Landslide identification method based on physical constraint machine learning

Abstract

Disclosed in the present invention is a landslide identification method based on physical constraint machine learning, comprising: first, collecting slope data information and storing same for later use; then, utilizing the collected data, and embedding physical constraints into a neural network for learning to obtain a prediction model, the physical constraints comprising an erosive landslide mobility dynamic equation, a boundary condition and an initial condition; and collecting slope data information of a slope to be predicted, and then using the prediction model obtained in step S2 to perform predictive analysis on the motion process of said slope, so as to implement the prediction of the landslide dynamic process of said slope. According to the present invention, more physical information, such as a mathematical physical equation, a boundary condition, an initial condition and the like related to landslide, are added into a neural network of machine learning, such that the neural network can better "understand" the physical significance of the landslide, and the neural network can predict the landslide more accurately in the absence of much observational data.

Description

A landslide identification method based on physical constraint machine learning Technical Field

[0001] The present invention belongs to the technical field of landslide identification and early warning, and in particular relates to a landslide identification method based on physical constraint machine learning. Background Art

[0002] Landslides are natural hazards that can cause severe damage to human settlements and infrastructure, resulting in loss of life and economic losses. Understanding and predicting landslides is a frontier in engineering geology and a crucial task in mitigating their impacts. Machine learning (ML) is a powerful tool for analyzing landslide data, identifying landslides, and predicting them. In recent years, the majority of published literature on the application of ML to landslide research falls into one of three broad categories: landslide identification, spatial prediction of landslide susceptibility, and temporal prediction of landslides. One application of ML in landslide research is to predict landslide occurrence by analyzing geospatial and geological data, such as slope, elevation, land cover, rainfall patterns, and drainage density, to estimate landslide susceptibility and risk. Another application of ML in landslide research is to analyze the dynamic processes of landslides, including their initiation, propagation, and stability. For example, ML has been used to analyze landslide deformation and motion based on satellite imagery and remote sensing data.

[0003] In machine learning, data collection methods used in landslide research primarily include satellite-based synthetic aperture radar (SAR) differential interferometry, satellite optical imagery, drone photogrammetry, and LiDAR (Light Detection and Ranging) (GLONASS). To develop a comprehensive landslide observation system, multiple technologies must be integrated to construct a multi-dimensional monitoring system encompassing satellite-based InSAR and high-resolution optical imagery, aerial drone photogrammetry and airborne LiDAR, and ground-based monitoring of slope surfaces and interiors. This system enables multi-layered, multi-angle, and multi-method all-weather monitoring of major landslide hazards—a collaborative sky-air-ground monitoring system. However, due to the long revisit periods of satellites or drones and the limited monitoring indicators, it is impossible to automatically monitor the rapid deformation processes of landslides and the related physical quantities that induce landslide deformation in real time. Relying solely on sky-air monitoring technologies makes it difficult to monitor and warn landslides in the short term and during emergency phases. Therefore, ground-based and soil-in-soil monitoring is also necessary. Furthermore, landslides are natural hazards that occur infrequently and are often localized, making obtaining sufficient data a challenge. Furthermore, due to a lack of data, trained machine learning models may overfit existing data, resulting in an inability to accurately predict unknown scenarios. Furthermore, due to the scarcity of data and the complexity of features in landslide research, the generalization ability of machine learning models is often limited, potentially leading to a decrease in performance in new scenarios. Many factors and variables in landslide research are uncertain, which also poses challenges for machine learning algorithms.

[0004] In summary, while landslide identification research has benefited significantly from advances in machine learning, temporal and spatial prediction still appears to have not benefited significantly from machine learning (ML) algorithms. This is primarily because landslide detection is essentially a computer vision (CV) problem, which does not require much physical understanding. However, the temporal and spatial prediction of landslides requires more information on physical mechanisms. Landslide development is the result of long-term evolutionary processes driven by dynamic forces both within and outside the slope's geologic structure. Scientific understanding of landslide evolution and its physical mechanisms is crucial for accurate landslide prediction and forecasting. However, most landslide machine learning algorithms are data-driven and disconnected from the landslide evolution process and physical and mechanical mechanisms. This has led to significant theoretical and technical bottlenecks in these prediction models, resulting in slow progress in theoretical research on landslide prediction and forecasting, and a low success rate for landslide prediction and forecasting.

[0005] Summary of the Invention

[0006] In response to the above shortcomings, the present invention discloses a landslide identification method based on physical constraint machine learning. Physical constraints such as mathematical physics equations, boundary conditions, and initial conditions related to landslides are added to the machine learning neural network, scientifically understanding the physical mechanism and evolution process of landslides, achieving accurate prediction and early warning of landslides, and solving the problem that landslides cannot be accurately predicted through the network when the amount of observation data is small.

