A method for predicting long-term consolidation settlement of foundations based on physical information neural networks

By constructing a long-term foundation consolidation settlement prediction method based on physical information neural networks, a loss function is defined using monitoring data and physical laws to invert the consolidation coefficient and make a positive prediction. This solves the problem of long-term foundation settlement prediction under the condition of unknown consolidation coefficient in the existing technology and achieves rapid and accurate settlement prediction.

CN118412049BActive Publication Date: 2026-11-17CHANGAN UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202410450366.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-04-15
Publication Date
2026-11-17
Estimated Expiration
2044-04-15

AI Technical Summary

Technical Problem

Existing technologies struggle to accurately predict long-term consolidation settlement of foundations using short-term and sparse excess pore water pressure monitoring data when the consolidation coefficient is unknown. Traditional methods are susceptible to human interference and rely on large-scale datasets.

Method used

A method for predicting long-term consolidation settlement of foundations based on physical information neural networks is adopted. By constructing a soil and rock parameter inversion model and a foundation excess pore water pressure forward prediction model, a loss function is defined using monitoring datasets and physical laws to invert the consolidation coefficient and make forward predictions.

Benefits of technology

It enables rapid and accurate prediction of long-term consolidation settlement of foundations using short-term sparse data under the condition of unknown consolidation coefficient, avoiding human error and dependence on large-scale datasets, and improving prediction efficiency and accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118412049B_ABST
    Figure CN118412049B_ABST
Patent Text Reader

Abstract

This invention discloses a method for predicting long-term consolidation settlement of foundations based on a physical information neural network. The method includes: acquiring a dataset, including a monitoring dataset, a residual set, and a test set; defining a physical-data loss function based on the monitoring dataset and physical laws, constructing a geotechnical parameter inversion model, training the inversion model using the residual set, and inverting the consolidation coefficient; updating the physical loss function based on the inverted consolidation coefficient, constructing a forward prediction model for excess pore water pressure of the foundation, training the forward model using the residual set, inputting the test set into the trained model, and outputting the predicted value of the excess pore water pressure of the foundation; and constructing a foundation consolidation settlement prediction model to predict long-term consolidation settlement of the foundation based on the predicted value of the excess pore water pressure. This method can accurately invert the consolidation coefficient using short-term and sparse monitoring data, and then predict the excess pore water pressure of the foundation, thereby achieving rapid and accurate prediction of long-term consolidation settlement of the foundation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of geotechnical mechanics and engineering technology, specifically relating to a method for predicting long-term consolidation settlement of foundations based on physical information neural networks. Background Technology

[0002] Consolidation refers to the process by which soil volume decreases under its own weight or external loads due to the expulsion of pore water. This process leads to foundation settlement, which in turn affects the stability of the foundation and its superstructure. A key parameter for assessing the degree of foundation consolidation is excess pore water pressure, the extent of which dissipates reveals the degree of consolidation. Therefore, accurately predicting excess pore water pressure and analyzing the patterns of foundation consolidation settlement is of great significance for guiding on-site construction practices, developing reasonable construction plans, and optimizing building design.

[0003] To develop methods for predicting excess pore water pressure in foundations, scholars both domestically and internationally have conducted extensive research. Terzaghi proposed a one-dimensional consolidation theory and established a one-dimensional consolidation model for the deformation of saturated soil layers at any time during the seepage consolidation process. Biot, by combining the effective stress principle, soil continuity, and equilibrium equations, developed a consolidation theory considering the relationship between pore pressure and soil skeleton deformation during soil consolidation. By solving the consolidation theory, excess pore water pressure in foundations can be predicted. The main methods include analytical and numerical approaches. While analytical methods offer high accuracy, they may be difficult to derive for complex engineering conditions; numerical methods, on the other hand, are widely used due to their versatility, but typically require significant computational resources and time. Furthermore, the aforementioned traditional methods are limited to the forward solution of the consolidation theory, requiring prior knowledge of the consolidation coefficient c. v However, in reality, the consolidation coefficient c v It is not directly measurable, but requires inversion using experimental data or on-site monitoring data. Currently commonly used inversion methods include logarithmic time fitting, square root time fitting, and inflection point methods. These methods can be categorized as graphical and empirical, and are easily affected by human factors, resulting in deviations between the inversion results and actual conditions. In recent years, artificial neural network technology has attracted much attention due to its advantage of effectively mitigating or even eliminating the influence of human factors. This technology achieves high-precision predictions through large-scale data training. However, the effective implementation of artificial neural network technology relies on long-term accumulated continuous observation data, requiring significant time and economic costs.

[0004] Therefore, how to invert the consolidation coefficient based on short-term and sparse excess pore water pressure monitoring data under the condition of unknown consolidation coefficient, and then positively predict the excess pore water pressure of the foundation, so as to predict the long-term consolidation settlement of the foundation, has become a key technical challenge that urgently needs to be solved. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides a method for predicting long-term consolidation settlement of foundations based on physical information neural networks.

[0006] The technical problem to be solved by this invention is achieved through the following technical solution:

[0007] This invention provides a method for predicting long-term consolidation settlement of foundations based on a physical information neural network, comprising:

[0008] Obtain the dataset, which includes a monitoring dataset, a residual set, and a test set;

[0009] Based on the monitoring dataset and physical laws, a physical-data loss function is defined, a soil and rock parameter inversion model is constructed, and the soil and rock parameter inversion model is trained using the residual set to invert the consolidation coefficient.

[0010] The physical loss function that conforms to physical law constraints is updated based on the inverted consolidation coefficient, and a forward prediction model of foundation excess pore water pressure is constructed. The foundation excess pore water pressure forward prediction model is trained using the residual set. The test set is input into the trained foundation excess pore water pressure forward prediction model, and the predicted value of foundation excess pore water pressure is output.

[0011] A foundation consolidation settlement prediction model is constructed, and the long-term consolidation settlement of the foundation is predicted based on the predicted value of the excess pore water pressure of the foundation.

