A highway differential settlement prediction and regulation system based on machine learning

Abstract

The application discloses a highway differential settlement prediction and regulation system based on machine learning and belongs to the technical field of intelligent traffic infrastructure construction and maintenance. The system aims at the differential settlement control problem in the highway reconstruction and expansion project in the alluvial plain deep soft soil area, and builds a closed-loop system including multi-source data fusion collection, space-time feature engineering, physical constraint space-time graph network prediction, reinforcement learning reverse regulation and visual decision warning. The system effectively solves the overfitting and physical consistency loss problems of the pure data-driven model under complex geological conditions by embedding the Biot consolidation theory constraint in the geological weighted graph network. At the same time, the reinforcement learning strategy network is used to replace artificial experience, and the global optimization regulation and control of the construction parameters are realized on the premise of meeting the differential settlement control requirements. The application is mainly used for the accurate prediction of settlement and intelligent construction management of the highway reconstruction and expansion project.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation infrastructure construction and maintenance technology, and in particular to a machine learning-based system for predicting and controlling differential settlement on highways. Background Technology

[0002] When carrying out highway reconstruction and expansion projects in the deep soft soil areas of alluvial plains, differential settlement control between the old and new roadbeds is a key factor determining the quality of the project and operational safety. The geological environment of this area is complex, with a wide distribution of marine sedimentary soft soil layers, characterized by high compressibility, low bearing capacity, and sensitivity to groundwater levels. In addition, the complex stratigraphic structure of alternating layers of underground silt and clay makes the settlement caused by roadbed widening exhibit significant spatial heterogeneity and asymmetry.

[0003] Existing settlement control methods mainly rely on traditional geotechnical calculations or regression analysis based on single monitoring data. However, purely theoretical calculation models struggle to accurately describe the nonlinear consolidation process under complex geological conditions, often resulting in significant prediction bias due to excessive simplification assumptions. While purely data-driven regression models can fit observational data, they are prone to overfitting when engineering sample sizes are limited and noise levels are high, and prediction results frequently violate physical consolidation laws, lacking interpretability. Furthermore, traditional construction control relies heavily on trial-and-error adjustments based on engineering experience, lacking global optimization methods for multi-objective constraints, making it difficult to achieve the optimal balance between construction period, cost, and settlement control effectiveness. Therefore, how to integrate physical mechanisms and monitoring data to achieve high-precision settlement prediction and automatically generate optimal construction control strategies is a pressing technical problem that needs to be solved. Summary of the Invention

[0004] The purpose of this invention is to provide a machine learning-based system for predicting and controlling differential settlement of highways, in order to solve the problems of insufficient accuracy in settlement prediction and lack of a global optimization mechanism in construction control under complex soft soil environments.

[0005] As a preferred technical solution of the present invention, a highway differential settlement prediction and control system based on machine learning includes a multi-source data fusion acquisition module, a spatiotemporal feature engineering module, a physical constraint spatiotemporal graph network prediction module, a reinforcement learning inverse control module, and a visualization decision and early warning module.

[0006] The multi-source data fusion acquisition module is used to acquire sensor data from the engineering site and outputs standardized spatiotemporal data tensors after preprocessing.

[0007] The spatiotemporal feature engineering module is used to construct a weighted adjacency graph containing geological consistency coefficients based on standardized spatiotemporal data tensors, and extract multi-scale temporal features and inject physical prior features to generate enhanced feature tensors.

[0008] The physical constraint spatiotemporal graph network prediction module is used to extract spatiotemporal dependencies by using enhanced feature tensors and weighted adjacency graphs through graph attention mechanisms and dilated causal convolutions, and outputs the settlement prediction value and uncertainty estimate for future time steps; during the training process, the module embeds the finite difference residual of the Biot consolidation equation as a regularization term into the loss function.

[0009] The reinforcement learning inverse control module is used to construct a state space containing settlement prediction values ​​and construction parameters, and to output control instructions for pile foundation parameters and compaction strategies using a strategy network.

[0010] The visualization decision-making and early warning module is used to display the three-dimensional settlement field and trigger multi-level early warnings based on the predicted differential settlement rate.

[0011] As a preferred embodiment of the present invention, the multi-source data fusion acquisition module includes a sensor access submodule, a data preprocessing submodule, and a spatiotemporal alignment submodule; the data preprocessing submodule is configured to: for long discontinuous missing data with a continuous missing length exceeding a set threshold, use truncated singular value decomposition to complete the low-rank matrix composed of sensor observations from the same cross-section; for sensor data with different sampling frequencies, resample to the target temporal resolution uniformly through moving average or linear interpolation.

[0012] As a preferred technical solution of the present invention, the spatiotemporal feature engineering module includes a spatial topology construction sub-module, which constructs an adjacency matrix of an undirected weighted graph. The weights of its elements are determined by the Euclidean distance attenuation term between nodes and the geological consistency coefficient. When two nodes belong to different geological zones, the geological consistency coefficient is taken as the reciprocal of the ratio of the comprehensive compression modulus of the foundation at the location of the two nodes, which is used to reduce the information propagation intensity between nodes across geological zones.

[0013] As a preferred technical solution of the present invention, the spatiotemporal feature engineering module further includes a physical prior feature injection submodule. This submodule uses Terzaghi's one-dimensional consolidation theory to calculate the theoretical degree of consolidation and theoretical settlement of each node at different times, and splices the theoretical calculated values ​​and the residuals between the theoretical values ​​and the measured values ​​as additional feature channels into the node feature vector.

[0014] As a preferred technical solution of the present invention, the physical constraint spatiotemporal graph network prediction module includes a graph attention spatial coding submodule. When calculating the attention coefficients between nodes, this submodule introduces the logarithmic term of the adjacency matrix elements as a bias, so that the attention weights follow the prior structure defined by the spatial topology and geological conditions. The module also includes an extended causal temporal convolution submodule, which uses multi-layer extended causal convolution to extract multi-scale temporal dependence features of the node spatial coding sequence.

[0015] As a preferred technical solution of the present invention, the physical regularization training submodule in the physical constraint spatiotemporal graph network prediction module defines a composite loss function, which is composed of a weighted sum of a data fitting term and a physical constraint term. The physical constraint term is constructed based on the finite difference scheme of the Biot consolidation equation, and calculates the mean square value of the discrete residual between the excess pore water pressure dissipation process predicted by the model and the law described by the consolidation equation. The weight coefficient of the physical constraint term is dynamically increased with the training rounds using an annealing strategy.

