Engineering settlement deformation monitoring and analyzing system and method based on digital twinning

Abstract

The invention discloses an engineering settlement deformation monitoring and analyzing system and method based on digital twinning, and relates to the technical field of intelligent engineering monitoring, target scene data and first-period monitoring data are collected through a sensor, in combination with historical source scene data, migration parameter prior distribution theta pri is obtained through a migration learning algorithm, and the model initialization period is shortened. Based on a physical correction consolidation equation and Bayesian regularization objective function inversion posterior parameter distribution theta post, a settlement prediction confidence interval is generated in combination with sampling calculation, and the model interpretability is enhanced. And dynamically comparing the maximum value of the confidence interval with an allowable settlement threshold value, and iteratively optimizing parameters according to a large amount of obtained data, so that the model training time is shortened. And feature vectors F are constructed to archive historical data, cross-scene knowledge reuse is supported, and the deployment cost of small and medium-sized projects is reduced. The problems that a traditional method is poor in model reusability, lags in early warning and the like are solved, and efficient and low-cost intelligent monitoring support is provided for complex geological engineering.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent engineering monitoring, in particular to an engineering settlement deformation monitoring analysis system and method based on digital twinning. BACKGROUND

[0002] In the intelligent transformation wave of civil engineering, digital twinning technology is gradually promoting engineering monitoring to cross from passive response to active prediction. Focusing on underground engineering and geotechnical engineering scenarios, the coupling of multi-source heterogeneous geological conditions and dynamic construction disturbance makes the settlement deformation present significant spatio-temporal heterogeneity. Traditional digital twinning systems mostly use single-scene independent modeling mode, which reconstructs the physical model for each project and optimizes the parameters from scratch, resulting in insufficient model generalization ability. In actual engineering, although similar scenarios have different geological parameters, they often imply cross-scene commonality rules. How to reuse historical engineering experience through knowledge transfer to quickly build a digital twinning model for a new scene has become a key challenge to improve monitoring efficiency.

[0003] The current settlement system faces two major defects in cross-scene application: poor model reusability and low parameter initialization efficiency. On the one hand, existing methods need to rebuild the physical model for each new project, ignoring the common parameters verified in historical projects, resulting in repeated calculation and waste of computing power. On the other hand, model initialization parameters rely on human experience and are difficult to quickly converge through sparse initial monitoring data of the target scene, causing long inversion optimization period and early warning lag.

[0004] The above defects directly lead to cross-scene knowledge fragmentation and a sharp increase in long-tail engineering costs. First, valuable parameters accumulated from historical projects become data islands due to the lack of transfer mechanism, and new projects are forced to start from zero to explore the parameter space of the model, increasing uncertainty risk. Second, in complex geological areas, there is no suitable historical data, and repeated parameter inversion calculations consume a lot of computing power and time, making the deployment cost of digital twinning systems high, which restricts their popularization in small and medium-sized projects. SUMMARY

[0005] To overcome the defects of the prior art, the present application provides an engineering settlement deformation monitoring analysis system and method based on digital twinning, which solves the problems mentioned in the background.

[0006] To achieve the above purpose, the present application is implemented by the following technical scheme: an engineering settlement deformation monitoring analysis system and method based on digital twinning, comprising the following steps:

[0007] S1, obtaining a target scene soil layer parameter characteristic set ΦTar and first period detection data D by field investigation on the target scene and sensor detection, extracting a source scene soil layer parameter characteristic set ΦSou from a historical database, combining the target scene soil layer parameter characteristic set ΦTar and the source scene soil layer parameter characteristic set ΦSou, designing a transfer learning algorithm, and obtaining a transfer parameter prior distribution θpri of the target scene;

[0008] S2, constructing a parameter inversion model according to the transfer parameter prior distribution θpri and the first period monitoring data D, combining a physical model and a Bayesian regularization objective function, and obtaining an inversion parameter posterior distribution θpost;

[0009] S3, extracting N groups of parameter samples based on the inversion parameter posterior distribution θpost to form an inversion parameter set calculating a hth group of settlement prediction value set based on the physical model and calculating a confidence interval [Slow, Shig];

[0010] S4, comparing the confidence interval [Slow, Shig] with a preset allowable settlement threshold Sall, generating a control suggestion according to a comparison result, reacquiring monitoring data Dnew and inputting the monitoring data Dnew into the parameter inversion model, updating the inversion parameter posterior distribution θnew, and reperforming confidence interval statistics and risk assessment, and training the physical model based on the acquired data;

[0011] S5, recording historical parameters involved in calculation of the inversion model of the target scene, constructing a feature vector F, archiving and storing the feature vector F into a database, and updating the database.

[0012] Preferably, S1 includes S11 and S12;

[0013] S11, obtaining the target scene soil layer parameter characteristic set ΦTar and the first period detection data D={Sobs(t), t} by field investigation on the target scene and detection of multiple types of sensors, and extracting the source scene soil layer parameter characteristic set ΦSou from the historical database;

[0014] The multiple types of sensors include a pressure sensor, a geological radar, a stress sensor, a pore water pressure sensor and a static water level gauge.

[0015] The target scene soil layer parameter feature set ΦTar includes a target scene soil layer compression coefficient C, a target scene initial void ratio E, a target scene soil layer thickness H, a target scene vertical effective stress σ, and a target scene consolidation rate coefficient k. The source scene soil layer parameter feature set ΦSou includes a source scene soil layer compression coefficient C_S, a source scene initial void ratio E_S, a source scene soil layer thickness H_S, a source scene vertical effective stress σ_S, and a source scene consolidation rate coefficient k_S. The first detection data D includes a measured settlement amount Sobs(t) at a time t and the time t.