[0012] Optionally, the monitoring dataset includes the burial depth of several pore pressure sensors, as well as the monitoring data of excess pore water pressure collected by the pore pressure sensors and the acquisition time. The residual set includes several residual points in a preset computational domain, and the residual points include the foundation depth and time. The test set includes the foundation depth and time that need to be predicted.

[0013] Optionally, the preset calculation domain includes a consolidation equation calculation domain, a boundary condition calculation domain, an initial condition calculation domain, and a monitoring data calculation domain.

[0014] Optionally, a physical-data loss function is defined based on the monitoring dataset and physical laws to construct a soil and rock parameter inversion model. The soil and rock parameter inversion model is trained using the residual set to invert the consolidation coefficient, including:

[0015] Based on the monitoring dataset and physical laws, a physical-data loss function is defined, and a soil and rock parameter inversion model is constructed. The first neural network of the soil and rock parameter inversion model includes two first input layers and one first output layer.

[0016] The foundation depth and time in the residual set are input into the two first input layers of the first neural network to estimate the excess pore water pressure during the forward propagation process, thereby obtaining the first estimated value of the excess pore water pressure.

[0017] The physical-data loss function is calculated using the chain rule based on the first estimated value of the excess pore water pressure, and the residual of the physical-data loss function is obtained.

[0018] The weights and biases of the first neural network are updated in reverse using the chain rule based on the residuals of the physical-data loss function, and an estimated value of the consolidation coefficient is generated. When the residuals of the physical-data loss function are less than a predetermined first standard, the training of the geotechnical parameter inversion model is completed, and the final consolidation coefficient is obtained directly by inversion.

[0019] Optionally, the physical-data loss function includes a loss function for the first monitoring data, a loss function for the first partial differential equation, a loss function for the first initial conditions, and a loss function for the first boundary conditions.

[0020] Optionally, the physical-data loss function is expressed as:

[0021]

[0022] in, Represents the physical-data loss function. The loss function representing the monitoring data, This represents the loss function of the first partial differential equation. The loss function represents the first initial condition. The loss function represents the first boundary condition;

[0023] The loss function of the first partial differential equation is expressed as:

[0024]

[0025] Where, λ P N represents the weights used to calculate the loss function for partial differential equations. P This indicates the number of residual points from the consolidation equation computation domain. This represents the i1th residual point within the computational domain of the consolidation equation. This represents the foundation depth of the i1th residual point within the calculation domain of the consolidation equation. Denotes the time taken to compute the i1-th residual point within the consolidation equation computation domain. c represents the first estimate of the excess pore water pressure at the i1th residual point within the computational domain of the consolidation equation. v Indicates the consolidation coefficient;

[0026] When the top of the foundation is a permeable boundary and the bottom is an impermeable boundary, the loss function of the first boundary condition is expressed as:

[0027]

[0028] When both the top and bottom of the foundation are permeable boundaries, the loss function of the first boundary condition is expressed as:

[0029]

[0030] in, The weights of the loss function representing the top boundary conditions. The weights of the loss function representing the bottom boundary conditions, N B This indicates the number of residual points derived from the boundary condition computation domain. This represents the i2th residual point within the boundary condition computation domain. This represents the foundation depth of the i2th residual point within the boundary condition computation domain. This represents the time taken to compute the i2th residual point within the boundary condition computation domain. This represents the first estimate of the excess pore water pressure at the i2th residual point within the boundary condition computation domain;

[0031] The loss function for the first initial condition is expressed as:

[0032]

[0033] Where, λ I N represents the weights of the loss function used to calculate the initial conditions. I This indicates the number of residual points derived from the initial condition computation domain. This represents the i-th residual point within the initial condition computation domain. This represents the foundation depth of the i3th residual point within the initial condition computation domain. This represents the pre-time of the i3th residual point within the initial condition computation domain. This represents the first estimated value of excess pore water pressure at the i3th residual point within the initial condition computation domain. This represents the initial pore water pressure at the i3th residual point within the initial condition computation domain;

[0034] The loss function of the monitoring data is expressed as:

[0035]

[0036] Where, λ D N represents the weights of the loss function used to calculate the monitoring data. D This indicates the number of residual points derived from the monitoring data calculation domain. This represents the i-4th residual point within the calculation domain of the monitoring data. This represents the foundation depth of the i4th residual point within the calculation domain of the monitoring data. This represents the time of the i4th residual point within the monitoring data calculation domain. This represents the excess pore water pressure monitoring data at the i4th residual point within the monitoring data calculation domain. This represents the first estimate of the i4th residual point within the computational domain of the monitoring data.

[0037] Optionally, a forward prediction model for excess pore water pressure in the foundation is constructed by updating the physical loss function that conforms to the physical law definition constraints based on the inverted consolidation coefficient, training the forward prediction model for excess pore water pressure in the foundation using the residual set, inputting the test set into the trained forward prediction model for excess pore water pressure in the foundation, and outputting the predicted value of excess pore water pressure in the foundation, including:

[0038] The partial differential equation constraints are updated using the inverted consolidation coefficients, a physical loss function conforming to physical law constraints is constructed, and a forward prediction model for excess pore water pressure in the foundation is constructed. The second neural network of the forward prediction model for excess pore water pressure in the foundation includes two second input layers and one second output layer.

[0039] The residual points of the consolidation equation calculation domain, the boundary condition calculation domain, and the initial condition calculation domain are respectively input into the two second input layers of the second neural network. During the forward propagation process, the excess pore water pressure is estimated to obtain a second estimated value of the excess pore water pressure.

[0040] The physical loss function is calculated using the chain rule based on the second estimate of the excess pore water pressure to obtain the residual of the physical loss function;

[0041] The weights and biases of the second neural network are updated in reverse using the chain rule based on the residual of the physical loss function. When the residual of the physical loss function is less than a predetermined second standard, the training of the forward prediction model of the excess pore water pressure of the foundation is completed.

[0042] The test set is input into the trained forward prediction model of excess pore water pressure in the foundation, and the predicted value of excess pore water pressure in the foundation is output.