[0016] As a preferred technical solution of the present invention, the reinforcement learning inverse control module includes an environment modeling submodule. This submodule defines the state space as a combination of the current predicted settlement vector, the differential settlement vector between adjacent nodes, the current composite foundation pile parameter vector, the current compaction parameter vector, and the residual budget vector; and defines the action space as a continuous space including the pile length increment, the pile spacing adjustment, and the compaction pass increment.

[0017] As a preferred technical solution of the present invention, the reinforcement learning reverse control module adopts a comprehensive reward function to evaluate the control strategy. The comprehensive reward function includes a differential settlement compliance penalty item, a material cost penalty item, a construction period penalty item, and a safety margin reward item. Among them, the weight coefficient of the differential settlement compliance penalty item is set to the highest to ensure that settlement control is the primary optimization objective.

[0018] As a preferred embodiment of the present invention, the reinforcement learning inverse control module further includes a control instruction generation submodule. This submodule performs post-processing operations after the policy network outputs the action vector: trimming the action components to the engineering feasible domain; applying a one-dimensional Gaussian smoothing filter to the parameter increments of adjacent control sections to eliminate parameter abrupt changes; and encoding the processed increments into construction control instructions for issuance.

[0019] As a preferred technical solution of the present invention, the visualization decision and early warning module includes a multi-level early warning engine sub-module, which sets a three-level early warning mechanism: when the predicted differential settlement rate exceeds the first threshold, a attention-level early warning is triggered; when the predicted differential settlement rate exceeds the second threshold or the prediction uncertainty exceeds the set proportion, a warning-level early warning is triggered and the reinforcement learning reverse control module is invoked; when the predicted differential settlement rate exceeds the third threshold, a disposal-level early warning is triggered and an emergency command is issued.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] This invention constructs a physically constrained spatiotemporal graph network prediction model, embedding classical Biot consolidation theory as a regularization term into the loss function of a deep learning network, while simultaneously injecting physical prior information into the topology and input features of the geological weighted map. This "physical-data" dual-driven mode significantly improves the model's generalization ability in small-sample, high-noise engineering environments, ensuring that the settlement prediction results conform to both the monitoring data trends and the physical laws of soil consolidation, effectively avoiding the non-physical prediction biases generated by pure data models.

[0022] This invention introduces a reinforcement learning-based inverse control mechanism, transforming the complex adjustment of construction parameters into a sequential decision-making problem under multi-objective constraints. Through a policy network, it automatically searches for the optimal combination of pile length, pile spacing, and number of compaction passes, while comprehensively considering settlement control, cost budgeting, and schedule constraints. This achieves a shift from "qualitative control based on manual experience" to "quantitative optimization using intelligent algorithms," significantly improving the efficiency and economy of differential settlement control. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall architecture of the system described in an embodiment of the present invention.

[0024] Figure labeling: 100, Multi-source data fusion acquisition module; 101, Sensor access submodule; 102, Data preprocessing submodule; 103, Spatiotemporal alignment submodule; 200, Spatiotemporal feature engineering module; 201, Spatial topology construction submodule; 202, Temporal feature extraction submodule; 203, Physical prior feature injection submodule; 300, Physically constrained spatiotemporal graph network prediction module; 301, Graph attention spatial encoding submodule; 302, Dilated causal temporal convolution submodule; 303, Physically regularized training submodule; 304, Multi-step rolling prediction submodule; 400, Reinforcement learning inverse control module; 401, Environmental modeling submodule; 402, Policy network submodule; 403, Control command generation submodule; 500, Visualized decision-making and early warning module; 501, 3D settlement field rendering submodule; 502, Multi-level early warning engine submodule; 503, Construction quality report generation submodule. Detailed Implementation

[0025] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described below are only for explaining the present invention and are not intended to limit the scope of protection of the present invention.

[0026] like Figure 1As shown, this invention provides a machine learning-based highway differential settlement prediction and control system. This system is designed to address the engineering needs of differential settlement control between old and new roadbeds in highway reconstruction and expansion projects in deep soft soil areas of alluvial plains. Its overall architecture uses a multi-source sensor network as the perception layer, a spatiotemporal graph network model with embedded physical constraints as the prediction core, and a reinforcement learning-driven inverse optimization solver as the control engine, constructing an intelligent closed-loop system covering the entire process of "perception-prediction-control-feedback". Taking the Bohai Bay alluvial plain area as a typical application scenario, this region has a widespread marine sedimentary soft soil layer below the surface, mainly composed of clayey and silty clay, with a high groundwater level, high compressibility, and low bearing capacity. At depths of 7 to 15 meters below the surface, there are complex alternating strata such as fine sand and silty clay. The differential settlement caused by roadbed widening exhibits spatially non-uniform and asymmetric characteristics. This system, targeting the above engineering environment, achieves full-chain intelligence from data acquisition, feature extraction, settlement prediction to construction control.

[0027] This system comprises five core modules: a multi-source data fusion and acquisition module 100, a spatiotemporal feature engineering module 200, a physically constrained spatiotemporal graph network prediction module 300, a reinforcement learning inverse regulation module 400, and a visualization decision-making and early warning module 500. These modules form a serial and feedback-coupled data flow path, jointly achieving a closed-loop signal transmission of "data acquisition → feature construction → settlement prediction → regulation decision → command execution → effect feedback". The specific structure and implementation of each module are described below.

[0028] The multi-source data fusion acquisition module 100 is responsible for acquiring raw observation data from various sensors and external data systems at the engineering site, and then cleaning and aligning the data to form a standardized spatiotemporal data tensor in a unified format for use by subsequent modules. This module contains three functional sub-units: a sensor access sub-module 101, a data preprocessing sub-module 102, and a spatiotemporal alignment sub-module 103.

[0029] The sensor access submodule 101 is responsible for accessing four types of data sources. The first type is surface and deep settlement data sources, specifically including GNSS high-precision positioning devices and deep stratified settlement meters distributed in the widened roadbed area. The GNSS devices form a grid-like observation array with 50-meter intervals along the longitudinal and transverse directions of the roadbed, with a sampling frequency set to once per hour and a positioning accuracy better than ±5mm. The deep stratified settlement meters are buried at several key depths in the soft soil layer, in a preferred embodiment, at 3m, 6m, 10m, and 15m below ground level, with a sampling frequency of once every 30 minutes and a measurement accuracy of ±0.1mm. The second type is construction process parameter data sources, including process parameters such as grouting pressure, grouting flow rate, drill rod lifting speed, and pile depth collected by the high-pressure jet grouting pile intelligent monitoring system, with a sampling frequency of 10Hz; and parameters such as vibration acceleration, number of compaction passes, travel speed, and real-time estimated compaction degree collected by the intelligent compaction equipment, with a sampling frequency of 5Hz. The third category is geological and environmental data sources, including parameters such as soil layer thickness, water content, void ratio, compression modulus, and consolidation coefficient provided by geological exploration data, as well as environmental factor data provided by temperature sensors and groundwater level monitoring wells, with a sampling frequency of once per hour. The fourth category is load and structural data sources, covering fill height, fill material unit weight, and graded loading time series information for each construction stage of the widened roadbed, as well as design parameters such as pile length, pile diameter, pile spacing, and area replacement ratio of the composite foundation. The signals from the above four types of data sources are transmitted to the edge computing node via a wireless communication link and then enter the data preprocessing submodule 102.