[0016] Preferably, S12, based on the target scene soil layer parameter feature set ΦTar={C, E, H, σ, k} and the source scene soil layer parameter feature set ΦSou={C_S, E_S, H_S, σ_S, k_S}, designs a transfer learning algorithm to calculate the maximum mean difference MMD between the target scene and the source scene. 2 ;

[0017] The transfer algorithm formula is as follows:

[0018]

[0019] In the formula, K represents a Gaussian kernel function, represents a Gaussian kernel function average value of all sample pairs in the target scene soil layer parameter feature set ΦTar, represents a Gaussian kernel function average value of all sample pairs in the source scene soil layer parameter feature set ΦSou, represents a cross-scene Gaussian kernel function average value of all sample pairs in the target scene soil layer parameter feature set ΦTar and all sample pairs in the source scene soil layer parameter feature set ΦSou, K represents a Gaussian kernel function, n_T represents a target scene sample number, n_S represents a source scene sample number, i1 and i2 represent traversal index values of target scene soil layer parameter feature set ΦTar sample pairs, ΦTar (i1) represents an index value of the outer loop that traverses target scene soil layer parameter feature set ΦTar sample pairs, ΦTar (i2) represents an index value of the inner loop that traverses target scene soil layer parameter feature set ΦTar sample pairs, j1 and j2 represent traversal index values of scene soil layer parameter feature set ΦSou sample pairs, ΦSou (j1) represents an index value of the outer loop that traverses source scene soil layer parameter feature set ΦSou sample pairs, ΦSou (j2) represents an index value of the inner loop that traverses source scene soil layer parameter feature set ΦSou sample pairs;

[0020] The maximum mean difference MMD is calculated by traversing and obtaining the minimum value of the maximum mean difference MMD 2 . 2Minimize the mean μTar and covariance ΣTar of the target scene soil layer parameter feature set ΦTar, combine the Adam gradient descent algorithm to generate the transfer parameter prior distribution θpri of the adapted target scene;

[0021] Wherein, the Adam gradient descent algorithm formula is as follows:

[0022]

[0023] In the formula, μadapt represents the mean value in the transfer parameter prior distribution, Σadapt represents the covariance in the transfer parameter prior distribution, η represents a preset learning rate, MMD represents the maximum mean difference 2 The partial derivative of the parameter distribution mean μ is MMD represents the maximum mean difference 2 The partial derivative of the parameter distribution covariance Σ is

[0024] The distribution of the transfer parameter prior distribution θpri is θpri~(μadapt, Σadapt).

[0025] Preferably, S2 includes S21 and S22.

[0026] S21, in combination with the target scene soil layer parameter feature set ΦTar={C, E, H, σ, k} and the first detection data D={Sobs(t), t}, constructs a physically modified consolidation equation to calculate the predicted settlement amount Spre(t) of the physical model at time t.

[0027] The physically modified consolidation equation is as follows:

[0028]

[0029] In the formula, g represents the gth soil layer, m represents the total number of soil layers, C_g represents the target scene gth soil compression coefficient C, E_g represents the target scene gth initial void ratio E, H_g represents the target scene gth soil layer thickness H, σ_g represents the target scene gth vertical effective stress σ, (σ_g, t) represents the target scene gth vertical effective stress σ at time t, e represents a natural constant, (k_g, t) represents the target scene gth consolidation rate coefficient k at time t, log represents a logarithmic function, α represents the target scene gth soil compression coefficient C and the target scene gth initial void ratio E contribution weight coefficient, β represents the target scene gth consolidation rate coefficient k at time t contribution weight coefficient, and γ represents the target scene gth vertical effective stress σ weight contribution coefficient.

[0030] Preferably, in step S22, based on the prior distribution of migration parameters θpri, a set of migration soil layer parameters Φpri that conform to the prior distribution of migration parameters θpri are generated. Based on the initial detection data D={Sobs(t), t} and the predicted settlement Spre(t) at time t, a Bayesian regularization objective function is designed, and the inversion parameter Φpost is calculated.

[0031] The Bayesian regularization objective function formula is as follows:

[0032]

[0033] In the formula, The L2 norm squaring operation is represented by ||·||1, which represents the L1 norm, and λ represents the preset regularization coefficient.

[0034] Based on the inversion parameter Φpost, the posterior distribution θpost of the inversion parameter is calculated.

[0035] Preferably, S3 includes S31;

[0036] S31. Based on the posterior distribution θpost of the inversion parameters, generate N sets of inversion parameter sets that conform to the posterior distribution θpost of the inversion parameters. The set of inversion parameters that conforms to the posterior distribution θpost of the inversion parameters Substitute the values ​​into the physically corrected consolidation equation in S21 to calculate the predicted settlement value Spre for each group. h Obtain a set of predicted settlement values.

[0037] Where N represents the total number of groups, and h represents the h-th group out of the total N groups;

[0038] Based on the set of settlement prediction values Calculate the confidence interval [Slow, Shigh]. The formula for calculating the confidence interval [Slow, Shigh] is as follows:

[0039] Slow = Quantile({Spre}) N ,0025);

[0040] Shigh = Quantile({Spre}) N ,0975);

[0041] Where Slow represents the minimum value in the confidence interval [Slow, Shigh], Shigh represents the maximum value in the confidence interval [Slow, Shigh], and Quantile represents the quantile function.

[0042] Preferably, S4 includes S41 and S42;

[0043] S41. Compare the maximum value Shigh in the obtained confidence interval [Slow, Shigh] with the preset allowable settlement threshold Sal;

[0044] If the maximum value Shigh in the confidence interval [Slow, Shigh] is less than or equal to the allowable settlement threshold Sal, then no warning will be triggered and no control recommendations need to be generated.

[0045] If the maximum value Shigh in the confidence interval [Slow, Shigh] is greater than the allowable settlement threshold Sall, an early warning is triggered and a control recommendation is generated.

[0046] Preferably, in step S42, new detection data Dnew is acquired again, and the physical correction consolidation equation and Bayesian regularization objective function are used again in combination with the current posterior parameter distribution θpost. Based on the updated posterior parameter distribution of the new detection data Dnew, the optimized posterior parameter distribution θnew is obtained.

[0047] Based on the optimized posterior parameter distribution θnew, repeat steps S3 and S4 until the KL dispersion of the optimized posterior parameter distribution θnew and the parameter distribution of the target scene soil layer parameter feature set ΦTar is less than the preset dispersion threshold.

[0048] The physical model is trained based on the large amount of optimized posterior parameter distribution θnew and the newly acquired detection data Dnew. The contribution weight coefficients α, β and γ in the physical correction consolidation equation are dynamically assigned according to the results of each training session.

[0049] Preferably, S5 includes S51;

[0050] S51. Construct a feature vector set F, which includes the following:

[0051] First, the posterior distribution set of inversion parameters {θopst_1,θnew_2,……,θnew_r} generated by the inversion model of the target scene in historical iterations is recorded in chronological order, where r represents the total number of iterations. Second, the comparison results between the confidence interval [Slow,Shigh] calculated in each iteration and the preset allowable settlement threshold Sal are recorded. Finally, the soil parameter feature set ΦTar of the input target scene and the soil parameter feature set ΦSou of the source scene are recorded.

[0052] The feature vector set F is stored in the knowledge base and the knowledge base is updated.

[0053] A digital twin-based engineering settlement deformation monitoring and analysis system and method includes a cross-scenario parameter transfer learning module, a physical inversion module, a settlement prediction module, a dynamic optimization feedback module, and a feature archiving module.

