Grouting slurry diffusion form characterization method and system based on deep learning

Abstract

The invention relates to the technical field of grouting simulation, in particular to a grouting slurry diffusion form characterization method and system based on deep learning, and the method comprises the steps: obtaining diffusion form data of slurry under different geological conditions, and taking the preprocessed data as a training set; constructing a grouting diffusion characterization model, and training the grouting diffusion characterization model based on the training set to obtain a trained grouting diffusion characterization model; a total loss function used in the training process is a weighted summation result of a data loss function, a physical constraint loss function, a boundary value condition loss function, an initial value condition loss function and an actual experience loss function; and inputting temperature and pressure data of a to-be-represented grouting area into the trained grouting diffusion representation model to obtain predicted values of the grouting diffusion flow velocity and the grouting diffusion distance. Model verification is carried out, and grouting intelligent control is carried out in an auxiliary mode according to the established model.

Description

Technical Field

[0001] The present invention relates to the technical field of grouting simulation, and particularly to a method and system for characterizing the diffusion morphology of grouting slurry based on deep learning. Background Art

[0002] With the gradual increase in the construction scale and difficulty of tunnels and underground projects, the geological environment traversed by the engineering construction area is becoming more and more complex, and the sudden disasters encountered are increasing, often causing large property losses and casualties. Among them, the sudden water inrush disaster and the excavation collapse problem caused by fractured rock mass seriously restrict the construction efficiency and quality of the project. As the most commonly used method for this type of disaster, grouting often uses a slurry with rheological properties, which gradually undergoes solidification phase change over time, and can effectively seal the sudden water inrush disaster and effectively reinforce the fractured rock mass. However, the complexity of the geotechnical medium in underground engineering makes the slurry diffusion process have strong concealment, and its diffusion situation is difficult to be visually judged. Therefore, how to effectively characterize the changes in the morphology and various parameters during the slurry diffusion process is very important for the efficient treatment of grouting disasters.

[0003] Existing grouting diffusion mostly focuses on grouting model tests and numerical simulation studies. However, in terms of grouting tests, it is difficult to effectively restore the on-site grouting medium, and limited by monitoring instruments, it is often difficult to obtain accurate and effective grouting process diffusion data. In terms of numerical simulation, through targeted numerical calculation methods, the visualization simulation of grouting diffusion can be realized. However, the existing grouting theory is difficult to consider the simulation research under complex medium conditions, and the diffusion theory of the interaction between slurry and water is not yet perfect, and there is still a lack of an efficient characterization method for the grouting slurry diffusion process. Summary of the Invention

[0004] In order to solve the deficiencies of the existing technology, the present invention provides a method and system for characterizing the diffusion morphology of grouting slurry based on deep learning; collecting grouting simulation test, on-site monitoring, and numerical simulation data, performing preprocessing, necessary annotation and feature extraction on the data, and dividing the processed data into a training set and a validation set. Establish a grouting diffusion characterization model, construct a neural network framework, define a loss function in the network that comprehensively considers data loss, physical constraint loss, boundary condition loss, initial value condition loss and actual experience loss, set the weights of each loss that are dynamically adjusted, accelerate the convergence of model training and improve the calculation efficiency, perform model verification and assist in carrying out intelligent grouting control based on the established model.

[0005] On the one hand, a method for characterizing the diffusion morphology of grouting slurry based on deep learning is provided, including:

[0006] Obtaining the diffusion morphology data of the slurry under different geological conditions, preprocessing the obtained data, and using the preprocessed data as the training set;

[0007] Build a grouting diffusion characterization model, which includes a spatial convolutional layer and a temporal convolutional layer. The spatial convolutional layer is used to extract spatial features; the temporal convolutional layer is used to extract temporal features;

[0008] Train the grouting diffusion characterization model based on the training set to obtain the trained grouting diffusion characterization model; among them, the total loss function used in the training process is the weighted sum result of the data loss function, the physical constraint loss function, the boundary condition loss function, the initial condition loss function and the actual experience loss function;

[0009] Input the temperature and pressure data of the grouting area to be characterized into the trained grouting diffusion characterization model to obtain the predicted values of the grouting diffusion velocity and the grouting diffusion distance.

