A method for predicting pavement settlement during tunnel construction based on Bayesian dynamic data assimilation
A tunnel construction pavement settlement prediction model was constructed by using the Bayesian dynamic data assimilation method. This model solved the problem of real-time dynamic prediction of on-site monitoring data during tunnel construction, and achieved the integration of engineering mechanism constraints and monitoring data, providing accurate settlement prediction and risk assessment.
Patent Information
- Application Number
- CN202611191854.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-07
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2046-08-07
AI Technical Summary
Existing technologies cannot effectively utilize on-site monitoring data for real-time dynamic prediction of road surface settlement during tunnel construction. They lack engineering mechanism constraints and uncertainty quantification, resulting in inaccurate prediction results and difficulty in meeting the rapid response requirements during the construction period.
A Bayesian dynamic data assimilation-based method is used to construct a priori settlement field model for the project. The residuals are probabilistically modeled using a Gaussian process, and the model is continuously updated using the Bayesian assimilation method to quantify the risk. A comprehensive risk index is then constructed for dynamic classification and evaluation.
It enables dynamic rolling prediction of road surface settlement during tunnel construction, which conforms to engineering mechanisms, provides probabilistic risk assessment, improves the accuracy and reliability of prediction, and meets the needs of rapid response during construction.
Smart Images

Figure CN122693496B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of geotechnical engineering monitoring and risk control technology, and in particular to a method for predicting pavement settlement during tunnel construction based on Bayesian dynamic data assimilation. Background Technology
[0002] Road surface settlement and associated differential deformation induced by tunnel construction under existing highways are core factors restricting project safety. Tunnel excavation releases stress in the overlying strata, and deformation is transmitted to the road surface structure. If not properly controlled, this will directly threaten the operational safety of the highway. Therefore, during construction, how to accurately predict the spatiotemporal evolution of road surface settlement using limited on-site monitoring data, and dynamically identify and evaluate potential risk areas, is a key technical challenge that urgently needs to be solved in the construction of such projects.
[0003] Currently, methods for predicting surface settlement caused by tunnel construction are mainly divided into three categories: empirical analysis method, numerical simulation method, and data-driven method.
[0004] Empirical analytical methods, represented by the Peck formula and its derivatives, describe the spatial morphology of surface settlement by assuming a normal distribution of settlement troughs. Based on this, researchers have developed settlement-time curve models that consider time effects, along with various modified formulas applicable to different geological formations and construction conditions. The advantages of these methods are their simplicity, the clear physical meaning of their parameters, and their ability to reflect the fundamental laws of settlement well. However, their limitations are equally significant: key parameters in the model (such as the settlement trough width coefficient and maximum settlement) are usually determined based on engineering experience or back-analysis after construction, making real-time adjustments based on on-site monitoring feedback impossible during construction; furthermore, their descriptive power is limited, making it difficult to accurately characterize the settlement response under the influence of complex factors such as heterogeneous formations and time-varying construction disturbances.
[0005] Numerical simulation methods (such as the finite element method and the finite difference method) can comprehensively simulate the deformation mechanism of strata under complex construction conditions by establishing a detailed numerical model that includes the constitutive properties of stratum materials, the simulation of tunnel excavation, and the role of support structures. These methods can consider the coupling effects of multiple factors and are suitable for scheme comparison and sensitivity analysis. However, numerical simulation faces bottlenecks such as numerous modeling parameters, long computation time, and difficulty in determining constitutive model parameters, making it difficult to achieve rapid and frequent rolling predictions during the construction period, and even more difficult to make timely corrections and updates based on field monitoring data.
[0006] In recent years, with the development of machine learning technology, data-driven methods such as random forests, support vector machines, and multilayer perceptrons have been gradually introduced into tunnel construction settlement prediction. The advantage of these methods lies in the fact that they do not require a pre-set physical model and can directly learn the complex mapping relationship between input features (such as burial depth, cross-sectional dimensions, construction parameters, etc.) and output settlement from historical monitoring data, exhibiting strong nonlinear fitting capabilities. However, purely data-driven methods have inherent drawbacks: First, the availability of high-quality monitoring samples for tunnel underpass projects is usually limited. Under sparse data conditions, machine learning models are prone to overfitting and have insufficient generalization performance. Second, the models lack constraints from engineering physical mechanisms, and the prediction results may exhibit phenomena that violate mechanical laws, such as distorted settlement trough morphology and spatial discontinuities, reducing the reliability and engineering acceptability of the predictions. Third, purely data-driven models are essentially "black box" models, making it difficult to provide quantitative information on the uncertainty of the prediction results, which is detrimental to risk assessment.
[0007] On-site monitoring data undoubtedly provides the most direct reflection of settlement during construction, but its value has not been fully realized. Existing monitoring data analysis methods mostly remain at the level of single-point settlement duration curve analysis or simple interpolation, failing to effectively integrate monitoring data with prior engineering knowledge. Monitoring points are typically sparsely distributed spatially, and monitoring data is gradually acquired as construction progresses. How to construct a complete spatiotemporal field of pavement settlement based on these limited, sequential data, and how to reasonably extrapolate the settlement of future unconstructed areas or unmonitored locations, remains an unresolved problem.
[0008] In summary, the main shortcomings of existing technologies can be summarized as follows:
[0009] (1) The empirical analysis method lacks the ability to dynamically assimilate data and cannot use on-site monitoring data to continuously correct prediction deviations as the construction progresses;
[0010] (2) Numerical simulation is computationally expensive and cannot meet the needs of rapid response during the construction period;
[0011] (3) Pure data-driven methods have poor prediction stability when the monitoring sample is limited, and lack engineering mechanism constraints and uncertainty quantification capabilities;
[0012] (4) A framework for probabilistically integrating engineering prior models with field monitoring data has not yet been established, nor has a systematic risk assessment method that comprehensively considers settlement amplitude, deformation morphology, development rate and prediction uncertainty been formed. Summary of the Invention
[0013] This invention provides a method for predicting pavement settlement during tunnel construction based on Bayesian dynamic data assimilation. This method can take into account both engineering mechanism constraints and monitoring data feedback, and achieve dynamic rolling prediction and risk quantification. It can improve the technical level of pavement settlement control during the construction period of tunnels passing under highways.
[0014] To solve the above-mentioned technical problems, the technical solution proposed by this invention is as follows:
[0015] A method for predicting pavement settlement during tunnel construction based on Bayesian dynamic data assimilation includes the following steps:
[0016] Step S1: Obtain pavement settlement monitoring data for the tunnel under the highway project, and construct an engineering prior settlement field model that integrates the morphology of the transverse settlement trough, the cumulative evolution over time, and non-stationary engineering corrections.
[0017] Step S2: Calculate the residual between the prior settlement field of the project and the field monitoring data. Use a Gaussian process to probabilistically model the residual. As construction monitoring data is continuously acquired, update the Gaussian process model to obtain the mean and posterior variance of the settlement prediction after Bayesian assimilation.
[0018] Step S3: Based on the posterior variance of the Gaussian process and the engineering correction uncertainty, calculate the confidence interval for pavement settlement prediction and the probability of settlement exceeding the limit;
[0019] Step S4: Based on the assimilated settlement prediction field, calculate the settlement gradient, settlement curvature and settlement change rate, and combine the settlement amplitude and the settlement exceedance probability to construct a comprehensive risk index for dynamic classification and evaluation of pavement construction risks.
[0020] Step S5 involves using a combination of engineering constraints, monitoring inversion, probabilistic training, and sensitivity analysis to determine and calibrate the model parameters involved in steps S1 to S4.
[0021] A further improvement to the above technical solution is as follows:
[0022] Preferably, the construction of the prior settlement field model in step S1 specifically includes:
[0023] A Gaussian-type settlement trough is used to describe the spatial distribution of lateral settlement of the road surface, and the shape of the settlement trough is controlled by the maximum settlement value and the settlement trough width parameter.
[0024] The Richards function is used to describe the development of road settlement over time, and the asymmetry of the settlement-time curve is adjusted by the growth rate parameter and the curve shape parameter.
[0025] By multiplying the transverse settlement trough model with a time evolution function, a foundation mechanism settlement field under homogeneous strata and stable construction conditions is established;
[0026] Non-stationary engineering corrections are introduced, including linear expansion of settlement trough width over time, periodic drift and local disturbance of settlement center position over time, and engineering correction coefficients that comprehensively reflect formation heterogeneity, changes during construction, and local random disturbances.
[0027] By incorporating non-stationary engineering corrections into the basic mechanism settlement field, an a priori settlement field for the engineering is constructed.
