A method for predicting the instability of soft rock embankments based on physical information neural networks.
By constructing a method based on physical information neural networks, dynamically updating constitutive equations and embedding them into loss functions, the problems of data sparsity and physical inconsistency in the prediction of instability of soft rock embankments are solved, achieving high-precision prediction with strong generalization ability and providing a scientific dynamic early warning mechanism.
Patent Information
- Application Number
- CN202611210937.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-11
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies for predicting the instability of soft rock embankments suffer from problems such as strong data dependence, poor generalization ability, and physical inconsistency. In particular, it is difficult to achieve high-precision predictions with strong generalization ability and strict physical consistency under conditions of sparse monitoring data.
A method based on physical information neural networks is constructed to dynamically update the constitutive equation by obtaining the nonlinear mapping relationship between the shear strength parameters of soft rock fill and environmental and construction factors. This relationship is then embedded as a physical constraint into the loss function of the neural network. The method is trained by combining response surface methodology and composite loss function to achieve high-precision prediction of the development trend of internal cracks and sliding surfaces in embankments.
It breaks through the limitations of sparse monitoring data, significantly improves the mechanical rationality and extrapolation generalization ability of the prediction results, finely characterizes the nonlinear degradation features of materials under multi-factor coupling, realizes high-fidelity dynamic early warning, and provides a scientific basis for decision-making on embankment instability disasters.
Smart Images

Figure CN123116783A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of geotechnical engineering and disaster prevention and mitigation, and in particular relates to a method for predicting the instability of soft rock embankments based on physical information neural networks. Background Technology
[0002] In the construction of infrastructure such as highways and railways in mountainous areas, local soft rocks such as carbonaceous mudstone are often used as embankment fillers. These materials exhibit significant softening, disintegration, and strength reduction characteristics upon contact with water. Their mechanical properties are highly nonlinearly coupled with multiple environmental and construction factors, including water content, stone content, and compaction degree. Under long-term operation or extreme weather conditions (such as heavy rainfall), progressive crack development is prone to occur within the embankment, eventually leading to overall shear instability and causing major engineering disasters. Therefore, achieving high-precision dynamic prediction of the crack development trend and macroscopic instability risk within soft rock embankments has significant engineering safety value.
[0003] Currently, prediction methods for embankment settlement, deformation, or instability risks can be mainly divided into two categories: numerical simulation methods based on physical mechanisms and pure data-driven machine learning methods.
[0004] Numerical simulation methods based on physical mechanisms (such as the finite element method (FEM) and the finite difference method (FDM) predict the mechanical behavior of soil and rock by solving the governing equations (such as equilibrium equations and constitutive equations). The prediction results of these methods have clear physical interpretability. However, their prediction accuracy heavily depends on the accuracy of the constitutive model and its parameters. For complex materials like carbonaceous mudstone, traditional empirical formulas (such as the linear Mohr-Coulomb criterion) are insufficient to accurately characterize their shear strength parameters (cohesion C, internal friction angle). This involves a nonlinear degradation process that dynamically changes with moisture content, compaction state, and other factors. Using simplified constitutive models with fixed parameters can lead to significant deviations when simulating complex responses such as material softening upon contact with water, resulting in prediction failure. Furthermore, high-fidelity numerical models are typically computationally expensive, making them unsuitable for real-time early warning systems in engineering projects.
[0005] To improve computational efficiency and adaptability, purely data-driven methods, such as artificial neural networks (e.g., backpropagation neural networks) and long short-term memory networks (LSTM), have been introduced into this field. These methods make predictions by mining statistical patterns in monitoring data (e.g., displacement, stress), without explicitly relying on physical equations. However, they have two inherent limitations:
[0006] Severe data dependence and weak generalization ability: These models are typical black boxes, with their predictions entirely based on the statistical relationships of the training data. In practical engineering, the number of monitoring devices that can be deployed, such as deep inclinometers and settlement gauges, is limited, resulting in extremely sparse monitoring data in both time and space. Under such small sample conditions, traditional data-driven models are prone to overfitting, meaning they excessively memorize limited monitoring data that may contain noise, failing to learn the underlying physical laws. This leads to models performing reasonably well within the training data range, but exhibiting extremely poor predictive ability for external conditions (such as stronger rainfall or longer durations), with predictions often deviating significantly from common physical sense, such as predicting displacement fields that violate the laws of mass or momentum conservation.
[0007] Physical inconsistency: Purely data-driven predictions are not constrained by any physical laws. Therefore, even with relatively sufficient data, the predicted fracture development paths and sliding surface morphologies may violate the basic principles of rock mechanics (such as strain localization mechanisms and shear failure criteria), lacking physical credibility and making them unreliable as a basis for engineering decisions.
[0008] In recent years, the emergence of Physical Information Neural Networks (PINNs) has provided a new approach to combining data with physical laws. PINNs embed partial differential equations (PDEs) describing physical processes as soft constraints into the loss function of the neural network, guiding the model to fit the data while satisfying physical laws. In some fluid mechanics and solid mechanics problems, standard PINNs have shown the potential to outperform purely data-driven methods when data is sparse. However, directly applying standard PINNs to predict the instability of soft rock embankments still faces key challenges. Standard PINNs typically treat physical parameters (such as material strength parameters) in the governing equations as constants. However, as mentioned earlier, the mechanical parameters of carbonaceous mudstone evolve dynamically with environmental and construction factors. If the physical equations embedded in PINNs are fixed in parameters, the model is essentially learning within an incorrect or static physical framework, failing to truly reflect the dynamic degradation of mechanical properties caused by material softening upon contact with water. Therefore, when simulating extreme conditions such as rapid changes in water content, its predictions will produce fundamental errors, rendering its early warning function ineffective.
