A soft soil foundation protection method based on geotechnical parameter inversion and deep learning

Abstract

This invention discloses a method for soft soil foundation protection based on soil and rock mass parameter inversion and deep learning, belonging to the field of soft soil foundation engineering protection technology. This invention collects in-situ test and real-time monitoring data of soft soil foundation engineering, preprocesses it to form a basic soil and rock mass dataset; uses a CVAE-ESMDA joint inversion framework of a first model to invert the parameters of the fractured zone medium; uses an improved DenseNet architecture of a second model to invert the core parameters of the soil and rock mass; constructs a dynamic soil and rock mass parameter set based on the two types of parameters, inputs it into a third model, and outputs multi-objective real-time prediction results for the soft soil foundation through an improved PI-DeepONet architecture embedding consolidation-deformation physical constraints; based on the multi-objective real-time prediction results of the soft soil foundation, combines the prediction results with the dynamic parameter set, completes regional division and hierarchical definition, and designs differentiated composite protection schemes by region and level after risk assessment. This invention achieves accurate and efficient protection of soft soil foundations through the collaborative work of three models, and has broad application value.

Description

Technical Field

[0001] This invention relates to the field of soft soil foundation engineering protection technology, specifically to a soft soil foundation protection method based on soil and rock mass parameter inversion and deep learning. Background Technology

[0002] As my country's infrastructure construction continues to expand into coastal and riverside areas with soft soil distribution, the scale of soft soil foundation engineering is constantly increasing, and the requirements for its engineering safety and stability are also continuously rising. Soft soil is characterized by high porosity, high compressibility, poor stability, uneven permeability, and a tendency to develop fissures. It is highly susceptible to engineering problems such as excessive settlement, uneven settlement, and soil instability during construction and operation, seriously affecting the service life and safety performance of infrastructure. Therefore, the engineering protection of soft soil foundations has always been a key focus and challenge in the field of civil engineering. Currently, there are several technological achievements in related fields dedicated to solving core problems such as soft soil foundation parameter inversion, working condition monitoring, and protection and remediation. However, the application of combining cutting-edge artificial intelligence technology for optimizing soft soil foundation protection parameters remains undeveloped.

[0003] Chinese patent (publication number CN222065487U) discloses a deep foundation pit support structure suitable for soft soil foundations. This method designs deep foundation pit support components with specific structural forms, and achieves soft soil foundation protection through the bearing and constraint effects of the components. Its effectiveness can only solve the support stability problem during the construction stage of deep foundation pits. The drawback is that it only focuses on the design of the foundation pit support structure, does not involve the inversion of soil and rock parameters and the prediction of working conditions, and cannot dynamically adjust the protection strategy according to the actual foundation parameters and working conditions. The protection is not targeted enough and has poor adaptability.

[0004] Most existing soft soil foundation protection methods are similar to CN222065487U, which do not integrate deep learning and regional layered protection concepts. They have low protection efficiency and accuracy, and cannot design differentiated solutions according to the differences in different regions and soil layers, which can easily lead to waste of protection resources or inadequate protection. Furthermore, they lack the ability to perform multi-model collaborative inversion and real-time prediction, making it difficult to adapt to the dynamic geological changes of soft soil foundations and cope with real-time fluctuations in parameters during construction.

[0005] In summary, existing technologies are generally insufficient to address the complex conditions of soft soil foundations with well-developed fissures and multi-field coupling, failing to meet the engineering requirements for precise, real-time, and targeted protection. Therefore, there is an urgent need for an integrated method that combines multi-model collaborative inversion, dynamic parameter construction, and regional layered protection to fill the gaps in existing technologies and achieve precise, efficient, and closed-loop protection for soft soil foundations. Summary of the Invention

[0006] To address the aforementioned technical problems, this application discloses a method for soft soil foundation protection based on soil and rock mass parameter inversion and deep learning, specifically including:

[0007] The soil and rock parameters of soft soil foundation engineering are collected and preprocessed to obtain the basic soil and rock dataset.

[0008] Based on the basic dataset of soil and rock mass, the parameters of the fracture zone medium are inverted and output through the first model, which is constructed through the CVAE-ESMDA joint inversion framework; the core parameters of soil and rock mass are inverted and output through the second model, which is the THT-NN model with the improved DenseNet architecture.

[0009] Based on the parameters of the fractured zone medium and the core parameters of the soil and rock mass, a dynamic set of soil and rock mass parameters is constructed.

[0010] Based on a dynamic set of soil and rock parameters, a third model is used to output multi-objective real-time prediction results for soft soil foundations. The third model is constructed using a branch architecture of an improved PI-DeepONet model.

[0011] Based on the multi-objective real-time prediction results of soft soil foundation, a composite protection scheme is generated in different regions and layers.

[0012] Preferably, the soil and rock mass basic dataset specifically includes an in-situ test parameter subset and a real-time monitoring parameter subset. The in-situ test parameter subset includes cone tip resistance and sidewall friction data obtained from static cone penetration tests. The real-time monitoring parameter subset includes time-series monitoring data of soft soil foundation stratified settlement, soil strain, and pore water pressure obtained from distributed fiber optic sensing monitoring.

[0013] Preferably, the first model specifically involves: importing the preprocessed soil and rock mass basic dataset, inputting the high-dimensional observation data in the soil and rock mass basic dataset into the CVAE algorithm, and outputting a low-dimensional latent feature vector through the encoder and decoder of the CVAE algorithm;

[0014] Based on the initial inversion values, the true values ​​of the medium parameters in the fracture zone are approximated through the ESMDA algorithm iteratively.