[0043] Optionally, the physical loss function includes the loss function of the second partial differential equation, the loss function of the second initial condition, and the loss function of the second boundary condition.

[0044] Optionally, the physical loss function is expressed as:

[0045]

[0046] in, Represents the physical loss function. This represents the loss function of the second partial differential equation. The loss function represents the second initial condition. The loss function representing the second boundary condition;

[0047] The loss function of the second partial differential equation is expressed as:

[0048]

[0049] Where, λ P N represents the weights used to calculate the loss function for partial differential equations. P This indicates the number of residual points from the consolidation equation computation domain. This represents the i5th residual point within the computational domain of the consolidation equation. This represents the foundation depth of the i5th residual point within the computational domain of the consolidation equation. This represents the time taken to reach the i5th residual point within the computation domain of the consolidation equation. c represents the second estimate of the excess pore water pressure at the i5th residual point within the computational domain of the consolidation equation. v Indicates the final consolidation coefficient;

[0050] When the top of the foundation is a permeable boundary and the bottom is an impermeable boundary, the loss function of the second boundary condition is expressed as:

[0051]

[0052] When both the top and bottom of the foundation are permeable boundaries, the loss function of the second boundary condition is expressed as:

[0053]

[0054] in, The weights of the loss function representing the top boundary conditions. The weights of the loss function representing the bottom boundary conditions, N B This indicates the number of residual points derived from the boundary condition computation domain. This represents the i-th residual point within the boundary condition computation domain. This represents the foundation depth of the i6th residual point within the boundary condition computation domain. This represents the time taken to calculate the i6th residual point within the boundary condition computation domain. This represents the second estimate of the excess pore water pressure at the i6th residual point within the boundary condition computation domain;

[0055] The loss function for the second initial condition is expressed as:

[0056]

[0057] Where, λ I N represents the weights of the loss function used to calculate the initial conditions. I This indicates the number of residual points derived from the initial condition computation domain. This represents the i7th residual point within the initial condition computation domain. This represents the foundation depth of the i7th residual point within the initial condition computation domain. This represents the time taken to compute the i7th residual point within the initial conditions calculation domain. This represents the second estimate of the excess pore water pressure at the i7th residual point within the initial condition computation domain. This represents the initial pore water pressure at the i7th residual point within the initial condition calculation domain.

[0058] Optionally, the foundation consolidation settlement prediction model is expressed as:

[0059] S t =m v (p0-u m )h

[0060] Among them, S t Indicates long-term consolidation settlement of the foundation, m v The coefficient of compression of the foundation soil is represented by h, and the thickness of the foundation is represented by u. m This represents the excess pore water pressure in a foundation with a thickness of h. u y This represents the predicted excess pore water pressure in the foundation.

[0061] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0062] This invention proposes a method for predicting long-term consolidation settlement of foundations based on a physical information neural network. This method defines a physical-data loss function based on physical laws and monitoring datasets, constructs a geotechnical parameter inversion model, trains the inversion model using a training set to obtain the consolidation coefficient, updates the physical laws based on the inverted consolidation coefficient, defines a physical loss function, constructs a forward prediction model for excess pore water pressure, trains the forward prediction model using a training set, inputs a test set into the trained model, and outputs the predicted value of excess pore water pressure. Then, a foundation consolidation settlement prediction model is constructed, and based on the predicted value of excess pore water pressure, long-term foundation consolidation settlement is predicted. This method can accurately invert the consolidation coefficient using short-term and sparse excess pore water pressure monitoring data and a geotechnical parameter inversion model, even when the consolidation coefficient is unknown. Then, the forward prediction model for excess pore water pressure is used to predict the excess pore water pressure, thus achieving rapid and accurate prediction of long-term foundation consolidation settlement.

[0063] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0064] Figure 1 This is a flowchart illustrating a method for predicting long-term consolidation settlement of a foundation based on a physical information neural network, provided in an embodiment of the present invention.

[0065] Figure 2 This is a schematic diagram of a geotechnical parameter inversion model provided in an embodiment of the present invention;

[0066] Figure 3 This is a graph showing the change of the loss function value of a geotechnical parameter inversion model with the number of training iterations, provided in an embodiment of the present invention.

[0067] Figure 4 This is a schematic diagram illustrating the inversion result of the consolidation coefficient obtained from short-term monitoring data, provided by an embodiment of the present invention.

[0068] Figure 5 This is a schematic diagram of a forward prediction model for excess pore water pressure in a foundation provided by an embodiment of the present invention;

[0069] Figure 6 This is a curve showing the change of the loss function value of a forward prediction model for excess pore water pressure in a foundation as a function of the number of training iterations, provided in an embodiment of the present invention.

[0070] Figure 7 This is a schematic diagram illustrating the result of using PINN (The physics-informed neural network) constructed from the inverted consolidation coefficient to predict excess pore water pressure, as provided in an embodiment of the present invention.

[0071] Figure 8 This is a schematic diagram of the excess pore water pressure result corresponding to an analytical solution provided in an embodiment of the present invention;

[0072] Figure 9 This is an absolute error diagram between the PINN ultrapore water pressure prediction result and the analytical solution provided in an embodiment of the present invention;

[0073] Figure 10 This is a comparison curve of PINN predicted settlement results and analytical solutions at different times provided in an embodiment of the present invention. Detailed Implementation

[0074] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0075] Example 1

[0076] Given the challenges faced by current foundation consolidation settlement prediction methods, such as insufficient reliability of inverted soil parameters, over-reliance on large-scale datasets, and low efficiency of forward prediction of excess pore water pressure, it is difficult to achieve rapid and accurate prediction of long-term foundation consolidation settlement in practical applications.

[0077] Therefore, please see Figure 1 , Figure 1 This is a flowchart illustrating a method for predicting long-term foundation consolidation settlement based on a physical information neural network, as provided in an embodiment of the present invention. The method includes:

[0078] Step 1: Obtain the dataset, which includes the monitoring dataset, residual set, and test set.