[0009] In summary, existing technologies have significant shortcomings: traditional data-driven methods (BP, LSTM, etc.) exhibit poor generalization ability with small sample sizes and their predictions are physically inconsistent; while the standard PINN model introduces physical constraints, its fixed material parameter assumptions cannot characterize the dynamic nonlinear degradation of soft rock's mechanical properties, leading to prediction distortion under critical operating conditions. Currently, there is a lack of a method for predicting the instability of soft rock embankments that can simultaneously guarantee high accuracy, strong generalization ability, strict physical consistency, and accurately reflect the dynamic changes in material properties, even under conditions of sparse monitoring data. This is precisely the core technical problem that this invention aims to solve. Summary of the Invention
[0010] The purpose of this invention is to provide a method for predicting the instability of soft rock embankments based on physical information neural networks, so as to achieve the prediction of internal cracks in embankments affected by multiple environmental and construction factors under the condition of sparse monitoring data.
[0011] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is a method for predicting the instability of soft rock embankments based on physical information neural networks, comprising:
[0012] Obtain the nonlinear mapping relationship between the shear strength parameters of soft rock filler and environmental and construction factors, and construct a dynamic constitutive equation to describe the dynamic changes of the shear strength parameters with the environmental and construction factors.
[0013] A physical information neural network is constructed. The input of the physical information neural network includes spatiotemporal coordinates and working condition variables of the environmental factors and the construction factors. The output includes displacement field and stress field. The dynamic constitutive equation is embedded as a physical constraint in the loss function of the physical information neural network so as to dynamically update the mechanical benchmark of the physical constraint based on the currently input working condition variables during the forward propagation of the physical information neural network.
[0014] The physical information neural network is trained, and after the model converges, real-time or predicted working condition parameters are input, and the global displacement field and stress field are output. Based on the output results, the embankment sliding zone is determined and instability warning is given.
[0015] Furthermore, the nonlinear mapping relationship between the shear strength parameters of the soft rock filler and environmental and construction factors is obtained, and a dynamic constitutive equation is constructed to describe the dynamic changes of the shear strength parameters with the environmental and construction factors, specifically including:
[0016] Large-scale direct shear tests with multiple coupled factors were conducted to obtain test data on the cohesion and internal friction angle of soft rock filler under different moisture content, stone content and compaction degree conditions;
[0017] Based on the response surface methodology, a nonlinear mapping relationship between moisture content, stone content, compaction degree and cohesion, and internal friction angle was established to obtain the dynamic shear strength parameter equation;
[0018] The dynamically updated Mohr-Coulomb yield criterion or stress-strain nonlinear constitutive equation is derived from the dynamic shear strength parameter equation and used as the dynamic constitutive equation.
[0019] Furthermore, the dynamic shear strength parameter equation is given by using water content ω, stone content P, and compaction degree K as independent variables, and cohesion C and internal friction angle as independent variables. The second-order nonlinear polynomial response surface model for the dependent variable is expressed as:
[0020]
[0021] Where i and j are index variables, Let i and j represent the i-th and j-th input variables, respectively. When the dependent variable y is cohesion C, the regression coefficients are... , , , This constitutes the first set of parameters; when the dependent variable y is the internal friction angle When, regression coefficient , , , This constitutes the second set of parameters; The intercept term is a constant. These are the linear effect regression coefficients, reflecting the basic impact of each univariate; These are the regression coefficients for the quadratic surface effect, reflecting the nonlinear attenuation or enhancement characteristics of each variable with respect to intensity. The interaction effect regression coefficient reflects the coupling mechanism between moisture content, stone content and compaction degree; This represents the residual term from the model fit.
[0022] Furthermore, the input layer of the physical information neural network adopts a dual-channel design, including a spatiotemporal coordinate channel and an environmental condition variable channel. The input of the spatiotemporal coordinate channel is spatial coordinates and time variables, and the input of the environmental condition variable channel is moisture content, rock content, and compaction degree. The output layer of the physical information neural network contains 5 neurons, corresponding to horizontal displacement, vertical displacement, and three stress components, respectively.
[0023] Furthermore, the loss function is a composite loss function, Loss. Total for:
[0024] Loss Total =w1Loss data +w2Loss bc +w3Loss PDE
[0025] Where w1, w2, and w3 are adaptive dynamic weight parameters; Loss data Data loss represents the mean square error between the displacement value predicted by the physical information neural network and the measured displacement value; Loss bc Loss is the error in the prediction result deviating from the given boundary conditions at the geometric boundary. PDE This represents the residual loss for partial differential equations.
[0026] Furthermore, the partial differential equation residual loss LossPDE includes the static equilibrium residual loss Loss. equilibrium and dynamic yield criterion residual loss yield ,Right now:
[0027] Loss PDE =Loss equilibrium +Loss yield .
[0028] Furthermore, the static equilibrium residual loss Loss equilibrium Specifically:
[0029]
[0030] in, This represents the partial differential residual loss value that indicates the deviation of the neural network prediction result from the two-dimensional static equilibrium equation of soil and rock. Represents horizontal normal stress Along the horizontal spatial coordinate x s First-order partial derivative in the direction, Represents vertical normal stress Along vertical spatial coordinates First-order partial derivative in the direction; Represents shear stress Along horizontal spatial coordinates First-order partial derivative in the direction, Represents shear stress Along the vertical spatial coordinate z s First-order partial derivative in the direction, The density of the carbonaceous mudstone filler is represented by g, and g represents the gravitational acceleration constant. Represents the coordinate points deployed within the physical computing domain;
[0031] Dynamic yield criterion residual loss Specifically:
[0032]
[0033] in, Represents the coordinate points deployed within the physical computing domain, where i is the index symbol. The internal shear stress predicted by the neural network during forward propagation at the i-th physical collocation point is... The normal stress predicted and derived by the neural network during forward propagation at the i-th physical collocation point is... This represents a dynamic function for calculating cohesion C. Represents the calculation of the internal friction angle The dynamic function, the and The moisture content ω, stone content p, and compaction degree k are used as input variables.