[0028] Preferably, in the non-stationary engineering correction:
[0029] The linear expansion of the settling trough width over time is controlled by the initial settling trough width and the settling trough expansion coefficient;
[0030] The change of the settlement center position over time is described by parameters including the initial settlement center position, periodic drift amplitude and frequency, local disturbance amplitude, local disturbance occurrence time, and local disturbance duration.
[0031] The engineering correction factor is composed of the superposition of the formation spatial heterogeneity correction term, the construction stage correction term, and the local random disturbance term, plus 1.
[0032] Preferably, step S2, which uses a Gaussian process to probabilistically model the residuals, specifically includes:
[0033] Calculate the residual between the measured settlement values at each monitoring point and the prior predicted values of the project, and form a residual vector;
[0034] The residuals are modeled as a Gaussian process with zero mean, and the combination of the squared exponential kernel function and the white noise term is used as the covariance function.
[0035] For any prediction point, calculate the posterior mean and posterior variance of the residuals;
[0036] The a priori predicted value of the project is added to the posterior mean of the residual to obtain the mean of the settlement prediction after Bayesian assimilation.
[0037] Preferably, the rolling update in step S2 is implemented as follows: whenever new monitoring data is obtained, the new data is added to the existing monitoring dataset, the residual vector is recalculated, the hyperparameters of the Gaussian process are re-optimized by maximizing the marginal log-likelihood function, and the posterior mean and posterior variance of the residuals of the prediction points are recalculated using the updated Gaussian process model to perform rolling updates of settlement prediction.
[0038] Preferably, step S3 specifically includes:
[0039] The confidence intervals for settlement prediction at monitoring points were constructed using the posterior standard deviation of a Gaussian process.
[0040] The engineering correction uncertainty is determined by comprehensively considering the posterior standard deviation of the Gaussian process, the normalized distance between the predicted point and the nearest monitoring point, the normalized settlement gradient, and the boundary extrapolation correction term.
[0041] Assuming that the actual settlement follows a normal distribution, the probability of settlement exceeding the limit at each location is calculated by combining the allowable settlement control value of the road surface.
[0042] Preferably, the specific method for calculating the settlement deformation characteristic index in step S4 is as follows:
[0043] Based on the assimilated settlement prediction field, the absolute value of the first-order partial derivative of the settlement along the transverse direction is calculated as the settlement gradient, which is used to reflect the degree of uneven settlement of the road surface.
[0044] The absolute value of the second-order partial derivative of the settlement along the transverse direction is calculated as the settlement curvature, which is used to reflect the local bending concentration effect of the pavement.
[0045] The absolute value of the first-order partial derivative of settlement with time is calculated as the settlement rate, which is used to reflect the speed of settlement development.
[0046] Preferably, the specific method for constructing the comprehensive risk index and classifying risk levels in step S4 is as follows: five indicators are used, namely settlement amplitude, settlement gradient, settlement curvature, settlement change rate and settlement exceedance probability. Each indicator is divided by its corresponding control value for normalization, and a comprehensive risk index is constructed by weighted summation. Based on the comprehensive risk index, the construction risk is divided into three levels: low risk, medium risk and high risk.
[0047] Preferably, the comprehensive risk index is:
[0048]
[0049] in, As a comprehensive risk index, This is the control value for settlement amplitude. This is the settlement gradient control value. This is the settlement curvature control value. This is the control value for the rate of change of settlement. These are the weights of each risk assessment indicator. The mean of the settlement predictions after Bayesian assimilation. For the settlement gradient, For settlement curvature, The rate of change of settlement, This represents the probability of settlement exceeding the limit.
[0050] Preferably, the determination and calibration of model parameters in step S5 specifically includes: using monitoring data from the early stage of construction, with the goal of minimizing the root mean square error between monitored settlement and model-predicted settlement, calibrating the physical parameters in the prior engineering model; and determining the hyperparameters of the Gaussian process from the residual data by maximizing the marginal log-likelihood function.
[0051] The tunnel construction pavement settlement prediction method based on Bayesian dynamic data assimilation provided by this invention has the following advantages compared with the prior art:
[0052] (1) The tunnel construction pavement settlement prediction method based on Bayesian dynamic data assimilation of the present invention constructs an engineering prior settlement field that integrates transverse settlement trough morphology, Richards time evolution, settlement trough width expansion, settlement center drift and comprehensive engineering correction. The method uses Bayesian residual assimilation to probabilistically integrate engineering mechanism constraints with real-time monitoring data with clear physical meaning. This ensures that the prediction results conform to mechanical laws and can continuously correct model deviations using field data.
[0053] (2) The tunnel construction pavement settlement prediction method based on Bayesian dynamic data assimilation of the present invention expands the prediction results from a single deterministic value to a probabilistic description including confidence interval and probability of exceeding limits through a dual quantification mechanism of Gaussian process posterior standard deviation and engineering correction uncertainty, realizing the transformation from deterministic prediction to probabilistic risk assessment. The constructed comprehensive risk index simultaneously incorporates five indicators: settlement amplitude, gradient, curvature, rate of change, and probability of exceeding limits, overcoming the limitation of traditional methods that rely solely on a single settlement threshold for judgment, and can more comprehensively reflect the stress and deformation state of the pavement structure.
[0054] (3) The tunnel construction pavement settlement prediction method based on Bayesian dynamic data assimilation of the present invention provides a complete parameter determination system. Physical parameters are calibrated by monitoring and inversion, statistical hyperparameters are automatically learned by maximizing likelihood, and risk weights are objectively determined by entropy weight method. The stability of the evaluation results is verified by weight perturbation sensitivity analysis, which effectively reduces the subjectivity of model parameter values and enhances the operability and reliability of the method in actual engineering. It can be widely applied to pavement settlement prediction and risk warning during the construction period of tunnels passing under highways. Attached Figure Description
[0055] Figure 1 This is a schematic diagram of a simplified model of a tunnel passing under a highway, as described in this invention.
[0056] Figure 2 This is a diagram showing the parameter sensitivity and risk weight analysis in the experimental verification, where (a) represents the parameter sensitivity of the engineering prior model and (b) represents the risk index weight determined by the entropy weight method.
[0057] Figure 3 The figure shows a comparison between the average settlement value of the monitored settlement field and the settlement value predicted by Bayesian assimilation in the experimental verification. (a) shows the settlement result of the monitored settlement field, and (b) shows the settlement result of Bayesian assimilation prediction.
[0058] Figure 4 This is a graph showing how the model prediction error changes with the update of monitoring data at different construction times during the experimental verification.
[0059] Figure 5 This is a distribution diagram of the standard deviation of the engineering correction uncertainty in the experimental verification.
[0060] Figure 6 The distribution diagrams of the probability of settlement exceeding the limit and the comprehensive risk index are shown in the experimental verification; where (a) is the probability of settlement exceeding the limit and (b) is the comprehensive risk index.
[0061] Figure 7 This is a comparison chart of the prediction results of different models in the experimental verification.
[0062] Figure 8 This is to verify the predicted results of the settlement trough during the final construction stage in the experiment. Detailed Implementation
[0063] The following provides a detailed description of specific embodiments of the present invention. It should be understood that the specific embodiments described herein are for illustrative and explanatory purposes only and are not intended to limit the scope of the invention.
[0064] like Figure 1 As shown, the present invention provides a method for predicting pavement settlement during tunnel construction based on Bayesian dynamic data assimilation. First, it establishes a priori engineering settlement field considering transverse settlement troughs, Richards time evolution, settlement trough expansion, settlement center drift, and heterogeneous engineering corrections. Second, it uses a Gaussian process to probabilistically model the residuals between the prior engineering model and field monitoring data, and dynamically assimilates them as the construction monitoring data is updated. Then, it constructs engineering correction uncertainties by combining the posterior standard deviation of the Gaussian process, monitoring constraints, settlement development, gradient concentration, boundary extrapolation, and risk proximity, and calculates the probability of settlement exceeding limits. Finally, it uses the entropy weight method to determine the weights of risk indicators, establishes a comprehensive risk assessment model including settlement amplitude, settlement gradient, settlement curvature, settlement change rate, and probability of exceeding limits, and establishes a method for pavement risk assessment throughout the entire construction process.
[0065] Specifically, the following steps are included:
[0066] Step S1: Data acquisition and construction of the engineering prior settlement field.
[0067] We acquired pavement settlement monitoring data for a tunnel passing under a highway and constructed a priori settlement field model that integrates transverse settlement trough morphology, time cumulative evolution, and non-stationary engineering corrections. Mathematical functions were used to quantitatively describe the basic distribution pattern of pavement settlement in the transverse direction, characterized by larger settlement in the middle and smaller settlement at the edges.