[0054] The cross-scene parameter transfer learning module obtains the target scene soil layer parameter feature set ΦTar and the initial detection data D by conducting on-site investigation and setting up sensor detection in the target scene. It extracts the source scene soil layer parameter feature set ΦSou from the historical database. Combining the target scene soil layer parameter feature set ΦTar and the source scene soil layer parameter feature set ΦSou, a transfer learning algorithm is designed to obtain the prior distribution θpri of the transfer parameters in the target scene.

[0055] Based on the prior distribution of migration parameters θpri and the initial monitoring data D, a parameter inversion model is constructed by combining the physical model and the Bayesian regularization objective function to obtain the posterior distribution of inversion parameters θpost.

[0056] The settlement prediction module extracts N sets of parameter samples based on the posterior distribution θpost of the inversion parameters to form the inversion parameter set. The h-th set of predicted settlement values ​​was calculated based on a physical model. And calculate the confidence interval [Slow, Shig];

[0057] The dynamic optimization feedback module compares the confidence interval [Slow, Shigh] with the preset allowable settlement threshold Sal, generates control suggestions based on the comparison results, reacquires monitoring data Dnew and inputs it into the parameter inversion model, updates the posterior distribution θnew of the inversion parameters, and re-performs the statistical analysis and risk assessment of the confidence interval. The physical model is then trained based on the acquired data.

[0058] The feature archiving module records the historical parameters used in the calculation of the inversion model of the target scene, constructs the feature vector F, archives it into the database, and updates the database.

[0059] This invention provides a digital twin-based engineering settlement deformation monitoring and analysis system and method, which has the following beneficial effects:

[0060] (1) During system operation, at the implementation level, the efficiency and reliability of engineering settlement monitoring are significantly improved by constructing a closed-loop process of "data acquisition - transfer learning - parameter inversion - dynamic optimization". First, the system utilizes cross-scenario transfer learning technology to combine the verified parameters of historical projects with the initial data of the target scenario, quickly generating a prior distribution of transfer parameters adapted to the target scenario, solving the problem of redundant calculations and wasted computing power caused by the traditional method of independently modeling each new project. Second, by combining the physical model with the Bayesian regularization objective function, the system can efficiently invert the posterior distribution of soil parameters based on sparse initial monitoring data, greatly shortening the model parameter initialization cycle and reducing the risk of lag in early warning. In addition, the system achieves quantitative assessment of settlement risk by dynamically extracting parameter samples and calculating the confidence interval of settlement prediction, providing a scientific basis for engineering decision-making. Overall, this method not only improves the generalization ability of the model, but also achieves accurate prediction and active control of settlement deformation under complex geological conditions through a closed-loop optimization mechanism, effectively reducing engineering costs and uncertainty risks.

[0061] (2) Collect the target scene soil layer parameter feature set ΦTar and the initial detection data D through multiple types of sensors, and combine them with the source scene soil layer parameter feature set ΦSou in the historical database, using the maximum mean difference MMD 2 The algorithm aligns parameter distributions across different scenarios. This design enables transfer learning to extract common patterns from historical engineering projects, providing a high-precision prior distribution θpri of transfer parameters for the target scenario, avoiding the "cold start" problem of traditional methods that ignore cross-scenario knowledge. Furthermore, by combining the physically modified consolidation equation and the Bayesian regularized objective function, the posterior distribution θpost of the inversion parameters is calculated by fusing prior knowledge with measured data. This process, through L1 and L2 regularization constraints, effectively balances model complexity and data fit, improving the robustness of parameter inversion. Multiple sets of parameter samples are generated based on the posterior distribution, and the predicted settlement values ​​are calculated using the physical model, with confidence intervals statistically analyzed to quantify the uncertainty of the prediction results. This series of interconnected technologies enables the system to achieve breakthroughs in both efficiency and accuracy in the three key stages of parameter initialization, model optimization, and risk prediction.

[0062] (3) The system's continuous optimization capability is further enhanced through dynamic feedback and knowledge archiving mechanisms. In step S4, the system compares the maximum value of the confidence interval Shig with the preset allowable settlement threshold Sal. If the threshold is exceeded, an early warning is triggered and control suggestions are generated. At the same time, the posterior distribution θnew is updated based on the new monitoring data Dnew. This dynamic optimization process gradually narrows the prediction confidence interval through iterative inversion and parameter correction until the KL dispersion of the parameter distribution reaches the preset threshold, thereby ensuring that the model always fits the actual engineering state. By constructing a feature vector set F, historical inversion parameters, confidence interval comparison results, and scene features are recorded, and these data are archived to the knowledge base. This design not only realizes the structured storage of data throughout the entire life cycle of the project, but also provides a rich experience base for the transfer learning of subsequent new scenarios. Through the closed loop of "monitoring-optimization-archiving", the system is promoted from passive response to a single scenario to active learning in multi-scenario collaboration, laying a sustainable technical foundation for the long-term application of digital twin technology in engineering monitoring. Attached Figure Description

[0063] Figure 1 This is a schematic diagram illustrating the steps of a digital twin-based engineering settlement deformation monitoring and analysis method according to the present invention.

[0064] Figure 2 This is a schematic diagram of a digital twin-based engineering settlement and deformation monitoring and analysis system according to the present invention.

[0065] Figure 3 This demonstrates the data processing flow of a transfer learning algorithm. Detailed Implementation

[0066] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0067] Example 1

[0068] This invention relates to an engineering settlement deformation monitoring and analysis system and method based on digital twins. Please refer to [link / reference]. Figure 1 This includes the following steps:

[0069] S1. By conducting on-site investigations and setting up sensor detection for the target scene, obtain the target scene soil layer parameter feature set ΦTar and the initial detection data D. Extract the source scene soil layer parameter feature set ΦSou from the historical database. Combine the target scene soil layer parameter feature set ΦTar and the source scene soil layer parameter feature set ΦSou to design a transfer learning algorithm and obtain the prior distribution θpri of the transfer parameters of the target scene.

[0070] S2. Based on the prior distribution of migration parameters θpri and the initial monitoring data D, construct a parameter inversion model by combining the physical model and the Bayesian regularization objective function, and obtain the posterior distribution of inversion parameters θpost.

[0071] S3. Based on the posterior distribution θpost of the inversion parameters, extract N sets of parameter samples to form the inversion parameter set. The h-th set of predicted settlement values ​​was calculated based on a physical model. And calculate the confidence interval [Slow, Shig];

[0072] S4. Based on the confidence interval [Slow, Shigh] and the preset allowable settlement threshold Sal, compare the results and generate control suggestions. Reacquire the monitoring data Dnew and input it into the parameter inversion model. Update the posterior distribution of the inversion parameters θnew and re-perform the statistics and risk assessment of the confidence interval. Train the physical model based on the acquired data.

