A ground collapse probability prediction method based on physical guidance neural spectrum operator

Abstract

The application discloses a ground collapse probability prediction method based on a physical guiding neural spectrum operator, and relates to the technical fields of municipal engineering and disaster prevention and reduction. The method firstly constructs a CFD-DEM fluid-solid coupling simulation model considering the coupling of double flow fields inside and outside the pipe, and obtains multi-physical field time sequence data in the soil erosion process; a physical guiding neural spectrum operator model is constructed, then, the velocity weighted normal velocity gradient of the local flow field is calculated, and a Beta random process is introduced to construct a probability fluidization function representing the mapping relationship between the hydraulic parameters and the soil fluidization time; finally, the soil fluidization probability is calculated according to the real-time hydraulic parameters. Through the fusion of the physical mechanism and the deep learning algorithm, the problems of data scarcity, poor physical consistency of the model and inability to quantify the random risk in the prior art are effectively solved, and high-precision and probabilistic prediction of the collapse risk caused by the damage of the drainage pipe is realized.

Description

Technical Field

[0001] This invention relates to the field of municipal engineering and urban disaster prevention and mitigation technology, specifically to a ground collapse probability prediction method based on a physical-guided neural spectrum operator. Background Technology

[0002] With the acceleration of urbanization, underground drainage pipe networks are becoming increasingly extensive and severely aging. Soil erosion caused by damage, corrosion, or misalignment of drainage pipes has become one of the main causes of urban ground collapse. This geological hazard is characterized by its suddenness, high degree of concealment, and great destructive power, seriously threatening urban public safety and the stable operation of infrastructure. Therefore, in-depth research into the evolution mechanism of pipe damage leading to collapse and effective risk prediction are of significant engineering importance.

[0003] Currently, research and prediction methods for this type of ground subsidence disaster mainly include physical model experiments, numerical simulation analysis, and traditional data-driven models. However, these existing technologies still have the following limitations in practical applications: Limitations of physical model tests: Existing studies often employ scaled-down model tests or geotechnical centrifuge tests to simulate soil erosion processes. However, physical tests are not only costly and time-consuming, but also limited by size effects, making it difficult to accurately reproduce the complex stress states at the prototype scale. Furthermore, physical tests often struggle to provide real-time, full-field, detailed observations of the micromechanical behavior within soil particles (such as the evolution of contact force chains between particles and the distribution of fluid drag forces), resulting in insufficient depth in understanding the mechanisms of disaster evolution.

[0004] Limitations of Traditional Numerical Simulation Methods: While computational fluid dynamics-discrete element coupled methods (CFD-DEM) can effectively simulate fluid-structure interaction processes from a microscopic perspective, their computational load is extremely large and time-consuming. Engineering risk assessments often require consideration of the stochasticity of various factors such as soil parameters and hydraulic conditions, necessitating massive amounts of repetitive calculations. Traditional CFD-DEM methods, due to computational efficiency limitations, struggle to meet the demands of such large-scale probabilistic analysis. This has led to existing research being largely limited to deterministic analysis under specific working conditions, neglecting the inherent stochasticity of soil fluidization processes and making it difficult to quantify the probabilistic risk of collapse.

[0005] The shortcomings of purely data-driven deep learning models: In recent years, prediction models based on deep learning (such as CNN, LSTM, etc.) have been widely used in geological disaster early warning. However, these purely data-driven models mainly have two problems: First, they are highly dependent on data, requiring massive amounts of high-quality labeled data for training. However, the acquisition cost of high-quality CFD-DEM fluid-structure interaction simulation data is extremely high, leading to the small sample dilemma and making the model prone to overfitting. Second, they have poor physical consistency. Traditional neural network models are usually black-box models, lacking the constraints of physical mechanisms. Their prediction results often violate the basic laws of physics, resulting in a significant reduction in the interpretability and credibility of the prediction results in engineering practice.