[0068] S1-1, Establish the transverse settling trough model.
[0069] Lateral surface settlement caused by tunnel excavation typically follows a normal distribution pattern, and a Gaussian settlement trough is used to represent the lateral settlement distribution:
[0070] (1)
[0071] in, Horizontal position Settlement value at the location (lateral settlement distribution). The horizontal position is defined with the tunnel centerline as the origin. The center position of the settling trough represents the distance of the maximum settlement point from the origin. For the width parameter of the settling trough, This represents the maximum settlement value, i.e., the maximum vertical displacement at the center of the settlement trough. Parameter Controlling the range of settlement influence means that the settlement value decreases from its maximum value to its minimum value. The lateral distance corresponding to (approximately 0.607 times). The larger the size, the wider the settlement trough, and the wider the range of impact from construction disturbance.
[0072] S1-2, Establish a time-cumulative evolution model.
[0073] During construction, road surface settlement typically goes through three stages: slow initial development, rapid intermediate growth, and eventual stabilization, reflecting the entire process from approaching, passing through, and moving away from the tunnel face. The Richards function is used to represent the settlement development over time at a specific point on the road surface:
[0074] (2)
[0075] in, Let be the time-cumulative evolution function. The construction time is calculated from the moment when tunnel excavation begins. This represents the time corresponding to the rapid development stage of settlement, signifying the moment when the settlement growth rate reaches its maximum value. This represents the timescale of settlement growth. ; For the growth rate parameter, The larger the value, the more intense the transition from slow to rapid settlement; The curve shape parameter is used to adjust the asymmetry of the settlement development curve. When, the function degenerates into a standard logistic curve; when or The curve exhibits different asymmetric shapes to adapt to the settlement time characteristics under different construction conditions (such as different tunneling speeds and different support timing). The Richards function can utilize curve shape parameters... Adjusting the asymmetry of the settlement development curve is suitable for describing the changes in the settlement growth rate at different construction stages.
[0076] S1-3, Establish the basic mechanism of the settlement field.
[0077] By combining the transverse settlement trough model and the time-cumulative evolution model, the foundation mechanism settlement field under homogeneous strata and stable construction conditions is obtained:
[0078] (3)
[0079] in, It is the basic mechanism settlement field, that is, the theoretical settlement value under homogeneous strata and stable construction conditions; The horizontal position For the construction period, The maximum settlement value, The center of the settling tank. Let be the width parameter of the settlement trough. Formula (3) gives the basic evolution framework of pavement settlement under homogeneous strata and stable construction conditions, that is, under ideal conditions (i.e., homogeneous strata and constant construction parameters), at any location At any moment The settlement value.
[0080] S1-4 introduces non-stationary engineering corrections.
[0081] In actual highway underpass projects, heterogeneous strata, tunnel face advancement, construction disturbance, and tunnel attitude adjustments can cause changes in the width of the settlement trough, the location of the settlement center, and the local settlement amplitude during construction. Therefore, it is necessary to introduce three non-stationary corrections into the foundation settlement field.
[0082] (1) Correction for widening of settling trough.
[0083] As the tunnel face advances, the area of construction disturbance gradually expands, and the width of the settlement trough also increases accordingly. Considering the gradual expansion of the impact range of construction disturbance over time, the width of the settlement trough is expressed as a function of time:
[0084] (4)
[0085] in, The settling trough width parameter varies over time. The initial width of the settlement trench, i.e., the width at the beginning of construction ( The width of the settling trough at (time); The settlement channel expansion coefficient represents the relative growth rate of the settlement channel width per unit time. It is determined by inversion of the half-width change of the settlement channel monitored at different construction stages and is used to describe the expansion process of the influence range of construction disturbance. This refers to the construction period.
[0086] (2) Correction for settling center drift.
[0087] During tunnel construction, the settlement center may shift due to adjustments in the tunnel construction posture, sudden changes in local strata, or bias pressure.
[0088] The location of the settlement center is represented as:
[0089] (5)
[0090] in, The location of the settlement center changes over time. The initial settlement center, i.e. The x-coordinate of the settlement center at any given time; The periodic center drift amplitude is used to describe the small reciprocating oscillation of the settlement center caused by repeated tunneling-support cycles or seasonal factors during construction. The periodic drift frequency controls the oscillation period; This refers to the local center offset amplitude, used to describe short-term, large-amplitude center offsets caused by local geological anomalies or construction accidents. Time of occurrence of local disturbance Construction time; The duration parameter of the local disturbance controls the range of the disturbance duration. Formula (5) is used to describe the slow shift of the settlement center and local abnormal disturbances during construction.
[0091] (3) Comprehensive engineering correction coefficient.
[0092] To reflect the combined effects of formation heterogeneity, construction stage variations, and random disturbances on settlement amplitude, a comprehensive engineering correction factor is introduced:
[0093] (6)
[0094] in, This is a comprehensive engineering correction factor; This is a correction term for spatial heterogeneity of the stratigraphy, along the lateral direction. The change is used to reflect the amplification or reduction effect of the settlement amplitude when the tunnel passes through different strata (such as alternating soft and hard strata, fault fracture zones, etc.). Its value is determined by combining geological survey data and monitoring inversion. These are revisions made during the construction phase, and will be made over time. The change is used to reflect the different impacts of different construction stages (such as excavation face advancement, initial support construction, secondary lining pouring, grouting reinforcement, etc.) on surface settlement. Its value is set according to the construction log and process parameters. It is a local random disturbance term that varies with space and time. It is used to capture local fluctuations that are difficult to predict completely (such as grouting pressure fluctuations, local over-excavation, etc.). It is usually regarded as a zero-mean random process when modeling. The value is obtained by the residual between statistical monitoring data and the prior model. When monitoring data is insufficient in the early stages of construction... Take 0 temporarily.
[0095] To ensure the physical rationality of the preliminary settlement field for the project, Should meet The constraints. In the actual parameter calibration process, if... Then take This is a preset positive lower limit value to avoid bulging phenomena that violate mechanical laws. Under normal construction conditions, The value is usually between 0.7 and 1.5.
[0096] Stratigraphic spatial heterogeneity correction term The specific form is as follows:
[0097] (7)
[0098] in, This is a left-right asymmetry correction factor, used to describe the left-right asymmetry effect of settlement troughs caused by tunnels passing through inclined strata or surface slopes; This is a correction factor for weak areas, used to describe the additional amplification effect of local weak strata (such as fault fracture zones and karst cave development areas) on the subsidence amplitude; The characteristic half-width of the settling trough is the settlement value that decreases from the center of the trough to the maximum settlement value. of (Approximately 60.7%) times; The center of the weak zone This is the parameter representing the range of influence of the weak zone.
[0099] and Determined based on geological survey data. Take the center position of the weak strata (such as fault fracture zones, karst cave development areas, and soil-rock interfaces) in the transverse projection of the road surface; Take the lateral influence range of the weak stratum Generally, it can be estimated based on the horizontal extension width of the weak layer in the geological profile. If there is no clearly defined weak layer, then... , and It is not included in the calculation.
[0100] S1-5, construct the final engineering preliminary settlement field.
[0101] Substituting the above non-stationary corrections into the fundamental mechanism settlement field, we obtain the final a priori engineering settlement field expression:
[0102] (8)
[0103] in, As a priori settlement field for engineering, this model unifies the transverse settlement trough, time cumulative evolution, settlement trough width expansion, settlement center drift, and formation heterogeneity correction into a single expression. Its role is not to replace on-site monitoring, but to provide a priori settlement field with engineering mechanism constraints for subsequent Bayesian residual assimilation. This is a comprehensive engineering correction factor; This represents the maximum settlement value. For time-cumulative evolution function; The location of the settlement center changes over time; The settling trough width parameter varies over time. The horizontal position This refers to the construction period.
[0104] Step S2: Bayesian residual assimilation and dynamic update.
[0105] By utilizing continuously acquired field monitoring data during construction, the prediction bias of the engineering prior model is probabilistically corrected and dynamically updated. While the engineering prior model has the advantage of mechanistic constraints, its parameters are based on historical experience or initial data calibration. As construction progresses, geological conditions are revealed, and disturbances accumulate, the prior model inevitably develops biases. Gaussian processes, as a nonparametric probabilistic model, can effectively utilize limited monitoring samples to provide optimal unbiased estimation of residuals. Moreover, with the addition of new data, its prediction distribution is naturally updated, eliminating the need to retrain the entire model.
