitoring points along highways in permafrost regions;

[0032] A horizontal spatial adjacency edge is established based on the actual spatial distribution relationship of the monitoring points, and a vertical interlayer coupling edge is established based on the thermal-hydraulic-mechanical coupling relationship between different depth layers of the same monitoring point. A weighted graph with weights is generated, wherein the horizontal spatial adjacency edge is negatively correlated with the distance between adjacent monitoring points, and the vertical interlayer coupling edge is negatively correlated with the magnitude of the difference in physical properties between layers.

[0033] The multidimensional monitoring data within the preset time window is cleaned, aligned, and standardized, and then combined into a node feature tensor.

[0034] Based on the weighted graph, a graph convolution operation is performed on the node feature tensor to obtain a spatial feature sequence;

[0035] Perform time encoding on the spatial feature sequence to obtain spatiotemporal coupled features;

[0036] Decoding is performed based on the aforementioned spatiotemporal coupling characteristics to obtain the evolution patterns along highways in permafrost regions, and the roadbed deformation prediction results are obtained by analyzing the evolution patterns.

[0037] Based on the difference between the predicted roadbed deformation and the actual settlement value, as well as the degree of deviation between the predicted roadbed deformation and physical laws, a loss function is constructed and the parameters are adjusted.

[0038] As a typical example, the prediction method proposed in this invention includes the following steps:

[0039] 1. Deploy various types of layered sensors along the highway subgrade, such as temperature needles, displacement gauges, distributed optical fibers, inertial inclinometers, satellite radar scattering points, or reflection angle reflectors, to collect multi-dimensional time-series monitoring data on temperature, moisture content, strain, and settlement of the surface layer, active layer, and permafrost layer in real time.

[0040] 2. Establish horizontal spatial adjacency edges based on the actual spatial distribution of monitoring points, and establish vertical interlayer coupling edges based on the thermal-hydraulic-mechanical coupling relationship between different depth layers at the same point, generating a weighted graph with weights.

[0041] 3. Clean, align, and standardize the multidimensional monitoring data within the preset time window, and combine them into a node feature tensor;

[0042] 4. Input the node feature tensor into a spatiotemporal graph neural network containing a spatial graph convolutional network and a time series encoder, extract spatiotemporal coupling features, and output multi-step hierarchical deformation prediction results;

[0043] 5. During model training, introduce a physical consistency constraint term based on the thermal-hydraulic-mechanical coupling mechanism to correct the physical rationality of the model output;

[0044] 6. Compare the prediction results with the set thresholds for each layer, output graded early warning information, and feed the latest monitoring data back into the model to achieve online updates.

[0045] In some implementations, the multidimensional time-series monitoring data includes surface data, active layer data, and permafrost layer data of the monitoring points; the surface data, active layer data, and permafrost layer data at least include the temperature, moisture content, strain, and settlement values ​​of the test points.

[0046] Based on the aforementioned spatiotemporal coupling characteristics, the system further includes a decoding module for mapping the encoded spatiotemporal coupling feature hidden state to a roadbed settlement prediction result for a future time period. The decoding module can employ an existing neural network decoding structure, for example:

[0047] (1) Sequence-to-sequence (Seq2Seq) decoder: It decodes the hidden state output by the encoder step by step through recurrent neural network, long short-term memory network (LSTM) or gated recurrent unit (GRU) to obtain the settlement prediction sequence for several future time steps; (2) Attention-based decoder: When predicting the settlement amount at a certain future time step, it uses attention weights to weight and fuse historical spatiotemporal features, which can capture long-term dependencies and improve prediction accuracy; (3) Convolutional decoder (1D-CNN): It uses one-dimensional convolution to extract local patterns in the time dimension to achieve batch prediction of settlement curves; (4) Multilayer perceptron (MLP) decoder: It directly maps the hidden state at the final moment to the settlement prediction value for several future time steps.

[0048] Of course, it is also possible that other decoders besides the exemplary decoders mentioned above can achieve the same effect.