[0015] Preferably, the process of outputting a low-dimensional latent feature vector through the encoder and decoder of the CVAE algorithm is as follows: Feature extraction and nonlinear dimensionality reduction are performed on the high-dimensional data to obtain the low-dimensional latent feature vector, as shown in the formula:

[0016]

[0017] in, This represents the posterior probability distribution of the encoder output. This is high-dimensional observation data from the basic dataset of soil and rock mass. For low-dimensional latent feature vectors, , These are the mean and standard deviation of the posterior distribution, respectively, derived from the encoder parameters. Decide, It is the identity matrix. It follows a normal distribution;

[0018] Based on the low-dimensional latent feature vector, the initial inversion values ​​of the fracture zone medium parameters are output through reconstruction using a CVAE decoder, as shown in the formula:

[0019]

[0020] in, Let be the likelihood probability distribution output by the decoder. , Let be the mean and standard deviation of the likelihood distribution, respectively. This represents the initial inversion values ​​of the fractured region medium parameters output by the decoder.

[0021] Preferably, the step of iteratively approximating the true values ​​of the fractured zone medium parameters using the ESMDA algorithm specifically involves: using the initial inversion values ​​of the fractured zone medium parameters output by the CVAE model as the initial input to the ESMDA model, constructing an observation operator, and establishing a mapping relationship between the fractured zone medium parameters and the observed values ​​of the basic soil and rock dataset, as shown in the formula:

[0022]

[0023] in, These are the predicted observation values ​​corresponding to the medium parameters in the fractured region. The observation operator is derived from the control equations for seepage and deformation in soil and rock masses. This is the observation error vector;

[0024] Based on the observation operator, the ESMDA model is used for multi-set iterative updates. Through multiple set smoothing iterations, the inverted values ​​of fracture zone medium parameters are continuously corrected, the inversion error is reduced, and the true values ​​of fracture zone medium parameters are gradually approximated. The formula is as follows:

[0025]

[0026] in, For the number of iterations, The sample number in the set. For the first The second iteration Inversion values ​​of a set of samples, For the first The covariance matrix of the variables and observations is retrieved in the next iteration. The covariance matrix of the observations, For the first The expansion coefficient of the next iteration. These are the actual observations in the soil and rock mass foundation dataset. For the first The second iteration Predicted observations for a set of samples, Let be the covariance matrix of the observation error. Let be a random vector derived from a standard normal distribution.

[0027] Preferably, the second model specifically involves: inputting the preprocessed soil and rock mass basic dataset into the improved DenseNet architecture; extracting data features through the improved DenseNet architecture; combining THM multi-field coupling constraints; and outputting preliminary soil and rock mass core parameter inversion values; and correcting the deviation of the preliminary inversion output core parameters through a linear correction method to obtain the target soil and rock mass core parameters.

[0028] Preferably, the improved DenseNet architecture specifically involves embedding a THM multi-field coupling loss term into the traditional DenseNet architecture, with the following formula:

[0029]

[0030] in, The loss function is a multi-field coupling function of thermo-hydraulic-mechanical systems. , , , These are the single-parameter losses for permeability, compression modulus, Poisson's ratio, and shear strength, respectively. - These are the weighting coefficients;

[0031] After pruning redundant convolutional layers in the architecture, the geotechnical basic dataset is divided into training, validation, and test sets. An adaptive learning rate optimization algorithm is used to train the model, and the hyperparameters are adjusted using the validation set. The formula is as follows:

[0032]

[0033] in, For the first The learning rate for the next iteration. The initial learning rate, The attenuation coefficient is... To minimize the learning rate, This represents the number of iterations.

[0034] The accuracy of the trained model is verified using a test set. The relative error of the inversion parameters is calculated to determine whether the model's inversion accuracy meets engineering requirements. If not, the model is returned to be retrained after adjusting the hyperparameters.

[0035] Preferably, obtaining the core parameters of the target soil and rock mass specifically involves: inputting the complete basic dataset of the soil and rock mass into the tested THT-NN model; the model extracts data features through the improved DenseNet architecture; and, combined with the THM multi-field coupling constraints, outputs preliminary inversion values ​​of the core parameters of the soil and rock mass, as shown in the formula:

[0036]

[0037] in, The core parameters of the soil and rock mass are output from the initial inversion. For the mapping function of the THT-NN model, , This represents the optimal weight matrix and bias vector after model training. Input the basic dataset of soil and rock mass;

[0038] By using a linear correction method to correct the deviation of the core parameters in the initial inversion output, systematic errors generated during model training are eliminated, and the core parameters of the target soil and rock mass are obtained. The formula is as follows:

[0039]

[0040] in, For the core parameters of the target soil and rock mass, , These are the correction coefficients obtained by fitting the test set data. After correction, the accuracy of the core parameter inversion is further improved, and the relative error is controlled within 3%, which meets the requirements of engineering inversion.

[0041] Preferably, the third model specifically comprises: constructing a branch structure of a PI-DeepONet model embedding physical constraints on the consolidation-deformation of soft soil foundations, wherein the physical constraint embedding formula is:

[0042]

[0043] in, This refers to the physical constraints of consolidation and deformation of soft soil foundations derived from Terzaghi's consolidation theory. This represents the foundation settlement vector. For time variables, This is the consolidation coefficient for soft soil foundations. For the Laplace operator;

[0044] The PI-DeepONet model was trained and its parameters optimized using historical engineering datasets and monitoring data samples from the current project.