[0106] The prior settlement field can describe the overall trend of pavement settlement, but due to the uncertainties in stratum parameters, construction disturbance, grouting effect, and monitoring errors, residuals still exist between the prior model and the field monitoring results. To make full use of the monitoring data continuously acquired during construction, this invention uses a Bayesian dynamic data assimilation method to probabilistically model and continuously update the residuals.
[0107] S2-1 defines the monitoring sample and residual.
[0108] During construction, several settlement monitoring points were set up laterally along the highway. As of [date missing], [data missing]. At each monitoring point, a total of [number] data points were obtained. The dataset consists of 10 monitored samples:
[0109] (9)
[0110] in, To monitor the dataset, including One sample; For the first The lateral position of each monitoring sample For the first The observation time for each monitoring sample For the first The measured settlement value of each monitoring sample For monitoring sample indexing, This represents the total number of currently monitored samples.
[0111] For each monitoring sample, the residual between the prior engineering settlement field and the measured settlement is defined as:
[0112] (10)
[0113] in, For the first The residuals of each monitoring sample For the first The measured settlement value of each monitoring sample For the preliminary settlement site of the project, at the monitoring point The predicted value at that location, For the first The lateral position of each monitoring sample For the first The observation time for each monitored sample. All residuals constitute a residual vector. , This is for the transpose operation.
[0114] S2-2, Establish the Gaussian process residual model.
[0115] The residuals are treated as random functions of space and time, and probabilistically modeled using Gaussian processes. Gaussian processes are powerful nonparametric Bayesian tools that do not require pre-defined functional forms of the residuals. Instead, they define the correlation of residuals at different locations and times through a covariance function, enabling the learning of the overall structure and local features of the residuals from a finite sample. The residuals are then treated as random functions of space and time. Consider it as a spatial coordinate and time The random function is used to establish a residual probability model using a Gaussian process:
[0116] (11)
[0117] in, For position and time The residual random function at the location, Input variables for space-time; It is the covariance function; To and Another different space-time input variable, It is a Gaussian process probability model.
[0118] Considering that the settlement residuals have spatial and temporal continuity, a covariance function is constructed using a radial basis function kernel and a white noise term:
[0119] (12)
[0120] in, Let covariance function be used. The standard deviation of the signal controls the overall magnitude of the residual function, reflecting the degree to which the residual deviates from zero. Input variables for space-time; To and A different spatial-temporal input variable; This is a transpose operation; The standard deviation of the observed noise reflects the random error inherent in the monitoring data itself; The space-time correlation length matrix is... The diagonal matrix represents the correlation length of the residuals in the spatial and temporal directions, respectively. The larger the correlation length, the stronger the correlation between the residuals of sample points that are far apart, and the smoother the prediction. For the Kronecker function, when The value is 1 if the condition is met, and 0 otherwise. This is used to add white noise terms only at the same input points. The correlation length reflects the spatial and temporal correlation range of the residuals; the closer the monitoring samples are, the stronger their effect on correcting the residuals at the prediction points.
[0121] S2-3, perform Bayesian inference.
[0122] Based on the obtained residual samples (i.e. monitoring data), infer the posterior distribution of the residuals at any new prediction point.
[0123] For any prediction point Given training samples (i.e., the obtained residual vectors) Under the condition that its posterior distribution is still Gaussian, the posterior mean and posterior variance of the residuals are expressed as follows:
[0124] (13)
[0125] (14)
[0126] in, The posterior mean of the residuals at the predicted points represents the best estimate of the residuals based on existing monitoring data. Input variables for the spatial-temporal dimensions of the point to be predicted. For prediction points The covariance vector between the training samples and the training samples To train the sample covariance matrix, To observe the standard deviation of noise, It is the identity matrix. It is the residual vector; The variance of the residual at the prediction point is the posterior variance. The larger the variance, the weaker the monitoring constraint or the more severe the local fluctuation at that point. This is the covariance between the predicted point and itself (i.e., the variance of the prior signal). This is for the transpose operation.
[0127] S2-4, calculate the predicted settlement value after assimilation.
[0128] The final assimilation prediction result is obtained by adding the engineering prior prediction with the Gaussian process residual correction.
[0129] The mean value of the pavement settlement prediction after Bayesian assimilation is expressed as follows:
[0130] (15)
[0131] in, The predicted settlement value after dynamic assimilation is provided by a prior model to provide the basic trend and overall shape of settlement, and by the posterior mean of Gaussian process residuals to provide local correction. The result of the superposition of the two not only preserves the engineering physical constraints, but also fully absorbs the feedback from field monitoring, thereby realizing the fusion prediction of mechanism and data. This is the preliminary settlement field (preliminary predicted value) for the project. For prediction points The posterior mean of the residuals.
[0132] Formula (15) shows that the model takes the prior settlement field of the engineering project as the basic trend and corrects the local deviation with the posterior mean of the Gaussian process residual, thus taking into account both the constraints of the engineering mechanism and the feedback of monitoring data. Near the monitoring point, the predicted value approaches the measured value; in the area far from the monitoring point, the predicted value gradually approaches the prior model, ensuring the rationality and continuity of the entire field.
[0133] S2-5, implement rolling updates.
[0134] During construction, monitoring data continuously increased over time. Let the... The newly added monitoring data for each construction phase are The updated monitoring dataset is then represented as:
[0135] (16)
[0136] in, For the first All monitoring datasets updated for each construction phase For the first All monitoring datasets updated after each construction phase (i.e., the previous phase), For the first A subset of monitoring data added during each construction phase. This is a sequence index for the construction phase.
[0137] After obtaining new monitoring data for each stage, repeat steps S2-1 to S2-4, i.e., recalculate the residual samples, update the Gaussian process hyperparameters, and obtain new settlement prediction mean and posterior variance. For any construction time... The predictions utilize only monitoring data obtained up to a given point in time, avoiding the involvement of future information. As the number of monitoring samples increases, the posterior variance gradually decreases, and the prediction results tend to converge and stabilize, reflecting the objective law that uncertainty decreases with the accumulation of information. The rolling update mechanism enables the model to continuously correct prediction errors as construction progresses, improving the ability to predict settlement in subsequent stages.
[0138] In practical engineering applications, when the amount of monitoring data is large, an update step size should be set, that is, every time a large amount of data is obtained... A complete hyperparameter re-optimization is performed after each new monitoring data point (or after each construction cycle), rather than triggering optimization for every data point. In this embodiment, [the following is taken]. On a conventional workstation, the calculation time for each re-optimization is no more than 3 seconds, which can meet the real-time prediction needs during the construction period.
[0139] Step S3: Uncertainty quantification and over-limit probability calculation.
[0140] In practical engineering, the predicted value itself does not reflect its reliability: the same predicted settlement value has completely different risk implications if accompanied by significant uncertainty compared to if accompanied by minimal uncertainty. This invention distinguishes between statistical prediction uncertainty and engineering-corrected uncertainty, and constructs an engineering-corrected uncertainty model to comprehensively reflect monitoring sparsity, settlement gradient concentration, and boundary extrapolation risk, ultimately achieving a key shift from deterministic prediction to probabilistic risk assessment.
[0141] The Bayesian dynamic data assimilation model can not only provide the mean of settlement prediction, but also output the prediction uncertainty. This invention classifies the uncertainty into two categories: one is the posterior standard deviation of the Gaussian process. One type primarily reflects the statistical prediction uncertainty of the residual model, used for confidence interval analysis of monitoring point predictions; the other type is engineering correction uncertainty. Taking into account the construction stage, settlement amplitude, settlement gradient and boundary extrapolation effect, it is mainly used for spatial over-limit probability and risk distribution analysis.
[0142] S3-1, Calculate the predicted confidence interval for the monitoring point.
[0143] For the predicted results of the monitoring points, confidence intervals are constructed using the posterior standard deviation of the Gaussian process. At a 95% confidence level, the prediction interval is expressed as:
[0144] (17)
[0145] in, The boundary values of the settlement prediction confidence interval at a 95% confidence level are given. The mean of the settlement predictions after Bayesian assimilation. Let be the posterior standard deviation of the Gaussian process.
[0146] Formula (17) is used to evaluate the consistency between the model prediction results and the measured settlement. The actual settlement value has a 95% probability of falling within this interval. If most of the measured settlement falls within the prediction interval, it indicates that the uncertainty estimation of the model has good reliability; conversely, if the measured value frequently exceeds the interval, it suggests that the hyperparameters of the Gaussian process may need to be retrained or that there is a systematic bias in the prior model.