[0049] After obtaining the settlement prediction sequence S(t), numerical analysis is performed on it to extract the evolution law of permafrost subgrade, including but not limited to the following steps:

[0050] ① Settling rate Used to reflect the rate of change in the frozen soil settlement process;

[0051] ② Settlement acceleration Used to identify the accelerating or decelerating trend of the settling process;

[0052] ③ The detection of settlement inflection points is used to identify the critical moments when a roadbed in permafrost regions transitions from stable to unstable or from unstable to stable. Based on the above evolutionary patterns, the system can output predicted roadbed settlement values ​​and their evolutionary trends for several future time periods. Furthermore, it can combine risk thresholds to assess and classify the roadbed deformation state, thereby achieving forward-looking early warning of roadbed settlement along highways in permafrost regions.

[0053] The training method constructs a loss function based on the difference between the predicted roadbed deformation and the actual settlement value, as well as the degree of deviation between the predicted roadbed deformation and the physical laws, and then adjusts the parameters accordingly.

[0054] Based on the above evolutionary patterns, the system can output the predicted values ​​of roadbed settlement and its evolutionary trends for several future time periods, and further combine risk thresholds to evaluate and classify the roadbed deformation state, so as to achieve forward-looking early warning of roadbed settlement along highways in permafrost areas.

[0055] In some implementations, the multidimensional time-series monitoring data is represented as a set of nodes, i.e., three-dimensional multi-type sensors are deployed along the highway subgrade to form a set of nodes covering the surface layer, active layer, and permafrost layer.

[0056]

[0057] Wherein, V represents the multidimensional time-series monitoring data; v represents a data sub-item; superscripts s, a, and p represent the surface layer, active layer, and permafrost layer, respectively; subscript i is a natural number representing the monitoring point number; N represents the total number of monitoring points;

[0058] Multidimensional observations (temperature τ, moisture content W) at each node within the time window [t-T+1, t] c After normalization, the nodal feature tensors (such as strain ε, settlement u, etc.) are stacked into a third-order tensor. The nodal feature tensor is represented as:

[0059] X∈R |V|*T*F

[0060] Where X represents the node feature tensor, which is formed by stacking the normalized multidimensional monitoring data V; *T represents the time step; and *F represents the number of features.

[0061] The weight of adjacent edges in the horizontal direction is determined based on the geographical distance between monitoring points. The closer the edges are, the higher their weight. More preferably, the weight can also be determined by combining the stratigraphic consistency, which is more similar in stratigraphic properties.

[0062] In some implementations, the weights of the horizontal spatial adjacent edges are represented as follows:

[0063]

[0064] in, The weight of the horizontally adjacent spatial edges is represented by d. ij σ represents the horizontal distance between monitoring point i and monitoring point j; d This represents the distance attenuation coefficient.

[0065] The weight of the vertical interlayer coupling edge is dynamically adjusted based on the differences in thermal conductivity, ice content and elastic modulus of different layer materials to reflect the seasonal changes in the intensity of interlayer interaction.

[0066] That is, in some implementations, the weight of the vertical interlayer coupling edge is represented as:

[0067]

[0068] in, The vertical interlayer coupling edge is represented; k represents the measured value of the physical property; α represents the adjustment coefficient;

[0069] The weighted graph is represented as an adjacency matrix:

[0070]

[0071] Where A represents the original adjacency matrix (composed of horizontal and interlayer coupling edges); Wh represents the horizontal spatial adjacency weight matrix; W v This represents the interlayer coupling weight matrix in the vertical direction; represents the adjacency matrix after adding a self-loop (ensuring the transmission of information between nodes); I represents the identity matrix; Representation degree matrix.

[0072] In addition, graph convolutional autoencoders based on the information of adjacent nodes of the same type can be used to interpolate and restore missing monitoring data to ensure the integrity of node feature tensors.

[0073] In some implementations, the graph convolution operation is represented as:

[0074]

[0075] Among them, H (l) This represents the l-th node, where l is the index of the graph convolutional layer. When l = 0, H... (0) The time slice X representing the node feature tensor of the preset time window ：t： t represents time; W (l) σ represents the trainable graph convolution weights; σ(·) represents the Sigmoid activation function.