[0007] The present invention is achieved by adopting the following technical solutions:

[0008] A landslide identification method based on physical constraint machine learning includes the following steps:

[0009] S1. Collecting slope data and storing it for future use. The slope data includes but is not limited to the location information of the slope, rock mass parameters of the slope, average thickness of the landslide, speed of the landslide, time for the landslide to move along the landslide surface, and distance the landslide moves along the landslide surface.

[0010] S2. Using the data collected in step S1, and embedding physical constraints into a neural network for learning to obtain a prediction model, wherein the physical constraints include conservation of mass and momentum during erosive landslide flow, boundary conditions, and initial conditions;

[0011] The conservation of mass and momentum in the erosional landslide flow process includes the following formulas:

[0012] Formula (1):

[0013] Formula (2):

[0014] Formula (3): Wherein formula (1) represents the mass conservation equation, formula (2) represents the momentum conservation equation, where h and u represent the average thickness and velocity of the landslide, t and x represent the time and distance the landslide moves along the landslide surface, E represents the erosion rate, and γ m is the density ratio between the fluid and the particles, α m is the volume fraction of solid particles, μ m is the friction coefficient, g x and g z represent the gravitational acceleration along and perpendicular to the landslide surface, respectively; K is the earth pressure coefficient; C DV is the viscous drag coefficient, b represents the surface of the eroded substrate, which is a function of time and space, and λ b is a parameter related to erosion velocity, i.e., erosion drift, and the superscripts m and b denote quantities related to the landslide mass and base, respectively;

[0015] The schematic diagram of the algorithm for embedding the dynamic model of erosive landslide mobility into a neural network is shown in FIG1 ;

[0016] The initial conditions include the following formula:

[0017] Formula (4): u| t＝0 =0

[0018] Formula (5): h| t＝0 =h0

[0019] The boundary conditions include the following formula:

[0020] Formula (6):

[0021] Formula (7): h| m =0

[0022] S3. Collect slope data information of the slope to be measured, and then use the prediction model obtained in step S2 to perform prediction analysis on the slope to be measured, so as to achieve landslide identification and early warning of the slope to be measured.

[0023] For further optimization, in step S2, the conservation of mass and momentum in the erosive landslide flow process is embedded in the neural network. That is, formulas (1) to (7) are added to the loss function. The gradient descent method is used to gradually reduce the value of the loss function. Through continuous training, the neural network is finally made to satisfy the equations. Since the velocity may not satisfy formulas (1) and (2) during the random initialization stage of the neural network, the gradient descent method is used to gradually reduce the value of the loss function. Through continuous training, the neural network is finally made to satisfy the equations.

[0024] For further optimization, the gradient descent method specifically selects an appropriate initial value u(0), iterates continuously, updates the value of u, and minimizes the objective function until convergence. The gradient descent method is as follows:

[0025] Formula (8): Where α represents the learning rate and J represents the loss function.

[0026] For further optimization, based on the comparison between experimental data and simulation results, the prediction model is modified to obtain an accurate early warning model.

[0027] Compared with the existing technology, this technical solution has the following beneficial effects:

[0028] This method imposes physical constraints on the neural network. Traditional numerical methods, such as the finite element method and the finite difference method, require meshing the computational domain. This extensive meshing process is time-consuming, and mesh generation is also difficult when the domain structure is complex. This method, however, does not require meshing and is a meshless method. Furthermore, using neural networks facilitates the derivation of equations because neural networks can perform backpropagation or automatic differentiation. This method can also incorporate other physical information to further complete the physical mechanisms of the neural network. As our understanding of the physical mechanisms of landslides increases, the equations describing them will continue to improve, allowing neural networks to keep pace with developments in the physical mechanisms of landslides. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] FIG1 is a schematic diagram of an algorithm for embedding the dynamic model of erosive landslide mobility into a neural network. DETAILED DESCRIPTION

[0030] The present invention is further illustrated by the following examples, which are not intended to limit the present invention. Specific experimental conditions and methods not specified in the following examples are conventional methods well known to those skilled in the art.

[0031] Example 1: A landslide identification method based on physical constraint machine learning, comprising the following steps:

[0032] S1. Collecting slope data and storing it for future use. The slope data includes but is not limited to the location information of the slope, rock mass parameters of the slope, average thickness of the landslide, speed of the landslide, time for the landslide to move along the landslide surface, and distance the landslide moves along the landslide surface.