[0079] Specifically, the monitoring dataset includes the burial depth of several pore pressure sensors, as well as the monitoring data of excess pore water pressure collected by the pore pressure sensors and the acquisition time. The residual set includes several residual points in the preset computational domain, and the residual points include the foundation depth and time. The test set includes the foundation depth and time that need to be predicted.

[0080] Furthermore, by burying several pore pressure sensors in the foundation at the site, short-term excess pore water pressure monitoring data can be collected. Therefore, each excess pore water pressure monitoring data corresponds to the burial depth of the pore pressure sensor and the data acquisition time. The excess pore water pressure monitoring data collected by the buried pore pressure sensors, along with the corresponding burial depth and acquisition time, constitute a dataset. Additionally, a foundation model is established based on the actual site conditions, and several residual points are randomly selected within this model. Each residual point includes the foundation depth and time.

[0081] For example, the method for generating the monitoring dataset is as follows: Assuming a soft soil foundation with a depth of 10m, a permeable boundary at the top, and an impermeable boundary at the bottom, the parameters are as follows:

[0082] h = 10m, u0 = 100kPa, c v =0.6m 2 / year,m v =0.5MPa -1

[0083] Where h is the thickness of the foundation; m vThis represents the compressibility coefficient of the foundation soil layer. Since the top of the foundation is a permeable boundary, the excess pore water pressure varies significantly. Therefore, pore pressure sensors were installed at distances of 0.5 meters and 1 meter from the top boundary, respectively. Short-term excess pore water pressure monitoring data were collected by the pore pressure sensors. The sensors performed 10 monitoring cycles at a rate of 0.01 years, for a total monitoring period of 0.1 years, obtaining 20 sets of monitoring data. These monitoring data were then organized chronologically to generate a monitoring dataset.

[0084] Furthermore, the preset computational domain includes the consolidation equation computational domain, the boundary condition computational domain, the initial condition computational domain, and the monitoring data computational domain. The consolidation equation computational domain is [t∈(0,1),z∈(0,10)]; the boundary condition computational domain is [t∈(0,1),z=0] and [t∈(0,1),z=10]; the initial condition computational domain is [t=0,z∈(0,10)]; and the monitoring data computational domain consists of the burial depth and acquisition time of 20 sets of pore pressure sensors.

[0085] Step 2: Define the physical-data loss function based on the monitoring dataset and physical laws, construct the soil and rock parameter inversion model, train the soil and rock parameter inversion model using the residual set, and invert the consolidation coefficient.

[0086] Here, the first neural network of the geotechnical parameter inversion model includes two first input layers and one first output layer. The two first input layers take depth and time as inputs, respectively, and the first output layer outputs the estimated excess pore water pressure. For example, the first neural network includes 5 hidden layers, each with 30 neural network nodes. The selected activation function is Tanh, the optimizer is Adam, and the initial value of the consolidation coefficient to be inverted is set to 1.

[0087] Therefore, the residual set is first input into the first neural network of the geotechnical parameter inversion model. During the forward propagation process, the excess pore water pressure is estimated, thus obtaining the first estimated value of the excess pore water pressure. Then, based on the first estimated value of the excess pore water pressure, the physical-data loss function is calculated using the chain rule, and the residual of the physical-data loss function is obtained. Then, based on the residual of the physical-data loss function, the weights and biases of the first neural network are continuously updated in reverse using the chain rule to generate the estimated value of the consolidation coefficient, thereby minimizing the residual of the physical-data loss function. The consolidation coefficient under this condition is then used as the final consolidation coefficient.

[0088] In one specific embodiment, step 2 may include:

[0089] Step 2.1: Define the physical-data loss function based on the monitoring dataset and physical laws, and construct the geotechnical parameter inversion model.

[0090] In this embodiment, the physical-data loss function includes a loss function for the first monitoring data, a loss function for the first partial differential equation, a loss function for the first initial conditions, and a loss function for the first boundary conditions. These loss functions are constructed based on different physical constraints.

[0091] Here, the physical-data loss function is expressed as:

[0092]

[0093] in, Represents the physical-data loss function. The loss function representing the monitoring data, This represents the loss function of the first partial differential equation. The loss function represents the first initial condition. This represents the loss function for the first boundary condition.

[0094] Furthermore, a loss function for the partial differential equation is constructed using Terzaghi's one-dimensional consolidation differential equation. This loss function This describes the settlement and deformation of soil caused by the expulsion of pore water from the soil under its own weight or external load, which can be characterized by the following formula:

[0095]

[0096] Among them, c v Let be the consolidation coefficient, z be the depth, t be the time, and u(z,t) be the excess pore water pressure at depth z and time t. Therefore, the loss function of the first partial differential equation... It can be represented as:

[0097]

[0098] Where, λ P The weights represent the weights used to calculate the loss function of a partial differential equation, for example, set to 1, N. P This represents the number of residual points from the consolidation equation computation domain, for example, set to 10000. This represents the i-th residual point within the computational domain of the consolidation equation, specifically the i-th residual point between the fitted neural network and the actual excess pore water pressure within the computational domain of the consolidation equation. This represents the foundation depth of the i1th residual point within the calculation domain of the consolidation equation. Denotes the time taken to compute the i1-th residual point within the consolidation equation computation domain. This represents the first estimate of the excess pore water pressure at the i1th residual point within the computational domain of the consolidation equation.