[0076] The time series encoder is a gated recurrent unit, a long short-term memory network, or other encoders with a multi-head attention mechanism, which can model short-term sharp fluctuations and long-term slow trends respectively.

[0077] Taking a gated loop unit as an example, in some implementations, the time encoding operation is performed using a gated feature loop, as follows:

[0078] z t =σ(W z h t-1 +U z x t )

[0079] r t =σ(W r h t-1 +U r x i )

[0080]

[0081] Where t represents the time step; Z t W represents updating the gate vector; z This indicates updating the gate input weights; h t-1 Indicates the hidden state at the previous moment; U zThis indicates updating the hidden weights of the gate; x t Represents the input vector; r t Represents the reset gate; σ(·) represents the Sigmoid activation function; W r U r This indicates resetting the gate weight; Represents the candidate hidden state; tanh(·) represents the activation function of the candidate state; W represents element-wise multiplication; h U h represents the candidate hidden state weight; ht represents the final hidden state.

[0082] In some implementations, the loss function includes a physical loss term, the calculation of which is expressed as follows:

[0083]

[0084] Where ρ represents soil density; c represents specific heat capacity; T represents temperature field; t represents time; and k represents thermal conductivity. Let represent the Laplace operator; T represent the temperature field; θ represent the water content; u represent the displacement; β and γ represent the adjustment coefficients (coupling parameters); and Lphys represent the physical loss term. This represents the predicted roadbed displacement (settlement) at node i and time t. This indicates the water content predicted by the model (reflecting the change in water content caused by the melting of frozen soil); This represents the temperature field predicted by the model.

[0085] In some implementations, the loss function is expressed as:

[0086] L total =L MSE +λ1L phys +λ2L reg

[0087] Among them, L total L represents the loss function; MSE L represents the mean square error between the predicted and actual settlement values. reg λ1 and λ2 represent regularization terms; λ1 and λ2 represent tradeoff coefficients. During training, the Adam optimizer can be used to iteratively update the network parameters until Ltotal converges. Of course, it is also possible that the customer has adopted other training iteration strategies.

[0088] Furthermore, in some implementations, the preset time window can be adaptively adjusted according to the typical freeze-thaw cycle of the area where the road is located. The window is shortened in winter to improve the accuracy of capturing extreme events, and the window is appropriately extended in summer to reduce model disturbances.

[0089] A second aspect of this invention provides a method for predicting roadbed settlement in permafrost regions based on graph neural networks, comprising the following steps:

[0090] Provides a spatiotemporal correlation prediction model for roadbed layering in permafrost regions obtained by the above training method;

[0091] Acquire multi-dimensional time-series monitoring data from multiple monitoring points along highways in permafrost regions;

[0092] Based on the actual spatial distribution of monitoring points, horizontal spatial adjacency edges are established, and vertical interlayer coupling edges are established based on the thermal-hydraulic-mechanical coupling relationship between different depth layers of the same monitoring point, generating a weighted graph with weights.

[0093] The multidimensional monitoring data within the preset time window is cleaned, aligned, and standardized, and then combined into a node feature tensor.

[0094] The node feature tensor is input into the spatiotemporal correlation prediction model of roadbed layering in the permafrost region to obtain the roadbed deformation prediction result.

[0095] In some implementations, the prediction method may further include the following steps:

[0096] Based on the predicted roadbed deformation, a graded early warning system is implemented. When the risk level is high, the monitoring frequency of the multi-dimensional time-series monitoring data is increased.

