Method, device, equipment and storage medium for rapid reconstruction of soil freeze-thaw hydrothermal field

Abstract

The present invention discloses a method, device, equipment and storage medium for rapid reconstruction of soil freeze-thaw hydrothermal field, the method comprising: establishing a one-dimensional soil column frost heave-thaw settlement calculation model through finite element software, simulating and calculating to obtain a one-dimensional freeze-thaw transient temperature field and moisture field of the soil, and obtaining a source domain data set containing position, time, temperature and saturation; using the position and time in the source domain data set as input, and the temperature and saturation as output, to establish a deep neural network, the loss function includes data loss and physical loss, and involves both the temperature field and the moisture field; training the source task model to obtain the neural network parameters; training the target task model. Based on numerical simulation and deep transfer learning, the present invention solves the multi-field coupling problem in the context of machine learning, and at the same time improves the generalization performance of the model, thereby realizing the rapid reconstruction of the transient temperature field and moisture field of the one-dimensional frost heave-thaw settlement test of the soil under sparse measuring point conditions.

Description

Technical Field

[0001] The present invention belongs to the technical field of multi-physical field reconstruction, and relates to a method, device, equipment and storage medium for rapid reconstruction of soil freeze-thaw hydrothermal field. Background Art

[0002] At present, the measuring instruments for temperature and unfrozen water content in one-dimensional frost heave-thaw settlement test of open system soil are all arranged in a point-like manner, and only partial measurement point data can be obtained, and the continuous distribution of temperature and moisture cannot be known. However, mastering the hydrothermal coupling process involving water migration and phase change is the key to studying the characteristics of frozen soil, so it is crucial to obtain the continuous and complete distribution characteristics of soil temperature field and moisture field. Thanks to the rapid development of machine learning, the physical field reconstruction method based on pure data drive has shown good results. However, when the number of known measurement points is small, its ill-posedness will cause the physical field reconstructed only by data to differ greatly from the real physical field; neural networks with embedded physical information have higher interpretability than traditional neural networks, but when faced with multi-physical field problems, the training cost is high and the dependence on initial values ​​is extremely strong; the existing single physical field reconstruction method based on transfer learning and physical constraints will face the problem of transient calculation of multi-physical fields and variable boundary conditions. Due to the increased complexity of the model, the network parameters obtained by the source model training of the existing method will face the risk of failure. Therefore, how to achieve rapid reconstruction of the one-dimensional freeze-thaw transient temperature field and moisture field of soil under sparse measurement point conditions is an urgent problem to be solved. Summary of the invention

[0003] In order to solve the above problems, the present invention provides a method, device, equipment and storage medium for rapid reconstruction of the freeze-thaw hydrothermal field of soil. The method is based on numerical simulation and deep transfer learning, which solves the multi-field coupling problem in the context of machine learning and improves the generalization performance of the model, thereby realizing the rapid reconstruction of the transient temperature field and moisture field of the soil in a one-dimensional freeze-thaw settlement test under sparse measuring point conditions.

[0004] The technical solution adopted by the present invention is a method for rapid reconstruction of soil freeze-thaw hydrothermal field, comprising the following steps:

[0005] S1: Source domain dataset construction;

[0006] A one-dimensional soil column frost heave-thaw settlement calculation model was established using finite element software. Time-varying temperature boundary conditions and water replenishment boundary conditions were applied. The one-dimensional freeze-thaw transient temperature field and moisture field of the soil were obtained through simulation calculation. The temperature and moisture saturation of each node were extracted to obtain a source domain data set containing position, time, temperature and saturation.

[0007] S2: training source task model;

[0008] Taking the location and time in the source domain data set as input and temperature and saturation as output, a deep neural network is established to automatically solve the water-thermal coupling partial differential equation through the embedded physical information of the deep neural network; the loss function includes data loss and physical loss, involving both temperature field and moisture field, and the physical loss is closely related to the water-thermal coupling equation, initial conditions and boundary conditions; the source task model is trained to obtain the neural network parameters;

[0009] S3: training target task model;

[0010] The trained neural network parameters are transferred to the neural network model of the target task as the initial values; the neural network parameters of the middle layer are frozen, and only the first and last network layers are retrained. The structure and training method of the neural network model are the same as those of the source task; after the training is completed, the coordinate position and time are input to realize the rapid reconstruction of the one-dimensional freeze-thaw transient temperature field and moisture field of the soil under sparse measurement point conditions.