[0034] Furthermore, the weight parameters w1, w2, w3 are updated using an adaptive dynamic weight adjustment algorithm, which includes:
[0035] In the current training iteration step t, the data loss is Loss data Using the baseline anchor point, calculate the dynamic weight target value for each physical constraint loss term. :
[0036]
[0037] in, For data loss terms, pair network parameters The gradient vector; Let be the gradient vector of the i-th physical or boundary constraint loss term with respect to the parameter θ; To find the absolute value of the maximum element in the gradient vector; To find the mean of the absolute values of all elements in the gradient vector; To prevent tiny positive numbers with a denominator of zero;
[0038] The weights are updated smoothly using an exponential moving average strategy:
[0039]
[0040] In the formula, w represents the dynamic weight target value calculated for the i-th physical or boundary constraint loss term at the current training iteration step t. i (t) This is the actual usage value after historical smoothing. To smooth out the annealing decay rate;
[0041] Furthermore, the physical information neural network contains 4 to 8 fully connected hidden layers, each with 50 to 120 neurons, and each hidden layer uses the tanh function or the Swish function as a non-linear activation function.
[0042] Furthermore, a joint optimization strategy is employed when training the physical information neural network. The first stage uses the Adam optimization algorithm for global gradient descent, and the second stage switches to the L-BFGS optimization algorithm for precise convergence. The model is considered complete when one of the following conditions is met:
[0043] The composite loss function value is less than the preset engineering tolerance cutoff value;
[0044] The relative rate of change of the loss function is less than the preset relative tolerance within a preset number of consecutive steps;
[0045] The mean square error of the verification set did not decrease or rebound within a preset number of iterations.
[0046] Compared with the prior art, the beneficial effects of the present invention include the following points:
[0047] 1. This invention overcomes the limitations of sparse monitoring data, significantly improving the mechanical rationality and extrapolation generalization ability of the prediction results. It breaks through the black-box limitations of traditional pure data-driven neural networks (such as BP and LSTM) by embedding partial differential equations (such as stress balance equations) controlling the deformation and failure of soil and rock masses as soft constraints into the loss function of the neural network. Even when monitoring data from inclinometers and settlement gauges are extremely sparse, the model can still "learn in the right direction" guided by physical laws, effectively avoiding overfitting. The predicted fracture development trend of carbonaceous mudstone and the embankment sliding surface strictly conform to the laws of rock mechanics, greatly improving the reliability of extrapolation predictions under extreme conditions.
[0048] 2. This invention meticulously characterizes the nonlinear degradation features of materials under multi-factor coupling, reproducing the true mechanical response. Instead of relying on traditional simplified empirical formulas, this invention is based on physical data from large-scale direct shear tests. By introducing the response surface methodology (RSM), the nonlinear mapping relationship between the three core variables—water content, rock content, and compaction degree—and shear strength parameters (cohesion and internal friction angle) is explicitly expressed mathematically. This method accurately quantifies the complex degradation mechanism of soft rock upon contact with water and the influence of filler components, making the constitutive parameters of the input model more consistent with engineering realities.
[0049] 3. A closed loop of physical experimentation, mathematical constitutive modeling, and deep learning has been achieved, constructing a high-fidelity dynamic early warning architecture. This invention innovatively and seamlessly embeds a multivariate nonlinear shear strength criterion obtained based on RSM fitting into the residual calculation of the physical partial differential equations of PINN in a dynamic and adaptive manner. This deeply coupled architecture enables the model to receive real-time changes in the external environment (such as a sharp increase in internal moisture content due to rainfall) and synchronously adjust the internal mechanical constraint benchmarks. This allows for high-precision, dynamic evolutionary deduction of the stress field, displacement field, and crack penetration risk within the embankment, providing a scientific and accurate decision-making basis for early warning and prevention of embankment instability disasters. Attached Figure Description
[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0051] Figure 1This is the process flow of this implementation method;
[0052] Figure 2 This is a schematic diagram of the Physical Information Neural Network (PINN) architecture and computational topology constructed for an embodiment of the present invention;
[0053] Figure 3 This is a schematic diagram of the physical geometric model and boundary condition settings for predicting embankment instability in this embodiment. Detailed Implementation
[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0055] like Figure 1 This embodiment provides a method for predicting the instability of soft rock embankments based on physical information neural networks, specifically including the following steps:
[0056] Step S1: Conduct large-scale direct shear tests to obtain a dataset of test results on the shear strength of carbonaceous mudstone under multi-factor coupling. Specifically, this includes:
[0057] A series of large-scale direct shear tests were designed and conducted for soft rock materials used in embankment construction. During the tests, the initial moisture content (ω), stone content (P), and compaction degree (K) of the fill material were used as control variables in cross-combinations. Different normal stresses (σ) were recorded. n Under the given conditions, the shear stress (τ) at which the specimen fails under shear stress. f Therefore, the macroscopic shear strength parameters under different working condition combinations are extracted, namely cohesion (C) and internal friction angle (C). ), construct the original dataset for multivariable mechanics experiments.
[0058] Step S2: Based on the response surface methodology (RSM) fitting of the experimental data, construct a macroscopic shear strength constitutive equation considering parametric nonlinear degradation. Specifically, this includes:
[0059] The large direct shear test dataset obtained in step S1 is processed. This step uses the response surface methodology (RSM) to establish a nonlinear mapping relationship between water content ω, stone content P, compaction degree K, and the shear strength parameters of carbonaceous mudstone. Specifically, water content ω, stone content P, and compaction degree K are used as independent variables, and cohesion C and internal friction angle are used as independent variables, respectively. We establish two independent second-order nonlinear polynomial response surface models, with x as the dependent variable. Let the input variable vector be x = [x1, x2, x3]. T =[ω,P,K] T T is the transpose symbol, and the output response target is y; x1, x2, and x3 are the moisture content variable, stone content variable, and compaction degree variable of the packing, respectively.