[0006] In summary, overcoming the high cost of physical experiments and the high computational barrier of numerical simulation, and constructing a ground subsidence prediction method that conforms to physical mechanisms and can accurately quantify random risks under small sample data conditions, is a technical problem that urgently needs to be solved. Summary of the Invention

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a ground collapse probability prediction method based on a physical-guided neural spectrum operator, characterized by comprising the following steps: Step S1: Construct a computational fluid dynamics-discrete element method (CFD-DEM) fluid-structure interaction simulation model considering the coupling effect of the two flow fields inside and outside the pipe; solve the fluid field based on the Navier-Stokes equations, solve the soil particle motion based on the discrete element method, use the Hertz-Mindlin with JKR contact model to describe the contact behavior between cohesive soil particles, and obtain multi-physics time series data of the soil erosion process; Step S2: Construct a physical-guided neural spectrum operator model, including a dual encoder module and a spectrum dynamics learning module. The dual encoder module is used to extract low-dimensional basic features of the input data, i.e., multi-physics time-series data, and encode the fluid dynamics physical parameters into physical enhancement feature vectors. The spectrum dynamics learning module is used to map the physical enhancement feature vectors to the spectrum space and separate them into macroscopic trend branches and microscopic fluctuation branches for dynamic feature learning. Step S3: Construct a physical-guided neural spectrum operator inversion model, including a physical constraint module and a time-domain reconstruction module; construct a comprehensive loss function of data fitting term and physical constraint term, and train the physical-guided neural spectrum operator model in Step S2; use the trained physical-guided neural spectrum operator model to predict the time series curve of soil particle loss rate and perform data propagation to generate a high-resolution dataset of the entire soil erosion process. Step S4: Calculate the local velocity-weighted normal velocity gradient of the flow field based on the simulation results of computational fluid dynamics-discrete element method (CFD-DEM); introduce a Beta stochastic process to describe the stochastic distribution characteristics of soil fluidization time based on the high-resolution full-process soil erosion dataset; construct the probability fluidization function between the velocity-weighted normal velocity gradient and the soil fluidization time. Step S5: Obtain real-time hydraulic parameters of the target area, input the probabilistic fluidization function, calculate the conditional probability of soil fluidization, and output the ground collapse risk assessment results.

[0008] Further, step S2, the specific steps are as follows: Step S21, the Physically Guided Neural Spectral Operator (PG-NSO) establishes a mapping function from the physical parameter space to the time series space: The input is a three-dimensional physical parameter vector. v represents flow velocity, h represents height, p represents pressure, and the output is a time-domain signal. Where T = 46 time points; The physical-guided neural spectrum operator employs an operator decomposition strategy, decomposing the mapping into four functionally defined sub-operators: (1); in Low-dimensional physical parameters are dual-encoded into high-dimensional features and injected into physical priors; Learning dynamics in the spectral space to achieve scale separation; Imposing physical constraints ensures the rationality of the solution; Reconstructing the time-domain signal from the spectral space; Step S22, First Encoding: Basic Feature Extraction: (2); in: and These are the weight matrices for the first and second layers; and These are the bias vectors for the first and second layers; For Swish activation function, The activation function for the rectified linear unit is... This is the encoded basic feature vector; Step S23, Second Coding: Physical Prior Injection (3); in, and The Reynolds number inside the damaged pipe. and The kinetic and pressure energy of the fluid inside the damaged pipe. This is the encoded physical enhancement feature vector; Step S24: To capture the "gradual-rapid-gradual" transition characteristics of the collapse curve, a new time-series feature prediction subnetwork is added: (4); (5); (6); in: The location and time of the end of the first turning point and the first collapse phase. This marks the start time of the second turning point, which is also the final collapse phase. Let be the slope parameter of segment s. For saturation level parameters, for During the calculation Weight parameters for During the calculation Weight parameters for During the calculation Weight parameters for During the calculation The weight parameters are the weight parameters. for During the calculation bias parameters, for During the calculation bias parameters, for During the calculation The bias parameters, for During the calculation The bias parameters, For the reason express Function identifier, For the reason express Function identifier, refer to function; The final encoder outputs the enhanced feature vector as follows: (7); in, The final encoder outputs an enhanced feature vector. These are the slope parameters for stages 1, 2, 3, and 4.