[0011] Furthermore, in S2, the number of network layers is set to 8, of which the number of hidden layers is 6, and each hidden layer contains 20 neurons; the activation function σ Select Tanh or ReLU function; the learning rate is adjusted using a fractional mitigation mechanism, and the learning rate gradually decreases as the number of iterations increases; the gradient descent method is used to update the neural network model parameters.

[0012] Furthermore, the loss function of the data loss is loss r for:

[0013] ;

[0014] Where: is the number of measuring points, is the number of time points; for The predicted value of soil temperature at time is a rich dataset from the source domain Real temperature at all times; for The predicted value of soil saturation at time for The real saturation of the moment; Indicates the location coordinates of the soil;

[0015] It represents the temperature loss term in the data loss during source model training, which is used to measure the error between the predicted value and the true value of soil temperature; It represents the moisture loss term in the data loss, which is used to measure the error between the predicted value and the true value of soil saturation.

[0016] Furthermore, the physical loss associated with the water-heat coupling equation is expressed by the physical equation residual loss function loss PDE1 express:

[0017] ;

[0018] The physical loss associated with the boundary conditions is expressed by the boundary condition loss function express:

[0019] ;

[0020] The physical loss associated with the initial condition is expressed by the initial condition loss function express:

[0021] ;

[0022] Where: and Respectively represent the number of sampling points of the physical equation loss term, boundary condition loss term and initial condition loss term generated by the random algorithm; Due to the hydrothermal coupling of frozen-thaw soil, there are two physical field control equations, so in the formula That is, the residual of the temperature field control equation and the residual of the moisture field control equation ; and are the pre-given model temperature and moisture boundary conditions respectively; is the initial value of soil temperature, is the initial value of soil saturation, and are the temperature and moisture initial conditions of the model, respectively;

[0023] for The predicted value of soil temperature at time for Predicted value of soil saturation at time t;

[0024] Loss function for model training loss 1 is represented by:

[0025] ;

[0026] Where: is the loss weight; It represents the temperature loss term in the data loss during source model training, which is used to measure the error between the predicted value and the true value of soil temperature; It represents the moisture loss term in the data loss, which is used to measure the error between the predicted value and the true value of soil saturation.

[0027] Furthermore, the residual of the temperature field control equation See the following formula:

[0028] ;

[0029] The residual of the moisture field control equation See the following formula:

[0030] ;

[0031] Hydrothermal coupling term: ;

[0032] Saturation S Unfrozen water content The relationship is: ;

[0033] in, is the density; is the specific heat capacity that varies with volumetric water content; It is the temperature T About Time t The derivative of is the thermal conductivity that varies with moisture content; is a differential operator, which for two-dimensional problems is in x , y Indicates the node location; L is the latent heat of phase change; is the volumetric ice content, is the unfrozen water content; is the density of ice, is the density of water; is the diffusion rate of water; is the soil permeability;

[0034] represents the solid-to-liquid ratio, is the freezing temperature; is a constant that varies with soil quality and salt content;

[0035] is the saturated water content; is the residual water content;

[0036] The residuals of the temperature field control equation and the moisture field control equation are connected through the hydrothermal coupling term, and the saturation is selected. S replace By combining the hydrothermal coupling term and training the neural network, the residuals of the temperature field control equation and the moisture field control equation are made close to 0, making the output data closer to the solution of the hydrothermal coupling partial differential equation.

[0037] Furthermore, the adaptive adjustment method of the loss weight is as follows: initialize the loss weight of each part to 1; then forward propagate to calculate the loss function value of each part and the proportion of each part in the total loss; redistribute the loss weight using the proportion as the loss weight adjustment coefficient; finally calculate the total loss function, and repeat this cycle to achieve adaptive adjustment of the loss weight.