[0073] S5. Record the historical parameters used in the calculation of the inversion model of the target scene, construct the feature vector F, archive and store it in the database, and update the database.

[0074] In this embodiment, based on the soil layer parameter feature set ΦTar of the target scene and the soil layer parameter feature set ΦSou of the source scene, a prior distribution θpri of transfer parameters is generated through a transfer learning algorithm. This effectively utilizes the common patterns of historical projects, avoiding the redundant calculations and wasted computing power caused by independent modeling in traditional methods, and shortening the model initialization cycle. Combining the initial monitoring data D and the Bayesian regularization objective function, the posterior distribution θpost of parameters is inverted. By using L1 and L2 regularization constraints to balance model complexity and data fit, rapid parameter convergence can be achieved with only sparse initial data, shortening the inversion optimization cycle. Based on the settlement prediction confidence interval [Slow, Shigh] generated by θpost, the maximum value Shigh in the confidence interval [Slow, Shigh] is dynamically compared with the allowable settlement threshold Sal, and new monitoring data Dnew is collected for iterative optimization, shortening the model training cycle and solving the problem of early warning lag. Finally, historical data is archived through feature vector F to build a cross-scene knowledge base, enabling new projects to directly reuse the historical parameter prior distribution θpri, reducing the deployment cost of small and medium-sized projects, and realizing a technological leap from data silos to multi-scene collaborative optimization.

[0075] Example 2

[0076] This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1 Specifically: S1 includes S11 and S12;

[0077] S11. By conducting on-site investigations of the target scene and setting up various types of sensors for detection, the target scene soil layer parameter feature set ΦTar and the initial detection data D={Sobs(t), t} are obtained, and the source scene soil layer parameter feature set ΦSou is extracted from the historical database.

[0078] Among them, various types of sensors include pressure sensors, ground-penetrating radar, stress sensors, pore water pressure sensors, and hydrostatic levels;

[0079] The target scene soil layer parameter feature set ΦTar includes the target scene soil layer compression coefficient C, the target scene initial void ratio E, the target scene soil layer thickness H, the target scene vertical effective stress σ, and the target scene consolidation rate coefficient k. The source scene soil layer parameter feature ΦSou includes the source scene soil layer compression coefficient C_S, the source scene initial void ratio E_S, the source scene soil layer thickness H_S, the source scene vertical effective stress σ_S, and the source scene consolidation rate coefficient k_S. The initial detection data D includes the measured settlement Sobs(t) at time t and time t.

[0080] Among them, the soil compression coefficient C was obtained by drilling undisturbed soil samples for consolidation tests and using pressure sensors to measure the load and record the compression deformation; the initial void ratio E was obtained by measuring the natural density, water content and specific gravity of the soil samples using a hydrometer method; the soil layer thickness H was obtained by analyzing the bottom interface using ground-penetrating radar; the consolidation rate coefficient k was obtained by detecting the dissipation process using pore water pressure sensors buried on site and combining it with time curves; the measured settlement Sobs was obtained by setting up static level nodes along the perimeter of the project to monitor the structural settlement data in real time.

[0081] S12. Based on the soil layer parameter feature set ΦTar={C,E,H,σ,k} of the target scene and the soil layer parameter feature set ΦSou={C_S,E_S,H_S,σ_S,k_S} of the source scene, design a transfer learning algorithm to calculate the maximum mean difference (MMD) between the target scene and the source scene. 2 ;

[0082] The migration algorithm formula is as follows:

[0083]

[0084] In the formula, K represents the Gaussian kernel function. This represents the average Gaussian kernel function of all sample pairs within the target scene's soil layer parameter feature set ΦTar, used to quantify the local consistency of the target scene's parameter distribution. The average Gaussian kernel function of all sample pairs within the source scene soil layer parameter feature set ΦSou represents the internal structure of the source scene parameter distribution. ΦTar represents the average Gaussian kernel function across all sample pairs within the target scene soil parameter feature set ΦTar and the source scene soil parameter feature set ΦSou, measuring the global distribution difference between the two scenes. K represents the Gaussian kernel function, n_T represents the number of samples in the target scene, n_S represents the number of samples in the source scene, and i1 and i2 represent the traversal index values ​​of the sample pairs in the target scene soil parameter feature set ΦTar. (i1) This represents the index value of the sample pair ΦTar in the outer loop traversing the soil layer parameter feature set of the target scene. (i2) j1 and j2 represent the index values ​​of the sample pairs in the target scene soil layer parameter feature set ΦTar during the inner loop traversal, and j1 and j2 represent the traversal index values ​​of the sample pairs in the scene soil layer parameter feature set ΦSou. (j1) This represents the index value of the sample pair ΦSou, which represents the feature set of soil layer parameters in the source scene that the outer loop iterates through. (j2) This represents the index value of the sample pair ΦSou, which represents the feature set of soil layer parameters in the source scene that is iterated through in the inner loop.

[0085] The objective of this formula is to achieve the Maximum Mean Difference (MMD) through a transfer learning algorithm. 2 This approach comprehensively captures the nonlinear differences in parameter distribution between the target and source scenes, rather than relying solely on single statistics such as mean or variance. This design is particularly suitable for complex soil parameters in engineering settlement monitoring, as they exhibit nonlinear and multimodal distribution characteristics that traditional linear measurement methods struggle to accurately characterize. Furthermore, the choice of the Gaussian kernel function endows the algorithm with robustness to changes in data scale. By adjusting the kernel width parameter, it can adapt to the alignment requirements of parameters with different dimensions. Through kernel space mapping, complex parameter distribution differences are transformed into optimizable numerical indicators, providing a clear optimization objective for transfer learning. This solves the problem of model cold-start failure caused by parameter distribution mismatch in traditional methods, laying a reliable foundation for subsequent physical inversion and dynamic optimization.