[0030] The data preprocessing submodule 102 performs three processes on the raw data in sequence: outlier detection and removal, missing value imputation, and multi-scale time synchronization.

[0031] During the outlier detection and removal phase, the original time series of each sensor channel is processed. Calculate the local median within the sliding window centered on the sampling point. and median absolute deviation A sampling point is identified as an outlier and marked for removal if it meets the following conditions:

[0032]

[0033] In the formula, the width of the sliding window The default sampling point is 120, and the coefficient 1.4826 is the theoretical conversion factor between the median absolute deviation and standard deviation under the assumption of normal distribution. The detection method based on the median absolute deviation, rather than the traditional mean-standard deviation method, has the advantage of stronger robustness to extreme values. In construction environments, the impact of occasional step failures or spike interference from the sensor on the detection threshold is effectively suppressed. Sampling points marked as outliers are treated as missing data during the missing value imputation stage.

[0034] During the missing value interpolation stage, different completion strategies are adopted based on the continuous length of the missing interval. When the continuous missing length is less than or equal to 6 sampling points, it is considered a short discontinuous missing value, and cubic spline interpolation is used to interpolate the missing interval. A smooth spline curve is constructed using the known data points at both ends of the missing interval to recover the missing sampling value in the middle. When the continuous missing length exceeds 6 sampling points, it is considered a long discontinuous missing value. Time interpolation of a single channel is no longer sufficient to guarantee accuracy. In this case, a matrix completion method based on the spatial correlation of adjacent sensors is introduced. Specifically, all data points on the same cross-section are interpolated. A matrix is ​​formed from the observations of each sensor within the same time period. ,in For time steps, Perform truncated singular value decomposition, retaining the first... There are singular values, among which Take the smallest positive integer that makes the cumulative singular value contribution rate exceed 95%, and use a low-rank approximation matrix. The missing locations are filled in. The physical basis of this method is that the settlement fields observed by different sensors on the same cross section have strong spatial correlation, and the settlement observation matrix itself is approximately low-rank. Therefore, truncated singular value decomposition can effectively utilize spatial redundancy information to recover missing data.

[0035] During the multi-scale time synchronization phase, due to significant differences in the sampling frequencies of various sensors—the sampling frequency of construction process parameters can reach 5Hz to 10Hz, while the sampling frequency of environmental data is only once per hour—it is necessary to unify all channels to the same time resolution. For high-frequency signals (construction process parameters with a sampling frequency greater than or equal to 5Hz), a moving average value is taken within each target time window as the representative value of that window, in order to preserve the overall trend of the signal and suppress high-frequency noise while downsampling the frequency. For low-frequency signals (environmental and geological data with a sampling frequency no higher than once per hour), linear interpolation is used to upsample them to the target time resolution. By default, the system uniformly resamples all channels to a time step of 10 minutes. It can be understood that the above-mentioned default time step of 10 minutes is an optimal value selected under the condition of balancing data volume and time resolution. It can be adjusted to other values ​​as needed in different engineering scenarios or construction stages. For example, it can be shortened to 5 minutes during periods of intensive construction loading to improve time sensitivity.

[0036] The spatiotemporal alignment submodule 103 maps all preprocessed sensor data to a unified spatiotemporal coordinate system. Spatially, it uses the starting point of the roadbed centerline as the origin and the station direction as... Axis and transverse are Axis and vertical are A three-dimensional Cartesian coordinate system is established using axes, and the installation position of each sensor is represented by a coordinate triplet. In terms of time, the project commencement date is used as the time origin, accurate to the minute. After all the above processing, the system outputs a standardized spatiotemporal data tensor. ,in The total number of monitoring nodes, The total number of time steps. This represents the original feature dimension for each node at each time step. This tensor serves as the unified input for subsequent spatiotemporal feature engineering and prediction modules.

[0037] Spatiotemporal Feature Engineering Module 200 uses standardized spatiotemporal data tensors Using this as input, the module performs spatial topological relationship encoding, multi-level extraction of temporal features, and injection of physical prior information based on consolidation theory, transforming the original observation data into an enhanced feature tensor with rich physical semantics, providing a high-quality input representation for subsequent graph network prediction. This module includes a spatial topology construction submodule 201, a temporal feature extraction submodule 202, and a physical prior feature injection submodule 203.

[0038] The function of the spatial topology construction submodule 201 is to integrate the distributed... The spatial relationship structure between monitoring nodes is explicitly expressed as a graph topology. Specifically, this involves using the entire road segment... An undirected weighted graph is constructed with each monitoring node as a vertex. ,in For a set of nodes, Let be the set of edges. This is an adjacency matrix. The elements of the adjacency matrix are calculated according to the following rules:

[0039]

[0040] In the formula For nodes With nodes The three-dimensional Euclidean distance between them This is the distance attenuation bandwidth parameter, with a default value of 50m. The cutoff distance is set to 200m by default. The reason for introducing the cutoff distance is that, for linear highway projects, the settlement correlation between monitoring points that are too far apart is close to zero. Keeping these edges not only has no physical meaning, but also increases the computational burden of the graph network. This is the geological consistency coefficient, used to encode the degree of similarity in geological conditions at the location of a node. When a node... and nodes When the location belongs to the same geological zone (i.e., the soil structure and main geotechnical parameters are similar), The value is 1.0; when the two belong to different geological zones, The reduction is made according to the reciprocal of the ratio of the comprehensive compression modulus of the foundation at the locations of the two nodes:

[0041]

[0042] In the formula and They are nodes and nodes The design considers the comprehensive compressibility modulus of the foundation at the location. The technical intent of this design is that areas with abrupt changes in geological conditions (such as stratigraphic boundaries with significant differences in compressibility modulus) are often high-risk areas for differential settlement. Encoding geological differences into the graph edge weights allows the subsequent graph attention mechanism to automatically reduce the information propagation intensity between nodes across geological zones during information aggregation, preventing the model from confusing significantly different settlement patterns. In an alternative implementation, the geological consistency coefficient can also be defined according to the matching degree of soil layer classification codes or other indicators characterizing geological similarity, as long as its value reflects the degree of difference in geological conditions between the two nodes.