[0038] Furthermore, in S3, the learning rate of the target task model is smaller than the learning rate of the source task model.

[0039] A soil freeze-thaw water thermal field rapid reconstruction device, using the soil freeze-thaw water thermal field rapid reconstruction method, comprising:

[0040] The source domain dataset construction module is used to establish a one-dimensional soil column frost heave-thaw settlement calculation model through finite element software to obtain a source domain dataset containing location, time, temperature and moisture information;

[0041] The source task model construction and training module is used to build a deep neural network model with the location and time in the source domain data set as input and the temperature and moisture information as output; automatically solve the water-heat coupling partial differential equation through the physical information embedded in the deep neural network; train the source task model to obtain the neural network parameters;

[0042] The training module of the target task model is used to transfer the neural network parameters to the neural network model of the target task as the initial values; the neural network parameters of the middle layer are frozen, and only the first and last network layers are retrained.

[0043] An electronic device, comprising:

[0044] a memory for storing instructions executable by a processor; and

[0045] The processor is used to execute the instructions to implement the above-mentioned method for rapid reconstruction of soil freeze-thaw water thermal field.

[0046] A computer storage medium storing computer program codes, wherein the computer program codes, when executed by a processor, implement the above-mentioned method for rapid reconstruction of soil freeze-thaw hydrothermal field.

[0047] The beneficial effects of the present invention are:

[0048] The present invention firstly uses numerical simulation to obtain the temperature and moisture data of each node in the model, constructs a rich source domain data set containing temperature and moisture information, and then uses coordinates to x , y ,time t As input, temperature T , Saturation SAs the output, a deep neural network is established to train the source task model. This method is based on the loss function in the transient calculation of hydrothermal coupled multi-physics field, and uses the loss weight adaptive adjustment mechanism to calculate the total loss. The object of the target task is the soil freeze-thaw test with sparse measurement point data. This method uses a neural network with the same structure as the source task to train the target task model, and uses the network structure parameters trained on the source task to improve the learning effect of the target task. Since the source task has fully learned the distribution characteristics of the one-dimensional frost heave temperature field and moisture field of the soil, for the target task with similar physical processes, there is no need for cumbersome numerical reproduction and time-consuming conventional neural network prediction. This method can be used to achieve rapid reconstruction of hydrothermal coupled multi-physics fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0050] Figure 1 This is the source task model training process of the embodiment of the present invention.

[0051] Figure 2 This is the adaptive loss weight adjustment process of the embodiment of the present invention.

[0052] Figure 3 This is the target task model training process of the embodiment of the present invention. DETAILED DESCRIPTION

[0053] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0054] Example 1

[0055] A method for rapid reconstruction of soil freeze-thaw hydrothermal field, comprising the following steps:

[0056] S1: Source domain dataset construction;

[0057] Comsol Multiphysics finite element software was used to establish a one-dimensional soil column frost heave-thaw settlement calculation model, and time-varying temperature boundary conditions and water replenishment boundary conditions were applied. The one-dimensional freeze-thaw transient temperature field and moisture field distribution of the soil were obtained through simulation calculation, and a large amount of training data was obtained. The temperature and moisture data of each node position were extracted to form a rich source domain data set containing position-time-temperature-moisture information (saturation S), as shown in Table 1.

[0058] Table 1 Source domain rich dataset

[0059]

[0060] S2: training source task model;

[0061] like Figure 1 As shown, based on the source domain dataset with coordinates x , y ,time t As input, temperature T , Saturation S As output, a deep neural network is established to train the source task model. The number of network layers is set to 8, of which the number of hidden layers is 6, and each hidden layer contains 20 neurons; the activation function σ Select Tanh or ReLU function; the learning rate adjustment adopts the fractional slowdown mechanism, and the learning rate gradually decreases with the increase of the number of iterations to accelerate convergence, prevent falling into the local optimum, improve training stability, and improve the final model performance; the gradient descent method is used to update the neural network model parameters. The number of neural network layers and neurons can be increased or decreased according to the decrease of the loss function when facing different models; Tanh (or Relu) is usually sufficient for general problems.