[0045] The validated set of total input parameters is input into the trained PI-DeepONet model. The model extracts parameter features through branch networks, realizes temporal evolution mapping through the backbone network, and outputs multi-objective prediction results for soft soil foundation by combining physical constraint correction.

[0046] Preferably, the layered and segmented composite protection scheme is specifically as follows:

[0047] Based on the output of the third model In addition to dynamic soil and rock mass parameter sets, combined with engineering geological survey data, the regional division and soil layer classification of soft soil foundations are completed, and the risk levels of each region and layer are clarified.

[0048] For each zone and soil layer after classification, a comprehensive risk assessment is conducted based on the uneven settlement difference, and quantified control indicators are determined using the following formula:

[0049]

[0050] in, For the first Class region, first The comprehensive risk index of the soil layers , , These are the weighting coefficients, , , These represent the stability coefficient, settlement, and mean value of uneven settlement difference for the corresponding region and level. , , These are, respectively, the allowable settlement, the predicted maximum settlement, and the allowable uneven settlement difference. for The predicted maximum value;

[0051] Based on the risk assessment index and combined with dynamic soil and rock parameters, differentiated composite protection schemes are designed for different regions and layers to achieve one policy for each region and one measure for each layer. In high-risk areas and deep soil layers, grouting and solid waste composite reinforcement parameters are optimized; in medium-risk areas and middle soil layers, solid waste improvement and drainage composite parameters are optimized; and in low-risk areas and shallow soil layers, surface protection and light improvement parameters are optimized.

[0052] Compared with the prior art, the technical solution of this application has the following technical effects:

[0053] This invention achieves accurate inversion of medium parameters in fractured zones through the collaborative inversion of the first and second models. It also improves the accuracy of core parameter inversion of soil and rock mass by embedding THM multi-field coupling constraints and redundancy pruning in the improved DenseNet architecture, thus adapting to the working conditions of complex soft soil foundations with multi-field coupling and fracture development.

[0054] This invention constructs a dynamic set of soil and rock parameters based on two types of inversion parameters, combines a third model to achieve real-time multi-objective prediction of soft soil foundations, and then generates composite protection schemes by region and layer based on the prediction results, thus opening up the entire process of parameter inversion, working condition prediction and protection implementation.

[0055] Based on prediction results and dynamic parameter sets, this invention completes the regional risk classification and soil layer level definition. For high, medium and low risk areas and different soil layers, it designs differentiated composite protection schemes to achieve one policy for one area and one measure for one layer, solving the problems of existing patent protection schemes being single, poorly adaptable, wasteful of resources or inadequate protection.

[0056] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0057] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0059] Based on the description of the figures and their corresponding technical content in the document, the titles of the figures are as follows:

[0060] Figure 1 This is a flowchart of a soft soil foundation protection method based on soil and rock mass parameter inversion and deep learning.

[0061] Figure 2 This is a general architecture diagram of a soft soil foundation protection method based on soil and rock mass parameter inversion and deep learning;

[0062] Figure 3 This is an architecture diagram of the first model in this application;

[0063] Figure 4 This is an architecture diagram of the second model in this application;

[0064] Figure 5 This is an architecture diagram of the third model in this application;

[0065] Figure 6 This is a data diagram of the first model iteration process in the embodiments of this application;

[0066] Figure 7 This is a data diagram of the second model iteration process in the embodiments of this application;

[0067] Figure 8 This is a data comparison chart of the third model in the embodiments of this application during 60 days of construction. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0069] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0070] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0071] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0072] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0073] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0074] Example 1 describes a soft soil foundation protection method based on soil and rock mass parameter inversion and deep learning. Figure 1 As shown, it specifically includes:

[0075] The soil and rock parameters of soft soil foundation engineering are collected and preprocessed to obtain the basic soil and rock dataset.

[0076] Based on the basic dataset of soil and rock mass, the parameters of the fracture zone medium are inverted and output through the first model, which is constructed through the CVAE-ESMDA joint inversion framework; the core parameters of soil and rock mass are inverted and output through the second model, which is the THT-NN model with the improved DenseNet architecture.

[0077] Based on the parameters of the fractured zone medium and the core parameters of the soil and rock mass, a dynamic set of soil and rock mass parameters is constructed.

[0078] Based on a dynamic set of soil and rock parameters, a third model is used to output multi-objective real-time prediction results for soft soil foundations. The third model is constructed using a branch architecture of an improved PI-DeepONet model.

[0079] Based on the multi-objective real-time prediction results of soft soil foundation, a composite protection scheme is generated in different regions and layers.

[0080] Furthermore, the basic dataset for soil and rock mass specifically includes a subset of in-situ test parameters and a subset of real-time monitoring parameters. The subset of in-situ test parameters includes data on cone tip resistance and sidewall friction obtained from static cone penetration tests. The subset of real-time monitoring parameters includes time-series monitoring data on layered settlement, soil strain, and pore water pressure of soft soil foundation obtained from distributed fiber optic sensing.

[0081] Furthermore, such as Figure 3 The first model architecture diagram shown is as follows: import the preprocessed soil and rock mass basic dataset, input the high-dimensional observation data in the soil and rock mass basic dataset into the CVAE algorithm, and output the low-dimensional latent feature vector through the encoder and decoder of the CVAE algorithm.

[0082] Based on the initial inversion values, an observation operator is constructed using the ESMDA algorithm, and multiple sets are iteratively updated to gradually approximate the true values ​​of the medium parameters in the fracture zone.