[0060] The two models share the same regression equation form, meaning the independent variables are the same (moisture content ω, stone content P, and compaction degree K), but the regression coefficients are independent of each other, and the expressions are both:
[0061]
[0062] Where i and j are index variables, Let i and j represent the i-th and j-th input variables, respectively. When the dependent variable y is cohesion C, the regression coefficients are... , , , This constitutes the first set of parameters; when the dependent variable y is the internal friction angle When the regression coefficients constitute the second set of parameters. The intercept term is a constant. These are the linear effect regression coefficients, reflecting the basic impact of each univariate; These are the regression coefficients for the quadratic surface effect, reflecting the nonlinear attenuation or enhancement characteristics of each variable with respect to intensity. The coefficients represent the interaction effect regression coefficients, reflecting the coupling mechanism between moisture content, stone content, and compaction degree (e.g., the weakening effect of compaction degree on shear strength under high moisture content). This represents the residual term from the model fit.
[0063] Subsequently, for the n sets of experimental data samples obtained in step S1, the above expansion is transformed into matrix form. = + X is the set of input variable vectors. It is an n×1 response column vector containing the target dependent variable data (i.e., the corresponding cohesion) measured in n sets of experiments. array or internal friction angle (array), and solve for the two sets of parameter vectors using the multivariate linear least squares method respectively. The matrix solution formula for its unbiased estimator is:
[0064]
[0065] Wherein, parameter vector An unbiased estimator; It is a transpose operator;
[0066] The coefficient matrix corresponding to the cohesion C and the coefficient matrix corresponding to the internal friction angle are calculated using this matrix operation. The coefficient matrix was calculated. To ensure the scientific validity and reliability of the established regression equation, further analysis of variance (ANOVA) and model significance tests were performed, and the coefficient of determination R of the model was calculated. 2 With the adjusted coefficient of determination R 2 adj When R 2 When the variance is greater than 0.90 and the p-value of the regression model is less than 0.05, the nonlinear regression model is considered highly significant and can effectively represent the true physical laws. Based on this, insignificant interaction terms in the analysis of variance (i.e., terms with p-values greater than 0.05) are removed, and the final output is the macroscopic deterioration parameter equation for the shear strength of carbonaceous mudstone with the highest goodness of fit:
[0067] C=f C (ω,P,K)
[0068] = (ω,P,K)
[0069] in, This represents the dynamic function for calculating cohesion C. It takes moisture content ω, stone content p, and compaction degree k as input variables and outputs the cohesion value under the current state. Represents the calculation of the internal friction angle The dynamic function, also with ω, P, and K as input variables, outputs the value of the internal friction angle in the current state.
[0070] The above equations can be used to further derive the dynamically updated Mohr-Coulomb yield criterion or the corresponding stress-strain nonlinear constitutive equation, which can be used to accurately characterize the dynamic deterioration of soft rock material strength with environmental factors (such as water infiltration) and construction quality (such as insufficient compaction). This explicit set of equations will serve as the benchmark formula for dynamic physical soft constraints, and will be directly nested into the partial differential equilibrium equation residual calculation node of the physical information neural network in the subsequent step S4, so as to achieve accurate simulation of the dynamic changes of material mechanical properties with multiple factors.
[0071] This leads to the dynamically updated Mohr-Coulomb yield criterion or the corresponding stress-strain nonlinear constitutive equation, which can be used to accurately characterize the deterioration of soft rock material strength with environmental factors (such as water infiltration) and construction quality (such as insufficient compaction).
[0072] Step S3: Establish the physical geometric model of the embankment, determine the boundary conditions, and collect discrete monitoring data from the field. Specifically, this includes:
[0073] like Figure 3As shown, a corresponding physical geometric model is established based on the typical cross-section of an actual embankment project. Specifically, in the spatial domain, a two-dimensional Cartesian coordinate system (x, y, z) is established based on the cross-sectional dimensions of the embankment. s ,z s or a three-dimensional rectangular coordinate system (x s ,y s ,z s Simultaneously, a time variable t is introduced in the time domain to characterize the time-varying characteristics of the embankment's mechanical response. Based on this, the geometric boundary conditions of the model are set: the contact surface between the bottom of the embankment and the bedrock is defined as a displacement-fixed boundary, meaning that the horizontal and vertical displacements at all points on this boundary are zero (Ux). s =0,Uz s =0); where the displacement-fixed boundary includes the bottom of the embankment (bedrock contact surface): defined as the displacement-fixed boundary. The explicit mathematical expression given in the document is that the horizontal and vertical displacements at all points on this boundary are zero, i.e., Ux s =0, Uz s =0. Lateral boundary (combined) Figure 3 (Note: Set as a lateral horizontal constraint, i.e., horizontal displacement is zero (Ux)) s =0), but free settling is allowed.
[0074] The top of the embankment and the surface of the slope are defined as stress-free boundaries, meaning that the outward normal traction force on these boundaries is zero. (Combined with...) Figure 3 The annotation, specifically the expression for the normal stress σ, is... n = 0, shear stress τ=0), to simulate the stress conditions of the embankment under natural conditions. In addition, measured data from various discrete displacement monitoring instruments deployed inside or on the surface of the embankment are collected. These monitoring instruments include, but are not limited to, deep inclinometer tubes and surface settlement gauges. The displacement observations obtained are recorded as the true displacement value U. true This data is used as supervisory data to constrain the training process of the neural network in subsequent steps.
[0075] Step S4: Construct a Physical Information Neural Network (PINN) and configure a composite loss function that includes material nonlinear constitutive constraints.