[0033] S2. Using the data collected in step S1, and embedding physical constraints into a neural network for learning to obtain a prediction model, wherein the physical constraints include conservation of mass and momentum during erosive landslide flow, boundary conditions, and initial conditions;

[0034] The conservation of mass and momentum in the erosional landslide flow process includes the following formulas:

[0035] Formula (1):

[0036] Formula (2):

[0037] Formula (3): Wherein formula (1) represents the mass conservation equation, formula (2) represents the momentum conservation equation, where h and u represent the average thickness and velocity of the landslide, t and x represent the time and distance the landslide moves along the landslide surface, E represents the erosion rate, and γ m is the density ratio between the fluid and the particles, α m is the volume fraction of solid particles, μ m is the friction coefficient, g x and g z represent the gravitational acceleration along and perpendicular to the landslide surface, respectively; K is the earth pressure coefficient; C DV is the viscous drag coefficient, b represents the surface of the eroded substrate, which is a function of time and space, and λ b is a parameter related to erosion velocity, i.e., erosion drift, and the superscripts m and b denote quantities related to the landslide mass and base, respectively;

[0038] The schematic diagram of the algorithm for embedding the dynamic model of erosive landslide mobility into a neural network is shown in FIG1 ;

[0039] The initial conditions include the following formula:

[0040] Formula (4): u| t＝0 =0

[0041] Formula (5): h| t＝0 =h0

[0042] The boundary conditions include the following formula:

[0043] Formula (6):

[0044] Formula (7): h| m =0

[0045] S3. Collect slope data information of the slope to be measured, and then use the prediction model obtained in step S2 to perform prediction analysis on the slope to be measured, so as to realize landslide identification and early warning of the slope to be measured; at the same time, the prediction model can be modified based on the comparison of experimental data and simulation results to obtain an accurate early warning model.

[0046] In the algorithm diagram shown in Figure 1, the NN part represents the forward neural network, x and z represent spatial coordinates, and u1~u4 can represent unknown speeds. In order to solve these unknowns, training is required and a loss function needs to be constructed. If a part of clear data information can be collected, such as u in Figure 1 1,target, then a commonly used loss function can be used, such as the root mean square error (MSE). In the random initialization stage of the neural network, the velocity of the landslide body may not satisfy Formula (1) and Formula (2). Formula (1) to Formula (7) are added to the loss function, and the gradient descent method is used to gradually reduce the value of the loss function. Through continuous training, the neural network will eventually satisfy the equation; the gradient descent method specifically selects an appropriate initial value u(0), continuously iterates, updates the value of u, and minimizes the objective function until convergence. The gradient descent method is as follows:

[0047] Formula (8): Where α represents the learning rate and J represents the loss function.

[0048] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A landslide identification method based on physical constraint machine learning, characterized by: The following steps are involved: S1. Collecting slope data and information and saving them for future use. The slope data and information include but are not limited to the location information of the slope, rock mass parameters of the slope, average thickness of the landslide body, speed of the landslide body, time for the landslide body to move along the landslide surface, distance for the landslide body to move along the landslide surface, density ratio between fluid and particles, volume fraction of solid particles, friction coefficient, earth pressure coefficient and viscous resistance coefficient; S2, using the data collected in step S1, and embedding physical constraints into a neural network for learning to obtain a prediction model, wherein the physical constraints include conservation of mass and momentum, boundary conditions, and initial conditions in the erosive landslide flow process; The conservation of mass and momentum in the erosional landslide flow process includes the following formulas: Formula (1): Formula (2): Formula (3): Wherein, formula (1) represents the mass conservation equation, and formula (2) represents the momentum conservation equation. Here, h and u represent the average thickness and velocity of the landslide body, respectively; t and x represent the time and distance of the landslide body moving along the landslide surface, respectively; E represents the erosion rate, and γ m is the density ratio between the fluid and the particles, α m is the volume fraction of solid particles, μ m is the friction coefficient, g x and g z They represent the gravitational acceleration along and perpendicular to the landslide surface, respectively, K is the earth pressure coefficient, C DV is the viscous drag coefficient, b represents the surface of the eroded substrate, which is a function of time and space, and λ b is a parameter related to erosion velocity, i.e., erosion drift, and the superscripts m and b denote quantities related to the landslide mass and base, respectively; The schematic diagram of the algorithm for embedding the dynamic model of erosive landslide mobility into a neural network is shown in FIG1 ; The initial conditions include the following formula: Formula (4): u| t＝0 =0 Formula (5): h| t＝0 =h0 The boundary conditions include the following formula: Formula (6): Formula (7): h| m =0 S3, collecting slope data information of the slope to be measured, and then using the prediction model obtained in step S2 to perform prediction analysis on the slope to be measured, so as to achieve landslide identification and early warning of the slope to be measured.

2. The landslide identification method based on physical constraint machine learning according to claim 1 is characterized in that: In step S2, the conservation of mass and momentum in the erosive landslide flow process is embedded into the neural network, that is, formula (1) to formula (7) are added to the loss function, and the gradient descent method is used to gradually reduce the value of the loss function. Through continuous training, the neural network finally satisfies the equation.

3. The landslide identification method based on physical constraint machine learning according to claim 2 is characterized in that: The gradient descent method specifically selects an appropriate initial value u(0), continuously iterates, updates the value of u, and minimizes the objective function until convergence. The gradient descent method is as follows: Formula (8): Where α represents the learning rate and J represents the loss function.

4. The landslide identification method based on physical constraint machine learning according to claim 1 is characterized in that: According to the comparison between the experimental data and the simulation results, the prediction model is modified and an accurate early warning model is obtained.

Images