[0043] Temporal feature extraction submodule 202 for each node time series Three types of time-series features were extracted.

[0044] The first category is multi-scale statistical characteristics. Statistical indicators of settlement are calculated at three different time scales: short-term window... The step corresponds to approximately 1 hour, a mid-term window. The step corresponds to approximately 1 day, and the long-term window is... Each step corresponds to approximately 7 days. At each scale, the mean, standard deviation, rate of change (first-order difference mean), and acceleration (second-order difference mean) of the settlement within that window are extracted, forming a 12-dimensional statistical feature vector across the three scales. This multi-scale statistical feature configuration allows the model to simultaneously perceive the short-term settlement response caused by construction loading, the daily consolidation process, and the long-term trend changes at the weekly scale.

[0045] The second category is trend-cycle decomposition features. The STL decomposition method is applied to the settlement time series of each node to separate the original sequence into trend components. Seasonal portion and residual components The model consists of three components. The trend component primarily reflects the evolution of long-term consolidation settlement; the seasonal component captures settlement fluctuations caused by periodic changes in temperature and groundwater level; and the residual component corresponds to the impact of random events such as construction disturbances. Inputting these three components as independent feature channels into the model, compared to directly using the original sequence, helps the model separate the settlement contributions from different physical mechanisms during learning.

[0046] The third category is construction status coding features. The current construction stage is uniquely coded. In a preferred embodiment, the construction stage includes four categories: foundation treatment period, subgrade filling period, preloading period, and pavement construction period, but this can be increased or decreased according to actual project needs. Furthermore, the time interval since the most recent graded loading is also included. The cumulative number of compaction passes and the most recent compaction operation The inclusion of construction state features allows the model to distinguish the differences in settlement response mechanisms under different construction conditions. For example, the settlement rate usually increases significantly in the short period after graded loading, while the settlement during the preloading period is mainly characterized by slow consolidation.

[0047] After the above three types of feature extraction, the feature dimension of each node at each time step is expanded to: ,in This represents the number of construction stage categories.

[0048] The physical prior feature injection submodule 203 injects analytical calculation results based on classical Biot consolidation theory as prior features into the feature vector of each monitoring node. The design rationale behind this submodule is that purely data-driven models are prone to overfitting in engineering scenarios with limited training data, while classical consolidation theory, despite its numerous simplification assumptions, provides a basic trend reference for settlement evolution. Using the theoretical solution as feature input is equivalent to providing the model with a "benchmark prediction," transforming the model's actual task into learning the deviation between theoretical and measured values, thereby reducing the learning difficulty and accelerating the convergence process.

[0049] Specifically, for nodes Based on the known load conditions and geological parameters, the theoretical degree of consolidation at the location is calculated using Terzaghi's one-dimensional consolidation theory at different times. and theoretical settlement The formula for calculating the theoretical degree of consolidation is:

[0050]

[0051] In the formula , As a time factor, For nodes The vertical consolidation coefficient at that location, Let be the length of the drainage path. The truncation order of the series summation. Take a value that makes the series residual term less than a set precision (e.g.) The smallest positive integer is required. The theoretical settlement is calculated as follows:

[0052]

[0053] In the formula This represents the final settlement of the node estimated based on the layered summation method or other empirical methods. The theoretical degree of consolidation will then be... Theoretical settlement and the residual between theoretical and measured values The three physical quantities are concatenated as additional features into the node feature vector. The residual term... It inherently contains information such as nonlinear formation response, construction disturbance effect, and geological parameter error that cannot be captured by theoretical models, and is a compensation quantity that the model needs to focus on learning.

[0054] The Physically Constrained Spatiotemporal Graph Network (PI-STGN) prediction module 300 is the core prediction engine of this system. Its task is to receive the enhanced feature tensor and graph topology, and output the predicted settlement values ​​and uncertainty estimates of each monitoring node in the entire road segment at multiple future time steps. This module uses a physically constrained spatiotemporal graph network as its core architecture. Unlike existing technologies that treat each monitoring point as an independent time series for separate prediction, PI-STGN models all monitoring points as nodes on a graph. In the spatial dimension, it learns the correlation weights between nodes through a graph attention mechanism, and in the temporal dimension, it extracts multi-scale temporal dependencies through dilated causal convolution. During training, the residuals of the partial differential equations of the Biot consolidation equation are embedded as regularization terms in the loss function to ensure the physical consistency of the prediction results. This module includes a graph attention spatial encoding submodule 301, a dilated causal temporal convolution submodule 302, a physical regularization training submodule 303, and a multi-step rolling prediction submodule 304.

[0055] The graph attention spatial coding submodule 301 learns the dynamic spatial correlation weights between monitoring nodes at each time step and updates the spatial embedding representation of the current node by aggregating neighbor node information. For each time step... ,node The input features are denoted as First, query vectors are generated through three sets of linear transformations. Key vector Sum value vector ,in For a learnable parameter matrix, For a single attention head, the dimension is [not specified]. Node Its neighboring nodes The attention coefficient is calculated as follows:

[0056]

[0057] In the formula The adjacency matrix elements generated by the aforementioned spatial topology construction submodule This is used to prevent overflow in logarithmic operations. It is introduced into the calculation of the attention coefficient. The technical significance of this method lies in incorporating prior information about the graph topology into the calculation of attention weights. For node pairs with larger adjacency weights, their attention coefficients receive a positive bias; conversely, for node pairs with adjacency weights close to zero, they receive a negative bias. This allows the attention mechanism to adaptively learn while adhering to the prior structure defined by spatial topology and geological conditions. Subsequently, the attention coefficients are normalized using softmax.

[0058]

[0059] In the formula For nodes The set of neighboring nodes in the graph. Node The updated spatial embedding calculation is as follows:

[0060]

[0061] In the formula The GELU activation function is used. The system defaults to a multi-head attention mechanism with 8 attention heads. Each head computes independently, and the outputs are concatenated along the feature dimension before being mapped back through a linear projection layer. dimension:

[0062]

[0063] Finally, after LayerNorm normalization and residual connection, the spatial encoding result is output:

[0064]

[0065] The aforementioned residual connections help alleviate the gradient degradation problem in deep networks while preserving the original feature information of the nodes themselves.