[0083] Furthermore, the process of outputting a low-dimensional latent feature vector through the encoder and decoder of the CVAE algorithm is as follows: Feature extraction and nonlinear dimensionality reduction are performed on the high-dimensional data to obtain the low-dimensional latent feature vector, as shown in the formula:

[0084]

[0085] in, This represents the posterior probability distribution of the encoder output. This is high-dimensional observation data from the basic dataset of soil and rock mass. For low-dimensional latent feature vectors, , These are the mean and standard deviation of the posterior distribution, respectively, derived from the encoder parameters. Decide, It is the identity matrix. It follows a normal distribution;

[0086] Based on the low-dimensional latent feature vector, the initial inversion values ​​of the fracture zone medium parameters are output through reconstruction using a CVAE decoder, as shown in the formula:

[0087]

[0088] in, Let be the likelihood probability distribution output by the decoder. , Let be the mean and standard deviation of the likelihood distribution, respectively. These are the initial inversion values ​​of the fractured region medium parameters output by the decoder;

[0089] Furthermore, the following loss function is used during the training of the CVAE model:

[0090]

[0091] in, The total loss of the CVAE model is... To reconstruct the loss and ensure that the decoded output accurately restores the input features, for Divergence, constraining the posterior distribution to approximate the prior distribution .

[0092] Furthermore, the true values ​​of the fractured zone medium parameters are iteratively approximated using the ESMDA algorithm. Specifically, the initial inversion values ​​of the fractured zone medium parameters output by the CVAE model are used as the initial input to the ESMDA model. An observation operator is constructed to establish a mapping relationship between the fractured zone medium parameters and the observed values ​​of the basic soil and rock dataset. The formula is as follows:

[0093]

[0094] in, These are the predicted observation values ​​corresponding to the medium parameters in the fractured region. The observation operator is derived from the control equations for seepage and deformation in soil and rock masses. This is the observation error vector;

[0095] Based on the observation operator, the ESMDA model is used for multi-set iterative updates. Through multiple set smoothing iterations, the inverted values ​​of fracture zone medium parameters are continuously corrected, the inversion error is reduced, and the true values ​​of fracture zone medium parameters are gradually approximated. The formula is as follows:

[0096]

[0097] in, For the number of iterations, The sample number in the set. For the first The second iteration Inversion values ​​of a set of samples, For the first The covariance matrix of the variables and observations is retrieved in the next iteration. The covariance matrix of the observations, For the first The expansion coefficient of the next iteration. These are the actual observations in the soil and rock mass foundation dataset. For the first The second iteration Predicted observations for a set of samples, Let be the covariance matrix of the observation error. Let be a random vector derived from a standard normal distribution.

[0098] Furthermore, such as Figure 4The second model architecture diagram shown is as follows: The preprocessed soil and rock mass basic dataset is input into the improved DenseNet architecture. Data features are extracted through the improved DenseNet architecture, and combined with THM multi-field coupling constraints, preliminary soil and rock mass core parameter inversion values ​​are output. The core parameters output by the preliminary inversion are corrected for deviation through the linear correction method to obtain the target soil and rock mass core parameters.

[0099] Furthermore, an inversion convergence threshold is set to determine whether the iterative inversion results of the ESMDA model have converged (convergence is determined when the mean square error of the inversion parameters in two adjacent iterations is less than the convergence threshold); after convergence, the final inversion values ​​of the fractured zone medium parameters are output, completing the inversion process of the first model. The formula is:

[0100]

[0101] in, The root mean square error of the inversion parameters between two consecutive iterations. The set contains the number of samples. When the RMSE is less than the preset inversion error threshold, the iteration stops and the inversion result is output.

[0102] Furthermore, the improved DenseNet architecture specifically involves embedding a THM (Thunder Memory Management) multi-field coupling loss term into the traditional DenseNet architecture, with the following formula:

[0103]

[0104] in, The loss function is a multi-field coupling function of thermo-hydraulic-mechanical systems. , , , These are the single-parameter losses for permeability, compression modulus, Poisson's ratio, and shear strength, respectively. - These are the weighting coefficients;

[0105] The redundant convolutional layers in the architecture are pruned using the following formula:

[0106]

[0107] in, For the first The pruning probability of a convolutional layer. Let be the weight matrix of the l-th convolutional layer. For the set of all convolutional layers of the model, when When the number of parameters is less than the preset pruning threshold, the redundant convolutional layer is deleted to reduce the number of model parameters and improve training efficiency.

[0108] The basic geotechnical dataset was divided into training, validation, and test sets in a 7:2:1 ratio. An adaptive learning rate optimization algorithm was used to train the model, and the hyperparameters were adjusted using the validation set. The formula is as follows:

[0109]

[0110] in, For the first The learning rate for the next iteration. The initial learning rate, The attenuation coefficient is... To minimize the learning rate, This represents the number of iterations.

[0111] The accuracy of the trained model is verified using a test set. The relative error of the inversion parameters is calculated to determine whether the model's inversion accuracy meets engineering requirements. If not, the model is returned to be retrained after adjusting the hyperparameters.

[0112] Furthermore, the core parameters of the target soil and rock mass are obtained. Specifically, the complete basic dataset of the soil and rock mass is input into the tested THT-NN model. The model extracts data features through the improved DenseNet architecture, combines THM multi-field coupling constraints, and outputs preliminary inversion values ​​of the core parameters of the soil and rock mass. The formula is as follows:

[0113]

[0114] in, The core parameters of the soil and rock mass are output from the initial inversion. For the mapping function of the THT-NN model, , This represents the optimal weight matrix and bias vector after model training. Input the basic dataset of soil and rock mass;

[0115] By using a linear correction method to correct the deviation of the core parameters in the initial inversion output, systematic errors generated during model training are eliminated, and the core parameters of the target soil and rock mass are obtained. The formula is as follows:

[0116]

[0117] in, For the core parameters of the target soil and rock mass, , These are the correction coefficients obtained by fitting the test set data. After correction, the accuracy of the core parameter inversion is further improved, and the relative error is controlled within 3%, which meets the requirements of engineering inversion.