[0076] like Figure 2As shown, this embodiment designs a parameterized physical constraint feedforward neural network, whose structure differs from the conventional PINN model. The conventional PINN model typically only uses spatiotemporal coordinates as input, and the parameters of the internal physical equations are fixed constants. The network structure of this embodiment has the following characteristics: the input layer adopts a dual-channel design, including not only the conventional spatiotemporal coordinates (x,z,t), but also a specially added channel for environmental condition variables, connecting water content ω, rock content P, and compaction degree K as auxiliary input nodes to the physical residual calculation graph. The hidden layers adopt a 4 to 8-layer fully connected structure, with 50 to 120 neurons in each layer. This depth-to-width ratio is specifically designed to address the gradient vanishing problem in cross-scale calculations between the large-scale range of the embankment and the extremely small scale of internal micro-cracks. Regarding the choice of activation function, this embodiment abandons the ReLU activation function commonly used in traditional deep learning. This is because the partial differential equations of embankment stress balance (Navier-Cauchy equations) involve the second-order partial derivatives of displacement with respect to spatial coordinates, while the second derivative of the ReLU function is zero, leading to the complete loss of physical information. This implementation uses tanh or Swish (i.e., x⋅σ(x)) as the nonlinear activation function to ensure that the network has infinite-order differentiability throughout the entire computational domain, thereby accurately transmitting the higher-order derivative information of the carbonaceous mudstone mechanical field. The output layer contains 5 neurons, corresponding to the horizontal displacement u, vertical displacement w, and three stress components σ at a certain point in space at a certain time. xs σ zs τ xszs .
[0077] The core improvement of this implementation lies in the deep physical nesting design of the composite loss function. This implementation improves the calculation of the partial differential equation residuals (Loss) PDE When performing this step, the underlying formulas were modified using an automatic differentiation framework (such as a Python-based computational graph), dynamically embedding the nonlinear laws extracted by the Response Surface Method (RSM). The composite loss function Loss... Total Defined as:
[0078] Loss Total =w1Loss data +w2Loss bc +w3Loss PDE
[0079] Where w1, w2, and w3 are adaptive dynamic weight parameters. This composite loss function consists of three parts: the first part is the data loss (Loss). data That is, the network-predicted displacement value and the sensor-measured displacement value U in step S3. true The mean square error between the two parts; the second part is the boundary condition loss. bcThe first part is the error in the prediction result deviating from the given boundary conditions at the geometric boundary; the second part is the loss of the partial differential equation residuals. PDE .
[0080] The improved LossPDE for this scenario is further divided into two parts: static equilibrium residuals and dynamic yield criterion residuals. Regarding the dynamic nonlinear yield residual constraint, the traditional yield criterion penalty term uses a constant cohesion C and the internal friction angle. In this embodiment, the response function is replaced by a nonlinear RSM model controlled by multiple variables. Its mathematical expression is:
[0081]
[0082] in, Represents the coordinate points deployed within the physical computing domain, where i is the index symbol. The internal shear stress predicted by the neural network during forward propagation at the i-th physical collocation point is... The normal stress predicted and derived by the neural network during forward propagation at the i-th physical collocation point is...
[0083] In the above formula, the internal shear stress τ is calculated in each forward propagation of the network. pred At any time, the current w, p, k states will be automatically read and called. and The function calculates the current shear strength threshold, thus enabling the network to intelligently identify local yielding caused by the softening of carbonaceous mudstone upon contact with water. Regarding the continuity partial differential residual of stress-displacement coupling, the network uses automatic differentiation techniques to calculate the gradient of the output stress with respect to spatial coordinates and substitutes it into the two-dimensional soil-rock equilibrium equations. The mathematical expression is as follows:
[0084]
[0085] in, This represents the partial differential residual loss value that indicates the deviation of the neural network prediction result from the two-dimensional static equilibrium equation of soil and rock. Represents horizontal normal stress Along the horizontal spatial coordinate x s First-order partial derivative in the direction, Represents vertical normal stress Along the vertical spatial coordinate z s First-order partial derivative in the direction; Represents shear stress Along the horizontal spatial coordinate x s First-order partial derivative in the direction, Represents shear stress Along the vertical spatial coordinate z s First-order partial derivative in the direction, denoted by , where represents the material density of the carbonaceous mudstone filler, and g represents the gravitational acceleration constant.
[0086] Ultimately, the physical residual term in this embodiment is configured as follows:
[0087] Loss PDE =Loss equilibrium +Loss yield
[0088] Through the above design, this embodiment directly embeds the dynamic constitutive equation established based on RSM in step S2 into the calculation process of the partial differential equation residual, thereby forcing the neural network to strictly abide by the mechanical deterioration law of the specific carbonaceous mudstone when deriving the mechanical field inside the entire embankment.
[0089] Regarding the design principles of the key structural parameters of the network, this implementation method is based on the following considerations. Embankment instability is essentially a cross-scale evolutionary process: its outer boundary belongs to the macro-scale (high embankments of tens of meters, i.e., i.e., x, z coordinates), while the load-bearing failure of the internal carbonaceous mudstone begins with the development of fissures at the micro / meso-scale (highly controlled by micro-component variables such as water content ω, stone content P, and compaction degree K). Regarding the number of neurons, preliminary experiments show that when a single layer has fewer than 50 neurons, the parameter space dimension of the network is insufficient to simultaneously analyze the macro-stress field distribution and the local fissure development trend under multivariate coupling. The model exhibits severe underfitting when approximating the complex second-order response surface nonlinear shear strength surface, leading to fuzzy predictions of potential slip surfaces. Furthermore, when the number of neurons exceeds 120, the network is highly prone to severe overfitting to extremely sparse sensor displacement anchor point data, and the floating-point computation cost of a single epoch forward propagation increases exponentially, weakening the model's real-time engineering early warning efficiency. Regarding the number of hidden layers, since the calculation of the second derivative requires back-chain differentiation across two computational graphs, if there are fewer than 4 hidden layers, the rank of the higher-order derivative matrix will rapidly decay, causing the stress prediction field to degenerate into a linear distribution, completely failing to capture the strain localization phenomenon within the shear zone. Although increasing the depth can improve the nonlinear expressive power, preliminary experimental data confirms that when there are more than 8 hidden layers, during the backpropagation of the calculation of the second physical derivative and the nonlinear Mohr-Coulomb yield residual, the gradient signal will dissipate exponentially in the deep topology, leading to the gradient vanishing problem across scales. This prevents the weights near the input layer from being effectively updated, and the dynamic influence of water content and rock content becomes ineffective. Therefore, 4 to 8 layers is the effective range for maintaining stable backpropagation of the higher-order physical residual gradient.