[0062] I Represents the output of the model, which is the temperature that needs to be solved in the equation T , Saturation S .

[0063] The input of the deep neural network model (coordinate x , y ,time t ) and output (temperature T , Saturation S ) is the mapping relationship between T ( x , y , t ), S ( x , y , t ). Equivalent to temperature T , SaturationS is a coordinate x , y ,time t function of different locations and times T , S A more specific connection is found in the equations of physics.

[0064] The loss function uses the mean square error loss (MSE), which consists of two parts and involves both the temperature field and the moisture field. One part is the data loss, which comes from the error between the predicted value and the true value. The data loss function loss r for:

[0065] (1)

[0066] Where: To enrich the number of measurement points, is the number of time points; for Time soil Temperature prediction value, is a rich dataset from the source domain Real temperature at all times; for The predicted value of soil saturation at time for The true saturation of the moment.

[0067] It represents the temperature loss term in the data loss during source model training, which is used to measure the error between the predicted value and the true value of soil temperature; Represents the moisture loss term in the data loss, which is used to measure the error between the predicted value and the true value of soil saturation. Both reflect the gap between the predicted value and the true value, and together constitute the data loss, participating in the training of the neural network. The data loss is combined with the subsequent physical loss to form the total loss.

[0068] i = 1, 2, 3, … , loop variable representing the measuring point, k = 1, 2, 3, … , A loop variable representing time. It can be expressed as i Measurement point, k a point in time.

[0069] The other part is physical loss, which is closely related to the hydrothermal coupling equation, initial conditions and boundary conditions. The loss term related to the hydrothermal coupling equation includes the residual of the temperature field control equation and the residual of the moisture field control equation, and the two parts are connected by the hydrothermal coupling term.

[0070] The water-heat coupling equation of frozen soil is:

[0071] (2)

[0072] in, is the density; is the specific heat capacity that varies with volumetric water content; It is the temperature T About Time t The derivative of is the thermal conductivity that varies with moisture content; is a differential operator, which for two-dimensional problems is ( x , y Indicates the node position); L is the latent heat of phase change; generally taken as 334.56 kJ / kg; is the volumetric ice content, is the unfrozen water content; is the density of ice, is the density of water; is the diffusion rate of water; is the soil permeability;

[0073] Hydrothermal coupling term: ;

[0074] Where: is the freezing temperature; B is a constant that varies with soil quality and salt content. Represents the solid-liquid ratio.

[0075] Saturation S Unfrozen water content The relationship is: ;

[0076] Where: is the saturated water content; is the residual water content.

[0077] Residual of temperature field control equation See the following formula:

[0078] ;

[0079] Residual of moisture field control equation See the following formula:

[0080] ;

[0081] Select Saturation S replace , combined with the hydrothermal coupling term, by training the neural network , Close to 0, the output data is closer to the solution of the physical equation.

[0082] The loss term associated with the hydrothermal coupling equation, i.e., the residual loss function of the physical equation loss PDE1 , see formula (3):

[0083] (3)

[0084] The loss term related to the boundary conditions, that is, the boundary condition loss function loss b1 , see formula (4):

[0085] (4)

[0086] With the initial conditions ( t = 0) related loss term, that is, the initial condition loss function loss I1 , see formula (5):

[0087] (5)

[0088] Where: N p1 , N b1 and N I1 They represent the number of sampling points of the physical equation loss term, boundary condition loss term and initial condition loss term generated by the random algorithm. Since there are two physical field control equations in the hydrothermal coupling of frozen-thaw soil, That is, the residual of the temperature field control equation and the residual of the moisture field control equation ; and are the pre-given model temperature and moisture boundary conditions respectively; is the initial value of soil temperature, is the initial value of soil saturation, and are the temperature and moisture initial conditions of the model, respectively.

[0089] for Moment soil ( x i , y i ) Temperature prediction value, for Moment soil ( xi , y i ) Saturation prediction value.