[0118] Furthermore, the structural parameters of the second model are as follows: based on the DenseNet architecture, a THT-NN is constructed with 3 to 5 densely connected blocks. Each densely connected block contains 6 to 10 convolutional layers with a kernel size of 3×3. The THM (thermal-hydraulic-mechanical) multi-field coupling loss function is embedded. The initial learning rate during training is 0.001 to 0.01, the number of iterations is set to 500 to 1000, and the weights of the THM loss function are set to 0.2 and 0.7, respectively.

[0119] Furthermore, two types of core input parameters are imported, and the parameters are standardized and unified to eliminate differences in dimensions and deviations in parameter ranges, ensuring that the two types of parameters can be calculated together, laying the foundation for subsequent calibration operations;

[0120] Bayesian theory is used to synergistically correct the two types of standardized parameters, eliminating coupling interference between parameters (such as the mutual influence between the permeability coefficient of the fractured zone and the overall permeability of the soil and rock mass), correcting inversion bias, and obtaining a corrected unified parameter set, as shown in the formula:

[0121]

[0122]

[0123] in, For the unified parameter set after collaborative correction, For the posterior probability distribution, Let be the likelihood probability. For prior probability, For evidence, This is the optimal parameter set after Bayesian collaborative correction.

[0124] Furthermore, based on real-time monitoring data from distributed fiber optic sensing, a dynamic update mechanism is established to iteratively optimize the corrected parameter set on an hourly basis, ensuring that the parameter set matches the actual working conditions of the soft soil foundation in real time. The formula is as follows:

[0125]

[0126] in, For the dynamic parameter set of the next update cycle, At the current update time, To update the step size, The gradient of the loss function with respect to the parameter set. This represents the real-time monitoring data of distributed fiber optic sensing at time t. To update the loss function.

[0127] Furthermore, such as Figure 5The third model architecture diagram shown is as follows: the third model is a branch structure that constructs a PI-DeepONet model with embedded physical constraints on the consolidation and deformation of soft soil foundations. The physical constraint embedding formula is as follows:

[0128]

[0129] in, This refers to the physical constraints of consolidation and deformation of soft soil foundations derived from Terzaghi's consolidation theory. This represents the foundation settlement vector. For time variables, This is the consolidation coefficient for soft soil foundations. For the Laplace operator;

[0130] The PI-DeepONet model was trained and its parameters optimized using historical engineering datasets and monitoring data samples from the current project.

[0131] The validated set of total input parameters is input into the trained PI-DeepONet model. The model extracts parameter features through branch networks, realizes temporal evolution mapping through the backbone network, and outputs multi-objective prediction results for soft soil foundation by combining physical constraint correction.

[0132] Furthermore, the accuracy of the multi-target prediction results is verified by real-time monitoring data (the latest set of monitoring values) using distributed optical fiber sensing. If the error exceeds the preset threshold, the prediction results are corrected using an error correction formula to ensure that the prediction accuracy meets engineering requirements.

[0133] Furthermore, the third model has a dual-branch architecture combining a branch network and a backbone network, adopts physical constraints based on Terzaghi consolidation theory, and uses the Adam optimization algorithm during training. The weight coefficients are set to 0.2~0.4 respectively, and the error threshold is set to a settlement amount ≤2mm.

[0134] Furthermore, a composite protection scheme is generated by region and layer, specifically as follows:

[0135] Based on the output of the third model In addition to dynamic soil and rock mass parameter sets, combined with engineering geological survey data, the regional division and soil layer classification of soft soil foundations are completed, and the risk levels of each region and layer are clarified.

[0136] For each zone and soil layer after classification, a comprehensive risk assessment is conducted based on the uneven settlement difference, and quantified control indicators are determined using the following formula:

[0137]

[0138] in, For the first Class region, first The comprehensive risk index of the soil layers , , These are the weighting coefficients, taken as 0.4, 0.3, and 0.3 respectively. , , These represent the stability coefficient, settlement, and mean value of uneven settlement difference for the corresponding region and level. , , These are, respectively, the allowable settlement, the predicted maximum settlement, and the allowable uneven settlement difference. for The predicted maximum value;

[0139] Based on the risk assessment index and combined with dynamic soil and rock parameters, differentiated composite protection schemes are designed for different regions and layers, achieving a "one policy per region, one measure per layer" approach. In high-risk areas and deep soil layers, grouting and solid-waste composite reinforcement parameters are optimized, using the following formula:

[0140]

[0141] in, For grouting pressure, Based on the grouting pressure, For industrial solid waste mixing dosage, To achieve the target permeability value after reinforcement, Original permeability;

[0142] In medium-risk areas, the solid waste improvement and drainage composite parameter optimization for the middle soil layer are calculated using the following formula:

[0143]

[0144] For drainage rate, Based on the drainage rate, To achieve the target value of the improved compression modulus, The original compression modulus;

[0145] For low-risk areas—shallow soil layers—parameter optimization for surface protection and minor soil improvement is achieved using the following formula:

[0146]

[0147] in, This refers to the thickness of the surface protective layer.