[0097] The tiered early warning thresholds include three levels: yellow, orange, and red, corresponding to three situations: single-point prediction confidence upper limit approaching the threshold, exceeding the threshold, and prediction mean significantly exceeding the threshold, respectively. Upon triggering an orange or higher-level warning, the system automatically enters a high-frequency monitoring mode, increasing the data collection frequency in key areas and dynamically refreshing the map structure to enhance local prediction accuracy. The tiered early warning method is, for example:

[0098] Run the model on the input window [t-T+1, t] and output the settlement / expansion / contraction prediction for the next K steps:

[0099] and confidence interval

[0100] when Trigger a yellow alert for the corresponding level; if A red alert was triggered, and high-frequency sampling and on-site verification were initiated.

[0101] In some implementations, the prediction method may further include the following steps:

[0102] For the missing monitoring points, interpolation is used to restore the corresponding multidimensional time-series monitoring data.

[0103] In some implementations, the model retains a differentiable physical consistency constraint module after training, and maintains physical consistency through mini-batch gradient updates when environmental parameters change significantly during the online inference phase.

[0104] An embodiment of the present invention also provides a system for monitoring and early warning of roadbed stratification in permafrost regions based on the above-described training and prediction methods, characterized in that the system comprises:

[0105] a) Data acquisition module;

[0106] b) Graph construction and dynamic update module;

[0107] c) Spatiotemporal graph neural network prediction module;

[0108] d) Physical consistency constraint training and online correction module;

[0109] e) Early warning and decision-making module;

[0110] f) Visualization and external interface module;

[0111] The modules are interconnected via a data bus and work collaboratively according to the steps in the above prediction method, outputting the prediction results of roadbed layered deformation and early warning information; the spatiotemporal graph neural network prediction module supports parallel computing based on graphics processors and automatically switches to degraded mode when encountering large-scale data loss, sensor failure, or network interruption; the graph construction and dynamic update module can automatically reconstruct the graph structure and synchronously adjust the adjacency weights when monitoring points are added, deleted, or fail, ensuring that the model topology is consistent with the on-site layout; the early warning decision module supports multiple alarm methods such as SMS, email, mobile application push and programmable API calls, and can be extended to access third-party intelligent maintenance systems.

[0112] Embodiments of the present invention also provide a computer device, including a processor, a memory, and a communication bus connected thereto, characterized in that the memory stores computer-executable instructions, which, when executed by the processor, cause the computer device to perform the above-described training and prediction methods.

[0113] Embodiments of the present invention also provide a non-transitory computer-readable storage medium storing a computer program that, when executed by a computer device, causes the device to perform the above-described training and prediction methods.

[0114] The model provided by this invention supports transfer learning, using model parameters already trained on similar permafrost road sections as initial weights to reduce reliance on large amounts of field data during the cold start phase. The model employs an incremental learning strategy, updating newly arrived monitoring data via a sliding window to adaptively track long-term roadbed degradation trends. The system uses visualization components to display the prediction curves, confidence intervals, and warning levels of each monitoring node in real time. It supports multi-dimensional retrieval by layer, mileage, and time. Alarm information is pushed to the road maintenance center via a communication interface and can interact with road information platforms and weather forecasting platforms for comprehensive analysis.

[0115] In the prediction method provided by this invention, when a node in a graph structure is detected to be continuously abnormal and may trigger a chain reaction in the region, the potential impact range is automatically calculated, and the warning level of adjacent nodes within that range is raised at the same time.

[0116] For a typical application example of the above technical solution, please refer to Figure 1 As shown in the accompanying drawings, specific embodiments of the present invention will be described in detail below for a better understanding of the invention. It should be understood that the following embodiments are merely illustrative and should not be construed as limiting the scope of protection of the present invention.

[0117] The permafrost-region highway subgrade layered monitoring and early warning system proposed in this invention can be divided into four functional layers from bottom to top: a sensing layer, a transmission layer, a computing layer, and an application layer. Efficient information flow and closed-loop control between layers are achieved through standardized data interfaces. The overall design follows a technical route of "layered data acquisition - centralized computing - intelligent decision-making - visualization services" to ensure long-term reliable operation and real-time early warning capabilities in high-altitude and remote areas.