[0090] To verify the rationality of the parameter selection, this implementation method conducted a grid search experiment for structural hyperparameters based on the Python deep learning framework during the initial stage of model construction. The controlled variable method was used to cross-validate the combination matrix of the number of hidden layers (2 to 12 layers) and the number of neurons (10 to 200). The results show that within the parameter domain of "4 to 8 layers and 50 to 120 neurons," the composite loss function LTotal can stably converge to the order of 10⁻⁴, and the stress distribution of the network output has the highest Pearson correlation coefficient (r > 0.92) with the physical response of a large direct shear test. Beyond this range, the model either gets trapped in local minima leading to persistently high physical residuals, or suffers from gradient explosion resulting in numerical overflow. This experimental data provides a solid quantitative basis for the parameter selection in this implementation method.
[0091] Constructing the composite loss function Loss Total =w1Loss data +w2Loss bc +w3Loss PDE Subsequently, regarding the displacement data residuals (Loss) in the embankment project... data (typically on the millimeter scale) and partial differential physical residuals (Loss PDE This addresses the inconsistency in physical dimensions and gradient magnitude differences between stress and higher-order derivatives (which are typically of extremely large magnitudes). Instead of using empirically fixed values, this implementation employs an adaptive dynamic weight adjustment algorithm based on backpropagation gradient statistics. During model training, this algorithm iteratively and automatically updates the weights w1, w2, and w3 based on the gradient representation of each loss in the network's hidden layers. The specific adjustment process and mathematical mechanism are as follows:
[0092] First, during the backpropagation process at each training iteration time t, the loss is calculated using displacement monitoring data. data Using the baseline anchor point, extract the maximum absolute value of its gradient with respect to the shared hidden layer parameter θ; simultaneously, extract other penalty terms (including the physical equation residual loss). data and boundary residual Loss b Loss i The average absolute value of the gradient relative to the parameters of the same hidden layer.
[0093] Secondly, to prevent the high-dimensional gradient of the physical residual from completely masking the gradient of the data residual during backpropagation, i.e., to prevent ill-conditioned gradient imbalance, the dynamic weight target values of each physical constraint term in the current iteration step are calculated. Its core calculation criterion is to balance the gradient magnitudes of each loss term, and the mathematical expression is:
[0094]
[0095] in, For data loss terms, pair network parameters The gradient vector; Let be the gradient vector of the i-th physical or boundary constraint loss term with respect to the parameter θ; To find the absolute value of the maximum element in the gradient vector; To find the mean of the absolute values of all elements in the gradient vector; To prevent tiny positive numbers with a denominator of zero (in this embodiment, it is set as follows): ).
[0096] Finally, to avoid drastic gradient fluctuations caused by single-iteration sampling, this implementation uses an exponential moving average strategy to perform smooth annealing updates on the weights. The actual weight w used in step t+1 is... i (t+1) The update formula is:
[0097]
[0098] In the formula, This represents the dynamic weight target value calculated for the i-th physical or boundary constraint loss term at the current training iteration step t. It is the original target value calculated in the current step, w i (t) This is the actual usage value after historical smoothing. The smoothing annealing decay rate is typically set between 0.1 and 0.2. Through this adaptive algorithm, in the early stages of training, the network is primarily guided by fitting sparse sensor data (w1 is dominant). As training progresses and the data fitting stabilizes, the algorithm automatically increases the weight of w2 and w3, forcing the network to output a high-fidelity approximation of the constitutive and stress equilibrium partial differential equations of carbonaceous mudstone's nonlinear mechanical structure.
[0099] Regarding the model training strategy, this implementation further sets detailed training parameters and convergence criteria. In terms of discretization and set partitioning of multi-source training data, the training data pool includes both supervised sensor data and unsupervised physical point mapping. For supervised sensor data, the field-measured displacement data obtained in step S3 (total...) is used... (Number of sample points) are randomly shuffled according to time series or spatial nodes and divided into training and validation sets in an 80%:20% ratio. The training set is used to calculate the loss. data The weights are updated in reverse order. The validation set does not participate in gradient backpropagation; it is only used to monitor the model's generalization ability and overfitting. For unsupervised physical domain collocations, the Latin hypercube sampling algorithm is used to randomly generate points within the embankment computational domain Ω. Internal collocation points (used to calculate the residual loss of the partial differential equation)PDE ), and generate at the physical boundary ∂Ω. One boundary point (used to calculate boundary residuals Loss) bc Physical collocations are dynamically resampled in each training round to enhance the model's ability to explore the spatial dynamics field across the entire domain.
[0100] Regarding optimizer selection, given that the partial differential equations contain complex second-order derivatives and the loss function is highly non-convex, a single Adam optimizer is prone to gradient oscillations in the later stages, while a single second-order optimizer such as L-BFGS is extremely sensitive to initial weights. Therefore, this implementation adopts a joint optimization strategy of "first global optimization with Adam, then precise convergence with L-BFGS". In the first stage, the Adam optimization algorithm is used for fast global gradient descent, with the initial learning rate set to [value missing]. Furthermore, a cosine annealing learning rate decay strategy is introduced to smoothly decay the learning rate over 10,000 epochs. This stage is mainly responsible for applying the composite loss function Loss Total Quickly pulled down to Order of magnitude. In the second stage, when the Adam optimization reaches the preset maximum number of iterations (e.g., 10,000), or when the validation set data loss no longer decreases after 500 consecutive epochs, it automatically switches to the L-BFGS quasi-Newton second-order optimizer. The maximum number of iterations is set to 50,000, the maximum number of line search steps to 50, and the number of historical information retained to 50. The approximate information of the Hessian matrix is used to optimize the physical residual loss. PDE Perform precise localized polishing.