[0086] A specific example of obtaining the average Gaussian kernel function KTT of all sample pairs within the target scene soil layer parameter feature set ΦTar:

[0087] The soil layer parameters of the target scene are shown in Table 1:

[0088] Table 1. Soil layer parameter characteristics of the target scenario:

[0089] Sample C E H σ k 1 0.65 0.60 0.83 0.50 0.40 2 0.75 0.55 0.75 0.60 0.35 3 0.55 0.65 0.85 0.40 0.45

[0090] The formula for the Gaussian kernel function is as follows:

[0091]

[0092] In the formula, ω represents the Gaussian kernel width, which controls the similarity sensitivity, ||·|| 2 This represents the squared Euclidean distance, specifically ω = 0.1;

[0093] With sample ΦTar (1) and sample ΦTar (2) The kernel function is used to calculate K(ΦTar(1), ΦTar (2) For example, the specific calculation is as follows:

[0094] Calculate item by item:

[0095] Cdifference = (0.65 - 0.75) 2 =(-0.10) 2 =0.01;

[0096] E difference = (0.60 - 0.55) 2 =(0.05) 2 =0.0025;

[0097] H difference = (0.083 - 0.85) 2 =(-0.02) 2 =0.0004;

[0098] σ difference = (0.50 - 0.40) 2 =(0.10) 2 =0.01;

[0099] k-difference = (0.40 - 0.45) 2 =(-0.05) 2 =0.0025;

[0100] ||xy|| 2 =0.01+0.0025+0.0064+0.01+0.0025=0.0314;

[0101] Substituting the Gaussian kernel function, the calculation is as follows:

[0102]

[0103] The kernel function calculation values ​​are summarized below:

[0104] K(ΦTar (1) ,ΦTar (2) ) = 0.208;

[0105] K(ΦTar (1) ,ΦTar (3) ) = 0.281;

[0106] K(ΦTar (2) ,ΦTar (3) ) = 0.0041;

[0107] Calculate the kernel mean of the target scene parameter features. The calculation is as follows:

[0108]

[0109] In the formula, all sample pairs need to be calculated, including self-pairing i=j, and the kernel value of each sample is K(x,x)=1;

[0110] Iterate through the data to obtain the maximum mean difference (MMD). 2 The minimum value, obtained through the maximum mean difference (MMD) 2 Minimize the mean μTar and covariance ΣTar of the soil layer parameter feature set ΦTar of the aligned target scene, and combine the Adam gradient descent algorithm to generate the prior distribution θpri of the transfer parameters of the adapted target scene;

[0111] The Adam gradient descent algorithm formula is as follows:

[0112]

[0113] In the formula, μadapt represents the mean of the prior distribution of the transfer parameters, Σadapt represents the covariance of the prior distribution of the transfer parameters, and η represents the preset learning rate. The maximum mean difference (MMD) represents the maximum mean difference. 2 The partial derivative with respect to the mean μ of the parameter distribution, The maximum mean difference (MMD) represents the difference between the two means. 2 Partial derivative with respect to the covariance Σ of the parameter distribution;

[0114] The prior distribution of the migration parameter θpri is θpri~(μadapt,Σadapt).

[0115] In this embodiment, multiple types of sensors are deployed to accurately collect the target scene soil layer parameter feature set ΦTar and the initial detection data D={Sobs(t), t}, and combined with the source scene soil layer parameter feature set ΦSou in the historical database, a maximum mean difference (MMD) is designed. 2 The transfer learning algorithm achieves efficient alignment of parameter distributions across scenes. This is achieved by minimizing the MMD (Missing Modulation Distribution) between the target scene and the source scene. 2The algorithm dynamically adjusts the mean μTar and covariance ΣTar of the target scene's soil layer parameter feature set ΦTar using the Adam gradient descent algorithm, generating a prior distribution of transfer parameters θpri ~ (μadapt, Σadapt) adapted to the target scene. This process not only significantly improves the accuracy of parameter initialization but also reduces reliance on human experience and redundant historical data through automated cross-scene knowledge transfer, shortening the initialization cycle of the new scene model. Furthermore, the transfer learning algorithm quantifies the deep feature differences between the target scene's soil layer parameter feature set ΦTar and the source scene's soil layer parameter feature set ΦSou, ensuring the robustness of the transfer parameter θpri under complex geological conditions. This lays a high-confidence prior foundation for subsequent physical inversion and dynamic optimization, solving the model cold-start failure problem caused by parameter distribution deviations in traditional methods.

[0116] Example 3

[0117] This embodiment is an explanation based on Embodiment 2. Please refer to it. Figure 1 Specifically: S2 includes S21 and S22;

[0118] S21. Combining the target scenario soil layer parameter feature set ΦTar={C,E,H,σ,k} and the initial detection data D={Sobs(t), t}, construct the physical correction consolidation equation and calculate the predicted settlement Spre(t) of the physical model at time t.

[0119] The physically modified consolidation equation is as follows:

[0120]

[0121] In the formula, g represents the g-th soil layer, m represents the total number of soil layers, C_g represents the compressibility coefficient C of the g-th soil layer in the target scenario, E_g represents the initial void ratio E of the g-th soil layer in the target scenario, H_g represents the thickness H of the g-th soil layer in the target scenario, σ_g represents the vertical effective stress σ of the g-th soil layer in the target scenario, (σ_g, t) represents the vertical effective stress σ of the g-th soil layer in the target scenario at time t, e represents the natural constant, (k_g, t) represents the consolidation rate coefficient k of the g-th soil layer in the target scenario at time t, log represents the logarithmic function, α represents the contribution weight coefficient of the compressibility coefficient C and the initial void ratio E of the g-th soil layer in the target scenario, β represents the contribution weight coefficient of the consolidation rate coefficient k of the g-th soil layer in the target scenario at time t, and γ represents the weight contribution coefficient of the vertical effective stress σ of the g-th soil layer in the target scenario.

[0122] The purpose of this formula is to retain the core framework of classical consolidation theory through a collaborative design that combines physical mechanisms and data-driven approaches. It also addresses the prediction distortion issues of traditional methods in multi-layered heterogeneous soils, dynamic construction scenarios, and small sample data scenarios by employing dynamic weighting, hierarchical modeling, and nonlinear function extensions. Its mathematical form strictly corresponds to the actual physical process, while possessing high accuracy, strong generalization, and interpretability, providing a reliable computational foundation for real-time monitoring and risk warning of engineering settlement.

[0123] S22. Based on the prior distribution of migration parameters θpri, generate a set of migration soil layer parameters Φpri that conform to the prior distribution of migration parameters θpri. Based on the initial detection data D={Sobs(t), t} and the predicted settlement Spre(t) at time t, design a Bayesian regularization objective function and calculate the inversion parameter Φpost.

[0124] The Bayesian regularization objective function formula is as follows:

[0125]

[0126] In the formula, represents the L2 norm square operation, which constrains the prediction error; ||·||1 represents the L1 norm, which constrains the parameter offset; and λ represents the preset regularization coefficient.