[0147] S3-2, Calculate engineering correction uncertainty.
[0148] A measure of uncertainty applicable to full-field risk analysis is constructed. This measure, based on the posterior standard deviation of the Gaussian process, further amplifies the risk in sparse, high-gradient, and boundary regions to avoid underestimating the probability of exceeding limits. For risk analysis across the entire field, especially in areas far from monitoring points or with large settlement gradients, relying solely on the posterior standard deviation of the Gaussian process may underestimate the actual risk.
[0149] Engineering correction uncertainty The following factors were considered when determining the Gaussian process posterior standard deviation, spatial distance between monitoring points, settlement gradient, and boundary extrapolation effect:
[0150] (18)
[0151] in, To correct for uncertainties in the project, The posterior standard deviation of the Gaussian process; For prediction points The normalized distance to the nearest monitoring point is defined as follows: ,in The coordinates of the monitoring sample closest to the prediction point. The spatial correlation length in the covariance function of the Gaussian process. Let be the time correlation length in the covariance function of the Gaussian process. ; A larger value indicates a greater lack of monitoring constraints on the prediction points and a higher degree of extrapolation uncertainty. Normalized settlement gradient; For the settlement gradient, The settlement gradient control value is determined according to highway engineering safety standards; This is a boundary extrapolation correction term. Its value is 0 inside the computational domain and gradually increases to 1 near the boundary (such as when the distance from the boundary is less than 1 times the correlation length). It is used to reflect the extrapolation uncertainty of the model at the edge where there is no monitoring data. This is the distance effect amplification factor, which controls the degree of amplification of uncertainty in sparse monitoring areas; It is the gradient effect amplification factor, which controls the degree of amplification of uncertainty in areas with large settlement gradients (i.e., areas of concentrated deformation); This is the boundary effect amplification factor, which controls the degree of amplification of uncertainty near the calculation boundary. Based on the uncertainty given by the statistical model, engineering factors (monitoring distance, gradient concentration, boundary) are used for amplification, thus preserving the objectivity of the statistical model while incorporating the conservatism of engineering judgment. The above amplification factor is determined through engineering experience or cross-validation; in this embodiment, it is taken as... , , .
[0152] For areas near monitoring points, the prediction uncertainty is mainly controlled by the posterior standard deviation of the Gaussian process; for areas far from monitoring points, with large settlement gradients, or near the calculation boundary, the engineering correction uncertainty is appropriately increased by formula (18). This amplification is a conservative approach that is on the safe side, and its purpose is to avoid underestimating the probability of exceeding the limit due to underestimating the uncertainty, thereby ensuring that the risk assessment results have sufficient safety redundancy.
[0153] S3-3, calculate the probability of settlement exceeding the limit.
[0154] After considering engineering correction uncertainties, the settlement at any location can be expressed as:
[0155] (19)
[0156] in, To account for uncertainties in engineering corrections, at any position The actual settlement random variable at the location; The mean of the sedimentation prediction after Bayesian assimilation; It follows a normal distribution; To correct for uncertainties in engineering.
[0157] Calculating the probability that settlement at each location exceeds the allowable control value is the core step in this invention's shift from deterministic prediction to probabilistic risk assessment.
[0158] Let the allowable settlement control value of the highway pavement be... The probability of settlement exceeding the limit is expressed as:
[0159] (20)
[0160] in, The probability of settlement exceeding the limit is between 0 and 1; The cumulative distribution function of the standard normal distribution; To account for uncertainties in engineering corrections, at any position The actual settlement random variable at the location; This is the control value for settlement amplitude. This represents the average sedimentation prediction after Bayesian assimilation.
[0161] As shown in formula (20), the probability of settlement exceeding the limit is controlled by both the predicted mean settlement and the engineering correction uncertainty. When the predicted settlement is close to the control value, the probability of exceeding the limit will increase significantly; when the uncertainty is large, even if the predicted mean has not yet reached the control value, a high risk may still occur. Therefore, the probability of exceeding the limit is a better indicator of the construction risk status than a single predicted settlement value.
[0162] Step S4, Road surface comprehensive risk assessment model.
[0163] A multi-indicator comprehensive risk index is constructed to achieve a comprehensive evaluation of road surface safety status. The safety of highway pavements depends not only on whether the settlement amplitude exceeds the limit, but also on the combined effects of uneven settlement (gradient), local bending effect (curvature), settlement development rate (rate of change), and prediction uncertainty. Judging solely by a single settlement threshold may overlook the potential risks posed by differential settlement, local bending, and rapid development stages. This invention calculates three derived deformation indicators—settlement gradient, curvature, and rate of change—and incorporates them, along with settlement amplitude and the probability of exceeding limits, into a weighted comprehensive index, achieving a crucial transformation from settlement prediction to risk identification. Finally, based on the comprehensive index value, three risk levels—low, medium, and high—are classified, providing intuitive and actionable early warning information for construction sites.
[0164] S4-1, Calculate the characteristic indices of settlement deformation.
[0165] Derivative indices reflecting pavement deformation morphology and dynamic changes were extracted from the assimilated settlement field, and settlement gradient, settlement curvature, and settlement rate of change were calculated respectively:
[0166] (21a)
[0167] (21b)
[0168] (21c)
[0169] in, The settlement gradient reflects the degree of uneven settlement of the pavement. The larger the gradient, the more significant the differential deformation, and the greater the additional shear force borne by the pavement structure. Settlement curvature reflects the local bending concentration effect of the pavement. The greater the curvature, the more severe the bending of the pavement structure, which may lead to pavement cracking. The settlement change rate reflects how fast the settlement develops. The larger the change rate, the more severe the construction disturbance, and the shorter the response time for engineers. The mean of the settlement predictions after Bayesian assimilation. For the construction period, This refers to the horizontal position.
[0170] S4-2, Construct a comprehensive risk index.
[0171] To eliminate the influence of different indicator dimensions, control values were used for normalization to establish a comprehensive risk index:
[0172] (twenty two)
[0173] in, This is a comprehensive risk index; The settlement amplitude control value is determined by design requirements; This is the settlement gradient control value, determined based on the requirements for highway pavement smoothness. This is the settlement curvature control value, determined based on the bending resistance of the pavement structure; This is the control value for the rate of settlement change, determined based on the allowable rate of construction disturbance. These are the weights of each indicator, and they satisfy... ,and ( ), For the first The weighting coefficients of each risk assessment indicator, . The mean of the settlement predictions after Bayesian assimilation. For the settlement gradient, For settlement curvature, The rate of change of settlement, This represents the probability of settlement exceeding the limit.
[0174] Formula (22) embodies the coupling relationship between settlement prediction, uncertainty quantification, and pavement deformation response. Settlement amplitude controls the risk benchmark level, settlement gradient and curvature reflect the degree of spatial deformation concentration, settlement change rate reflects the speed of risk development, and the probability of settlement exceeding limits introduces the impact of prediction uncertainty on risk assessment. Through the unified normalization and weighted coupling of the above indicators, the transformation from settlement prediction to dynamic pavement risk assessment can be achieved.
[0175] S4-3, Determine the risk level.
[0176] According to the comprehensive risk index The construction risk status is divided into three levels, providing on-site construction management personnel with intuitive early warning signals and corresponding handling suggestions.
[0177] when When the risk level is low, it indicates that the road surface deformation is within a controllable range and normal construction and monitoring frequency can be maintained.
[0178] when If the risk level is determined to be medium, it indicates that the settlement or deformation index in a local area is close to the control standard. The monitoring frequency should be increased and construction parameters should be adjusted appropriately, such as reducing the tunneling speed and optimizing the grouting pressure.
[0179] when If the situation is deemed high-risk, it indicates that the construction disturbance may have a significant impact on the operational safety of the highway. Measures such as reducing tunneling disturbance, optimizing grouting parameters, strengthening support, or speed limit monitoring should be taken in a timely manner.
[0180] The values of 0.4 and 0.7 correspond to the weighted combination of the normalized values of each individual indicator reaching the level of "approaching the control standard" and "reaching or exceeding the control standard", respectively, which are corresponding to the "attention value" and "warning value" of the engineering early warning.
[0181] The thresholds of 0.4 and 0.7 are determined as follows: (1) When all individual indicators are below 40% of the control value, the weighted combination of settlement amplitude, gradient, curvature, rate of change and probability of exceeding the limit usually does not exceed 0.4. At this time, the engineering risk is acceptable and is defined as a low-risk threshold; (2) When any key indicator is close to 70% of its control value, considering the coupling effect of each indicator, the weighted comprehensive index will reach above 0.7. At this time, the project is close to the warning state and is defined as a high-risk threshold. The above values are based on the management provisions on construction monitoring and early warning classification in the "Technical Specification for Highway Tunnel Construction" (JTG / T 3660—2020) and are determined in combination with engineering practice experience. In actual engineering applications, the thresholds can also be appropriately adjusted according to the safety management requirements of specific projects.