[0101] Regarding model convergence determination, this implementation does not rely on a single upper limit for the number of iterations, but instead adopts a multi-dimensional composite convergence determination criterion. An early stopping mechanism is triggered when any of the following conditions are met, determining that the model training is complete and outputting a fixed network: First, absolute threshold determination, i.e., the total composite loss function Loss under joint optimization. Total and its sub-items Loss PDE Loss data All strictly less than The first is the preset engineering tolerance cutoff value; the second is the relative tolerance determination, that is, the relative rate of change of the loss function in 100 consecutive L-BFGS line searches. This indicates that the model has reached the theoretical minimum basin of the physical residuals; third, the generalization performance degradation judgment, that is, the mean square error of the validation set has not shown a substantial decrease within 1000 consecutive epochs, or even shows signs of a reverse rebound. In order to prevent overfitting to sparse sensor noise, the training is forcibly terminated.
[0102] Step S5: Input real-time environmental parameters to dynamically deduce the internal stress-displacement field and slip zone expansion trend of the embankment, and realize instability early warning; specifically including:
[0103] After completing model training based on the above training strategy, this implementation method enters the real-time prediction and instability warning stage. First, the composite loss function is minimized using a gradient descent algorithm (such as Adam or L-BFGS optimizer), and the neural network weights are continuously updated until the residuals tend to converge, thus obtaining a trained and converged fixed network model.
[0104] Subsequently, extreme working condition parameters that may be encountered in actual engineering are input as external conditions into the trained and converged model. These extreme working condition parameters include, but are not limited to, the increase in filler moisture content ω due to heavy rainfall, changes in stone content P, and fluctuations in compaction degree K. Upon receiving these environmental variables, the model dynamically adjusts the physical residual benchmark, that is, it automatically calls the dynamic constitutive equation based on RSM established in step S2 to update the shear strength threshold under the current working condition, and then outputs a high-resolution global displacement field and stress field (including horizontal displacement u, vertical displacement w, and three stress components σ). x σ z τ xz 。 ).
[0105] Finally, based on the output mechanical state distribution, combined with the strain localization criterion or the Mohr-Coulomb yield criterion threshold, the development location and connectivity of shear zones or microcracks within the embankment are determined. Specifically, when the stress state in a certain region reaches or exceeds the current dynamic yield threshold, the region is determined to have entered a yielding state; through connectivity analysis of the yielding region, the depth and morphology of potential sliding surfaces are predicted, thereby providing quantitative safety warnings for disaster prevention and control.
[0106] To intuitively verify the superiority of the technical solution of this invention compared with the prior art, this embodiment systematically compares and verifies the invention with Comparative Example 1 (traditional pure data-driven model, such as LSTM / conventional BP) and Comparative Example 2 (standard PINN model, fixed strength parameters, ReLU activation) from four dimensions: displacement prediction accuracy under sparse monitoring data, strength response under water content surge conditions, prediction ability of internal cracks and macroscopic sliding surfaces of embankments, and calculation of higher-order mechanical derivatives and residual convergence. The results are shown in Table 1.
[0107] Table 1. Multidimensional performance comparison of the present invention and comparative examples.
[0108]
[0109] The technical solution adopted in this embodiment overcomes the inherent limitations of existing traditional prediction models that heavily rely on massive amounts of measured data. Even in engineering scenarios where on-site monitoring samples are scarce and data is extremely limited, the model's prediction accuracy can be improved by approximately 88.6% compared to conventional prediction methods. The solution accurately reproduces the intrinsic evolution mechanism of carbonaceous mudstone disintegrating and gradually softening its mechanical properties after water immersion by embedding the nonlinear constitutive relationship of the response surface method (RSM). The early warning system built based on the above architecture has clear physical mechanism support and can complete the visualization and early warning output of the entire geological disaster process, significantly reducing the probability of missed disaster warnings under extreme engineering conditions. At the same time, this application optimizes the design from the bottom-level topology architecture of the model, improves the defects of high-order gradient dispersion and gradient vanishing, and ensures that the physical field results output by the model strictly satisfy the laws of mechanical conservation, achieving full-domain mechanical self-consistency.
[0110] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.
[0111] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.
Claims
1. A soft rock embankment instability prediction method based on a physical information neural network, characterized in that, include: Obtain the nonlinear mapping relationship between the shear strength parameters of soft rock filler and environmental and construction factors, and construct a dynamic constitutive equation to describe the dynamic changes of the shear strength parameters with the environmental and construction factors. A physical information neural network is constructed. The input of the physical information neural network includes spatiotemporal coordinates and working condition variables of the environmental factors and the construction factors. The output includes displacement field and stress field. The dynamic constitutive equation is embedded as a physical constraint in the loss function of the physical information neural network so as to dynamically update the mechanical benchmark of the physical constraint based on the currently input working condition variables during the forward propagation of the physical information neural network. The physical information neural network is trained, and after the model converges, real-time or predicted working condition parameters are input, and the global displacement field and stress field are output. Based on the output results, the embankment sliding zone is determined and instability warning is given.
2. The soft rock embankment instability prediction method based on physical information neural network according to claim 1, characterized in that, The nonlinear mapping relationship between the shear strength parameters of soft rock filler and environmental and construction factors is obtained, and a dynamic constitutive equation is constructed to describe the dynamic changes of the shear strength parameters with the environmental and construction factors. Specifically, this includes: Large-scale direct shear tests with multiple coupled factors were conducted to obtain test data on the cohesion and internal friction angle of soft rock filler under different moisture content, stone content and compaction degree conditions; Based on the response surface methodology, a nonlinear mapping relationship between moisture content, stone content, compaction degree and cohesion, and internal friction angle was established to obtain the dynamic shear strength parameter equation; The dynamically updated Mohr-Coulomb yield criterion or stress-strain nonlinear constitutive equation is derived from the dynamic shear strength parameter equation and used as the dynamic constitutive equation.