[0127] Based on the inversion parameter Φpost, the posterior distribution θpost of the inversion parameter is calculated.

[0128] In this embodiment, based on the soil layer parameter feature set ΦTar of the target scene and the initial monitoring data D={Sobs(t), t}, combined with the prior distribution of migration parameters θpri, a physical correction consolidation equation is designed to calculate the predicted settlement Spre(t). By introducing the soil layer compression coefficient weight α, the consolidation rate coefficient weight β, and the vertical effective stress weight γ, the traditional consolidation theory is dynamically corrected, embedding domain knowledge into the model and enhancing the model's dynamic adjustment capability. Furthermore, using the Bayesian regularization objective function, L2 norm squared constraint on prediction error, and L1 norm constraint on parameter offset, the posterior distribution of parameters θpost is inverted, effectively balancing model complexity and data fit, and avoiding the risk of overfitting due to initial data sparsity. By fusing the statistical characteristics of the prior distribution θpri with the measured data D, the inversion process only requires 10% of the conventional monitoring data of the target scene to achieve rapid convergence of the posterior distribution θpost. This technological breakthrough not only solves the shortcomings of traditional physical models, which rely on manual parameter tuning and have weak generalization ability, but also provides a high-confidence parameter benchmark for subsidence prediction under complex geological conditions, laying a precise model foundation for subsequent dynamic optimization and risk warning.

[0129] Example 4

[0130] This embodiment is an explanation based on Embodiment 3. Please refer to it. Figure 1 Specifically: S3 includes S31;

[0131] S31. Based on the posterior distribution θpost of the inversion parameters, generate N sets of inversion parameter sets that conform to the posterior distribution θpost of the inversion parameters. The set of inversion parameters that conforms to the posterior distribution θpost of the inversion parameters Substitute the values ​​into the physically corrected consolidation equation in S21 to calculate the predicted settlement value Spre for each group. h Obtain a set of predicted settlement values.

[0132] Where N represents the total number of groups, and h represents the h-th group out of the total N groups;

[0133] Based on the set of settlement prediction values Calculate the confidence interval [Slow, Shigh]. The formula for calculating the confidence interval [Slow, Shigh] is as follows:

[0134] Slow = Quantile({Spre}) N ,0025);

[0135] Shigh = Quantile({Spre}) N ,0975);

[0136] Where Slow represents the minimum value in the confidence interval [Slow, Shigh], Shigh represents the maximum value in the confidence interval [Slow, Shigh], and Quantile represents the quantile function.

[0137] S4 includes S41 and S42;

[0138] S41. Compare the maximum value Shigh in the obtained confidence interval [Slow, Shigh] with the preset allowable settlement threshold Sal;

[0139] If the maximum value Shigh in the confidence interval [Slow, Shigh] is less than or equal to the allowable settlement threshold Sal, then no warning will be triggered and no control recommendations need to be generated.

[0140] If the maximum value Shigh in the confidence interval [Slow, Shigh] is greater than the allowable settlement threshold Sall, an early warning is triggered and a control recommendation is generated.

[0141] S42. Reacquire new detection data Dnew, and combine it with the current posterior parameter distribution θpost. Then, apply the physical correction consolidation equation and the Bayesian regularization objective function again. Based on the updated posterior parameter distribution of the new detection data Dnew, obtain the optimized posterior parameter distribution θnew.

[0142] Based on the optimized posterior parameter distribution θnew, repeat steps S3 and S4 until the KL dispersion of the optimized posterior parameter distribution θnew and the parameter distribution of the target scene soil layer parameter feature set ΦTar is less than the preset dispersion threshold.

[0143] The physical model is trained based on the large amount of optimized posterior parameter distribution θnew and the newly acquired detection data Dnew. The contribution weight coefficients α, β and γ in the physical correction consolidation equation are dynamically assigned according to the results of each training session.

[0144] S5 includes S51;

[0145] S51. Construct a feature vector set F, which includes the following:

[0146] First, the posterior distribution set of inversion parameters {θopst_1,θnew_2,……,θnew_r} generated by the inversion model of the target scene in historical iterations is recorded in chronological order, where r represents the total number of iterations. Second, the comparison results between the confidence interval [Slow,Shigh] calculated in each iteration and the preset allowable settlement threshold Sal are recorded. Finally, the soil parameter feature set ΦTar of the input target scene and the soil parameter feature set ΦSou of the source scene are recorded.

[0147] The feature vector set F is stored in the knowledge base and the knowledge base is updated.

[0148] In this embodiment, based on the posterior distribution θpost of the inversion parameters, N sets of parameter samples {Φpt}N are generated using Monte Carlo sampling. The set of predicted settlement values ​​{Spre}N is calculated through the physical correction consolidation equation, and the quantile function is used to statistically determine the 95% confidence interval [Slow, Shigh]. This design incorporates the spatial heterogeneity of parameters such as the soil compressibility coefficient C and the consolidation rate coefficient k into the probabilistic risk assessment, expanding the prediction results from deterministic numerical values ​​to risk distributions, providing a quantitative tolerance basis for engineering decisions, and solving the "pseudo-accuracy" problem of traditional single-point prediction in complex strata. Through a dynamic closed-loop optimization mechanism, the upper limit of the confidence interval Shigh and the allowable settlement threshold Sall are compared in real time, triggering graded early warning and model iteration. Then, the system automatically collects new monitoring data Dnew, updates the posterior distribution to θnew using the Bayesian regularization objective function, and uses KL dispersion as the convergence criterion to ensure that the parameter distribution dynamically adapts to the soil layer characteristics ΦTar of the target scenario. This process causes the model prediction error to decrease exponentially with the number of iterations. A feature vector F is constructed, and historical iteration parameters {θpost_1, θnew_2, ... θnew_r}, confidence interval comparison results, and scene features are structurally archived to form a cross-engineering knowledge base. This base allows new scenes to directly call the high-confidence prior distribution θpri through similarity retrieval, reducing the computational load of transfer learning and improving the deployment efficiency of small and medium-sized projects. At the same time, the knowledge base continuously injects optimized θnew and Dnew data, breaking through the limitations of traditional historical data silos and promoting the evolution of the model from "single-scene optimization" to "multi-scene symbiosis".

[0149] Example 5

[0150] A digital twin-based engineering settlement deformation monitoring and analysis system and method, please refer to... Figure 2 Specifically, it includes a cross-scene parameter transfer learning module, a physical inversion module, a settlement prediction module, a dynamic optimization feedback module, and a feature archiving module;

[0151] The cross-scene parameter transfer learning module obtains the target scene soil layer parameter feature set ΦTar and the initial detection data D by conducting on-site investigation and setting up sensor detection in the target scene. It extracts the source scene soil layer parameter feature set ΦSou from the historical database. Combining the target scene soil layer parameter feature set ΦTar and the source scene soil layer parameter feature set ΦSou, a transfer learning algorithm is designed to obtain the prior distribution θpri of the transfer parameters in the target scene.