[0090] loss PDE1 , loss b1 , loss I1 The three together constitute the physical loss. In summary, the loss function loss 1 is represented by:

[0091] (6)

[0092] Where: is the loss weight.

[0093] The loss weight is adopted as follows Figure 2 The adaptive adjustment strategy shown in the figure. First, initialize the loss weight of each part to 1, then perform forward propagation to calculate the loss function value of each part and its proportion of the total loss, and use the proportion as the loss weight adjustment coefficient to redistribute the loss weight, and finally calculate the total loss function, so as to realize the adaptive adjustment of the loss weight in this cycle.

[0094] In the embodiment of the present invention, the residual term of the frozen soil hydrothermal coupling equation, the loss term of the boundary conditions and the initial conditions are embedded into the neural network as physical constraints, and the loss function is formed together with the data loss. The specific form of the loss function in the case of frozen soil hydrothermal coupling is clarified, and a weight adjustment mechanism based on the loss ratio is adopted, so that the model pays more attention to the loss terms with a large loss ratio, which will increase the penalty for large loss terms, and quickly optimize the large loss terms so that the internal points of the model satisfy the physical equation as much as possible. The loss function adopts the form of mean square error (MSE). In the embodiment of the present invention, the loss function, i.e., equation (6), is minimized through the gradient descent algorithm, thereby approaching the true solution. After the training is completed, the final neural network parameters (weights and biases) are obtained.

[0095] Figure 2 middle, Indicates j In the training iteration i The weight adjustment coefficient of the loss term, Indicates j In the iteration i The loss function value, Indicates j +1 iteration i The weight of the loss term, Indicates i The initial weights of the loss terms, Indicates j In the iteration iThe weight of the loss term, Indicates j +The total loss function value for 1 iteration.

[0096] S3: training target task model;

[0097] The target task is to reconstruct the one-dimensional soil column freeze-thaw transient temperature field and moisture field under sparse measurement point conditions. This method uses deep transfer learning. The neural network model of the target task adopts the same structure as the source task model, and the neural network parameters (weights and biases) obtained by source task training are transferred to the target task as its initial values, such as Figure 3 shown.

[0098] Since the middle layer of the neural network reflects the general law of the hydrothermal coupling process, out of comprehensive consideration of training cost and accuracy, the neural network parameters of the middle layer are frozen, and only the first and last network layers are retrained, and the training method is the same as the source task. In addition, due to the adoption of a model-based migration strategy, the learning rate of the target task model is appropriately reduced to prevent over-updates, maintain useful features of pre-training, help the weights of the training layer slowly adapt to the needs of the target task, and maintain the stability of the underlying features. The degree of reduction is determined by factors such as the number of measurement points of the target model and the changes in the loss function.

[0099] The loss function is composed as follows:

[0100] (7)

[0101] Where: N p is the number of sparse measurement points in the one-dimensional soil column freeze-thaw test; for t k Moment soil ( x j , y j ) Temperature prediction value, is the actual measuring point temperature from the one-dimensional soil column freeze-thaw test; for t k The predicted value of soil saturation at time for t k The true saturation of the moment.

[0102] It represents the temperature loss term in the data loss during the training of the target model, which is used to measure the error between the predicted value and the true value of the soil temperature; Represents the moisture loss term in the data loss, which is used to measure the error between the predicted value and the true value of soil saturation. The two together constitute the data loss of the target model. . With the above The difference lies in the different objects. The former object is the source task model, and the measurement points are abundant; the object here is the target task model, and the measurement points are sparse.

[0103] (8)

[0104] (9)

[0105] (10)

[0106] N p2 , N b2 and N I2 Respectively represent the number of sampling points of the physical equation loss term, boundary condition loss term and initial condition loss term generated by the random algorithm in the target model, and the collected point coordinates ( x j , y j ) is used as the input of the target model. N p1 , N b1 and N I1 The corresponding number is the number of sampling points in the source model. Both can be generated by the same random algorithm or random seed.