3. The method for predicting the instability of soft rock embankments based on a physical information neural network according to claim 2, characterized in that, The dynamic shear strength parameter equation is a second-order nonlinear polynomial response surface model with water content ω, stone content P, and compaction degree K as independent variables, and cohesion C and internal friction angle φ as dependent variables, and the expression is as follows: The dynamic shear strength parameter equation is a second-order nonlinear polynomial response surface model with water content ω, stone content P, and compaction degree K as independent variables, and cohesion C and internal friction angle φ as dependent variables, and the expression is as follows: ; where i, j are index variables, respectively represent the i-th and j-th input variable, when the dependent variable y is cohesion C, the regression coefficients , , , constitute the first group of parameters; when the dependent variable y is the internal friction angle , the regression coefficients , , , constitute the second group of parameters; is the constant intercept term; is the linear effect regression coefficient, reflecting the basic influence of each single variable; is the quadratic surface effect regression coefficient, reflecting the nonlinear attenuation or enhancement characteristics of each variable on the strength; is the interaction effect regression coefficient, reflecting the coupling mechanism between water content, stone content and compaction degree; is the residual term of model fitting.
4. The soft rock embankment instability prediction method based on physical information neural network according to claim 1, characterized in that, The input layer of the physical information neural network adopts a dual-channel design, including a spatiotemporal coordinate channel and an environmental condition variable channel. The input of the spatiotemporal coordinate channel is spatial coordinates and time variables, and the input of the environmental condition variable channel is moisture content, rock content and compaction degree. The output layer of the physical information neural network contains 5 neurons, corresponding to horizontal displacement, vertical displacement and three stress components, respectively.
5. The method for predicting the instability of soft rock embankments based on a physical information neural network according to claim 1, characterized in that, The loss function is a composite loss function Loss Total is: Loss Total =w1Loss data +w2Loss bc +w3Loss PDE where w1, w2, w3 are adaptive dynamic weight parameters; Loss data is the data loss, representing the mean square error between the displacement value predicted by the physical information neural network and the measured displacement value; Loss bc is the error of the prediction result deviating from the established boundary condition at the geometric boundary; Loss PDE is the partial differential equation residual loss.
6. The method for predicting the instability of soft rock embankments based on a physical information neural network according to claim 5, characterized in that, The partial differential equation residual loss LossPDE comprises a static equilibrium residual loss Loss equilibrium and a dynamic yield criterion residual loss Loss yield i.e.: Loss PDE =Loss equilibrium +Loss yield .
7. The method for predicting the instability of soft rock embankments based on a physical information neural network according to claim 6, characterized in that, The static force balance residual loss Loss equilibrium Specifically: ; in, This represents the partial differential residual loss value that indicates the deviation of the neural network prediction result from the two-dimensional static equilibrium equation of soil and rock. Represents horizontal normal stress Along the horizontal spatial coordinate x s First-order partial derivative in the direction, Represents vertical normal stress Along vertical spatial coordinates First-order partial derivative in the direction; Represents shear stress Along horizontal spatial coordinates First-order partial derivative in the direction, Represents shear stress Along the vertical spatial coordinate z s First-order partial derivative in the direction, The density of the carbonaceous mudstone filler is represented by g, and g represents the gravitational acceleration constant. Represents the coordinate points deployed within the physical computing domain; Dynamic yield criterion residual loss Specifically: ; in, Represents the coordinate points deployed within the physical computing domain, where i is the index symbol. The internal shear stress predicted by the neural network during forward propagation at the i-th physical collocation point is... The normal stress predicted and derived by the neural network during forward propagation at the i-th physical collocation point is... This represents a dynamic function for calculating cohesion C. Represents the calculation of the internal friction angle The dynamic function, the and The moisture content ω, stone content p, and compaction degree k are used as input variables.
8. The method for predicting the instability of soft rock embankments based on a physical information neural network according to claim 5, characterized in that, The weight parameters w1, w2, w3 are updated using an adaptive dynamic weight adjustment algorithm, which includes: In the current training iteration step t, the data loss is Loss data Using the baseline anchor point, calculate the dynamic weight target value for each physical constraint loss term. : ; in, For data loss terms, pair network parameters The gradient vector; Let be the gradient vector of the i-th physical or boundary constraint loss term with respect to the parameter θ; To find the absolute value of the maximum element in the gradient vector; To find the mean of the absolute values of all elements in the gradient vector; To prevent tiny positive numbers with a denominator of zero; The weights are updated smoothly using an exponential moving average strategy: ; In the formula, w represents the dynamic weight target value calculated for the i-th physical or boundary constraint loss term at the current training iteration step t. i (t) This is the actual usage value after historical smoothing. To smooth the annealing decay rate.
9. The method for predicting the instability of soft rock embankments based on a physical information neural network according to claim 1, characterized in that, The physical information neural network contains 4 to 8 fully connected hidden layers, each with 50 to 120 neurons, and each hidden layer uses the tanh function or the Swish function as a non-linear activation function.
10. The method for predicting the instability of soft rock embankments based on a physical information neural network according to claim 1, characterized in that, A joint optimization strategy is employed when training the physical information neural network. The first stage uses the Adam optimization algorithm for global gradient descent, and the second stage switches to the L-BFGS optimization algorithm for precise convergence. Model training is considered complete when one of the following conditions is met: The composite loss function value is less than the preset engineering tolerance cutoff value; The relative rate of change of the loss function is less than the preset relative tolerance within a preset number of consecutive steps; The mean square error of the verification set did not decrease or rebound within a preset number of iterations.