[0099] Furthermore, the boundaries of a foundation are typically divided into permeable and impermeable boundaries, corresponding to Dirichlet boundary conditions and Newman boundary conditions, respectively, which represent the values ​​of solutions to the differential equations at the boundaries or the values ​​of their derivatives. If the foundation boundary is permeable, the corresponding boundary conditions are expressed as Newman boundary conditions, i.e.:

[0100] (top) or (bottom)

[0101] When the foundation boundary is impermeable, the corresponding boundary condition is the Dirichlet boundary condition, i.e.:

[0102] u(z=0,t)=0 (top) or u(z=H,t)=0 (bottom)

[0103] Therefore, when the top of the foundation is a permeable boundary (i.e., top drainage) and the bottom is an impermeable boundary, the loss function of the first boundary condition is expressed as:

[0104]

[0105] When both the top and bottom of the foundation are permeable boundaries (i.e., both the top and bottom are drainable), the loss function of the first boundary condition is expressed as:

[0106]

[0107] in, The weights of the loss function representing the top boundary conditions. The weights of the loss function represent the bottom boundary conditions. and For example, all are set to 1, N B This indicates the number of residual points from the boundary condition computation domain, for example, set to 1000. This represents the i-th residual point within the boundary condition computation domain, specifically the i-th residual point between the fitted neural network and the actual excess pore water pressure within the boundary condition computation domain. This represents the foundation depth of the i2th residual point within the boundary condition computation domain. This represents the time taken to compute the i2th residual point within the boundary condition computation domain. This represents the first estimate of the excess pore water pressure at the i2th residual point within the boundary condition computation domain.

[0108] Furthermore, the initial conditions are related to the initial pore water pressure, and are as follows:

[0109] u(z,t=0)=u0(z)

[0110] Where u0(z) is the initial pore water pressure.

[0111] Correspondingly, the loss function of the first initial condition It can be represented by the following formula:

[0112]

[0113] Where, λ I This represents the weights of the loss function used to calculate the initial conditions; for example, it can be set to 1 or N. I This indicates the number of residual points from the initial condition computation domain, for example, set to 1000. This represents the i-th residual point within the initial condition computational domain, specifically the i-th residual point between the fitted neural network and the actual excess pore water pressure within the initial condition computational domain. This represents the foundation depth of the i3th residual point within the initial condition computation domain. This represents the pre-time of the i3th residual point within the initial condition computation domain. This represents the first estimated value of excess pore water pressure at the i3th residual point within the initial condition computation domain. The initial pore water pressure at the i3th residual point in the initial condition calculation domain is 100 kPa.

[0114] Furthermore, the loss function for monitoring data The loss function of the monitoring data is constructed from the excess pore water pressure monitoring data collected by the sensor. for:

[0115]

[0116] Where, λ D This represents the weights of the loss function used to calculate the monitoring data; for example, set to 1, N. D This indicates the number of residual points from the monitoring data calculation domain, for example, set to 20. This represents the i-4th residual point within the calculation domain of the monitoring data. This represents the foundation depth of the i4th residual point within the calculation domain of the monitoring data. This represents the time of the i4th residual point within the monitoring data calculation domain. This represents the excess pore water pressure monitoring data at the i4th residual point within the monitoring data calculation domain. This represents the first estimated value of the i4th residual point within the monitoring data calculation domain. Since the residual points within the monitoring data calculation domain correspond one-to-one with the positions of the pore pressure sensors in the actual foundation, the i4th residual point within the monitoring data calculation domain is also the i4th sampling point in the actual foundation. This refers to the excess pore water pressure monitoring data collected by the i4th pore pressure sensor embedded in the foundation.

[0117] Step 2.2: Input the foundation depth and time from the residual set into the two first input layers of the first neural network, and estimate the excess pore water pressure during the forward propagation process to obtain the first estimated value of the excess pore water pressure.

[0118] Step 2.3: Calculate the physical-data loss function using the chain rule based on the first estimated value of excess pore water pressure, and obtain the residual of the physical-data loss function.

[0119] Step 2.4: Based on the residuals of the physical-data loss function, use the chain rule to back-update the weights and biases of the first neural network and generate an estimated value of the consolidation coefficient. When the residuals of the physical-data loss function are less than the predetermined first standard, complete the training of the geotechnical parameter inversion model and directly invert to obtain the final consolidation coefficient.

[0120] Specifically, the weights and biases of the first neural network are updated using the first inverse calculation result obtained in step 2.4, and the consolidation coefficient in the physical-data loss function is calculated as an estimate of the consolidation coefficient. This process continues until the residual of the physical-data loss function is less than a predetermined first standard. The consolidation coefficient in the physical-data loss function is then calculated as the final consolidation coefficient. It should be noted that the first standard can be set according to actual conditions; this embodiment does not specifically limit it. For example, the first standard could be e. -2 .

[0121] In other words, please see Figure 2 In each training iteration of the geotechnical parameter inversion model, excess pore water pressure is estimated during the forward propagation process. Then, the residual of the physical-data loss function is calculated using the chain rule, followed by backpropagation using the chain rule. Based on the backpropagation results, the Adam optimizer is used to update the hidden layer weights and biases, and then the consolidation coefficient c is directly generated. v The estimated value; when the residual of the loss function of the geotechnical parameter inversion model reaches the predetermined first standard ε1, the training of the geotechnical parameter inversion model is completed. Figure 3 The curves showing the change of the loss function of the geotechnical parameter inversion model with the number of training iterations are presented. It can be seen that the loss of the geotechnical parameter inversion model continuously decreases with the number of training iterations, and finally the error of the total loss function reaches the order of e. -3 This training process was conducted on a T7825 workstation equipped with an NVIDIA 3090 graphics card, and the training time was only 121 seconds. Furthermore, during the training process, the soil parameter inversion model saved the consolidation coefficient c once every 1000 training iterations. v The estimated value is taken as the consolidation coefficient c from the last training iteration. v The final result of the inversion. To qualitatively measure the error between the true value and the inverted / predicted value, L is selected. wThe relative error is used as a metric, and its calculation formula is as follows:

[0122]

[0123] In the formula, and y i These represent the inverted / predicted value and the actual value, respectively. This can represent the inverted / predicted and actual values ​​of excess pore water pressure, or the inverted / predicted and actual values ​​of the consolidation coefficient; N represents the number of inverted / predicted values. For example... Figure 4 As shown, after 40,000 training iterations, the consolidation coefficient c of the soil and rock parameter inversion model is... v The inversion error is only 0.002, which indicates that the geotechnical parameter inversion model constructed in this invention can quickly and accurately invert the consolidation coefficient c based on short-term pore water pressure monitoring data. v .