[0107] That is, the residual of the temperature field control equation and the residual of the moisture field control equation ; is the initial value of the soil temperature of the target model, is the initial value of soil saturation of the target model, g T2 and g S2 are the model temperature and moisture boundary conditions pre-given for the target model respectively; h T2 and h S2 are the temperature and moisture initial conditions of the target model, respectively. Their form and physical meaning are consistent with those in the source model.

[0108] The final form of the loss function is:

[0109] (11)

[0110] The meanings of the subscripts in the formula are consistent with those in the previous text.

[0111] After the training is completed, the neural network parameters (weights and biases) of the target task model that characterize the general process of soil hydrothermal coupling are obtained. By inputting the coordinate position and time, the one-dimensional freeze-thaw transient temperature field and moisture field of the soil under sparse measurement point conditions can be quickly reconstructed.

[0112] The innovation of the embodiment of the present invention compared to the existing method is that it constructs a specific form of the loss function in the case of transient calculation of soil hydrothermal coupling multi-physics field. By training the deep neural network, the soil hydrothermal coupling process can be learned, and then the reconstruction of the soil hydrothermal coupling multi-physics field can be realized. Even if the boundary conditions and initial conditions of the target task model change, since the physical process is always the coupling of soil temperature and moisture, the source model has been fully learned, that is, the loss function is reduced through training, and the back propagation is used to update the neural network parameters. At the same time, the loss weight adaptive adjustment mechanism is introduced, so the source model parameters can still be used to quickly realize the reconstruction of the soil hydrothermal field, which improves efficiency and accuracy and further enhances the generalization ability of the model. It can be applied to the rapid reconstruction of the hydrothermal field of frozen soil in cold regions under any conditions.

[0113] The embodiment of the present invention realizes efficient and accurate reconstruction of one-dimensional soil freeze-thaw transient temperature field and moisture field by integrating numerical simulation, supervised learning, physical constraints, loss weight adaptive adjustment mechanism and transfer learning. Strong supervised learning is introduced into the source task model while physical information is embedded (strong supervised learning corresponds to formula (1), and physical information constraints correspond to formulas (3)-(5)), ensuring that the neural network can accurately reflect the complex hydrothermal coupling physical process of the soil. Subsequently, deep transfer learning technology is used to fine-tune the target task. This process effectively utilizes the knowledge accumulation in the source task, reduces the sample size and time required for target task training, and enhances the adaptability and generalization ability of the neural network model. Even under the constraints of sparse measurement point data, the reconstruction of the soil hydrothermal field can be completed quickly and accurately.

[0114] The embodiment of the present invention automatically solves the hydrothermal coupling partial differential equation through the physical information embedded in the deep neural network (PINN), and forward solves the soil temperature field and moisture field; due to the increase in the dimension of the physical field, the loss function becomes more complex, which in turn increases the difficulty of solving. The contribution of each loss to the total loss is different and constantly changes during training. It is inappropriate to use fixed loss weights or manual parameter adjustment methods. In this application, the stability of the model calculation is ensured by the adaptive adjustment mechanism of the loss weights, and the calculation efficiency is improved by transfer learning (TL).

[0115] The PDE term of the loss function in the PINN model in the embodiment of the present invention corresponds to a hydrothermal coupling multi-physics field equation, which has many parameters and involves coupling terms, and has a relatively complex structure.

[0116] Example 2

[0117] A soil freeze-thaw hydrothermal field rapid reconstruction device, comprising:

[0118] The source domain dataset construction module is used to establish a one-dimensional soil column frost heave-thaw settlement calculation model through finite element software to obtain a source domain dataset containing location, time, temperature and moisture information;

[0119] The source task model construction and training module is used to build a deep neural network model with the location and time in the source domain data set as input and the temperature and moisture information as output; automatically solve the water-heat coupling partial differential equation through the physical information embedded in the deep neural network; train the source task model to obtain the neural network parameters;

[0120] The training module of the target task model is used to transfer the neural network parameters to the neural network model of the target task as the initial values; the neural network parameters of the middle layer are frozen, and only the first and last network layers are retrained, and the training method is the same as that of the source task.