[0010] On the other hand, a grouting slurry diffusion morphology characterization system based on deep learning is provided, including:

[0011] An acquisition module, which is configured to: acquire the diffusion morphology data of the slurry under different geological conditions, preprocess the acquired data, and use the preprocessed data as the training set;

[0012] A construction module, which is configured to: build a grouting diffusion characterization model, which includes a spatial convolutional layer and a temporal convolutional layer. The spatial convolutional layer is used to extract spatial features; the temporal convolutional layer is used to extract temporal features;

[0013] A training module, which is configured to: train the grouting diffusion characterization model based on the training set to obtain the trained grouting diffusion characterization model; among them, the total loss function used in the training process is the weighted sum result of the data loss function, the physical constraint loss function, the boundary condition loss function, the initial condition loss function and the actual experience loss function;

[0014] An output module, which is configured to: input the temperature and pressure data of the grouting area to be characterized into the trained grouting diffusion characterization model to obtain the predicted values of the grouting diffusion velocity and the grouting diffusion distance.

[0015] The above technical solution has the following advantages or beneficial effects:

[0016] Collect the diffusion data of the slurry under different geological conditions through laboratory tests, on-site monitoring, and obtain it. At the same time, conduct numerical simulation of grouting as a supplement to the actual test data. Perform preprocessing of the data, and conduct necessary annotation and feature extraction. At the same time, divide the processed data into a training set and a validation set. Establish a grouting diffusion characterization model, construct a neural network architecture including a spatial convolutional layer and a temporal convolutional layer, perform non-linear activation processing, define a loss function in the network that comprehensively considers data loss, physical constraint loss, boundary condition loss, initial condition loss, and actual experience loss. Consider the influence of the comprehensive loss function and physical constraint loss on network training, set the weights of various losses with dynamic adjustment, so that the neural network focuses on optimizing specific objectives at different stages. Optimize the weights of the neural network through the Adam optimizer and introduce regularization processing, perform adaptive dynamic adjustment of time steps, accelerate the convergence of model training and improve computational efficiency, conduct model verification, and assist in the intelligent control of grouting based on the established model. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The accompanying drawings forming a part of this invention are used to provide a further understanding of the invention. The schematic embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0018] Figure 1 It is a flowchart of the method for the first embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which this invention belongs.

[0020] In the first embodiment, as Figure 1 shown, this embodiment provides a method for characterizing the diffusion morphology of grouting slurry based on deep learning, including:

[0021] S101: Obtain the diffusion morphology data of the slurry under different geological conditions, preprocess the obtained data, and use the preprocessed data as the training set;

[0022] S102: Construct a grouting diffusion characterization model. The grouting diffusion characterization model includes a spatial convolutional layer and a temporal convolutional layer. The spatial convolutional layer is used to extract spatial features; the temporal convolutional layer is used to extract temporal features;

[0023] S103: Train the grouting diffusion characterization model based on the training set to obtain the trained grouting diffusion characterization model; among them, the total loss function used in the training process is the weighted sum result of the data loss function, physical constraint loss function, boundary condition loss function, initial condition loss function, and actual experience loss function;

[0024] S104: Input the temperature and pressure data of the grouting area to be characterized into the trained grouting diffusion characterization model to obtain the predicted values of the grouting diffusion flow rate and the grouting diffusion distance.

[0025] Further, in S101: Obtain the diffusion morphology data of the grout under different geological conditions. The diffusion morphology data includes: two-dimensional grout diffusion images, three-dimensional grout diffusion images, grout diffusion pressure, grout flow rate, and grout temperature change.

[0026] Further, the obtaining of the diffusion morphology data of the grout under different geological conditions can be carried out by laboratory tests or on-site collection in the tunnel.

[0027] Further, preprocess the obtained data. The preprocessing includes: grouting numerical simulation, data denoising, data standardization processing, and data annotation processing.

[0028] It should be understood that grouting numerical simulation refers to: based on existing fluid mechanics numerical theories such as vof, double Euler method, etc., simulate the grouting process through existing calculation software.

