Landslide deformation prediction method and landslide early warning system based on data and physical driving

Abstract

The invention provides a landslide deformation prediction method and a landslide early warning system based on data and physical driving, belongs to the technical field of geological engineering and machine learning, and aims to improve the accuracy of landslide deformation prediction, and a landslide deformation prediction model for predicting landslide displacement has relatively high interpretability and universality. The method comprises the following steps: constructing a landslide data set; adopting the trained physical mechanism matching model to determine a target landslide physical mechanism model matched with each landslide data sample in the landslide data set; taking each landslide data sample in the landslide data set as an input value, taking real landslide deformation data corresponding to each landslide data sample as an observation value, and training a landslide deformation prediction model by adopting a machine learning algorithm; the loss function of the prediction model comprises a physical loss function generated based on a target landslide physical mechanism model; and inputting the landslide data of the to-be-detected area into the landslide deformation prediction model to predict the landslide deformation data of the to-be-detected area.

Description

Technical Field

[0001] The embodiments of the present application belong to the technical fields of geological engineering and machine learning, and relate to a landslide deformation prediction method and a landslide warning system based on data and physical driving. Background Technique

[0002] With the changes of natural conditions and geological conditions, rock and soil masses will gradually deform. When the deformation reaches a certain degree, landslides may occur, which will cause huge losses to agricultural production and the safety of life and property, and even bring devastating disasters. Therefore, it is of great significance to predict the landslide displacement of rock and soil masses in advance to know the landslide situation and give necessary warnings.

[0003] Currently, among many techniques for predicting landslide displacement based on machine learning methods, the effects of machine learning models for predicting landslide displacement all depend on data driving, that is, a large amount of landslide data is required for model training. In the case of limited data volume, the accuracy of the predicted landslide displacement is not high.

[0004] Furthermore, in the existing methods, it is impossible to intuitively understand how to obtain the prediction result from the input data through a machine learning model, that is, the existing machine learning models lack interpretability; and due to the influence of different geological conditions, climate conditions and human activities, the machine learning models do not have universality, that is, the generalization ability of the machine learning models is weak. Summary of the Invention

[0005] The present application proposes a landslide deformation prediction method and a landslide warning system based on data and physical driving, and introduces a landslide physical mechanism model that can reflect the physical processes and laws of the occurrence and development of landslides to train the landslide deformation prediction model, improving the prediction effect of the landslide deformation prediction model, thereby improving the accuracy of the landslide deformation prediction result, and then accurately warning of landslides based on the landslide deformation prediction result. Furthermore, since the landslide physical mechanism model is introduced into the landslide deformation prediction model, the landslide deformation prediction model has strong interpretability and generalization ability. Further,...

[0006] In order to achieve the above object, the present application adopts the following technical solutions:

[0007] In a first aspect, the present application provides a landslide deformation prediction method based on data and physical driving, including: constructing a landslide data set, which includes a plurality of landslide data samples and the corresponding true landslide deformation data of the plurality of landslide data samples. A landslide data sample includes landslide data of a plurality of sampling points in a monitoring area at a plurality of time points; wherein, the landslide data includes geological parameters, terrain and geological structure data, and environmental factor data, and the landslide deformation data includes the landslide displacements of the plurality of sampling points at a plurality of time points; using a trained physical mechanism matching model to determine a target landslide physical mechanism model that matches each landslide data sample in the landslide data set. The input of the physical mechanism matching model is each landslide data sample in the landslide data set, and the output of the physical mechanism matching model is used to indicate the target landslide physical mechanism model, and the target landslide physical mechanism model is used to describe the physical laws followed by the landslide data in a monitoring area; then, taking each landslide data sample in the landslide data set as an input value, and taking the corresponding true landslide deformation data of each landslide data sample as an observation value, and training a landslide deformation prediction model using a machine learning algorithm; wherein, the loss function for optimizing the landslide deformation prediction model includes a data loss function and a physical loss function, and the physical loss function is generated based on the target landslide physical mechanism model; finally, inputting the landslide data of the area to be measured into the trained landslide deformation prediction model to predict the landslide deformation data of the area to be measured.