[0121] Example 3

[0122] Corresponding to the above method embodiment, this embodiment also provides a soil freeze-thaw water thermal field rapid reconstruction device. The soil freeze-thaw water thermal field rapid reconstruction device described below and the soil freeze-thaw water thermal field rapid reconstruction method described above can refer to each other.

[0123] Example 4

[0124] Corresponding to the above method embodiment, this embodiment further provides a readable storage medium. The readable storage medium described below and the soil freeze-thaw hydrothermal field rapid reconstruction method described above can correspond to each other.

[0125] A readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the soil freeze-thaw hydrothermal field rapid reconstruction method of the above method embodiment.

[0126] If the soil freeze-thaw water thermal field rapid reconstruction method described in the embodiment of the present invention is implemented in the form of a software function module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the soil freeze-thaw water thermal field rapid reconstruction method described in the embodiment of the present invention. The aforementioned readable storage specific medium can be a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk, and other readable storage media that can store program codes.

[0127] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. A method for rapid reconstruction of soil freeze-thaw hydrothermal field, characterized in that: The following steps are involved: S1: Source domain dataset construction; A one-dimensional soil column frost heave-thaw settlement calculation model was established using finite element software. Time-varying temperature boundary conditions and water replenishment boundary conditions were applied. The one-dimensional freeze-thaw transient temperature field and moisture field of the soil were obtained through simulation calculation. The temperature and moisture saturation of each node were extracted to obtain a source domain data set containing position, time, temperature and saturation. S2: training source task model; Taking the location and time in the source domain data set as input and temperature and saturation as output, a deep neural network is established to automatically solve the hydrothermal coupling partial differential equation through the embedded physical information of the deep neural network. The loss function includes data loss and physical loss, involving both temperature field and moisture field. The physical loss is closely related to the hydrothermal coupling equation, initial conditions and boundary conditions. Train the source task model to obtain the neural network parameters; The physical loss associated with the water-heat coupling equation is expressed by the physical equation residual loss function loss PDE1 express: N p1 represents the number of sampling points of the loss term of the physical equation generated by the random algorithm; since there are two physical field control equations for the hydrothermal coupling of frozen-thaw soil, n = 1, 2; f n (x i ,y i ,t i ) is the residual f1(x i ,y i ,t i ) and the residual f2(x i ,y i ,t i ); The residual f1(x i ,y i ,t i ) See the following formula: The residual error f2(x i ,y i ,t i ) See the following formula: Hydrothermal coupling term: Saturation S and unfrozen water content θ u The relationship is: Where ρ is the density; c(θ) is the specific heat capacity that varies with volumetric water content; is the derivative of temperature T with respect to time t; λ(θ) is the thermal conductivity that varies with moisture content; is a differential operator, which for two-dimensional problems is Where x and y represent the node positions; L is the latent heat of phase change; θ I is the volumetric ice content, θ u is the unfrozen water content; ρ I is the ice density, ρ w is the density of water; D(θ u ) is the diffusion rate of water; k(θ u ) is the soil permeability; B(T) represents the solid-liquid ratio, T f is the freezing temperature; B is a constant that changes with soil quality and salt content; θ s is the saturated water content; θ r is the residual water content; The residual of the temperature field control equation and the residual of the moisture field control equation are connected through the hydrothermal coupling term, and the saturation S is selected instead of θ u , combined with the hydrothermal coupling term, the residuals of the temperature field control equation and the moisture field control equation are made close to 0 by training the neural network, so that the output data is closer to the solution of the hydrothermal coupling partial differential equation; S3: training target task model; The trained neural network parameters are transferred to the neural network model of the target task as the initial values; the neural network parameters of the middle layer are frozen, and only the first and last network layers are retrained. The structure and training method of the neural network model are the same as those of the source task; after the training is completed, the coordinate position and time are input to realize the rapid reconstruction of the one-dimensional freeze-thaw transient temperature field and moisture field of the soil under sparse measurement point conditions.