[0029] Data denoising refers to: denoise the obtained image data by bilateral filtering.

[0030] Data standardization processing refers to: unify the units of the data of diffusion pressure, grout flow rate, and grout temperature change and perform decimal scaling standardization processing to reduce the data scale to a set range.

[0031] Data annotation processing refers to: label and extract features for set key points and abnormal points during the grouting process.

[0032] Further, use the preprocessed data as the training set. The training set includes: the diffusion pressure and grout flow rate with known grout temperature change during the grouting process.

[0033] It should be understood that collect the diffusion morphology data of the grout under different geological conditions through laboratory tests, on-site monitoring, including two-dimensional or three-dimensional diffusion images, pressure, flow rate, and temperature change, etc. Conduct grouting numerical simulation, and collect the grout diffusion data under the same working conditions as those of the laboratory tests and on-site monitoring geological conditions as a supplement to the actual test data. Conduct preprocessing of the data, including cleaning the data, removing noise, standardizing the diffusion morphology, and performing necessary annotation and feature extraction, and at the same time determine the coverage of the training data under sufficient initial conditions and boundary conditions. Divide the data into a training set and a validation set.

[0034] Further, S102: Construct a grouting diffusion characterization model, which includes a spatial convolutional layer and a temporal convolutional layer. The spatial convolutional layer is used to extract spatial features; the temporal convolutional layer is used to extract temporal features; among them, the grouting diffusion characterization model is implemented using a three-dimensional convolutional neural network; the network structure of the three-dimensional convolutional neural network includes, connected in sequence: an input layer, a spatial convolutional layer, a temporal convolutional layer, a fully connected layer, and an output layer.

[0035] Based on the two-dimensional convolution of the original convolutional neural network, a convolution operation in the time dimension is added. For example, the original two-dimensional 3×3 convolution is increased to 3×3×3. Among them, 3×3×3 is decomposed into a spatial convolutional layer of 1×3×3 and a temporal convolutional layer of 3×1×1. Further, the temporal convolutional layer of 3×1×1 is decomposed into three convolutional layers of size 1×1, which are respectively used to learn the features of the previous time step, the next time step, and the overlapping features of the previous and next time steps.

[0036] Further, the working process of the spatial convolutional layer includes:

[0037] The value at (x, y) on the γ-th feature map of the β-th convolutional layer Can be expressed as:

[0038]

[0039] Among them Is the activation function, b β,γ Is the added bias, N is the total number of feature map inputs in the (β - 1)-th layer connected to the current feature map, Is the value of the kernel at (φ, ψ), and Φ and Ψ are the height and width of the convolutional kernel respectively.

[0040] Further, the working process of the temporal convolutional layer includes:

[0041] Assume that a one-dimensional time series T is input, and its convolution operation can be expressed as

[0042] O(t) = ∑T(t, a)·W(a)

[0043] O(t) represents the output feature after convolution, T(t, a) represents the data in the time series, and W(a) represents the convolution weight value.

[0044] It should be understood that to establish a grouting diffusion characterization model and construct a three-dimensional convolutional neural network, including a spatial convolutional layer and a temporal convolutional layer, the input is spatial position, time, and initial and boundary conditions, and the output is the diffusion state variables of the slurry, including pressure, flow rate, temperature, and slurry-water multiphase distribution. The spatial convolutional layer is used to extract spatial features, and the one-dimensional temporal convolutional layer is used to extract temporal features. In the fully connected layer, the output data in the convolutional layer is non-linearly transformed by introducing a non-linear activation function.

[0045] Furthermore, in S103: the total loss function is the weighted sum of a data loss function, a physical constraint loss function, a boundary condition loss function, an initial condition loss function, and an actual experience loss function; where the formula expression of the total loss function is:

[0046] L t = α d (t s )L d + α p (t s )L p + α b (t s )L b + α i (t s )L i + α en (t s )L e

[0047] Among them, L t is the total loss function, α d represents the weight of the data loss function, α p represents the weight of the physical constraint loss function, α b represents the weight of the boundary condition loss function, α i represents the weight of the initial condition loss function, α e represents the weight of the actual experience loss function, t s represents the time step;

[0048] n represents different empirical losses under different medium conditions. If there are multiple media acting together, the empirical loss is re-solved by weighted averaging according to the distribution of the multiple media, which can be expressed as Among them, θ represents different media, and γ θ represents the weight coefficient of the action of each different medium.