[0124] Step 3: Update the physical loss function defined by the physical law based on the inverted consolidation coefficient, construct a forward prediction model for excess pore water pressure in the foundation, train the forward prediction model for excess pore water pressure in the foundation using the residual set, input the test set into the trained forward prediction model for excess pore water pressure in the foundation, and output the predicted value of excess pore water pressure in the foundation.

[0125] Here, the second neural network of the forward prediction model for excess pore water pressure in the foundation includes two second input layers and one second output layer. The two second input layers take depth and time as inputs, respectively, and the second output layer outputs the estimated excess pore water pressure. For example, the second neural network includes 5 hidden layers, each with 30 neural network nodes, using Tanh as the activation function and Adam as the optimizer.

[0126] In one specific embodiment, step 3 may include:

[0127] Step 3.1: Update the partial differential equation constraints using the inverted consolidation coefficients, define a physical loss function that conforms to physical law constraints, and construct a forward prediction model for excess pore water pressure in the foundation.

[0128] In this embodiment, the physical loss function includes the loss function of the second partial differential equation, the loss function of the second initial condition, and the loss function of the second boundary condition.

[0129] Here, the physical loss function is expressed as:

[0130]

[0131] in, Represents the physical loss function. This represents the loss function of the second partial differential equation. The loss function represents the second initial condition. This represents the loss function for the second boundary condition.

[0132] Furthermore, the loss function of the second partial differential equation is expressed as:

[0133]

[0134] in, This represents the i5th residual point within the computational domain of the consolidation equation. This represents the foundation depth of the i5th residual point within the computational domain of the consolidation equation. This represents the time taken to reach the i5th residual point within the computation domain of the consolidation equation. c represents the second estimate of the excess pore water pressure at the i5th residual point within the computational domain of the consolidation equation. v This represents the final consolidation coefficient.

[0135] Furthermore, when the top of the foundation is a permeable boundary and the bottom is an impermeable boundary, the loss function of the second boundary condition is expressed as:

[0136]

[0137] When both the top and bottom of the foundation are permeable boundaries, the loss function of the second boundary condition is expressed as:

[0138]

[0139] in, This represents the i-th residual point within the boundary condition computation domain. This represents the foundation depth of the i6th residual point within the boundary condition computation domain. This represents the time taken to calculate the i6th residual point within the boundary condition computation domain. This represents the second estimate of the excess pore water pressure at the i6th residual point within the boundary condition computation domain.

[0140] Furthermore, the loss function for the second initial condition is expressed as:

[0141]

[0142] in, This represents the i7th residual point within the initial condition computation domain. This represents the foundation depth of the i7th residual point within the initial condition computation domain. This represents the time taken to compute the i7th residual point within the initial conditions calculation domain. This represents the second estimate of the excess pore water pressure at the i7th residual point within the initial condition computation domain. This represents the initial pore water pressure at the i7th residual point within the initial condition calculation domain.

[0143] Step 3.2: Input the foundation depth and time of the consolidation equation calculation domain, boundary condition calculation domain, and initial condition calculation domain into the two second input layers of the second neural network, respectively, and use the forward propagation process to estimate the excess pore water pressure to obtain the second estimated value of the excess pore water pressure.

[0144] Specifically, the foundation depth and time of the consolidation equation calculation domain, the foundation depth and time of the boundary condition calculation domain, and the foundation depth and time of the initial condition calculation domain are respectively input into the two second input layers of the second neural network, and the second estimated values ​​of excess pore water pressure corresponding to the consolidation equation calculation domain, the boundary condition calculation domain, and the initial condition calculation domain are output from the second output layer respectively.

[0145] Step 3.3: Calculate the physical loss function using the chain rule based on the second estimate of the excess pore water pressure to obtain the residual of the physical loss function.

[0146] Step 3.4: Update the weights and biases of the second neural network in reverse using the chain rule based on the residuals of the physical loss function. When the residuals of the physical loss function are less than the predetermined second standard, the training of the forward prediction model for excess pore water pressure in the foundation is completed.

[0147] Specifically, the weights and biases of the second neural network are updated using the second inverse calculation result obtained in step 3.3. Then, the above steps are continued until the residual of the physical loss function is less than a predetermined second criterion, completing the training of the forward prediction model for excess pore water pressure in the foundation. It should be noted that the second criterion can be set according to actual conditions; this embodiment does not specifically limit it. For example, the second criterion could be e. -2 .

[0148] Step 3.5: Input the test set into the trained forward prediction model of excess pore water pressure in the foundation, and output the predicted value of excess pore water pressure in the foundation.

[0149] In other words, such as Figure 5 As shown, a second estimate of excess pore water pressure is generated using the forward propagation process. Then, the physical loss function constrained by physical laws is calculated, and the chain rule is used for back calculation. Based on the back calculation results, the Adam optimizer is used to update the hidden layer weights and biases. When the residual of the physical loss function of the foundation excess pore water pressure forward prediction model reaches the predetermined second standard ε2, the training of the foundation excess pore water pressure forward prediction model is completed. Figure 6The curves showing the change of the loss function value of the forward prediction model for excess pore water pressure in the foundation with the number of training iterations are presented. Finally, the error of the total loss function reaches the order of e. -3 This indicates that after 20,000 training iterations, the forward prediction model for excess pore water pressure in the foundation exhibits good predictive capabilities, and this training process only requires 82 seconds. Then, location and time information are input into the trained forward prediction model for excess pore water pressure in the foundation to predict the excess pore water pressure. Figure 7 The results of using PINN constructed from the inverted consolidation coefficients to predict excess pore water pressure are presented. Figure 8 This is the analytical solution corresponding to the true consolidation coefficient. Figure 9 The absolute error between these two values ​​is given. It can be seen that the prediction results of PINN are in high agreement with the analytical solution, indicating that PINN can accurately predict the excess pore water pressure in the foundation. Finally, the depth and time information are input into the trained forward prediction model for the excess pore water pressure in the foundation to predict the excess pore water pressure, thus obtaining the predicted excess pore water pressure.