[0009] Step S25: The spectrum dynamics learner consists of a macroscopic branch and a microscopic branch; the macroscopic branch processes low-frequency, global trend components, while the microscopic branch processes high-frequency, local oscillation components. (8); (9); in: The first layer of weights for the macro branch. The second layer of weights for the macro branch; First-level bias for macroscopic branch This is the second-level bias for the macro branch; The first layer weight of the micro-branch, The second layer weights for the micro-branch; For the first layer bias of micro-branch and The second layer bias is used for micro-branching; This is a macroscopic vector of spectral coefficients. This represents the vector of microscopic spectral coefficients. Macroeconomic trends are based on a piecewise cubic polynomial base: (10); in, , , and For polynomial coefficients, Based on macroeconomic trends, For the first stage of polynomial basis, For the polynomial basis of the second and third stages, For the fourth stage of polynomial basis; Step S26: Based on the full-process fluctuations that realize time-varying changes, use a globally supported periodic basis as the micro-trend basis function: (11); in As the micro-trend base, t is time. and These are frequency and phase, respectively. To ensure that the intensity of fluctuations is aligned with macroeconomic trends, the amplitude function should be correlated with the macroeconomic trends. (12); in, Let be the amplitude function. Based on amplitude, This is the trend sensitivity coefficient. This represents the instantaneous slope of the macro trend.

[0010] Further, step S3, the specific steps are as follows: Step S31, integrate the loss function It is a weighted summation form: (13); in, It is a hyperparameter that balances various losses. , , , and These are data fitting loss, nonnegativity constraint, monotonicity constraint, multi-scale separation constraint, and inflection point smoothing constraint. , , and These are the trade-off hyperparameters for the above constraints, used to control the strength of the physical constraints; All parameters of the physical-guided neural spectrum operator inversion model are updated using gradient descent. Step S32, Temporal Reconstruction Module The function is responsible for converting the learned spectral coefficients back into physically meaningful time-domain signals. The complete mathematical expression is: (14); Where y is the reconstructed time-domain signal, and These are the spectral coefficients of each macroscopic and microscopic basis. and For each macroscopic basis function and microscopic basis function, For the number of macroscopic basis functions, denoted as the number of microscopic basis functions.

[0011] Further, step S4, the specific steps are as follows: Step S41, the soil collapse fluidization process is a typical solid-liquid two-phase coupled dynamic problem in geotechnical engineering. In its later stage (i.e., the second stage of collapse), the system enters a fully fluidized state. The start time of the later stage... It exhibits significant random characteristics and is strongly influenced by some hydraulic parameters.

[0012] The instantaneous soil particle loss rate curves were multiplied using the Physically Guided Neural Spectrum Operator (PG-NSO), and the sliding window comparison method was used to determine the start time of the second stage of the multiplied curve group. The identification process is performed; if position i satisfies equation (15), it is identified as a turning point: (15); in: for The average slope of the interval, for The average slope of the interval, Half the width of the window For threshold; Step S42, the velocity-weighted normal velocity gradient, which describes the local characteristics of the flow field, is defined as: (16); in, For velocity-weighted normal velocity gradient, The average fluid velocity at the location of soil particles, For soil particle size, The fluid velocity variation rate is within the range of soil particle size. Step S43, Normalized weighted normal velocity gradient and start time of the second stage: (17); (18); Step S44, start of the second stage The conditional distribution follows a parameterized Beta distribution: ; (19); Probabilistic fluidization function: under given hydraulic conditions The conditional probability of fluidization occurring before time t is defined as follows: ; Analytical expression of the probability flow function: Based on the conditional Beta distribution assumption, the above equation can be expressed as follows: (20); In the formula, , This is a Beta function with incomplete regularization. Step S45, start of the second stage The conditional expectation can be parsably expressed as: (twenty one); Step S46, due to the start time of the second stage Given the non-linear decreasing trend, we choose the exponential function as the basic form: (twenty two); The above formula , , , , and The parameter to be estimated; The model parameters are solved using constrained optimization methods: (twenty three).

[0013] The beneficial effects of this invention are as follows: By constructing a Physically Guided Neural Spectrum Operator (PG-NSO) model, this invention utilizes a small amount of high-fidelity CFD-DEM simulation data for physical enhancement feature extraction and data propagation, effectively resolving the contradiction between the high computational cost of traditional numerical simulation and the small sample dilemma faced by deep learning. Simultaneously, this invention explicitly injects physical priors such as Reynolds number and Euler number into the model and applies physical constraints such as nonnegativity and monotonicity to the loss function. Combined with the spectrum dynamics module for scale separation modeling of macroscopic trends and microscopic fluctuations, this not only eliminates the black-box defects of purely data-driven models and ensures the physical consistency and precision of prediction results, but also achieves a leap from deterministic prediction to probabilistic risk assessment by introducing a Beta stochastic process to construct a probabilistic flow function, providing a scientific and reliable decision-making basis for urban disaster prevention and mitigation. Attached Figure Description