[0066] The function of the dilated causal temporal convolutional submodule 302 is to extract multi-scale temporal dependent features from the spatially encoded node sequence. For nodes... , in continuous Spatial coding results at each historical time step The input is fed into a temporal convolutional network. This network consists of... The system consists of stacked dilated causal convolutional layers, with each layer having a dilation rate set to [value missing]. That is, the kernel sizes are 1, 2, 4, and 8 respectively. . No. The convolution calculation of a layer can be expressed as:

[0067]

[0068] in This serves as the input for the first layer. The causal convolutional architecture ensures that the time steps... The output depends only on the input values ​​at that moment and those prior, strictly avoiding the leakage of future information. The exponentially increasing dilation rate allows the network's receptive field to expand exponentially with the number of layers: the total receptive field of a 4-layer dilated convolution is... Each time step, with a temporal resolution of 10 minutes per step, covers approximately 5 hours of historical information, which is sufficient to capture the typical delay characteristics between construction loading and settlement response.

[0069] A gating mechanism is introduced after each convolutional layer to enhance the network's non-linear expressive power:

[0070]

[0071] In the formula and The first The layer output channel is divided into the first and second halves by average along the feature dimension. This represents element-wise multiplication. The sigmoid function acts as a "gate," controlling the proportion of information from the tanh transform that is passed to subsequent layers. Residual connections maintain open information pathways between layers, and Dropout regularization (with a dropout rate of 0.2) is applied to suppress overfitting. The network takes the gated output of the last time step. As a node spatiotemporal joint embedding vector This vector simultaneously encodes the spatial context information and temporal evolution characteristics of the node, and can be mapped to the settlement prediction value through a linear decoding layer.

[0072] The Physical Regularization Training Submodule 303 defines the composite loss function for the PI-STGN model, which is a weighted sum of a data fitting term and a physical constraint term.

[0073]

[0074] In the formula These are the physical constraint weight coefficients. During training, Dynamically adjust the initial value according to the annealing strategy. Every 50 epochs of training The incremental approach focuses the model's initial training on fitting the observed data to quickly establish a basic mapping relationship. Subsequently, the weight of physical constraints is gradually increased to correct non-physical fitting biases. This progressive constraint embedding strategy avoids the problem of training instability caused by imposing excessively strong physical constraints before the model parameters have reached a reasonable range.

[0075] The data fitting term is defined as follows:

[0076]

[0077] In the formula For the model to nodes In the future Predicted settlement values ​​at each time step The corresponding measured value, To predict the step size.

[0078] The physical constraint terms are constructed based on a simplified form of Biot's three-dimensional consolidation equations. For saturated soft soil foundations, the vertical excess pore water pressure... The dissipation follows the following partial differential equation:

[0079]

[0080] The equation is discretized using a finite difference scheme on the spatiotemporal grid of the monitoring nodes. The settlement rate output by the prediction model is then used in conjunction with the effective stress principle. After calculating the change in excess pore water pressure at each node and time step, the physical constraint term is defined as the mean square value of the discrete residuals of the finite difference:

[0081]

[0082] In the formula For nodes calculated based on a rigorous mapping of the predicted settlement rate At time step The excess pore water pressure value is calculated as follows: According to the effective stress principle, during the consolidation stage after the application of the roadbed filling load, the total stress of the soil element remains constant. Therefore, the dissipation rate of the excess pore water pressure is equal to the increase rate of the effective stress, i.e. Combining one-dimensional compression theory, nodes The effective stress increment of the representative soil element and the resulting vertical volumetric strain Proportional, that is ,in This represents the overall compression modulus of the soil layer; the local vertical strain can be predicted by the model, and the node represents the soil layer thickness. Internal compression settlement Calculated, i.e. Substituting the strain expression above into the effective stress formula and differentiating it with respect to time, we can establish the relationship between the rate of change of pore water pressure and the settlement rate predicted by the model. The mathematical relationship between them: At the discrete time steps specified in the model Within the framework, nodes At time step The excess pore water pressure value The predicted settlement rate is obtained by discrete integration and summation: In the formula The initial excess pore water pressure generated at the beginning of the load is applied. Furthermore, The vertical spacing between adjacent deep sedimentation meters. For nodes The vertical consolidation coefficient at a given location. This physical regularization term requires that the settlement prediction output by the model macroscopically satisfy the excess pore water pressure dissipation law described by the consolidation equation, thereby effectively suppressing non-physical settlement abrupt changes or oscillating predictions that may arise from a purely data-driven model. The consolidation coefficient for different soft soil layers... It can be obtained from geological exploration data, or it can be used as a learnable parameter for automatic model calibration during training.

[0083] In terms of training configuration, the AdamW optimizer is used, and the initial learning rate is [value missing]. Decay to cosine annealing strategy The training batch size is set to 64, and the total training epochs are 300. An early stopping strategy is used—training is terminated when the validation set loss no longer decreases for 30 consecutive epochs to prevent overfitting. The historical input window length is also specified. The prediction window is set to 144 time steps, or approximately one day, of historical data, for a single forward propagation. The forecast is for 432 time steps, or approximately 3 days, of future settlement. For long-term forecasts exceeding 3 days, a rolling forecast mechanism extends the forecast period to 30 days.

[0084] The multi-step rolling prediction submodule 304 performs multi-step rolling prediction during the inference phase to calculate long-term subsidence trends. A single forward propagation can output future... The predicted settlement value for each step. For those exceeding... For long-term prediction, an autoregressive rolling strategy is employed: the output of the previous prediction is appended to the end of the current input window, while historical data of equal length at the beginning of the input window is discarded to maintain the window length. This forms a new input window, which is then used for forward propagation again. During each rolling process, if new actual monitoring data arrives within the predicted time range, the predicted value at the corresponding position is replaced with the measured data (measured data priority principle), thus continuously correcting accumulated errors during the rolling prediction process. Through this mechanism, the accuracy of long-term predictions does not rapidly decline due to the gradual accumulation of errors in the autoregressive chain. At each prediction time step, the model outputs the predicted settlement value... In addition, it also outputs an uncertainty estimate. Uncertainty estimation is obtained using the Monte Carlo Dropout method—the Dropout layer is kept active during the inference phase, 20 forward propagations are performed on the same input, and the standard deviation of each output is taken as the uncertainty measure for that prediction point, characterizing the width of the prediction confidence interval. The uncertainty estimation information will be passed to downstream control and early warning modules for use.

[0085] The reinforcement learning-based inverse control module 400 automatically generates pile parameter adjustment schemes and compaction strategy adjustment instructions for variable stiffness composite foundations based on the settlement field prediction results output by the prediction module 300. This module employs a reinforcement learning paradigm to model the construction control problem as a sequential decision-making process. Through a policy network, it searches for the globally optimal control scheme under multi-objective constraints, replacing the traditional sequential mode that relies on manual experience. This module includes an environment modeling submodule 401, a policy network submodule 402, and a control instruction generation submodule 403.