[0150] Step 4: Construct the foundation consolidation settlement prediction model and predict the long-term consolidation settlement of the foundation based on the predicted value of the excess pore water pressure of the foundation.

[0151] Here, the foundation consolidation settlement prediction model is expressed as:

[0152] S t =m v (p0-u m )h

[0153] Among them, S t Indicates long-term consolidation settlement, m v The coefficient of compression of the foundation soil is represented by h, and the thickness of the foundation is represented by u. m This represents the excess pore water pressure in a foundation with a thickness of h. u y This represents the predicted excess pore water pressure in the foundation.

[0154] Therefore, the predicted excess pore water pressure of the foundation is input into the constructed foundation consolidation settlement prediction model to predict the long-term consolidation settlement of the foundation.

[0155] For example, the excess pore water pressure of the foundation at 2, 4, 6, 8 and 10 years, predicted by the above model, can be used as inputs to the foundation consolidation settlement prediction model to predict the long-term consolidation settlement of the foundation. Figure 10The comparison curves between the settlement prediction results of PINN and the analytical solution at different times are presented, with an error value of only 0.003. It can be seen that the prediction results of the foundation settlement prediction model established in this invention are basically consistent with the analytical solution, proving that the method of this invention can accurately invert the consolidation coefficient using short-term and sparse excess pore water pressure monitoring data and a soil-rock parameter inversion model under the condition of unknown consolidation coefficient. Then, it updates the physical constraints to construct a positive model to predict the excess pore water pressure of the foundation, thereby achieving rapid and accurate prediction of long-term consolidation settlement of the foundation.

[0156] This invention proposes a method for predicting long-term consolidation settlement of foundations based on the forward and inverse functions of a physical information neural network. Compared with existing forward calculation methods (analytical and numerical methods), this method effectively solves the inverse problem and avoids potential human errors in traditional inverse analysis methods (time logarithmic fitting, time square root fitting, and inflection point methods). It also overcomes the limitation of artificial neural network methods relying on large-scale datasets. Under conditions where the consolidation coefficient is unknown, this method accurately inverts the consolidation coefficient using short-term and sparse excess pore water pressure monitoring data and a geotechnical parameter inversion model. This updated physical constraints are then used to construct a forward model to predict excess pore water pressure in the foundation, thus achieving rapid and accurate prediction of long-term consolidation settlement. This method can be applied to geotechnical engineering construction, providing a more reliable theoretical basis for its design and decision-making.

[0157] It should be noted that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined as "first" or "second" may explicitly or implicitly include one or more features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0158] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0159] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings and the disclosure in carrying out the claimed invention. In this specification, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality. While certain measures are described in different embodiments, this does not mean that these measures cannot be combined to produce good results.

[0160] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for predicting long-term consolidation settlement of foundations based on physical information neural networks, characterized in that, include: Acquire a dataset, which includes a monitoring dataset, a residual set, and a test set; the monitoring dataset includes the burial depth of several pore pressure sensors and the monitoring data of excess pore water pressure collected by the pore pressure sensors and the collection time; the residual set includes several residual points in a preset computational domain, and the residual points include the foundation depth and time; the test set includes the foundation depth and time that need to be predicted. The preset calculation domain includes the consolidation equation calculation domain, the boundary condition calculation domain, the initial condition calculation domain, and the monitoring data calculation domain; Based on the monitoring dataset and physical laws, a physical-data loss function is defined, a soil and rock parameter inversion model is constructed, and the soil and rock parameter inversion model is trained using the residual set to invert the consolidation coefficient. The physical loss function defined by the physical law is updated based on the inverted consolidation coefficient. A forward prediction model of excess pore water pressure in the foundation is constructed. The model is trained using the residual set. The test set is input into the trained model and the predicted value of excess pore water pressure in the foundation is output. A foundation consolidation settlement prediction model is constructed, and the long-term consolidation settlement of the foundation is predicted based on the predicted value of the excess pore water pressure. The foundation consolidation settlement prediction model is expressed as follows: in, This indicates long-term consolidation and settlement of the foundation. This represents the compressibility coefficient of the foundation soil layer. Indicates the thickness of the foundation. Indicates thickness as Excess pore water pressure in the foundation , This represents the predicted excess pore water pressure in the foundation.

2. The method for predicting long-term consolidation settlement of foundations according to claim 1, characterized in that, Based on the monitoring dataset and physical laws, a physical-data loss function is defined, a soil and rock parameter inversion model is constructed, and the model is trained using the residual set to invert the consolidation coefficient, including: Based on the monitoring dataset and physical laws, a physical-data loss function is defined, and a soil and rock parameter inversion model is constructed. The first neural network of the soil and rock parameter inversion model includes two first input layers and one first output layer. The foundation depth and time in the residual set are input into the two first input layers of the first neural network to estimate the excess pore water pressure during the forward propagation process, thereby obtaining the first estimated value of the excess pore water pressure. The physical-data loss function is calculated using the chain rule based on the first estimated value of the excess pore water pressure, and the residual of the physical-data loss function is obtained. The weights and biases of the first neural network are updated in reverse using the chain rule based on the residuals of the physical-data loss function, and an estimated value of the consolidation coefficient is generated. When the residuals of the physical-data loss function are less than a predetermined first standard, the training of the geotechnical parameter inversion model is completed, and the final consolidation coefficient is obtained directly by inversion.

3. The method for predicting long-term consolidation settlement of foundations according to claim 2, characterized in that, The physical-data loss function includes the loss function of monitoring data, the loss function of the first partial differential equation, the loss function of the first initial condition, and the loss function of the first boundary condition.