[0014] Figure 1 This is the overall flowchart of the invention of this method. Detailed Implementation

[0015] 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. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0016] Example 1: A ground collapse probability prediction method based on physically guided neural spectrum operators: This embodiment provides a ground collapse probability prediction method based on the Physically Guided Neural Spectrum Operator (PG-NSO). This method integrates high-fidelity computational fluid dynamics-discrete element coupling mechanism with deep learning algorithms and incorporates stochastic process analysis to quantify the risk of drainage pipe rupture leading to collapse. The specific implementation steps are as follows: Step S1: Constructing a multiphysics coupled simulation model and acquiring data First, a computational fluid dynamics-discrete element method (CFD-DEM) fluid-structure interaction simulation model considering the coupling effect of the two flow fields inside and outside the pipe is constructed.

[0017] In the fluid domain solution, the Navier-Stokes equations are used to solve the fluid field, accurately simulating the pressure distribution of high-velocity water flow inside the pipe and the seepage field outside the pipe. In the solid domain solution, the discrete element method (DEM) is used to solve the soil particle motion. To accurately describe the contact behavior between cohesive soil particles, this embodiment uses the Hertz-Mindlin with JKR contact model, which considers the effects of van der Waals forces and surface energy between particles.

[0018] By co-simulating, we can obtain multi-physics time-series data (such as velocity field, pressure field, particle loss rate, and cavity expansion morphology) during soil erosion, which can be used as training data and physical ground truth for subsequent deep learning models.

[0019] Step S2: Construct a Physically Guided Neural Spectrum Operator (PG-NSO) model The core model constructed in this step—the Physically Guided Neural Spectral Operator (PG-NSO)—is essentially a nonlinear mapping function from the physical parameter space to the time series space: ; The input is a three-dimensional physical parameter vector. (representing flow rate, height, and pressure respectively), Output For time series (e.g., soil loss rate curves), which contain A specific point in time.

[0020] PG-NSO employs an operator decomposition strategy, breaking down the complex mapping process into four functionally defined sub-operators, whose mathematical expressions are shown in formula (1): (1); in: It is a dual encoder used to dual-encode low-dimensional physical parameters into high-dimensional features and inject physical prior knowledge; This is a spectrum dynamics learning module used to learn the dynamic evolution laws in the spectrum space, achieving separation of macroscopic and microscopic scales; This is the physical constraint module, used to apply physical constraints (such as nonnegativity and monotonicity) to ensure the rationality of the solution; This is a time-domain reconstruction module used to reconstruct the coefficients back to the time-domain signal from the spectral space.

[0021] The specific implementation details of the model structure are as follows: 1. Dual encoder module (steps S21-S23) First level of encoding: Basic feature extraction Low-dimensional basic features of the input data are extracted using a multilayer perceptron. The calculation formula is shown in formula (2): (2); in: and These are the weight matrices for the first and second layers; and These are the bias vectors for the first and second layers; The Swish activation function is chosen to ensure gradient propagation in deep networks. The Rectified Linear Unit (ReLU) activation function is selected to increase feature sparsity; This is the encoded basic feature vector.

[0022] Second-level encoding: Physical prior injection To enhance the physical interpretability of the model, this embodiment explicitly injects the dimensionless fluid dynamics number and energy parameters into the eigenvector. The calculation formula is shown in formula (3): (3); Where Re is the Reynolds number and Eu is the Euler number (characterizing the ratio of pressure difference to dynamic pressure). The kinetic energy of the fluid inside the damaged pipe. This is pressure energy (potential energy); This is the encoded physical enhancement feature vector.

[0023] Temporal Feature Prediction Subnet To address the typical "gradual-rapid-gradual" transition characteristics of soil collapse loss rate curves, this embodiment designs a dedicated subnetwork to predict key time-series parameters. The calculation formulas are shown in formulas (4)-(6): (4); (5); (6); in: and The time positions of the two key inflection points of the curve (corresponding to the fluidization start and steady-state phases, respectively); For the first The slope parameter of the segment; This is the parameter for the final saturation loss level; For weight parameters, This is the bias parameter.