[0118] The sensing layer is responsible for direct interaction with the roadbed engineering entity. Its core task is to conduct high-density, multi-layered observations of the temperature, strain, and settlement status of the surface layer, active layer, and permafrost layer. To this end, distributed fiber optic temperature-strain composite cables can be laid at a shallow burial depth of 0.2m along the road shoulder to achieve continuous profile monitoring with meter-level resolution. Embedded temperature probes and earth pressure cells are deployed at three key depths of 0m, 1.5m, and 3.0m to calibrate fiber optic data and acquire high-precision point information. An inertial tiltmeter is installed at the bottom boundary of the active layer to capture shear slip trends. Simultaneously, a radar corner reflector is installed at the center of the profile to cooperate with spaceborne InSAR to acquire a large-scale settlement field on the road surface.

[0119] The transmission layer adopts an "edge buffer + multi-channel backhaul" architecture: the roadside edge base station integrates an industrial Ethernet switch, a 4G / LTE-M module, and a Beidou short message terminal to collect and initially screen multi-source sensor data in real time. High-frequency data (sampling period ≤ 10 min) is pushed to the data center via the MQTT protocol; when communication is blocked or enters a blind zone, the system automatically degrades, only redundantly backhauling key variables via Beidou short messages to ensure uninterrupted monitoring links.

[0120] The computing layer is the intelligent core of the system, consisting of a data preprocessing server, an inference-training server, a time-series database, and a message queue. The data preprocessing server is responsible for denoising, time synchronization, normalization, and missing data interpolation of the raw monitoring sequences. The inference-training server deploys multiple GPUs and runs a physically constrained graph neural network model using the PyTorch and DGL frameworks to complete the spatiotemporal prediction and early warning discrimination of roadbed layered deformation. The time-series database (InfluxDB) is used to store raw and predicted data, while the message queue (Kafka) manages high-frequency data streams, early warning events, and operation and maintenance logs, enabling reliable communication in a high-concurrency environment.

[0121] The application layer provides visualization and collaborative services for decision-makers and maintenance personnel. The Web-GIS platform, built on OpenLayers and Vue, can display real-time 3D profiles of the roadbed, temperature cloud maps, and early warning heat maps, and supports historical playback and trend analysis. A RESTful API interfaces with the traffic management system; once an orange or higher level warning is triggered, the platform automatically generates work orders for speed limits, maintenance, and material dispatch, achieving closed-loop management of monitoring, prediction, and intervention. Through the collaboration of these four layers, this system can operate stably for extended periods in harsh, cold environments, providing accurate and timely intelligent early warning guarantees for the safe operation of highway roadbeds in permafrost regions.

[0122] In the above embodiments, several key technical points include:

[0123] 1. Layered-spatiotemporal integrated expression, accurately revealing the coupling mechanism of permafrost.

[0124] Traditional methods often focus on point data from road surfaces, neglecting the dynamic evolution of deep active layers and permafrost. This invention incorporates both horizontal spatial adjacency and vertical interlayer coupling into a weighted graph topology, utilizing graph convolution and temporal coding for joint modeling, thus achieving the overall extraction of three-dimensional coupling features between the surface layer, active layer, and permafrost. Experiments show that compared to classic LSTM or ARIMA, the prediction mean square error is reduced by approximately 30%, enabling earlier identification of the potential risk of deep melt subsidence propagating to the surface.

[0125] 2. Physical consistency constraints significantly improve extrapolation and confidence capabilities.

[0126] Most data-driven models are prone to distortion due to a lack of physical support in extreme scenarios. This invention embeds the thermo-hydraulic-mechanical coupling equation into the loss function, imposing constraints on the prediction gradient in a differential form to ensure that the model output conforms to the deformation mechanism of permafrost. When simulating extreme temperature rise scenarios over the past decade, the prediction results still maintain an error limit of <5mm, and the false alarm rate decreases by 40%, providing a reliable confidence interval for engineering control.

[0127] 3. Multi-source data fusion and self-healing for missing data enhance the robustness of the monitoring network.