4. The method for predicting long-term consolidation settlement of foundations according to claim 3, characterized in that, The physical-data loss function is expressed as follows: in, Represents the physical-data loss function. The loss function representing the monitoring data, This represents the loss function of the first partial differential equation. The loss function represents the first initial condition. The loss function represents the first boundary condition; The loss function of the first partial differential equation is expressed as: in, The weights represent the weights used to calculate the loss function for partial differential equations. This indicates the number of residual points from the consolidation equation computation domain. Represents the first consolidation equation within the computational domain. One residual point, Represents the first consolidation equation within the computational domain. The foundation depth of each residual point Represents the first consolidation equation within the computational domain. The time for a residual point Represents the first consolidation equation within the computational domain. The first estimate of the excess pore water pressure at each residual point. Indicates the consolidation coefficient; When the top of the foundation is a permeable boundary and the bottom is an impermeable boundary, the loss function of the first boundary condition is expressed as: When both the top and bottom of the foundation are permeable boundaries, the loss function of the first boundary condition is expressed as: in, The weights of the loss function representing the top boundary conditions. The weights of the loss function represent the bottom boundary conditions. This indicates the number of residual points derived from the boundary condition computation domain. The first boundary condition within the computational domain represents the boundary condition. One residual point, Indicates the boundary conditions within the computational domain. The foundation depth of each residual point Indicates the boundary conditions within the computational domain. The time for a residual point The first boundary condition within the computational domain represents the boundary condition. The first estimate of excess pore water pressure at each residual point; The loss function for the first initial condition is expressed as: in, This represents the weights of the loss function used to calculate the initial conditions. This indicates the number of residual points derived from the initial condition computation domain. Represents the first condition within the computational domain. One residual point, Represents the first condition within the computational domain. The foundation depth of each residual point Represents the first condition within the computational domain. Pre-time for each residual point Represents the first condition within the computational domain. The first estimate of the excess pore water pressure at each residual point. Represents the first condition within the computational domain. Initial pore water pressure at each residual point; The loss function of the monitoring data is expressed as: in, This represents the weights of the loss function used to calculate the monitoring data. This indicates the number of residual points derived from the monitoring data calculation domain. Indicates the first [data] in the monitoring data calculation domain One residual point, Indicates the first [data] in the monitoring data calculation domain The foundation depth of each residual point Indicates the first [data] in the monitoring data calculation domain The time for a residual point Indicates the first [data] in the monitoring data calculation domain Monitoring data of excess pore water pressure at each residual point Indicates the first [data] in the monitoring data calculation domain The first estimate of each residual point.

5. The method for predicting long-term consolidation settlement of foundations according to claim 1, characterized in that, Based on the inverted consolidation coefficient, the physical loss function defined by the physical law is updated. A forward prediction model for excess pore water pressure in the foundation is constructed. The model is trained using the residual set. The test set is input into the trained model, and the predicted value of excess pore water pressure in the foundation is output, including: The consolidation coefficients obtained from the inversion are used to update the constraints of the partial differential equations, and a physical loss function that conforms to the constraints of physical laws is constructed to build a forward prediction model for excess pore water pressure in the foundation. The second neural network of the forward prediction model for excess pore water pressure in the foundation includes two second input layers and one second output layer. The residual points of the consolidation equation calculation domain, the boundary condition calculation domain, and the initial condition calculation domain are respectively input into the two second input layers of the second neural network. During the forward propagation process, the excess pore water pressure is estimated to obtain a second estimated value of the excess pore water pressure. The physical loss function is calculated using the chain rule based on the second estimate of the excess pore water pressure to obtain the residual of the physical loss function; The weights and biases of the second neural network are updated in reverse using the chain rule based on the residual of the physical loss function. When the residual of the physical loss function is less than a predetermined second standard, the training of the forward prediction model of the excess pore water pressure of the foundation is completed. The test set is input into the trained forward prediction model of excess pore water pressure in the foundation, and the predicted value of excess pore water pressure in the foundation is output.

6. The method for predicting long-term consolidation settlement of foundations according to claim 5, characterized in that, The physical loss function includes the loss function of the second partial differential equation, the loss function of the second initial condition, and the loss function of the second boundary condition.

7. The method for predicting long-term consolidation settlement of foundations according to claim 6, characterized in that, The physical loss function is expressed as: in, Represents the physical loss function. This represents the loss function of the second partial differential equation. The loss function represents the second initial condition. The loss function representing the second boundary condition; The loss function of the second partial differential equation is expressed as: in, The weights represent the weights used to calculate the loss function for partial differential equations. This indicates the number of residual points from the consolidation equation computation domain. Represents the first consolidation equation within the computational domain. One residual point, Represents the first consolidation equation within the computational domain. The foundation depth of each residual point Represents the first consolidation equation within the computational domain. The time for a residual point Represents the first consolidation equation within the computational domain. The second estimate of the excess pore water pressure at each residual point Indicates the final consolidation coefficient; When the top of the foundation is a permeable boundary and the bottom is an impermeable boundary, the loss function of the second boundary condition is expressed as: When both the top and bottom of the foundation are permeable boundaries, the loss function of the second boundary condition is expressed as: in, The weights of the loss function representing the top boundary conditions. The weights of the loss function represent the bottom boundary conditions. This indicates the number of residual points derived from the boundary condition computation domain. The first boundary condition within the computational domain represents the boundary condition. One residual point, The first boundary condition within the computational domain represents the boundary condition. The foundation depth of each residual point The first boundary condition within the computational domain represents the boundary condition. The time for a residual point The first boundary condition within the computational domain represents the boundary condition. The second estimate of excess pore water pressure at each residual point; The loss function for the second initial condition is expressed as: in, This represents the weights of the loss function used to calculate the initial conditions. This indicates the number of residual points derived from the initial condition computation domain. Represents the first condition within the computational domain. One residual point, Represents the first condition within the computational domain. The foundation depth of each residual point Represents the first condition within the computational domain. The time for a residual point Indicates the initial conditions within the computational domain. The second estimate of the excess pore water pressure at each residual point Indicates the initial conditions within the computational domain. The initial pore water pressure at each residual point.

Citation Information

Patent Citations

  • Rapid sub-grade settlement predicting method based on static sounding and BP (Back Propagation) neural network

    CN102425148A

  • Handheld deep foundation pitground surface subsidence predicting device and working procedures thereof

    CN109083118A