[0152] Based on the prior distribution of migration parameters θpri and the initial monitoring data D, a parameter inversion model is constructed by combining the physical model and the Bayesian regularization objective function to obtain the posterior distribution of inversion parameters θpost.

[0153] The settlement prediction module extracts N sets of parameter samples based on the posterior distribution θpost of the inversion parameters to form the inversion parameter set. The h-th set of predicted settlement values ​​was calculated based on a physical model. And calculate the confidence interval [Slow, Shig];

[0154] The dynamic optimization feedback module compares the confidence interval [Slow, Shigh] with the preset allowable settlement threshold Sal, generates control suggestions based on the comparison results, reacquires monitoring data Dnew and inputs it into the parameter inversion model, updates the posterior distribution θnew of the inversion parameters, and re-performs the statistical analysis and risk assessment of the confidence interval. The physical model is then trained based on the acquired data.

[0155] The feature archiving module records the historical parameters used in the calculations of the target scene's inversion model, constructs a feature vector F, archives it into the database, and updates the database.

[0156] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A digital-twin-based engineering settlement deformation monitoring analysis system and method, characterized in that: The method comprises the following steps: S1, obtaining a target scene soil layer parameter characteristic set ΦTar and first period detection data D by field investigation and sensor detection on the target scene, extracting a source scene soil layer parameter characteristic set ΦSou from a historical database, combining the target scene soil layer parameter characteristic set ΦTar and the source scene soil layer parameter characteristic set ΦSou, designing a transfer learning algorithm, and obtaining a transfer parameter prior distribution θpri of the target scene; S2, constructing a parameter inversion model according to the transfer parameter prior distribution θpri and the first period monitoring data D, combining a physical model and a Bayesian regularization objective function, and obtaining an inversion parameter posterior distribution θpost; S3, extract N groups of parameter samples based on the inversion parameter posterior distribution θpost to form an inversion parameter set calculate the hth group of settlement prediction value set based on the physical model and statistical confidence interval [Slow, Shig]; S4, comparing the confidence interval [Slow, Shigh] and a preset allowable settlement threshold Sall, generating a control suggestion according to a comparison result, reacquiring monitoring data Dnew and inputting the monitoring data Dnew into the parameter inversion model, updating the inversion parameter posterior distribution θnew and reperforming confidence interval statistics and risk assessment, and training the physical model based on the obtained data; S5, recording historical parameters involved in calculation of the inversion model of the target scene, constructing a feature vector F, archiving and storing the feature vector F into a database, and updating the database.

2. The engineering settlement deformation monitoring analysis system and method based on digital twinning according to claim 1, characterized in that: S1 comprises S11 and S12; S11, obtaining a target scene soil layer parameter characteristic set ΦTar and first period detection data D={Sobs(t), t} by field investigation and detection of multiple types of sensors on the target scene, and extracting a source scene soil layer parameter characteristic set ΦSou from a historical database; The multiple types of sensors comprise a pressure sensor, a geological radar, a stress sensor, a pore water pressure sensor and a static water level gauge. The target scene soil layer parameter characteristic set ΦTar comprises a target scene soil layer compression coefficient C, a target scene initial void ratio E, a target scene soil layer thickness H, a target scene vertical effective stress σ and a target scene consolidation rate coefficient k, the source scene soil layer parameter characteristic set ΦSou comprises a source scene soil layer compression coefficient C_S, a source scene initial void ratio E_S, a source scene soil layer thickness H_S, a source scene vertical effective stress σ_S and a source scene consolidation rate coefficient k_S, and the first period detection data D comprises a measured settlement amount Sobs(t) at a time t and the time t.

3. The system and method for monitoring and analyzing engineering settlement deformation based on digital twinning according to claim 2, characterized in that: S12, based on the target scene soil layer parameter feature set ΦTar={C, E, H, σ, k} and the source scene soil layer parameter feature set ΦSou={C_S, E_S, H_S, σ_S, k_S}, a transfer learning algorithm is designed to calculate the maximum mean difference MMD of the target scene and the source scene 2 ; The transfer algorithm formula is as follows: In the formula, K represents a Gaussian kernel function, represents the Gaussian kernel function average value of all sample pairs inside the target scene soil parameter feature set ΦTar, represents the Gaussian kernel function average value of all sample pairs inside the source scene soil parameter feature set ΦSou, represents the cross-scene Gaussian kernel function average value of all sample pairs inside the target scene soil parameter feature set ΦTar and all sample pairs inside the source scene soil parameter feature set ΦSou, K represents a Gaussian kernel function, n_T represents the number of target scene samples, n_S represents the number of source scene samples, i1 and i2 represent the traversal index values of the sample pairs of the target scene soil parameter feature set ΦTar, (i1) represents the index value of the outer loop for traversing the sample pairs of the target scene soil parameter feature set ΦTar, (i2) represents the index value of the inner loop for traversing the sample pairs of the target scene soil parameter feature set ΦTar, j1 and j2 represent the traversal index values of the sample pairs of the scene soil parameter feature set ΦSou, (j1) represents the index value of the outer loop for traversing the sample pairs of the source scene soil parameter feature set ΦSou, (j2) represents the index value of the inner loop for traversing the sample pairs of the source scene soil parameter feature set ΦSou; The maximum mean difference MMD is obtained by traversing calculation 2 The minimum value of the maximum mean difference MMD 2 The average number muTar and the covariance SigmaTar of the parameter feature set of the aligned target scene soil layer parameter are minimized, and the transfer parameter prior distribution thetaPri of the adapted target scene is generated in combination with the Adam gradient descent algorithm. The Adam gradient descent algorithm formula is as follows: wherein μadaptdenotes the mean in the prior distribution of the transfer parameters, Σadaptdenotes the covariance in the prior distribution of the transfer parameters, η denotes a preset learning rate, denotes the maximum mean discrepancy MMD 2 the partial derivative with respect to the mean μ of the parameter distribution, denotes the maximum mean discrepancy MMD 2 the partial derivative with respect to the covariance Σ of the parameter distribution; The distribution of the transfer parameter prior distribution θpri is θpri~(μadapt, Σadapt).