[0128] In response to the frequent sensor failures and unstable communication in high-altitude and cold regions, this invention utilizes a graph convolutional autoencoder to interpolate missing data and fuses multi-source information such as InSAR, fiber optics, and embedded temperature probes, enabling the monitoring network to have self-healing capabilities. Even if 25% of the sensor nodes fail, the overall prediction accuracy of the model fluctuates only slightly (<8%), ensuring the continuity of early warning.

[0129] 3. Online incremental learning and GPU-accelerated inference enable real-time early warning in uninhabited areas.

[0130] This method designs a differentiable physical constraint module and a sliding window incremental update framework, combined with the parallel inference capability of GPUs, to achieve a closed loop of "one-hour monitoring - second-level prediction - real-time early warning". In a field test on a 100km test section of the Qinghai-Tibet Highway, after a five-minute data update for a single cross-section, the inference time was <0.5s, meeting the emergency response requirements under extreme weather conditions.

[0131] 4. Topology and parameters are portable, reducing deployment costs across lines.

[0132] By employing a formulaic edge weight calculation and graph structure reconstruction strategy, this invention can transfer the weights of trained models to permafrost highways with similar geological conditions, requiring only a small amount of local data for adaptation. Compared to remodeling, deployment time is reduced by more than 60%, significantly lowering the economic cost of upgrading and expanding monitoring systems.

[0133] 5. Tiered early warning and proactive enhancement mode to support intelligent management and maintenance decision-making.

[0134] Based on the confidence intervals output by the model, this invention sets three threshold levels: yellow, orange, and red. This automatically triggers high-frequency monitoring and on-site verification, and intelligently adjusts the thresholds and sampling frequencies for key road sections, achieving a closed loop of "monitoring-prediction-intervention." Actual operational cases show that after adopting this invention, the lead time for disaster precursor identification increased from an average of 7 days to 14 days, providing maintenance departments with ample maintenance windows.

[0135] It should be understood that the above embodiments are merely illustrative of the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A training method for a spatiotemporal correlation prediction model of roadbed layering in permafrost regions based on graph neural networks, characterized in that, include: Provides multi-dimensional time-series monitoring data from multiple monitoring points along highways in permafrost regions; A horizontal spatial adjacency edge is established based on the actual spatial distribution relationship of the monitoring points, and a vertical interlayer coupling edge is established based on the thermal-hydraulic-mechanical coupling relationship between different depth layers of the same monitoring point. A weighted graph with weights is generated, wherein the horizontal spatial adjacency edge is negatively correlated with the distance between adjacent monitoring points, and the vertical interlayer coupling edge is negatively correlated with the magnitude of the difference in physical properties between layers. The multidimensional time-series monitoring data within the preset time window are cleaned, aligned, and standardized, and combined into a node feature tensor. Based on the weighted graph, a graph convolution operation is performed on the node feature tensor to obtain a spatial feature sequence; Perform time encoding on the spatial feature sequence to obtain spatiotemporal coupled features; Decoding is performed based on the aforementioned spatiotemporal coupling characteristics to obtain the evolution patterns along highways in permafrost regions, and the roadbed deformation prediction results are obtained by analyzing the evolution patterns. Based on the difference between the predicted roadbed deformation and the actual settlement value, and the degree of deviation between the predicted roadbed deformation and physical laws, a loss function is constructed for parameter adjustment; The weights of the horizontal spatial adjacent edges are represented as follows: ； in, The weight of the horizontally adjacent spatial edges is represented by d. ij σ represents the horizontal distance between monitoring point i and monitoring point j; d The constant represents the distance attenuation coefficient; The weight of the vertical interlayer coupling edge is represented as follows: ； in, The weight of the vertical interlayer coupling edge is represented; k represents the measured value of the physical property; e and f represent different layers; α represents the adjustment coefficient. The weighted graph is represented as an adjacency matrix: ； Where A represents the original adjacency matrix consisting of horizontal and interlayer coupling edges; This represents the spatial adjacency weight matrix in the horizontal direction; This represents the interlayer coupling weight matrix in the vertical direction; This represents the adjacency matrix after adding the self-loop, used to ensure the transmission of information between nodes; I represents the identity matrix. Representation degree matrix.