[0049]

[0050] Among them, m can respectively correspond to the five loss functions d, p, b, i, e. There is a floating threshold for the weight of each loss function, which is expressed as α m (t s ) ∈ [m 1 , m 2 , α m (t s ) represents the loss weight at each time step, L M represents the sum of the values of each loss function at each time step, L M = Ld +L p +L b +L i +L e 。

[0051] Furthermore, the data loss function is expressed as:

[0052]

[0053] where L d is the data loss function, N d is the number of training samples, represents the corresponding true value, y j represents the predicted value of the model, and j represents the j-th training sample.

[0054] Furthermore, the physical constraint loss function is expressed as:

[0055]

[0056] where L p is the physical constraint loss function, N p is the number of sampling points for physical constraints, represents the differential operation at the sampling points, and f o (x o ,t o ) represents the source term of the physical equation. For example, for the continuity equation its is exactly and its f o (x o ,t o ) is 0. Therefore,

[0057] its is exactly (the expanded form of the continuity equation), where x, y, and z are physical independent variables, are the components of the velocity in the x, y, and z directions at each sampling point.

[0058] can be expanded separately as where t o represents the time step at this sampling point, and its f o (x o ,t o ) is 0. Therefore,

[0059] The sampling points for physical constraints are sampled at different spatial positions using the numerical solution of the equation, and regions with large changes in physical parameters are mainly selected, such as the boundary layer, the turbulent occurrence region, etc.

[0060] Furthermore, the boundary value condition loss function is expressed as:

[0061]

[0062] where L b represents the boundary value condition loss function, N b represents the number of boundary points, represents the boundary condition, v(x k , t k ) represents the predicted value on the boundary; for example, for a boundary temperature that fluctuates with time, when x = L, it follows the boundary condition, then there is

[0063] Furthermore, the initial value condition loss function is expressed as:

[0064]

[0065] where L i represents the initial value condition loss function, N i represents the number of initial condition sampling points, represents the initial condition function, w(x l , 0) represents the predicted value on the initial condition; for example, for an initial boundary temperature, when t = 0, it follows the boundary condition, then there is

[0066] Furthermore, the actual empirical loss function is expressed as:

[0067]

[0068] where L e is the actual empirical loss, N e is the number of actual empirical loss constraint sampling points, represents the miscibility loss measure; ε(x q , t q ) is the change in void fraction, and the change in void fraction is the difference between the predicted and actual porosity at the sampling points. μ(x q , t q ) is the change in slurry viscosity; v q is the change in velocity, β is the adjustment coefficient, r is the pressure coefficient, is the pressure gradient, is the gradient operator.

[0069]

[0070] where T is the total time, cin is the slurry injection concentration, c out (t) is the predicted slurry concentration over time.

[0071] μ(x q ,t q )=∫ 0 T |μ yc (t)-μ sj (t)| 2 dt

[0072] Among them, μ yc (t) is the predicted viscosity of the slurry over time, μ sj (t) is the actual viscosity of the slurry obtained over time.

[0073] Considering the importance of physical equation constraints, the total loss function error and the physical constraint loss error in model training should be within the threshold for comprehensive judgment.

[0074] Dynamically adjust the weights of different loss terms so that the neural network can focus on optimizing specific goals at different stages. According to the diffusion of slurry in different media, physical constraints are gradually added to avoid the problem of convergence time being affected by the difficulty of physical constraints in convergence.

[0075] The Adam optimizer is used to optimize the neural network weights, minimize the loss function, and train the model. At the same time, regularization is used to prevent network overfitting, and gradient clipping is introduced to avoid gradient explosion.