[0182] Step S5: Model parameter determination and model calibration method.
[0183] To avoid arbitrariness in the values of model parameters, this step determines various model parameters according to the approach of engineering constraints, monitoring inversion, probability training, and risk verification. The model parameters include engineering prior settlement field parameters, Gaussian process hyperparameters, engineering corrected uncertainty parameters, and risk assessment parameters. The engineering corrected uncertainty parameters are determined by formula (18).
[0184] S5-1, Calibration of parameters for the prior model of the project.
[0185] By using monitoring data from the early stages of construction, the physical parameters in the prior model of the project are determined through inversion, so that the prior model can be as close as possible to the actual site conditions while maintaining the mechanistic constraints.
[0186] Preliminary settlement field parameters for the project include the maximum settlement value. Initial settling trough width Settling trough expansion coefficient Richards function parameters , , The parameters include settlement center drift and local disturbance parameters. First, the initial range of parameters is determined based on tunnel depth, cross-sectional dimensions, geological conditions, and construction methods; then, inversion calibration is performed using monitoring data from the initial construction phase.
[0187] The calibrated objective function is to minimize the root mean square error between the monitored settlement and the model-predicted settlement.
[0188] (twenty three)
[0189] in, This is the optimal parameter vector after calibration. The parameters of the prior engineering model to be inverted are... This represents the value of the independent variable that minimizes the objective function within the parentheses. For the first The measured settlement value of each monitoring sample For a given parameter vector Below, the engineering prior model is at the monitoring point. The predicted value at that location, For the first The lateral position of each monitoring sample For the first The observation time for each monitoring sample For monitoring sample indexing, This represents the total number of currently monitored samples.
[0190] Settlement center drift related parameters (periodic drift amplitude) ,frequency Local offset amplitude Time of occurrence Duration The calibration method for () is as follows: The time interval is determined based on the tunnel construction cycle, such as the time interval between each ring excavation. ,but ; and The determination is based on the time points and duration of impact of abnormal events (such as crossing faults, machine shutdown for maintenance, grouting accidents, etc.) recorded in the construction log; if no significant abnormal events occur during construction, then... Take 0, and It is not included in the calculation.
[0191] This inversion process allows the prior settlement field of the project to be matched with the morphology and settlement time history of the settlement trough monitored on site while maintaining mechanistic constraints.
[0192] S5-2, Gaussian process hyperparameter training.
[0193] The optimal hyperparameters of the Gaussian process are automatically learned from the residual data, enabling the residual model to have the best predictive performance.
[0194] Hyperparameters of a Gaussian process include the signal standard deviation. Standard deviation of observation noise The space-time related length. The hyperparameters of a Gaussian process are automatically determined by maximizing the marginal log-likelihood:
[0195] (twenty four)
[0196] in, This is the calibrated optimal Gaussian process hyperparameter vector. For the hyperparameters of the Gaussian process, This indicates the objective function within the parentheses that achieves its maximum value. The value of ; To train the sample covariance matrix, To observe the standard deviation of noise, It is the identity matrix. For the residual vector, This represents the total number of existing monitoring samples. This is for the transpose operation.
[0197] As construction monitoring data is continuously updated, the hyperparameters of the Gaussian process are continuously optimized along with the assimilation process, thereby gradually adapting to the changes in the residual field at different construction stages.
[0198] S5-3, Determination of risk assessment parameters.
[0199] Risk assessment parameters include settlement control values, gradient control values, curvature control values, settlement change rate control values, and the weights of each indicator. Each control value is determined based on design control standards and on-site monitoring and early warning standards.
[0200] Determine the weights of each indicator in the comprehensive risk index to avoid biases caused by subjective weighting. Comprehensive Risk Index Weights in This invention employs the entropy weight method to objectively determine the risk level, thereby reducing subjectivity. The basic idea of the entropy weight method is that the greater the dispersion of an indicator, the stronger its ability to distinguish risk states, and therefore it should be assigned a greater weight. The specific steps are as follows:
[0201] (1) Each risk assessment indicator ( , , , , The values taken at different spatial locations and construction times constitute an evaluation matrix, which is then normalized to obtain standardized values. ( As the indicator number, (Sample number).
[0202] (2) Calculate the first Information entropy of each indicator : ,in In the entropy weight method calculation, this refers to the total number of samples participating in the statistical evaluation; when At that time, it was agreed This is to ensure the validity of entropy calculation.
[0203] (3) Calculate the first Coefficient of difference of each indicator The larger the difference coefficient, the greater the variation of the indicator among different samples, and the stronger its ability to distinguish risks.
[0204] (4) Calculate the weights ,in For the first The coefficient of difference among risk assessment indicators This is the growth rate parameter.
[0205] In this embodiment, the weights calculated using the entropy weight method are as follows: (Settlement amplitude) (gradient), (curvature) (Rate of change) (Probability of exceeding limits). Among them, the probability of exceeding limits has the highest weight (0.415), indicating that it has the strongest ability to distinguish risk states under different working conditions.
[0206] S5-4, Model robustness test.
[0207] This invention verifies the sensitivity of the risk assessment results to the weight values, and assesses the reliability and stability of the results through perturbation analysis. To test the sensitivity of the comprehensive risk assessment results to the weight values, this invention specifies performing perturbation sensitivity analysis on the determined weights. Specifically, this involves: perturbing each weight... Perturbations of ±10% and ±20% are applied respectively, while other weights are normalized proportionally to make the sum equal to 1. The overall risk index is then recalculated, and the changes in the following three key evaluation indicators are compared: area of high-risk zone ( The spatial distribution of risk levels, including the regional proportion, the maximum risk index value, and the risk level.
[0208] If the location and risk level of the main risk areas do not change significantly under the above disturbances, the risk assessment results are considered to have good robustness, and the method can be applied to practical engineering decisions. If the method is sensitive to weight disturbances, it suggests that the rationality of the control values should be re-examined or more monitoring data should be added to reduce uncertainty.
[0209] Experimental verification:
[0210] To verify the applicability of this invention in tunneling projects under highways, the Zhaibeishan Tunnel project of the Guangdong section of the Wuhan-Meizhou Railway was used as an example. During the tunneling process under the existing highway, the excavation disturbance may cause deformation of the overlying strata and transmit it to the road surface, thereby affecting the safety of the highway pavement. Using pavement settlement monitoring data collected on-site, dynamic prediction, uncertainty quantification, and risk assessment of pavement settlement during the construction period were conducted.
[0211] Engineering generalization model such as Figure 1 As shown. Based on the cross-sectional characteristics of the underpass section, the tunnel excavation profile is equivalent to a two-dimensional cross-sectional model. The equivalent tunnel span is taken as... m, tunnel height taken m, soil cover thickness is taken m. Road surface settlement monitoring points are deployed laterally along the highway. Monitoring data includes the lateral location of the monitoring points, monitoring time, and measured settlement value, i.e. .in, For the first The lateral position of each monitoring sample; For the first The observation time for each monitoring sample, i.e., the construction monitoring time. For the first The measured settlement value of each monitoring sample, that is, the road surface settlement value collected on site.
[0212] The parameters of the prior settlement field of the project are inverted using field monitoring data, and the residuals between the prior model and the measured settlement are dynamically assimilated using a Gaussian process, thereby realizing the rolling prediction of pavement settlement during the construction period.
[0213] To avoid directly assigning values to model parameters manually, the parameter value range was first determined based on tunnel depth, cross-sectional dimensions, settlement control standards, and on-site monitoring data. Then, with the goal of minimizing the root mean square error between monitored settlement and model-predicted settlement, the parameters of the engineering prior model were calibrated. The main parameter values are shown in Table 1.
[0214] Table 1. Values of key parameters in the engineering prior model.
[0215]
[0216] Road surface settlement amplitude control value is taken as mm, settlement gradient control value taken mm / m, settlement curvature control value is taken as The settlement change rate control value is taken as follows: mm / d. In engineering uncertainty correction, the amplification factors for distance effect, gradient effect, and boundary effect are respectively calibrated as... , , .