2. According to claim 1, a method for rapid reconstruction of soil freeze-thaw water thermal field, characterized in that: In S2, the number of network layers is set to 8, of which the number of hidden layers is 6, and each hidden layer contains 20 neurons; the activation function σ selects Tanh or ReLU function; the learning rate adjustment adopts a fractional mitigation mechanism, and the learning rate gradually decreases with the increase in the number of iterations; the gradient descent method is used to update the neural network model parameters.

3. According to claim 1, a method for rapid reconstruction of soil freeze-thaw hydrothermal field, characterized in that: The loss function loss of the data loss r for: Where: N r is the number of measuring points, N t is the number of time points; T(x i ,y i ,t k ) is t k Predicted soil temperature at time T i k is t from the rich dataset in the source domain k The actual temperature at the moment; S(x i ,y i ,t k ) is t k The predicted value of soil saturation at time t k The actual saturation at the moment; (x i ,y i ) represents the position coordinates of the soil body; It represents the temperature loss term in the data loss during source model training, which is used to measure the error between the predicted value and the true value of soil temperature; It represents the moisture loss term in the data loss, which is used to measure the error between the predicted value and the true value of soil saturation.

4. According to claim 1, a method for rapid reconstruction of soil freeze-thaw hydrothermal field, characterized in that: The physical loss associated with the boundary condition is expressed by the boundary condition loss function loss b1 express: The physical loss associated with the initial condition is expressed by the initial condition loss function loss I1 express: Where: N b1 With N I1 They represent the number of sampling points of the boundary condition loss term and the initial condition loss term generated by the random algorithm; g T1 With g S1 are the pre-given model temperature and moisture boundary conditions; T(x i ,y i ,0) is the initial value of soil temperature, S(x i ,y i ,0) is the initial value of soil saturation, h T1 With h S1 are the temperature and moisture initial conditions of the model, respectively; T(x i ,y i ,t i ) is t i The predicted value of soil temperature at time S(x i ,y i ,t i ) is t i Predicted value of soil saturation at time t; The loss function loss1 of model training is expressed as: Where: λ1, λ2, λ3, λ4, λ5 are loss weights; It represents the temperature loss term in the data loss during source model training, which is used to measure the error between the predicted value and the true value of soil temperature; It represents the moisture loss term in the data loss, which is used to measure the error between the predicted value and the true value of soil saturation.

5. According to claim 4, a method for rapid reconstruction of soil freeze-thaw hydrothermal field is characterized in that: The adaptive adjustment method of the loss weight is as follows: initialize the loss weight of each part to 1; then forward propagate to calculate the loss function value of each part and the proportion of each part in the total loss; redistribute the loss weight by using the proportion as the loss weight adjustment coefficient; finally calculate the total loss function, and repeat this cycle to achieve adaptive adjustment of the loss weight.

6. According to claim 1, a method for rapid reconstruction of soil freeze-thaw hydrothermal field, characterized in that: In S3, the learning rate of the target task model is smaller than the learning rate of the source task model.

7. A soil freeze-thaw water thermal field rapid reconstruction device, characterized in that: A method for rapid reconstruction of soil freeze-thaw hydrothermal field as claimed in any one of claims 1 to 6 is adopted, comprising: The source domain data set construction module is used to establish a one-dimensional soil column frost heave-thaw settlement calculation model through finite element software to obtain a source domain data set containing location, time, temperature and saturation; The source task model construction and training module is used to build a deep neural network model with the location and time in the source domain data set as input and temperature and saturation as output; automatically solve the water-heat coupling partial differential equation through the embedded physical information of the deep neural network; train the source task model to obtain the neural network parameters; The training module of the target task model is used to transfer the neural network parameters to the neural network model of the target task as the initial values; the neural network parameters of the middle layer are frozen, and only the first and last network layers are retrained.

8. An electronic device, characterized in that: include: a memory for storing instructions executable by a processor; as well as A processor is used to execute the instructions to implement a method for rapid reconstruction of soil freeze-thaw hydrothermal field as described in any one of claims 1-6.

9. A computer storage medium storing computer program code, characterized in that: When the computer program code is executed by a processor, the computer program code implements a method for rapid reconstruction of soil freeze-thaw hydrothermal field as described in any one of claims 1 to 6.

Images