[0148] This embodiment details a soft soil foundation protection method based on soil and rock mass parameter inversion and deep learning. It collects in-situ test and real-time monitoring data from soft soil foundation engineering projects, preprocesses this data to form a basic soil and rock mass dataset, and then inverts the parameters of the fractured zone using a CVAE-ESMDA joint inversion architecture of the first model. It then inverts the core parameters of the soil and rock mass using an improved DenseNet architecture of the second model. Based on these two types of parameters, a dynamic soil and rock mass parameter set is constructed and input into a third model. Through an improved PI-DeepONet architecture embedding consolidation-deformation physical constraints, it outputs multi-objective real-time prediction results for the soft soil foundation. Based on these multi-objective real-time prediction results, and combining the prediction results with the dynamic parameter set, regional division and hierarchical definition are completed. After risk assessment, differentiated composite protection schemes are designed for different regions and levels.

[0149] Example 2, based on Example 1, details an experiment using this method to protect soft soil foundations in a coastal soft soil municipal road project. The soft soil layer in the project area is 12-15m thick, mainly composed of silty clay, with a void ratio of 1.8-2.2 and a compression modulus of 3.0-4.5MPa. Locally, there are areas with well-developed cracks, making excessive and uneven settlement prone to occur during construction. The specific implementation process is as follows:

[0150] The 1.2km long and 24m wide main road was divided into three geological zones: Zone I, a low-fracture development zone; Zone II, a medium-fracture development zone; and Zone III, a high-fracture development zone. Distributed fiber optic sensors were used to monitor stratified settlement, soil strain, and pore water pressure, and in-situ static cone penetration tests were conducted at a frequency of once per hour. The collected data were processed to remove outliers and normalize, resulting in a basic geotechnical dataset containing 86 sets of in-situ test data and 12,438 real-time monitoring time-series data.

[0151] The preset engineering accuracy requirements are: relative error of soil and rock parameters inversion ≤5%, settlement prediction error ≤2mm, and soil stability coefficient ≥1.2.

[0152] The preprocessed soil and rock mass dataset was input into the first model, and then subjected to CVAE feature extraction and ESMDA multi-set iterative inversion (26 iterations, convergence threshold). The final output includes the medium parameters for three fractured zones. The inversion results are compared with the field borehole verification data in Table 1 below:

[0153] Table 1 Comparison of the inversion results of the first model with the field drilling verification data

[0154] According to Table 1 and Figure 6As shown in the data graph of the first model iteration process, the inversion data of all regions are highly fitted with the measured verification data, and the relative inversion error is less than 5%, which verifies the inversion accuracy of the first model of this method.

[0155] The basic dataset of soil and rock mass was input into the improved DenseNet architecture THT-NN model. After THM multi-field coupling training, redundant layer pruning, and linear bias correction, the core parameters of soil and rock mass were output. The inversion results are compared with the laboratory test data as shown in Table 2 below:

[0156] Table 2 compares the inversion results of the second model with the laboratory experimental verification data.

[0157] According to Table 2 and Figure 7 As shown in the data graph of the second model iteration process, the inversion data of all regions are highly fitted with the measured verification data, and the relative error of the core parameter inversion is less than 3%, which verifies the inversion accuracy of the second model of this method.

[0158] By integrating the inversion results of the first and second models, and through Bayesian collaborative correction and hourly dynamic iterative updates, a dynamic set of soil and rock parameters was constructed. The parameter values ​​were verified to be within the reasonable range for engineering purposes through physical and mechanical property constraints.

[0159] The dynamic soil and rock mass parameter set plus key engineering parameters (solid waste content, grouting pressure, construction speed) are input into a PI-DeepONet model embedded with Terzaghi consolidation-deformation physical constraints. The model is then trained using the Adam optimization algorithm with a loss convergence threshold. (Weight coefficient β=0.3), feature extraction and temporal mapping, outputting prediction results for three major objectives of soft soil foundation: total settlement, uneven settlement difference, and soil stability coefficient. The predicted and measured values ​​within 60 days of construction are compared using real-time monitoring data, as shown in Table 3 below.

[0160] Table 3. Comparison of predicted and measured average values ​​of the third model over 60 days of construction.

[0161] According to Table 3 and Figure 8 The data comparison chart of the third model over 60 days of construction shows that the average absolute error of the prediction method is ≤2mm, which meets the engineering accuracy requirements; among them, Zone III is a high-risk area with a stability coefficient slightly lower than 1.2, and a protection plan needs to be designed.

[0162] Based on the prediction results of the third model and combined with engineering geological survey data, soft soil foundations are divided into three risk levels: high (Zone III), medium (Zone II), and low (Zone I). They are also divided into three soil layers based on depth: shallow (0m~3m), medium (3m~8m), and deep (8m~15m). A comprehensive risk assessment (risk index weighting) is then conducted. , , Differentiated composite protection schemes are designed for different regions and levels. The core design parameters and protection objectives are as follows:

[0163] Referring to the zonal and stratified composite protection scheme described in Example 1, in the high-risk zone (Zone III), grouting and solid waste composite reinforcement are carried out in the deep and middle soil layers. The grouting pressure is 2.1 MPa (foundation pressure 1.5 MPa × 1.4), the industrial solid waste content is 24% (0.15 + 0.1 × 0.9), and the target permeability value after reinforcement is 1.79 × 1.4 MPa. m / s (original value × 0.2);

[0164] In the medium-risk area (zone II), solid waste improvement and drainage composite treatment are carried out in the middle and shallow soil layers. The industrial solid waste content is 16.4% (0.10 + 0.08 × 0.8), the drainage rate is 0.9 m³ / d (basic rate 0.5 × 1.8), and the target value of the improved compression modulus is 5760 kPa (original value × 1.6).