2. The training method according to claim 1, characterized in that, The multidimensional time-series monitoring data includes surface data, active layer data, and permafrost layer data of the monitoring points; the surface data, active layer data, and permafrost layer data at least include the temperature, moisture content, strain, and settlement value of the test points.

3. The training method according to claim 2, characterized in that, The multidimensional time-series monitoring data is represented as a set of nodes: ； Wherein, V represents the multidimensional time-series monitoring data; v i The data item is represented by the superscripts s, a, and p, which represent the surface layer, active layer, and permafrost layer, respectively. The subscript i is a natural number representing the monitoring point number. N represents the total number of the monitoring points. The node feature tensor is represented as follows: ； Where X represents the node feature tensor, which is formed by stacking the normalized multidimensional time-series monitoring data V; *T represents the number of time steps; and *F represents the number of features.

4. The training method according to claim 1, characterized in that, The graph convolution operation is represented as follows: ； in, This represents the l-th node, where l is the index of the graph convolutional layer. When l=0, The time slice X representing the node feature tensor of the preset time window : t : t represents time; σ represents the trainable graph convolution weights; σ(·) represents the Sigmoid activation function; This represents the adjacency matrix after adding the self-loop; Representation degree matrix.

5. The training method according to claim 1, characterized in that, The time encoding operation is performed using a gated feature loop, as follows: ； ； ； ； Where t represents the time step; Z t W represents updating the gate vector; z This indicates updating the gate input weights; h t-1 Indicates the hidden state at the previous moment; U z This indicates updating the hidden weights of the gate; x t Represents the input vector; r t Represents the reset gate; σ(·) represents the Sigmoid activation function; W r、 U r This indicates resetting the gate weight; λ1 represents the candidate hidden state; tanh(·) represents the activation function of the candidate state; W represents element-wise multiplication; h、 U h h represents the candidate hidden state weights. t This indicates the final hidden state.

6. The training method according to claim 1, characterized in that, The loss function includes a physical loss term, and the calculation process of the physical loss term is expressed as follows: ； ； ； Where ρ represents soil density; c represents specific heat capacity; T represents temperature field; t represents time; and k represents thermal conductivity. Represents the Laplace operator; T represents the temperature field; θ represents the water content; u represents the displacement; β and γ represent the adjustment coefficients; L phys This represents the physical loss term; This represents the value of the roadbed displacement predicted by the model at node 𝑖 and time 𝑡. This indicates the water content predicted by the model. This represents the temperature field predicted by the model.

7. The training method according to claim 6, characterized in that, The loss function is expressed as: ； Among them, L total L represents the loss function; MSE L represents the mean square error between the predicted and actual settlement values. reg λ represents the regularization term. 1、 λ2 represents the tradeoff coefficient.

8. A method for predicting roadbed settlement in permafrost regions based on graph neural networks, characterized in that, include: Provides a spatiotemporal correlation prediction model for roadbed layering in permafrost regions obtained by the training method described in any one of claims 1-7; Acquire multi-dimensional time-series monitoring data from multiple monitoring points along highways in permafrost regions; Based on the actual spatial distribution of monitoring points, horizontal spatial adjacency edges are established, and vertical interlayer coupling edges are established based on the thermal-hydraulic-mechanical coupling relationship between different depth layers of the same monitoring point, generating a weighted graph with weights. The multidimensional time-series monitoring data within the preset time window are cleaned, aligned, and standardized, and combined into a node feature tensor. The node feature tensor is input into the spatiotemporal correlation prediction model of roadbed layering in the permafrost region to obtain the roadbed deformation prediction result.

9. The prediction method according to claim 8, characterized in that, Also includes: Based on the predicted roadbed deformation, a graded early warning system is implemented. When the risk level is high, the monitoring frequency of the multi-dimensional time-series monitoring data is increased. And / or, for the partially missing monitoring points, interpolation is used to generate corresponding multidimensional time-series monitoring data.

Images