[0024] Finally, the enhanced feature vector output by the encoder is: (7); 2. Spectral Dynamics Learning Module (Steps S24-S25) This module consists of macroscopic and microscopic branches, which handle dynamic characteristics at different frequencies respectively. The former specifically handles low-frequency, global trend components, while the latter specifically handles high-frequency, local oscillation components. The calculation formulas are shown in formulas (8)-(9): (8); (9); in: The weights and biases of the macro branches; The weights and biases of the micro-branches; and These are the macroscopic and microscopic spectral coefficient vectors, respectively.

[0025] The macroeconomic trend base uses a piecewise cubic polynomial base to accurately fit the nonlinear trend, as shown in formula (10): (10); Where a, b, c, and d are polynomial coefficients, obtained by network learning.

[0026] The micro-trend base uses a globally supported periodic base to achieve full-range fluctuations based on time changes. As shown in formula (11): (11); in and These are frequency and phase, respectively.

[0027] To ensure that the intensity of fluctuations is consistent with the macro trend (i.e., large fluctuations when the loss rate changes drastically, and small fluctuations when it is stable), the amplitude function... It is designed to correlate with the derivative of macroeconomic trends, as shown in Equation (12): (12); in, Based on amplitude, This is the trend sensitivity coefficient. This represents the instantaneous slope of the macro trend.

[0028] Step S3: Training and Prediction of the Physically Guided Neural Spectral Operator Inversion Model 1. Construction of physical constraint loss function (step S31) To ensure that the prediction results conform to physical laws, this embodiment constructs a comprehensive loss function that includes data fitting terms and multiple physical constraint terms. The calculation formula is shown in formula (13). (13); in: The hyperparameters for balancing various losses; For data fitting loss (MSE); As a non-negativity constraint, it penalizes negative churn rate predictions; To constrain monotonicity, ensure that the cumulative loss of quality increases monotonically; Multi-scale separation constraints ensure effective decoupling of macroscopic and microscopic features; Smoothing constraints are applied to inflection points to ensure that the piecewise function is continuously differentiable at the connection points.

[0029] All parameters of the model are updated using gradient descent.

[0030] 2. Temporal reconstruction and data augmentation (step S32) The time-domain reconstruction module is responsible for converting the learned spectral coefficients back into physically meaningful time-domain signals. Its complete mathematical expression is shown in equation (14): (14); Using the trained model, we can quickly predict and augment the time-series curves of soil particle loss rate under different hydraulic conditions, generating a high-resolution dataset of the entire soil erosion process.

[0031] Step S4: Construction of the probability fluidization function This step introduces stochastic process theory to quantify the uncertainty of soil fluidization time.

[0032] 1. Identification of the start time of the second stage (fluidization time) (step S41) The later stage (second stage) of the soil collapse and fluidization process is characterized by the system entering a fully fluidized state. In this embodiment, the instantaneous soil particle loss rate curve was propagated using PG-NSO, and the sliding window comparison method was used to determine the start time of the second stage of the propagated curve. Perform identification. At location If the conditions of formula (15) are met, then it is identified as a turning point: (15); in: The average slope of the interval. Half the width of the window The threshold value is set.

[0033] 2. Calculation of velocity-weighted normal velocity gradient (step S42) Define the velocity-weighted normal velocity gradient to describe the local characteristics of a flow field. , as the main physical driving variable for collapse. The calculation formula is shown in formula (16): (16); This parameter comprehensively reflects the collapse-causing factors based on fluid kinetic energy and shear intensity.

[0034] 3. Data normalization (step S43) For the calculated and Normalization is performed, as shown in formulas (17)-(18); (17); (18); 4. Construction of the probabilistic fluidization function (step S44) Assume that the conditional distribution at the start of the second stage follows a parameterized Beta distribution: Its probability density function is shown in equation (19): (19); Furthermore, we define the probabilistic fluidization function, which is the function of time under given hydraulic conditions. The conditional probability of fluidization occurring before the fluidization occurs. Its analytical expression is a regularized incomplete Beta function, as shown in formula (20): (20); In the formula, .

[0035] 5. Solving for conditional expectation and parameters (steps S45-S46) The start of the second phase The conditional expectation can be analytically expressed as formula (21): (twenty one); Considering that the fluidization time decreases non-linearly with increasing hydraulic intensity, an exponential function is chosen as the Beta distribution parameter. The basic form is shown in formula (22): (twenty two); The above formula These are the parameters to be estimated.