[0165] In the low-risk area (Zone I), surface protection and slight improvement are carried out on the shallow soil layer, with an industrial solid waste content of 9% (0.05 + 0.05 × 0.8) and a surface protection layer thickness of 0.46m (0.3 + 0.2 × 0.8).

[0166] After the above-mentioned regional and layered composite protection scheme was implemented in this project, real-time monitoring was conducted on the soft soil foundation at 90 days and 120 days after construction. The measured data of the protection effect are shown in Table 4 below:

[0167] Table 4. Measured data on the protective effect in each region

[0168] As shown in Table 4, the settlement in all areas remained stable for 30 and 60 days after the implementation of the layered composite protection scheme. The soil stability coefficient was ≥1.25. There were no engineering problems such as excessive settlement, uneven settlement, or soil instability. The stability and safety of the soft soil foundation met the requirements for the construction and operation of municipal road engineering.

[0169] This embodiment details an experiment demonstrating the application of this method for soft soil foundation protection in coastal soft soil municipal road engineering. Through the collaborative use of three models—CVAE-ESMDA, the improved DenseNet-THT-NN, and the improved PI-DeepONet—accurate inversion of soil and rock parameters and real-time multi-objective prediction of soft soil foundations are achieved. The constructed dynamic soil and rock parameter set can adapt to changes in engineering conditions in real time. Finally, the composite protection scheme generated by region and layer achieves precise protection with a tailored approach for each region and a specific measure for each layer. The application results of this embodiment fully demonstrate the accuracy and effectiveness of this method, solving the engineering protection challenges caused by the development of fissures, high compressibility, and poor stability of soft soil foundations. It has good engineering practicality and promotional value.

[0170] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. A method for soft soil foundation protection based on soil and rock mass parameter inversion and deep learning, characterized in that, include: The soil and rock parameters of soft soil foundation engineering are collected and preprocessed to obtain the basic soil and rock dataset. Based on the basic dataset of soil and rock mass, the parameters of the fracture zone medium are inverted and output through the first model, which is constructed through the CVAE-ESMDA joint inversion framework; the core parameters of soil and rock mass are inverted and output through the second model, which is the THT-NN model with the improved DenseNet architecture. Based on the parameters of the fractured zone medium and the core parameters of the soil and rock mass, a dynamic set of soil and rock mass parameters is constructed. Based on a dynamic set of soil and rock parameters, a third model is used to output multi-objective real-time prediction results for soft soil foundations. The third model is constructed using a branch architecture of an improved PI-DeepONet model. Based on the multi-objective real-time prediction results of soft soil foundation, a composite protection scheme is generated in different regions and layers.

2. The soft soil foundation protection method based on soil and rock mass parameter inversion and deep learning according to claim 1, characterized in that, The soil and rock mass basic dataset specifically includes a subset of in-situ test parameters and a subset of real-time monitoring parameters. The subset of in-situ test parameters includes cone tip resistance and sidewall friction data obtained from static cone penetration tests. The subset of real-time monitoring parameters includes time-series monitoring data of soft soil foundation stratified settlement, soil strain, and pore water pressure obtained from distributed fiber optic sensing.

3. The soft soil foundation protection method based on soil and rock mass parameter inversion and deep learning according to claim 1, characterized in that, The first model specifically involves: importing the preprocessed basic soil and rock dataset, inputting the high-dimensional observation data from the basic soil and rock dataset into the CVAE algorithm, and outputting a low-dimensional latent feature vector through the encoder and decoder of the CVAE algorithm. Based on the initial inversion values, the true values ​​of the medium parameters in the fracture zone are approximated through the ESMDA algorithm iteratively.

4. The soft soil foundation protection method based on soil and rock mass parameter inversion and deep learning according to claim 3, characterized in that, The process of outputting a low-dimensional latent feature vector through the encoder and decoder of the CVAE algorithm is as follows: Feature extraction and nonlinear dimensionality reduction are performed on the high-dimensional data to obtain the low-dimensional latent feature vector, as shown in the formula: in, This represents the posterior probability distribution of the encoder output. This is high-dimensional observation data from the basic dataset of soil and rock mass. For low-dimensional latent feature vectors, , These are the mean and standard deviation of the posterior distribution, respectively, derived from the encoder parameters. Decide, It is the identity matrix. It follows a normal distribution; Based on the low-dimensional latent feature vector, the initial inversion values ​​of the fracture zone medium parameters are output through reconstruction using a CVAE decoder, as shown in the formula: in, Let be the likelihood probability distribution output by the decoder. , Let be the mean and standard deviation of the likelihood distribution, respectively. This represents the initial inversion values ​​of the fractured region medium parameters output by the decoder.