4. The engineering settlement deformation monitoring analysis system and method based on digital twinning according to claim 3, characterized in that: S2 comprises S21 and S22; S21, combining the target scene soil layer parameter characteristic set ΦTar={C, E, H, σ, k} and the first period detection data D={Sobs(t), t}, constructing a physical modified consolidation equation, and calculating a predicted settlement amount Spre(t) of the physical model at the time t; The physical modified consolidation equation is as follows: In the formula, g represents the gth layer of soil, m represents the total number of soil layers, C_g represents the target scene gth layer soil compression coefficient C, E_g represents the target scene gth layer initial void ratio E, H_g represents the target scene gth layer soil layer thickness H, σ_g represents the target scene gth layer vertical effective stress σ, (σ_g, t) represents the target scene gth layer vertical effective stress σ at time t, e represents a natural constant, (k_g, t) represents the target scene gth layer consolidation rate coefficient k at time t, log represents a logarithmic function, α represents the target scene gth layer soil compression coefficient C and the target scene gth layer initial void ratio E contribution weight coefficient, β represents the target scene gth layer consolidation rate coefficient k at time t contribution weight coefficient, and γ represents the target scene gth layer vertical effective stress σ weight contribution coefficient.

5. The engineering settlement deformation monitoring analysis system and method based on digital twinning according to claim 4, characterized in that: S22, based on the migration parameter prior distribution θpri, a set of migration soil layer parameters Φpri conforming to the migration parameter prior distribution θpri is generated, a Bayesian regularization objective function is designed according to the first detection data D={Sobs(t), t} and the predicted settlement amount Spre(t) at time t, and the inversion parameter Φpost is calculated; The Bayesian regularization objective function is as follows: In the formula, represents the L2 norm square operation, ||·||1 represents the L1 norm, and λ represents a preset regularization coefficient. Based on the inversion parameter Φpost, the inversion parameter posterior distribution θpost is calculated.

6. The engineering settlement deformation monitoring analysis system and method based on digital twinning according to claim 5, characterized in that: S3 includes S31; S31, generating N sets of inversion parameters in accordance with the inversion parameter posterior distribution θpost based on the inversion parameter posterior distribution θpost The sets of inversion parameters in accordance with the inversion parameter posterior distribution θpost The sets of inversion parameters in accordance with the inversion parameter posterior distribution θpost h The sets of inversion parameters in accordance with the inversion parameter posterior distribution θpost Wherein, N represents the total number of groups, h represents the hth group in the total number N groups; Based on a set of settlement predictions A confidence interval [Slow, Shigh] is calculated, the confidence interval [Slow, Shigh] being calculated according to the following formula: Slow = Quantile({Spre N , 0.025); Shigh = Quantile({Spre N , 0.975); Wherein, Slow represents the minimum value in the confidence interval [Slow, Shigh], Shigh represents the maximum value in the confidence interval [Slow, Shigh], and Quantile represents the quantile function.

7. The engineering settlement deformation monitoring analysis system and method based on digital twinning according to claim 6, characterized in that: S4 includes S41 and S42; S41, compare the maximum value Shigh in the obtained confidence interval [Slow, Shigh] with the preset allowable settlement threshold Sall; If the maximum value Shigh in the confidence interval [Slow, Shigh] is less than or equal to the allowable settlement threshold Sall, no warning is triggered, and no control suggestion needs to be generated; If the maximum value Shigh in the confidence interval [Slow, Shigh] is greater than the allowable settlement threshold Sall, a warning is triggered and a control suggestion is generated.

8. The engineering settlement deformation monitoring analysis system and method based on digital twinning according to claim 7, characterized in that: S42, reacquire new detection data Dnew, and combine the current posterior parameter distribution θpost to use the physical correction consolidation equation and the Bayesian regularization objective function again, update the posterior parameter distribution based on the new detection data Dnew, and obtain the optimized posterior parameter distribution θnew; Repeat steps S3 and S4 based on the optimized posterior parameter distribution θnew until the KL divergence of the optimized posterior parameter distribution θnew and the parameter distribution of the target scene soil layer parameter feature set ΦTar is less than the preset divergence threshold; Based on the obtained large number of optimized posterior parameter distributions θnew and the obtained new detection data Dnew, the physical model is trained, and the contribution weight coefficients α, β and γ in the physical correction consolidation equation are dynamically allocated according to the training result each time.

9. The engineering settlement deformation monitoring analysis system and method based on digital twinning according to claim 1, characterized in that: S5 includes S51; S51, a feature vector set F is constructed, and the feature vector set F includes the following: Firstly, record the inversion parameter posterior distribution set {θopst_1, θnew_2, …, θnew_r} generated by the inversion model of the target scene in the historical iterations in chronological order, where r represents the total number of iterations, secondly, record the comparison results of the confidence interval [Slow, Shigh] calculated in each iteration and the preset allowable settlement threshold Sall, and finally, record the input target scene soil layer parameter feature set ΦTar and the source scene soil layer parameter feature set ΦSou; The feature vector set F is stored in the knowledge base and the knowledge base is updated.

10. The digital-twin-based engineering settlement deformation monitoring analysis system and method of any one of claims 1-9, wherein: It includes a cross-scene parameter transfer learning module, a physical inversion module, a settlement prediction module, a dynamic optimization feedback module, and a feature archiving module. The cross-scene parameter transfer learning module obtains the target scene soil layer parameter feature set ΦTar and the first period detection data D by field investigation and sensor detection of the target scene, extracts the source scene soil layer parameter feature set ΦSou from the historical database, combines the target scene soil layer parameter feature set ΦTar and the source scene soil layer parameter feature set ΦSou, designs a transfer learning algorithm, and obtains the transfer parameter prior distribution θpri of the target scene; According to the transfer parameter prior distribution θpri and the first period monitoring data D, the parameter inversion model is constructed by combining the physical model and the Bayesian regularization objective function, and the inversion parameter posterior distribution θpost is obtained; The settlement prediction module extracts N groups of parameter samples based on the inversion parameter posterior distribution θpost to form an inversion parameter set The hth group of settlement prediction value sets is calculated based on the physical model And the confidence interval [Slow, Shig] is counted; The dynamic optimization feedback module compares the confidence interval [Slow, Shigh] and the preset allowable settlement threshold Sall, generates control suggestions according to the comparison results, reacquires monitoring data Dnew and inputs it into the parameter inversion model, updates the inversion parameter posterior distribution θnew, and re-performs confidence interval statistics and risk assessment based on the obtained data to train the physical model; The feature archiving module records the historical parameters calculated by the inversion model of the target scene, constructs the feature vector F, archives and stores it in the database, and updates the database.

Images