[0076] Perform adaptive time step dynamic adjustment, and calculate the error value and gradient change rate of each time step during model training. If the error is large and changes dramatically, reduce the time step to capture a more accurate state; otherwise, increase the time step appropriately to improve calculation efficiency.

[0077] The model is validated on different boundary conditions and data to ensure that its predicted results conform to physical laws. Small random perturbations are imposed on the input data to verify the stability and robustness of the model's prediction results. The model is also tested under different geological conditions and parameter changes to ensure its applicability in a wide range of engineering scenarios.

[0078] The grouting diffusion characterization model is used to predict the diffusion state of the slurry in different media, assist in determining the optimal grouting position, pressure and flow rate, monitor the grouting parameters in real time and use the grouting diffusion characterization model for prediction, dynamically adjust the pressure and flow rate to achieve the optimal grouting effect, and realize intelligent grouting control.

[0079] Embodiment 2

[0080] This embodiment provides a grouting slurry diffusion morphology characterization system based on deep learning, including:

[0081] An acquisition module, which is configured to: acquire the diffusion morphology data of the slurry under different geological conditions, preprocess the acquired data, and use the preprocessed data as a training set;

[0082] A construction module, which is configured to: construct a grouting diffusion characterization model. The grouting diffusion characterization model includes a spatial convolutional layer and a temporal convolutional layer. The spatial convolutional layer is used to extract spatial features; the temporal convolutional layer is used to extract temporal features;

[0083] A training module, which is configured to: train the grouting diffusion characterization model based on the training set to obtain a trained grouting diffusion characterization model; wherein, the total loss function used in the training process is the weighted sum result of the data loss function, the physical constraint loss function, the boundary condition loss function, the initial condition loss function, and the actual experience loss function;

[0084] An output module, which is configured to: input the temperature and pressure data of the grouting area to be characterized into the trained grouting diffusion characterization model to obtain the predicted values of the grouting diffusion flow rate and the grouting diffusion distance.

[0085] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A grouting slurry diffusion morphology characterization method based on deep learning, characterized by: include: Obtain the diffusion morphology data of slurry under different geological conditions, preprocess the acquired data, and use the preprocessed data as a training set; Construct a grouting diffusion characterization model, which includes a spatial convolution layer and a temporal convolution layer. The spatial convolution layer is used to extract spatial features. The temporal convolution layer is used to extract temporal features; The grouting diffusion characterization model is trained based on the training set to obtain the trained grouting diffusion characterization model; wherein the total loss function used in the training process is the weighted sum of the data loss function, the physical constraint loss function, the boundary condition loss function, the initial condition loss function and the actual experience loss function; The temperature and pressure data of the grouting area to be characterized are input into the trained grouting diffusion characterization model to obtain the predicted values ​​of the grouting diffusion velocity and grouting diffusion distance.

2. The grouting slurry diffusion morphology characterization method based on deep learning as claimed in claim 1, characterized in that: A grouting diffusion characterization model is constructed, which includes a spatial convolution layer and a temporal convolution layer. The spatial convolution layer is used to extract spatial features; the temporal convolution layer is used to extract temporal features. The grouting diffusion characterization model is implemented using a three-dimensional convolutional neural network. The three-dimensional convolutional neural network has a network structure including: an input layer, a convolution layer, a spatial convolution layer, a temporal convolution layer and an output layer connected in sequence.

3. The grouting slurry diffusion morphology characterization method based on deep learning as claimed in claim 1, characterized in that: The spatial convolution layer works as follows: the value at (x, y) on the γth feature map of the βth convolution layer It is expressed as: in, is the activation function, b β,γ is the added bias, N is the total number of feature map inputs in the β-1th layer connected to the current feature map, is the value of the kernel at (φ, ψ), Φ and Ψ are the height and width of the convolution kernel respectively.

4. The grouting slurry diffusion morphology characterization method based on deep learning as claimed in claim 1, characterized in that: The working process of the temporal convolution layer includes: assuming a one-dimensional time series T is input, its convolution operation can be expressed as: O(t)=∑T(t,a)·W(a); O(t) represents the output feature after convolution, T(t,a) represents the data in the time series, and W(a) represents the convolution weight value.