5. The soft soil foundation protection method based on soil and rock mass parameter inversion and deep learning according to claim 3, characterized in that, The iterative approximation of the true values ​​of the fractured zone medium parameters using the ESMDA algorithm specifically involves: using the initial inversion values ​​of the fractured zone medium parameters output by the CVAE model as the initial input to the ESMDA model; constructing an observation operator; and establishing a mapping relationship between the fractured zone medium parameters and the observed values ​​of the basic soil and rock dataset, as shown in the formula: in, These are the predicted observation values ​​corresponding to the medium parameters in the fractured region. The observation operator is derived from the control equations for seepage and deformation in soil and rock masses. This is the observation error vector; Based on the observation operator, the ESMDA model is used for multi-set iterative updates. Through multiple set smoothing iterations, the inverted values ​​of fracture zone medium parameters are continuously corrected, the inversion error is reduced, and the true values ​​of fracture zone medium parameters are gradually approximated. The formula is as follows: in, For the number of iterations, The sample number in the set. For the first The second iteration Inversion values ​​of a set of samples, For the first The covariance matrix of the variables and observations is retrieved in the next iteration. The covariance matrix of the observations, For the first The expansion coefficient of the next iteration. These are the actual observations in the soil and rock mass foundation dataset. For the first The second iteration Predicted observations for a set of samples, Let be the covariance matrix of the observation error. Let be a random vector derived from a standard normal distribution.

6. The soft soil foundation protection method based on soil and rock mass parameter inversion and deep learning according to claim 1, characterized in that, The second model specifically involves: inputting the preprocessed soil and rock mass basic dataset into the improved DenseNet architecture; extracting data features through the improved DenseNet architecture; combining THM multi-field coupling constraints to output preliminary soil and rock mass core parameter inversion values; and using a linear correction method to correct the deviation of the preliminary inversion output core parameters to obtain the target soil and rock mass core parameters.

7. The soft soil foundation protection method based on soil and rock mass parameter inversion and deep learning according to claim 6, characterized in that, The improved DenseNet architecture specifically involves embedding a THM multi-field coupling loss term into the traditional DenseNet architecture, with the following formula: in, The loss function is a multi-field coupling function of thermo-hydraulic-mechanical systems. , , , These are the single-parameter losses for permeability, compression modulus, Poisson's ratio, and shear strength, respectively. - These are the weighting coefficients; After pruning redundant convolutional layers in the architecture, the geotechnical basic dataset is divided into training, validation, and test sets. An adaptive learning rate optimization algorithm is used to train the model, and the hyperparameters are adjusted using the validation set. The formula is as follows: in, For the first The learning rate for the next iteration. The initial learning rate, The attenuation coefficient is... To minimize the learning rate, This represents the number of iterations. The accuracy of the trained model is verified using a test set. The relative error of the inversion parameters is calculated to determine whether the model's inversion accuracy meets engineering requirements. If not, the model is returned to be retrained after adjusting the hyperparameters.

8. The soft soil foundation protection method based on soil and rock mass parameter inversion and deep learning according to claim 6, characterized in that, The process of obtaining the core parameters of the target soil and rock mass is as follows: The complete basic dataset of the soil and rock mass is input into the tested THT-NN model. The model extracts data features using the improved DenseNet architecture, combines THM multi-field coupling constraints, and outputs preliminary inversion values ​​of the core parameters of the soil and rock mass. The formula is: in, The core parameters of the soil and rock mass are the initial output of the inversion. For the mapping function of the THT-NN model, , This represents the optimal weight matrix and bias vector after model training. Input the basic dataset of soil and rock mass; By using a linear correction method to correct the deviation of the core parameters in the initial inversion output, systematic errors generated during model training are eliminated, and the core parameters of the target soil and rock mass are obtained. The formula is as follows: in, For the core parameters of the target soil and rock mass, , These are the correction coefficients obtained by fitting the test set data. After correction, the accuracy of the core parameter inversion is further improved, and the relative error is controlled within 3%, which meets the requirements of engineering inversion.

9. The soft soil foundation protection method based on soil and rock mass parameter inversion and deep learning according to claim 1, characterized in that, The third model specifically involves constructing a branch structure for the PI-DeepONet model that embeds physical constraints on the consolidation and deformation of soft soil foundations. The physical constraint embedding formula is as follows: in, This refers to the physical constraints of consolidation and deformation of soft soil foundations derived from Terzaghi's consolidation theory. This represents the foundation settlement vector. For time variables, This is the consolidation coefficient for soft soil foundations. For the Laplace operator; The PI-DeepONet model was trained and its parameters optimized using historical engineering datasets and monitoring data samples from the current project. The validated set of total input parameters is input into the trained PI-DeepONet model. The model extracts parameter features through branch networks, realizes temporal evolution mapping through the backbone network, and outputs multi-objective prediction results for soft soil foundation by combining physical constraint correction.

10. The soft soil foundation protection method based on soil and rock mass parameter inversion and deep learning according to claim 1, characterized in that, The layered and regionally generated composite protection scheme is specifically as follows: Based on the output of the third model In addition to dynamic soil and rock mass parameter sets, combined with engineering geological survey data, the regional division and soil layer classification of soft soil foundations are completed, and the risk levels of each region and layer are clarified. For each zone and soil layer after classification, a comprehensive risk assessment is conducted based on the uneven settlement difference, and quantified control indicators are determined using the following formula: in, For the first Class region, first The comprehensive risk index of the soil layers , , These are the weighting coefficients, , , These represent the stability coefficient, settlement, and mean value of uneven settlement difference for the corresponding region and level. , , These are, respectively, the allowable settlement, the predicted maximum settlement, and the allowable uneven settlement difference. for The predicted maximum value; Based on the risk assessment index and combined with dynamic soil and rock parameters, differentiated composite protection schemes are designed for different regions and layers to achieve one policy for each region and one measure for each layer. In high-risk areas and deep soil layers, grouting and solid waste composite reinforcement parameters are optimized; in medium-risk areas and middle soil layers, solid waste improvement and drainage composite parameters are optimized; and in low-risk areas and shallow soil layers, surface protection and light improvement parameters are optimized.

Images