[0217] The results of parameter sensitivity and risk weight analysis are as follows: Figure 2 As shown. By Figure 2 (a) It can be seen that, The fact that the model prediction error is most affected indicates that identifying the rapid development stage of settlement is the key to improving prediction accuracy. , , and Secondly, this indicates that the settlement time history parameters and settlement trough morphology parameters jointly control the pavement settlement prediction results. Figure 2 (b) It can be seen that the risk index weights determined by the entropy weight method are as follows: settlement amplitude 0.159, settlement gradient 0.136, settlement curvature 0.120, settlement change rate 0.170, and exceedance probability 0.415. Among them, the exceedance probability has the highest weight, indicating that it has the strongest ability to distinguish risk states.
[0218] Using the calibrated engineering prior model and field monitoring data, the Bayesian residual assimilation method is employed to dynamically predict the road surface settlement field. Figure 3 The results of comparing the monitored settlement field with the Bayesian assimilation prediction of the mean settlement are presented.
[0219] Depend on Figure 3 (a) It is evident that road surface settlement was relatively small in the early stages of construction, gradually increasing as tunnel construction progressed, and forming a significant settlement concentration zone near the tunnel centerline in the later stages of construction. Figure 3 (b) It is evident that the Bayesian assimilation prediction results can reproduce the spatiotemporal evolution of settlement in the field relatively well, and the predicted location, range, and development trend of high settlement areas are basically consistent with the reference settlement field. This result indicates that the engineering prior model can describe the basic morphology of the settlement trough, while Bayesian residual assimilation can further utilize field monitoring data to correct local deviations and improve the adaptability of settlement field prediction.
[0220] To analyze the impact of continuous monitoring data updates on prediction accuracy, different construction times were selected as assimilation cutoff times, and the model was updated on a rolling basis. Figure 4 The data shows that as monitoring data gradually increases, the overall prediction error decreases, indicating that dynamic assimilation can continuously correct biases in the prior engineering model. The prediction error in the future stage increases in stages during the middle of construction, mainly because settlement enters a rapid development phase during this period, with increased settlement amplitude and gradient changes, making prediction more difficult. However, as subsequent monitoring data is further supplemented, the prediction error in the future stage decreases significantly.
[0221] Based on the average settlement prediction, the engineering correction uncertainty is further calculated. Figure 5 The standard deviation distribution of engineering correction uncertainty is given.
[0222] Depend on Figure 5It is evident that uncertainty is low in the initial stages of construction, gradually increasing as construction progresses and settlement develops; uncertainty is relatively high near the tunnel centerline and in areas with significant settlement impact. This indicates that prediction uncertainty is not only related to statistical prediction uncertainty, but also to the degree of settlement development, the strength of monitoring constraints, and the location of risk areas. These results can provide a basis for intensified monitoring and identification of key risk areas during construction.
[0223] Based on the predicted settlement mean and engineering correction uncertainty, the probability of settlement exceeding the limit and the comprehensive risk index were further calculated, and the results are as follows: Figure 6 As shown. By Figure 6 (a) It can be seen that the probability of settlement exceeding the limit is mainly concentrated in the later stage of construction and near the centerline of the tunnel, indicating that the settlement in this area is likely to be close to or exceed the control value. Figure 6 (b) shows that the high-value area of the comprehensive risk index is slightly wider than the over-limit probability area, indicating that pavement risk is not only controlled by whether settlement exceeds the limit, but also by the settlement amplitude, gradient, curvature and rate of change. This result realizes the transformation from single settlement prediction to probabilistic risk assessment.
[0224] To verify the effectiveness of the method of this invention, the Bayesian assimilation model was compared with the engineering prior model, the random forest model, and the multilayer perceptron model. The prediction errors of different models are shown in Table 2, and the final settlement trough prediction results during the construction stage are as follows: Figure 8 As shown in Table 2, “Future Stage RMSE” is defined as: (The rest of the text appears to be incomplete and requires further context for accurate translation.) All monitoring data prior to the first construction phase were used as training samples for the second phase. The road surface settlement within 5 days after each construction stage is predicted, and the root mean square error between the predicted and measured values is calculated. This index is used to evaluate the model's extrapolation prediction capability, that is, its ability to predict future settlement that has not yet occurred using existing historical data. In Table 2, "future stage" refers to the 5-day time window after each prediction time.
[0225] Table 2 Comparison of prediction accuracy of different models
[0226]
[0227] As shown in Table 2, the overall RMSE of the Bayesian assimilation model is 0.282 mm, which is lower than that of the engineering prior model, the random forest model, and the multilayer perceptron model, indicating that assimilation of field monitoring data can significantly improve the accuracy of settlement prediction. The RMSE in the future stage is 0.915 mm, which also remains at a low level, indicating that the model has good subsequent prediction capabilities.
[0228] Figure 8Further analysis shows that the engineering prior model can reflect the basic morphology of the settlement trough, but there is some deviation near the peak value; the random forest model curve is step-shaped, lacking spatial continuity; the multilayer perceptron model is weak in characterizing the settlement peak value and the settlement trough width; the Bayesian assimilation model is relatively close to the reference value in terms of settlement peak value, settlement trough width, and attenuation trend on both sides. The figure shows the comparison results of the spatial distribution of the settlement trough between different prediction models and the reference value. Since field monitoring data is usually only deployed at limited cross-sections and characteristic locations, it is difficult to directly obtain a continuous full-field settlement distribution. Therefore, this verification establishes a settlement analysis model based on engineering geological conditions, construction process parameters, and field monitoring data, and uses the simulation results verified by monitoring data as a reference value to evaluate the ability of different prediction models to characterize the spatial distribution characteristics of the settlement trough. Figure 8 It can be seen that the engineering prior model can reflect the overall trend of settlement trough changes well, but there is a certain deviation in the peak region; the random forest model, due to the use of discrete decision rules, shows a step-like change in the prediction curve, and its spatial continuity is relatively insufficient; the multilayer perceptron model can describe the overall settlement trend, but its ability to identify settlement peaks and settlement trough widths is limited. In contrast, the Bayesian assimilation model, after fully integrating field monitoring information, has a better ability to describe the location of settlement peaks, settlement trough widths, and attenuation trends on both sides, and maintains a high consistency with the reference values. This indicates that dynamic data assimilation can effectively correct the uncertainty of the engineering prior model and improve the reliability of pavement settlement spatial prediction.
[0229] In summary, the case study of the Zhaibeishan Tunnel under the expressway demonstrates that the Bayesian dynamic data assimilation model can effectively utilize on-site collected pavement settlement monitoring data to dynamically correct and predict pavement settlement during the construction period. The engineering correction uncertainty and exceedance probability reflect the prediction reliability and spatial distribution of risk; the entropy-weighted comprehensive risk index can identify key risk areas in the later stages of construction. This method can provide quantitative evidence for pavement settlement prediction, risk warning, and construction control during the construction of tunnels under expressways.
[0230] This invention addresses the problem of pavement settlement prediction and risk assessment during the construction of tunnels passing under highways. It constructs a Bayesian dynamic data assimilation model for pavement settlement prediction and risk assessment, and verifies it using pavement settlement monitoring data collected on-site, taking the Zhaibeishan Tunnel project of the Wumei Railway in Guangdong Province as an example. The main findings of this verification case are as follows:
[0231] (1) The constructed prior settlement field can comprehensively describe the spatial distribution, temporal evolution, and non-stationary construction effects of pavement settlement. The model unifies the Gaussian settlement trough, Richards time development function, settlement trough width expansion, settlement center drift, and formation heterogeneity correction into the same framework, which can provide prior prediction results with engineering mechanism constraints for subsequent Bayesian residual assimilation.
[0232] (2) Bayesian residual assimilation can effectively utilize field monitoring data to correct biases in the prior engineering model. The case of the Zhaibeishan Tunnel passing under the expressway shows that as construction monitoring data is continuously updated, the overall prediction error decreases, and the Bayesian assimilation model can better reproduce the spatiotemporal evolution of pavement settlement from slow initial development, rapid growth in the middle stage, to a stable state in the later stage. Compared with the prior engineering model, random forest model, and multilayer perceptron model, the Bayesian assimilation model is closer to the field monitoring response in terms of settlement trough peak value, settlement trough width, and attenuation trend on both sides.
[0233] (3) Parameter calibration and sensitivity analysis show that the time parameters during the rapid settlement development stage are... The shape parameters of the Richards function have the greatest impact on model prediction error. Settlement growth timescale Maximum settlement value and the initial settling trough width It also exhibits high sensitivity, indicating that the rapid settlement development stage during construction and the morphology of the settlement trough are key factors in controlling prediction accuracy. Determining parameters through engineering constraints, monitoring inversion, probability training, and sensitivity analysis can reduce the subjectivity of parameter selection.