[0086] The environmental modeling submodule 401 encapsulates the output of the prediction module 300 and the construction state information into a Markov Decision Process (MDP) environment. State space The definition of is:

[0087]

[0088] In the formula This represents the current predicted settlement vector for all nodes. This represents the differential settlement vector between adjacent nodes. This is the current composite foundation pile parameter vector (including pile length, pile diameter, and pile spacing for each control section). This is the current compaction parameter vector (containing the number of compaction passes and vibration frequency for each section). The residual budget vector contains three pieces of information: the proportion of remaining material budget, the proportion of remaining construction period, and the proportion of completed work. The introduction of the residual budget vector enables the strategy network to perceive project progress and resource consumption, automatically balancing control effectiveness with cost constraints in regulatory decisions.

[0089] Action space Defined as a continuous space:

[0090]

[0091] In the formula This represents the increment in pile length for each control zone, with positive values ​​indicating longer piles, and the range of values ​​is [value missing]. ; This represents the pile spacing adjustment; negative values ​​indicate denser pile foundation layout, and the range is [value missing]. ; The increment for the number of compaction passes, with a value range of [value range missing]. . To control the total number of sections, the system divides the entire road segment into 50m sections. The value range of each action component is set based on the operable range of pile foundation adjustment and compaction operations in actual engineering, which ensures that the exploration space of the strategy network is large enough to search for effective control schemes, while avoiding the generation of extreme parameters that are not feasible in engineering.

[0092] The design of the reward function is crucial to the effectiveness of reinforcement learning regulation. This system defines a comprehensive reward function comprising four components:

[0093]

[0094] The specific definitions of each item are as follows. The differential settlement compliance penalty item measures the degree to which differential settlement at each node exceeds the allowable limit under the current control scheme:

[0095]

[0096] In the formula mm / m represents the allowable differential settlement rate. The material cost penalty applies to material consumption resulting from increased pile length and decreased pile spacing.

[0097]

[0098] In the formula and These are the normalized cost coefficients corresponding to the increase in unit pile length and the decrease in unit pile spacing, respectively. The schedule penalty term penalizes delays caused by control measures.

[0099]

[0100] In the formula This represents the construction delay calculated based on changes in pile length, pile spacing, and number of compaction passes. The safety margin bonus provides positive incentives for sections where differential settlement is significantly below the allowable value.

[0101]

[0102] The weighting coefficients for each sub-item are set as follows: , , , The weight of the differential settlement compliance penalty item is significantly higher than that of other items, reflecting a multi-objective priority strategy that prioritizes differential settlement control, followed by minimizing costs and shortening the construction period while meeting control objectives. In practical engineering applications, the above weight coefficients can be adjusted according to the specific project's management needs and control preferences.

[0103] Policy network submodule 402 uses the Proximal Policy Optimization (PPO) algorithm to train the policy network. The policy network structure is as follows: the state vector... The input is a three-layer fully connected network with hidden layer dimensions of 512, 256, and 128 respectively. The activation function is ReLU, and the output layer generates the mean vector of the actions. Sum of logarithmic and standard deviation vectors The motion follows a Gaussian distribution. The values ​​are sampled and mapped to a bounded action space using the tanh function to satisfy the range constraints of each action component. Value function network By sharing the parameters of the first two layers with the policy network and only outputting scalar value estimates in the last layer branch, this parameter-sharing structure is beneficial for accelerating feature learning in the early stages of training.

[0104] The objective function for PPO is defined as follows:

[0105]

[0106] In the formula The importance sampling ratio between the old and new strategies. The value of the dominance function is calculated using the generalized dominance estimation (GAE) method. This is for pruning parameters. The pruning operation limits the magnitude of each parameter update, ensuring the stability of the strategy update process.

[0107] The policy network is trained through simulation. Each training round corresponds to a complete road construction simulation process. In the simulation, prediction module 300 serves as the environmental dynamics model. Given the current pile foundation and compaction parameters, the evolution of the future settlement field is obtained through forward inference by prediction module 300. Based on this, the reward is calculated and the system transitions to the next state. The network parameters are updated every 2048 steps of accumulated interaction data, for a total of 5000 training rounds, to ensure that the policy converges to a stable control strategy.

[0108] The control command generation submodule 403 operates online during actual construction. Whenever new monitoring data arrives at the system and the settlement prediction is updated by the prediction module, this submodule will generate the current state vector. Input the trained policy network to obtain the control actions. Then, perform three post-processing operations in sequence.

[0109] The first step is feasibility constraint trimming. Components in the action vector that do not meet engineering constraints are trimmed to the feasible region. Engineering constraints include, but are not limited to: pile length not exceeding the bearing layer depth, pile spacing not less than 3 times the pile diameter, and the number of compaction passes not exceeding the single-operation limit of the road roller equipment. This step ensures that the control commands are executable at the engineering level.

[0110] The second step is smoothing filtering. A one-dimensional Gaussian smoothing filter is applied to the parameter increments of adjacent control sections, with a filter kernel width of three sections, to eliminate unreasonable abrupt changes in parameters between adjacent sections caused by local noise in the prediction model. The physical basis for smoothing is that in actual engineering, the design parameters of pile foundations should maintain a certain degree of gradual change along the longitudinal direction. Drastic changes in parameters are not conducive to construction implementation and do not meet the requirement of continuity in the transfer of foundation loads.

[0111] The third item is instruction encoding and distribution. The incremental control parameters, after being trimmed and smoothed, are encoded into construction control instructions in standard JSON format. The instructions include fields such as target section number, pile length adjustment value, pile spacing adjustment value, compaction pass adjustment value, and confidence rating, and are distributed to the construction management platform and intelligent compaction equipment terminal through the system communication interface.

[0112] The Visual Decision and Early Warning Module 500 undertakes three functions: intuitive presentation of system operation status, real-time monitoring of construction safety, and periodic assessment of construction quality. It includes a 3D settlement field rendering submodule 501, a multi-level early warning engine submodule 502, and a construction quality report generation submodule 503.

[0113] The 3D settlement field rendering submodule 501 visualizes the settlement field data output by the prediction module as a 3D heatmap on the GIS platform. Using the 3D geometric model of the roadbed as a base map, this submodule renders the measured settlement value, predicted settlement value, and differential settlement value of each monitoring node at its corresponding spatial location using different color levels. It supports dragging along the time axis to view the continuous dynamic changes of the settlement field from historical measurements to future predictions. On top of this, a pile foundation layout layer is overlaid, where pile length is represented by the height of a bar chart and pile spacing by color density; simultaneously, a compaction trajectory layer is overlaid, with the number of compaction passes represented by element transparency and compaction degree by color levels. Through the multi-layered display of "geology—settlement—construction," on-site managers and technical decision-makers can intuitively grasp the correspondence between settlement evolution trends and construction control measures within a single view.