5. The grouting slurry diffusion morphology characterization method based on deep learning as claimed in claim 1, characterized in that: The total loss function is the weighted sum of the data loss function, the physical constraint loss function, the boundary condition loss function, the initial condition loss function and the actual experience loss function; wherein, the total loss function is expressed as follows: L t =a d (t s )L d +a p (t s )L p +a b (t s )L b +a i (t s )L i +a en (t s )L e ; Among them, L t is the total loss function, α d represents the weight of the data loss function, α p represents the weight of the physical constraint loss function, α b represents the weight of the boundary condition loss function, α i represents the weight of the initial condition loss function, α e represents the weight of the actual empirical loss function, t s represents the time step; n represents different empirical losses under different medium conditions, where θ represents different media, γ θ Indicates the weight coefficient of different media effects; m corresponds to five loss functions d, p, b, i, and e respectively. There is a floating threshold for the weight of each loss function, expressed as α m (t s )∈[m1,m2],α m (t s ) represents the loss weight at each time step, L M Indicates the sum of the loss function values ​​at each time step, L M =L d +L p +L b +L i +L e .

6. The grouting slurry diffusion morphology characterization method based on deep learning as claimed in claim 1 or 5, characterized in that: The data loss function is expressed as: Among them, L d is the data loss function, N d is the number of training samples, Represents the corresponding true value, y j Represents the predicted value of the model, j represents the jth training sample; The physical constraint loss function is expressed as: Among them, L p is the physical constraint loss function, N p is the number of physical constraint sampling points, represents the differential operation at the sampling point, f o (x o ,t o ) represents the source term of the physical equation.

7. The grouting slurry diffusion morphology characterization method based on deep learning as claimed in claim 1 or 5, characterized in that: The boundary condition loss function is expressed as: Among them, L b represents the boundary condition loss function, N b represents the number of boundary points, represents the boundary condition, v(x k ,t k ) represents the predicted value on the boundary.

8. The method for characterizing the diffusion morphology of grouting slurry based on deep learning according to claim 1 or 5, characterized in that: The initial value condition loss function is expressed as: Among them, L i represents the initial condition loss function, N i represents the number of initial condition sampling points, represents the initial condition function, w(x l ,0) represents the predicted value under the initial conditions.

9. The grouting slurry diffusion morphology characterization method based on deep learning as claimed in claim 1 or 5, characterized in that: The actual experience loss function is expressed as: Among them, L e is the actual experience loss, N e is the number of sampling points constrained by the actual empirical loss, represents the miscibility loss measure; ε(x q ,t q ) is the porosity change, which is the porosity difference between the predicted and actual porosity at the sampling point, μ(x q ,t q ) is the change in slurry viscosity; v q is the velocity change, β is the adjustment coefficient, r is the pressure coefficient, is the pressure gradient, is the gradient operator; Where T is the total time, c in is the slurry injection concentration, c out (t) is the predicted concentration of the slurry over time; Among them, μ yc (t) is the predicted viscosity of the slurry over time, μ sj (t) is the actual viscosity of the slurry obtained over time.

10. A grouting slurry diffusion morphology characterization system based on deep learning, characterized by: include: An acquisition module is configured to: acquire diffusion morphology data of slurry under different geological conditions, preprocess the acquired data, and use the preprocessed data as a training set; A construction module is configured to: construct a grouting diffusion characterization model, the grouting diffusion characterization model includes a spatial convolution layer and a temporal convolution layer, the spatial convolution layer is used to extract spatial features; the temporal convolution layer is used to extract temporal features; A training module is configured to: train the grouting diffusion characterization model based on the training set to obtain the trained grouting diffusion characterization model; wherein the total loss function used in the training process is a weighted sum of a data loss function, a physical constraint loss function, a boundary condition loss function, an initial condition loss function, and an actual experience loss function; The output module is configured to: input the temperature and pressure data of the grouting area to be characterized into the trained grouting diffusion characterization model to obtain the predicted values ​​of the grouting diffusion flow rate and the grouting diffusion distance.

Images