[0234] (4) Engineering correction uncertainty can reflect the relationship between settlement development, monitoring constraints, and risk areas. Case results show that the prediction uncertainty is low in the early stage of construction, but gradually increases with the progress of construction and settlement development, and shows a high level near the tunnel centerline and in areas with significant settlement impact. This result indicates that the prediction uncertainty is not only related to the posterior standard deviation of the Gaussian process, but is also affected by monitoring sample constraints, settlement amplitude, settlement gradient, boundary extrapolation, and the degree of proximity to control values, which can provide a basis for intensified monitoring and identification of key risk areas during construction.
[0235] (5) The evaluation results based on the combined risk index of over-limit probability and entropy weight can realize the transformation from settlement prediction to risk identification. In the case of Zhaibeishan Tunnel, the over-limit probability of settlement is mainly concentrated in the later stage of construction and near the centerline of the tunnel; the high value area of the combined risk index is wider than the single over-limit probability area, indicating that the pavement risk is not only controlled by the settlement amplitude, but also by the combined influence of settlement gradient, curvature, rate of change and prediction uncertainty. This method can provide a quantitative reference for settlement early warning, risk zoning and construction parameter adjustment during the construction of tunnels passing under highways.
[0236] Overall, the method of this invention takes into account engineering prior constraints, on-site monitoring feedback, dynamic updates, and uncertainty quantification, and can effectively serve the prediction and risk assessment of pavement settlement in tunnels passing under highways. Future research could further incorporate more on-site construction parameters, geological exploration information, and multi-source monitoring data to improve the model's applicability under complex geological conditions and multi-condition construction scenarios.
[0237] The above embodiments are merely preferred examples of the present invention and are not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Therefore, any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention should fall within the protection scope of the present invention.
Claims
1. A method for predicting pavement settlement during tunnel construction based on Bayesian dynamic data assimilation, characterized in that, Includes the following steps: Step S1: Obtain pavement settlement monitoring data for the tunnel under the highway project, and construct an engineering prior settlement field model that integrates the morphology of the transverse settlement trough, the cumulative evolution over time, and non-stationary engineering corrections. The construction of the prior settlement field model for the project specifically includes: A Gaussian-type settlement trough is used to describe the spatial distribution of lateral settlement of the road surface, and the shape of the settlement trough is controlled by the maximum settlement value and the settlement trough width parameter. The Richards function is used to describe the development of road settlement over time, and the asymmetry of the settlement-time curve is adjusted by the growth rate parameter and the curve shape parameter. By multiplying the transverse settlement trough model with a time evolution function, a foundation mechanism settlement field under homogeneous strata and stable construction conditions is established; Non-stationary engineering corrections are introduced, including linear expansion of settlement trough width over time, periodic drift and local disturbance of settlement center position over time, and engineering correction coefficients that comprehensively reflect formation heterogeneity, changes during construction, and local random disturbances. By incorporating non-stationary engineering corrections into the basic mechanism settlement field, an a priori settlement field for the engineering is constructed. Step S2: Calculate the residual between the prior settlement field of the project and the field monitoring data. Use a Gaussian process to probabilistically model the residual. As construction monitoring data is continuously acquired, update the Gaussian process model to obtain the mean and posterior variance of the settlement prediction after Bayesian assimilation. The step S2, which uses a Gaussian process to probabilistically model the residuals, specifically includes: Calculate the residual between the measured settlement values at each monitoring point and the prior predicted values of the project, and form a residual vector; The residuals are modeled as a Gaussian process with zero mean, and the combination of the squared exponential kernel function and the white noise term is used as the covariance function. For any prediction point, calculate the posterior mean and posterior variance of the residuals; The a priori predicted value of the project is added to the posterior mean of the residual to obtain the mean settlement prediction after Bayesian assimilation. Step S3: Based on the posterior variance of the Gaussian process and the engineering correction uncertainty, calculate the confidence interval for pavement settlement prediction and the probability of settlement exceeding the limit; Step S4: Based on the assimilated settlement prediction field, calculate the settlement gradient, settlement curvature and settlement change rate, and combine the settlement amplitude and the settlement exceedance probability to construct a comprehensive risk index for dynamic classification and evaluation of pavement construction risks. Step S5 involves using a combination of engineering constraints, monitoring inversion, probabilistic training, and sensitivity analysis to determine and calibrate the model parameters involved in steps S1 to S4.
2. The method for predicting pavement settlement during tunnel construction based on Bayesian dynamic data assimilation according to claim 1, characterized in that, In the correction of non-stationary engineering: The linear expansion of the settling trough width over time is controlled by the initial settling trough width and the settling trough expansion coefficient; The change of the settlement center position over time is described by parameters including the initial settlement center position, periodic drift amplitude and frequency, local disturbance amplitude, local disturbance occurrence time, and local disturbance duration. The engineering correction factor is composed of the superposition of the formation spatial heterogeneity correction term, the construction stage correction term, and the local random disturbance term, plus 1.
3. The method for predicting pavement settlement during tunnel construction based on Bayesian dynamic data assimilation according to claim 2, characterized in that, The specific implementation of the rolling update in step S2 is as follows: whenever new monitoring data is obtained, the new data is added to the existing monitoring dataset, the residual vector is recalculated, the hyperparameters of the Gaussian process are re-optimized by maximizing the marginal log-likelihood function, and the posterior mean and posterior variance of the residuals of the prediction points are recalculated using the updated Gaussian process model to perform rolling updates of the settlement prediction.
4. The method for predicting pavement settlement during tunnel construction based on Bayesian dynamic data assimilation according to claim 3, characterized in that, Step S3 specifically includes: The confidence intervals for settlement prediction at monitoring points were constructed using the posterior standard deviation of a Gaussian process. The engineering correction uncertainty is determined by comprehensively considering the posterior standard deviation of the Gaussian process, the normalized distance between the predicted point and the nearest monitoring point, the normalized settlement gradient, and the boundary extrapolation correction term. Assuming that the actual settlement follows a normal distribution, the probability of settlement exceeding the limit at each location is calculated by combining the allowable settlement control value of the road surface.
5. The method for predicting pavement settlement during tunnel construction based on Bayesian dynamic data assimilation according to claim 4, characterized in that, The specific method for calculating the settlement deformation characteristic index in step S4 is as follows: Based on the assimilated settlement prediction field, the absolute value of the first-order partial derivative of the settlement along the transverse direction is calculated as the settlement gradient, which is used to reflect the degree of uneven settlement of the road surface. The absolute value of the second-order partial derivative of the settlement along the transverse direction is calculated as the settlement curvature, which is used to reflect the local bending concentration effect of the pavement. The absolute value of the first-order partial derivative of settlement with time is calculated as the settlement rate, which is used to reflect the speed of settlement development.
6. The method for predicting pavement settlement during tunnel construction based on Bayesian dynamic data assimilation according to claim 5, characterized in that, The specific method for constructing the comprehensive risk index and classifying risk levels in step S4 is as follows: five indicators are used, namely settlement amplitude, settlement gradient, settlement curvature, settlement change rate and settlement exceedance probability. Each indicator is divided by its corresponding control value for normalization, and the comprehensive risk index is constructed by weighted summation. Based on the comprehensive risk index, construction risks are divided into three levels: low risk, medium risk, and high risk.
7. The method for predicting pavement settlement during tunnel construction based on Bayesian dynamic data assimilation according to claim 6, characterized in that, The comprehensive risk index is: in, As a comprehensive risk index, This is the control value for settlement amplitude. This is the settlement gradient control value. This is the settlement curvature control value. This is the control value for the rate of change of settlement. These are the weights of each risk assessment indicator. The mean of the settlement predictions after Bayesian assimilation. For the settlement gradient, For settlement curvature, The rate of change of settlement, This represents the probability of settlement exceeding the limit.
8. The method for predicting pavement settlement during tunnel construction based on Bayesian dynamic data assimilation according to claim 7, characterized in that, The determination and calibration of model parameters in step S5 specifically includes: using monitoring data from the early stage of construction, with the goal of minimizing the root mean square error between monitored settlement and model-predicted settlement, calibrating the physical parameters in the prior engineering model; and determining the hyperparameters of the Gaussian process from the residual data by maximizing the marginal log-likelihood function.
Citation Information
Patent Citations
Intelligent inversion system for three-dimensional deformation field of land subsidence in ultra-large water level drawdown area
CN122149400A
Uncertainty quantification method for sea-land intersection mutual soil layer liquefaction risk evaluation
CN122197722A