[0114] The multi-level early warning engine submodule 502 establishes a three-level progressive early warning mechanism. The triggering conditions and response measures for each level of early warning are as follows: Blue warning is the level of concern; when the predicted differential settlement rate of any monitoring node... Exceeding 60% of the allowable value When the settlement rate reaches 0.5 mm / m, the system will highlight the area on the visualization interface and push a notification to the construction management terminal to remind relevant personnel to pay attention to the settlement trend of that section. An orange alert is a warning level; if the predicted differential settlement rate at any node exceeds 80% of the allowable value, the system will issue a warning. mm / m, or model prediction uncertainty When the predicted value exceeds 30%, the system automatically invokes the reinforcement learning inverse control module 400 to generate a control plan suggestion and pushes it to the project's technical lead for review. High prediction uncertainty is one of the triggering conditions for an orange alert. This is because even if the predicted value itself has not exceeded the limit, a low reliability of the prediction result means there is a potential risk of actual settlement exceeding the limit; early warning helps to gain a window of opportunity for control. A red alert is a response level alert; it is triggered when the predicted differential settlement rate at any node exceeds 95% of the allowable value. When the speed reaches mm / m, the system triggers a control command directly to the intelligent compaction equipment terminal and generates an emergency response plan for on-site management personnel to make decisions. Emergency measures options include partial work stoppage, increased monitoring frequency, and additional pile foundations, requiring on-site management personnel to confirm implementation within 30 minutes. The three-level early warning mechanism's threshold gradient design balances the sensitivity and practicality of the warnings—the blue warning threshold is set relatively low to ensure potential risks are detected as early as possible, while the red warning threshold is close to the engineering control limit to avoid frequent false alarms interfering with normal construction.

[0115] The construction quality report generation submodule 503 automatically generates a daily construction quality report. The report includes a comparison chart of measured and predicted settlement values ​​for the entire road section, a statistical table of differential settlement compliance rates for each controlled section, a tracking record of the implementation of control instructions, and evaluation indicators of model prediction accuracy (including mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination). The report includes a forecast of settlement trends and risk warnings for the next 7 days. It is generated in PDF format and incorporated into the project's full lifecycle database, providing data support for subsequent project acceptance and long-term performance evaluation.

[0116] The complete workflow of this system forms a closed-loop signal transmission path of "data acquisition → feature construction → settlement prediction → control decision → command execution → effect feedback".

[0117] During the data acquisition and fusion phase, multiple sensors deployed in the widened roadbed area, including GNSS high-precision positioning devices, deep-layer settlement meters, a high-pressure jet grouting pile intelligent monitoring system, an intelligent compaction equipment sensor cluster, temperature sensors, and groundwater level monitoring wells, continuously acquire raw observation data. The raw data is transmitted wirelessly to the edge computing node and then enters the multi-source data fusion acquisition module 100. This module performs outlier detection and removal, missing value imputation, and multi-scale time synchronization processing. After aligning all data to a unified spatiotemporal coordinate system, a standardized spatiotemporal data tensor is output. .

[0118] In the spatiotemporal feature engineering phase, standardized data tensors are entered into the spatiotemporal feature engineering module 200. The spatial topology construction submodule 201 constructs a weighted adjacency graph based on the spatial location and geological condition similarity of each monitoring node. The temporal feature extraction submodule 202 extracts multi-scale statistical features, trend-period decomposition features, and construction status coding features from the time series of each node; the physical prior feature injection submodule 203 incorporates the analytical calculation results of Terzaghi consolidation theory as physical prior features into the feature vectors of each node, forming an enhanced feature tensor together.

[0119] In the settlement prediction stage, the enhanced feature tensor and graph topology are input into the physically constrained spatiotemporal graph network prediction module 300. The graph attention spatial encoding submodule 301 aggregates neighbor node information at each time step through a multi-head attention mechanism to learn the spatial propagation pattern of the settlement field. The spatial encoding sequence is fed into the dilated causal temporal convolution submodule 302, where it extracts multi-scale temporal dependent features through four layers of dilated causal convolution to form a node-level spatiotemporal joint embedding vector. The embedding vector is then processed by a linear decoding layer to output the settlement prediction values ​​and uncertainty estimates for multiple future time steps. During training, the physical regularization training submodule 303 uses the finite difference residuals of the Biot consolidation equation as a regularization term to embed the loss function. In the inference stage, a 30-day long-term prediction is achieved through the multi-step rolling prediction submodule 304.

[0120] During the control and decision-making phase, the predicted settlement and differential settlement of each node output by the prediction module, together with the current pile foundation parameters, compaction parameters, and residual budget information, constitute a state vector, which is input into the reinforcement learning inverse control module 400. The policy network outputs the pile length increment, pile spacing adjustment, and compaction pass increment for each control section according to the PPO algorithm. The control instructions are encoded into standardized construction control instructions after feasibility constraint pruning and smoothing filtering of adjacent sections.

[0121] During the instruction execution phase, the generated control instructions are sent to two execution terminals via the system interface. One is the construction management platform, which receives the adjustment plan for pile length and pile spacing and pushes it to the on-site technical management personnel, who then organize and implement the subsequent adjustment of pile foundation construction parameters. The other is the vehicle-mounted terminal of the intelligent compaction equipment, which receives the adjustment instructions for the number of compaction passes and compaction path and displays them directly on the roller's operating interface. The operator then executes the compaction reinforcement operation according to the instructions.

[0122] During the feedback and model update phase, after the control commands are executed, multi-source sensors continuously collect new settlement monitoring data. This new data re-enters the data acquisition phase, initiating a new closed-loop cycle. Simultaneously, the system incrementally fine-tunes the PI-STGN prediction model weekly using the latest accumulated measured data—training for five epochs with the most recent week's data while freezing all network parameters except the last two layers, allowing the model to adapt to dynamic changes in geological conditions as construction progresses. The strategy network updates its strategy online every two weeks using the latest environmental interaction data to continuously optimize the control strategy. The visualization decision-making and early warning module 500 operates throughout the entire construction process, rendering the three-dimensional settlement field in real time, executing three-level early warning logic, and generating daily construction quality reports, thus forming a complete "perception-prediction-control-feedback" intelligent closed-loop management system.