[0036] Finally, the constrained optimization method is used to solve for the above model parameters, and the objective function is shown in formula (23): (twenty three).

[0037] Step S5: Ground Collapse Risk Assessment The system acquires real-time hydraulic parameters of the target area, calculates the corresponding velocity-weighted normal velocity gradient S, inputs the constructed probability fluidization function, calculates the conditional probability of soil fluidization at the current moment, and outputs the risk assessment result of ground collapse.

[0038] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any person skilled in the art, without departing from all the technical content described in the present invention, using equivalent substitutions or transformations, falls within the protection scope of the present invention.

Claims

1. A ground collapse probability prediction method based on a physically guided neural spectrum operator, characterized in that, Includes the following steps: Step S1: Construct a computational fluid dynamics-discrete element method fluid-structure interaction simulation model that considers the coupling effect of the two flow fields inside and outside the pipe; solve the fluid field based on the Navier-Stokes equations, solve the soil particle motion based on the discrete element method, use the Hertz-Mundling contact model combined with JKR to describe the contact behavior between cohesive soil particles, and obtain multi-physics time series data of the soil erosion process. Step S2: Construct a physical-guided neural spectrum operator model, including a dual encoder module and a spectrum dynamics learning module. The dual encoder module is used to extract low-dimensional basic features of the input data, i.e., multi-physics time-series data, and encode the fluid dynamics physical parameters into physical enhancement feature vectors. The spectrum dynamics learning module is used to map the physical enhancement feature vectors to the spectrum space and separate them into macroscopic trend branches and microscopic fluctuation branches for dynamic feature learning. Step S3: Construct a physical-guided neural spectrum operator inversion model, including a physical constraint module and a time-domain reconstruction module; construct a comprehensive loss function of data fitting term and physical constraint term, and train the physical-guided neural spectrum operator model in Step S2; use the trained physical-guided neural spectrum operator model to predict the time series curve of soil particle loss rate and perform data propagation to generate a high-resolution dataset of the entire soil erosion process. Step S4: Calculate the local velocity-weighted normal velocity gradient of the flow field based on the simulation results of computational fluid dynamics-discrete element method; introduce a Beta stochastic process to describe the stochastic distribution characteristics of soil fluidization time on the basis of high-resolution soil erosion full process dataset; construct the probabilistic fluidization function between velocity-weighted normal velocity gradient and soil fluidization time. Step S5: Obtain real-time hydraulic parameters of the target area, input the probabilistic fluidization function, calculate the conditional probability of soil fluidization, and output the ground collapse risk assessment results.