[0123] It should be noted that the specific numerical parameters in the above embodiments (such as GNSS deployment spacing of 50m, sampling frequency, sliding window width of 120 points, cutoff distance of 200m, network hidden layer dimension, training rounds, early warning threshold, etc.) are all example values ​​in preferred embodiments. Those skilled in the art can reasonably adjust the above parameters according to factors such as geological conditions, road length, monitoring accuracy requirements, and computing resource configuration of specific projects. As long as they do not deviate from the core technical concept of the present invention, they should be considered to fall within the protection scope of the present invention. Similarly, specific algorithm components used in the system (such as GELU activation function, AdamW optimizer, PPO algorithm, etc.) can be replaced by alternatives with equivalent functions in the art. For example, the GELU activation function can be replaced by ReLU or Swish function, and the PPO algorithm can be replaced by other policy gradient reinforcement learning algorithms such as SAC (SoftActor-Critic). These substitutions do not change the technical essence of the present invention.

[0124] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A machine learning-based system for predicting and controlling differential settlement on highways, characterized in that, include: The multi-source data fusion acquisition module is used to acquire sensor data from the engineering site and output standardized spatiotemporal data tensors after preprocessing. The spatiotemporal feature engineering module is used to construct a weighted adjacency graph containing geological consistency coefficients based on the standardized spatiotemporal data tensor, and extract multi-scale temporal features and inject physical prior features to generate an enhanced feature tensor. The physically constrained spatiotemporal graph network prediction module is used to extract spatiotemporal dependencies by utilizing the enhanced feature tensor and the weighted adjacency graph through graph attention mechanism and dilated causal convolution, and output the settlement prediction value and uncertainty estimate for future time steps; during the training process, the physically constrained spatiotemporal graph network prediction module embeds the finite difference residual of the Biot consolidation equation as a regularization term into the loss function. The reinforcement learning inverse control module is used to construct a state space containing the settlement prediction value and construction parameters, and to output control instructions for pile foundation parameters and compaction strategy using a strategy network. The visualization decision-making and early warning module is used to display the three-dimensional settlement field and trigger multi-level early warnings based on the predicted differential settlement rate.

2. The machine learning-based highway differential settlement prediction and control system according to claim 1, characterized in that, The multi-source data fusion acquisition module includes a sensor access submodule, a data preprocessing submodule, and a spatiotemporal alignment submodule. The data preprocessing submodule is configured to: for long, discontinuous missing data with a continuous missing length exceeding a set threshold, use truncated singular value decomposition to complete the low-rank matrix composed of sensor observations from the same cross-section; and for sensor data with different sampling frequencies, uniformly resample to the target temporal resolution through moving average or linear interpolation.

3. The machine learning-based highway differential settlement prediction and control system according to claim 1, characterized in that, The spatiotemporal feature engineering module includes a spatial topology construction submodule, which constructs an adjacency matrix of an undirected weighted graph. The weights of its elements are determined by the Euclidean distance attenuation term between nodes and the geological consistency coefficient. When two nodes belong to different geological zones, the geological consistency coefficient is taken as the reciprocal of the ratio of the comprehensive compression modulus of the foundation at the location of the two nodes, which is used to reduce the information propagation intensity between nodes across geological zones.

4. The machine learning-based highway differential settlement prediction and control system according to claim 1, characterized in that, The spatiotemporal feature engineering module also includes a physical prior feature injection submodule. This submodule uses Terzaghi's one-dimensional consolidation theory to calculate the theoretical degree of consolidation and theoretical settlement of each node at different times, and splices the theoretical calculated values ​​and the residuals between the theoretical values ​​and the measured values ​​as additional feature channels into the node feature vector.

5. The machine learning-based highway differential settlement prediction and control system according to claim 1, characterized in that, The physical constraint spatiotemporal graph network prediction module includes a graph attention spatial encoding submodule. When calculating the attention coefficients between nodes, this submodule introduces the logarithmic term of the adjacency matrix elements as a bias, so that the attention weights follow the prior structure defined by the spatial topology and geological conditions. The module also includes an extended causal temporal convolution submodule, which uses multi-layer extended causal convolution to extract multi-scale temporal dependency features of the node spatial encoding sequence.

6. The highway differential settlement prediction and control system based on machine learning according to claim 1, characterized in that, The physical regularization training submodule in the physical constraint spatiotemporal graph network prediction module defines a composite loss function, which is composed of a weighted sum of data fitting terms and physical constraint terms. The physical constraint terms are constructed based on the finite difference scheme of the Biot consolidation equation, and calculate the mean square value of the discrete residual between the excess pore water pressure dissipation process predicted by the model and the law described by the consolidation equation. The weight coefficients of the physical constraint terms are dynamically increased with the training rounds using an annealing strategy.

7. The machine learning-based highway differential settlement prediction and control system according to claim 1, characterized in that, The reinforcement learning inverse control module includes an environment modeling submodule. This submodule defines the state space as a combination of the current predicted settlement vector, the differential settlement vector between adjacent nodes, the current composite foundation pile parameter vector, the current compaction parameter vector, and the residual budget vector; and defines the action space as a continuous space containing the pile length increment, the pile spacing adjustment, and the compaction pass increment.

8. The machine learning-based highway differential settlement prediction and control system according to claim 7, characterized in that, The reinforcement learning inverse control module uses a comprehensive reward function to evaluate the control strategy. The comprehensive reward function includes a differential settlement compliance penalty, a material cost penalty, a construction period penalty, and a safety margin reward. The weight coefficient of the differential settlement compliance penalty item is set to the highest level to ensure that settlement control is the primary optimization objective.

9. The machine learning-based highway differential settlement prediction and control system according to claim 7, characterized in that, The reinforcement learning inverse control module also includes a control instruction generation submodule. After the policy network outputs the action vector, the submodule performs post-processing operations: trimming the action components to the engineering feasible region; applying a one-dimensional Gaussian smoothing filter to the parameter increments of adjacent control sections to eliminate parameter abrupt changes; and encoding the processed increments into construction control instructions for issuance.

10. The machine learning-based highway differential settlement prediction and control system according to claim 1, characterized in that, The visualization decision-making and early warning module includes a multi-level early warning engine sub-module, which sets a three-level early warning mechanism: when the predicted differential settlement rate exceeds the first threshold, a attention-level early warning is triggered. When the predicted differential settlement rate exceeds the second threshold or the prediction uncertainty exceeds the set ratio, an alert level warning is triggered and the reinforcement learning inverse control module is invoked; when the predicted differential settlement rate exceeds the third threshold, a disposal level warning is triggered and an emergency command is issued.

Images