2. The ground collapse probability prediction method based on a physical-guided neural spectrum operator according to claim 1, characterized in that: Step S2, the specific steps are as follows: Step S21, the physical-guided neural spectrum operator establishes a mapping function from the physical parameter space to the time series space: ; The input is a three-dimensional physical parameter vector. v represents flow velocity, h represents height, p represents pressure, and the output is a time-domain signal. Where T = 46 time points; The physical-guided neural spectrum operator employs an operator decomposition strategy, breaking down the mapping into four functionally defined sub-operators: (1)； in Low-dimensional physical parameters are dual-encoded into high-dimensional features and injected into physical priors; Learning dynamics in the spectral space to achieve scale separation; Imposing physical constraints ensures the rationality of the solution; Reconstructing the time-domain signal from the spectral space; Step S22, First Encoding: Basic Feature Extraction: （2）； in: and These are the weight matrices for the first and second layers; and These are the bias vectors for the first and second layers; For Swish activation function, The activation function for the rectified linear unit is... This is the encoded basic feature vector; Step S23, Second Coding: Physical Prior Injection （3）； in, and The Reynolds number inside the damaged pipe. and The kinetic and pressure energy of the fluid inside the damaged pipe. This is the encoded physical enhancement feature vector; Step S24: To capture the "gradual-rapid-gradual" transition characteristics of the collapse curve, a new time-series feature prediction subnetwork is added: （4）； （5）； （6）； in: The location and time of the end of the first turning point and the first collapse phase. This marks the start time of the second turning point, which is also the final collapse phase. Let be the slope parameter of segment s. For saturation level parameters, for During the calculation Weight parameters for During the calculation Weight parameters for During the calculation Weight parameters for During the calculation The weight parameters are the weight parameters. for During the calculation bias parameters, for During the calculation bias parameters, for During the calculation The bias parameters, for During the calculation The bias parameters, For the reason express Function identifier, For the reason express Function identifier, refer to function; The final encoder outputs the enhanced feature vector as follows: （7）； in, The final encoder outputs an enhanced feature vector. These are the slope parameters for stages 1, 2, 3, and 4; Step S25: The spectrum dynamics learner consists of a macroscopic branch and a microscopic branch; the macroscopic branch processes low-frequency, global trend components, while the microscopic branch processes high-frequency, local oscillation components. （8）； （9）； in: The first layer of weights for the macro branch. The second layer of weights for the macro branch; First-level bias for macroscopic branch This is the second-level bias for the macro branch; The first layer weight of the micro-branch, The second layer weights for the micro-branch; For the first layer bias of micro-branch and The second layer bias is used for micro-branching; This is a macroscopic vector of spectral coefficients. This represents the vector of microscopic spectral coefficients. Macroeconomic trends are based on a piecewise cubic polynomial base: （10）； in, , , and For polynomial coefficients, Based on macroeconomic trends, For the first stage of polynomial basis, For the polynomial basis of the second and third stages, For the fourth stage of polynomial basis; Step S26: Based on the full-process fluctuations that realize time-varying changes, use a globally supported periodic basis as the micro-trend basis function: （11）； in As the micro-trend base, t is time. and These are frequency and phase, respectively. To ensure that the intensity of fluctuations is aligned with macroeconomic trends, the amplitude function should be correlated with the macroeconomic trends. （12）； in, Let be the amplitude function. Based on amplitude, This is the trend sensitivity coefficient. This represents the instantaneous slope of the macro trend.

3. The ground collapse probability prediction method based on a physical-guided neural spectrum operator according to claim 1, characterized in that: Step S3, the specific steps are as follows: Step S31, integrate the loss function It is a weighted summation form: （13）； in, It is a hyperparameter that balances various losses. , , , and These are data fitting loss, nonnegativity constraint, monotonicity constraint, multi-scale separation constraint, and inflection point smoothing constraint, respectively. , , and These are the trade-off hyperparameters for the above constraints, used to control the strength of the physical constraints; All parameters of the physical-guided neural spectrum operator inversion model are updated using gradient descent. Step S32, Temporal Reconstruction Module The function is responsible for converting the learned spectral coefficients back into physically meaningful time-domain signals. The complete mathematical expression is: （14）； Where y is the reconstructed time-domain signal, and These are the spectral coefficients of each macroscopic and microscopic substrate. and For each macroscopic basis function and microscopic basis function, For the number of macroscopic basis functions, denoted as the number of microscopic basis functions.

4. The ground collapse probability prediction method based on a physical-guided neural spectrum operator according to claim 1, characterized in that: Step S4, the specific steps are as follows: Step S41: The instantaneous soil particle loss rate curve is multiplied using a physically guided neural spectrum operator, and the sliding window comparison method is used to determine the start time of the second stage of the multiplied curve group. The identification process is performed; if position i satisfies equation (15), it is identified as a turning point: （15）； in: for The average slope of the interval, for The average slope of the interval, Half the width of the window For threshold; Step S42, the velocity-weighted normal velocity gradient, which describes the local characteristics of the flow field, is defined as: （16）； in, For velocity-weighted normal velocity gradient, The average fluid velocity at the location of soil particles, For soil particle size, The fluid velocity variation rate is within the range of soil particle size. Step S43, Normalized weighted normal velocity gradient and start time of the second stage: （17）； （18）； Step S44, start of the second stage The conditional distribution follows a parameterized Beta distribution: ; （19）； Probabilistic fluidization function: under given hydraulic conditions The conditional probability of fluidization occurring before time t is defined as follows: ; Analytical expression of the probability flow function: Based on the conditional Beta distribution assumption, the above equation can be expressed as follows: （20）； In the formula, , This is a Beta function with incomplete regularization. Step S45, start of the second stage The conditional expectation can be parsably expressed as: （21）； Step S46, due to the start time of the second stage Given the non-linear decreasing trend, we choose the exponential function as the basic form: （22）； The above formula , , , , and The parameter to be estimated; The model parameters are solved using constrained optimization methods: （23）。

Images