[0008] The core of the landslide deformation prediction method based on data and physical driving provided by the embodiments of the present application lies in that, in the process of training a landslide deformation prediction model for predicting landslide displacement, a landslide physical mechanism model that can reflect the physical processes and laws of the occurrence and development of landslides is introduced to construct a physical loss function, so that the prediction result of the model conforms to the landslide physical mechanism. In this way, the prediction effect of the landslide deformation prediction model can be improved, thereby improving the accuracy of the landslide deformation prediction result, and further enabling accurate landslide warning based on the landslide deformation prediction result. Moreover, since the landslide physical mechanism is used as prior knowledge to train the model, the limitation of insufficient landslide data is overcome, and even when the data is small, the landslide deformation prediction model can still have good performance.

[0009] Furthermore, since the physical laws followed by the landslide data are used as constraints in the training process of the landslide deformation prediction model, the prediction process of the landslide deformation prediction model is more intuitive and transparent, facilitating understanding of how the model uses physical laws for prediction, and improving the interpretability of the landslide deformation prediction model.

[0010] In addition, since the physical laws that the landslide data conforms to are used as constraints during the training process of the landslide deformation prediction model, and physical laws usually have generalization ability, the landslide deformation prediction model has good adaptability to different landslide data under different geological conditions, climate conditions, and human activities, improving the generalization ability of the landslide deformation prediction model, thereby promoting the prediction reliability of the model.

[0011] In a possible implementation, the geological parameters include one or more of the following: cohesion, internal friction angle, normal stress, yield strength, plasticity index, compression coefficient, porosity ratio, volumetric strain, or pore water pressure; the terrain and geological structure data include one or more of the following: geometric shape data of the landslide body, position and shape data of the slip surface, joint fracture data; the environmental factor data includes rainfall or geological activity data.

[0012] In a possible implementation, the physical mechanism matching model is a graph neural network GNN. The physical mechanism matching model has a preliminary matching strategy and a fine matching strategy, and the preliminary matching strategy and the fine matching strategy are trained during the process of training the physical mechanism matching model.

[0013] In a possible implementation, the above method of using the trained physical mechanism matching model to determine the target landslide physical mechanism model matching each landslide data sample includes: inputting each landslide data sample in the landslide data set into the trained physical mechanism matching model, and performing preliminary matching based on the preliminary matching strategy of the physical mechanism matching model to determine the landslide physical mechanism pattern corresponding to each landslide data sample; each landslide physical mechanism pattern corresponds to multiple landslide physical mechanism models;

[0014] Performing fine matching based on the fine matching strategy of the physical mechanism matching model, and determining the target landslide physical mechanism model matching each landslide data sample from the multiple landslide physical mechanism models corresponding to the landslide physical mechanism pattern.

[0015] In a possible implementation, the above landslide physical mechanism patterns include at least one of the following: shear failure mode, plastic deformation mode, flow slide mode, or rock-soil coupling mode; among them, the landslide deformation corresponding to the shear failure mode is the deformation that the rock and soil undergo under the action of shear stress, the landslide deformation corresponding to the plastic deformation mode is the deformation that the rock and soil undergo under the action of plastic deformation, the landslide deformation corresponding to the flow slide mode is the flow deformation that the rock and soil undergo under the action of pore water pressure, and the landslide deformation corresponding to the rock-soil coupling mode is the deformation that the rock and soil undergo under the interaction of the rock and soil mass.

[0016] In a possible implementation, the landslide physical mechanism model corresponding to the shear failure mode at least includes: Mohr-Coulomb model and Drucker-Prager model.

[0017] In a possible implementation, the landslide physical mechanism model corresponding to the plastic deformation mode at least includes: the Cambridge model and the critical state soil model.

[0018] In a possible implementation, the landslide physical mechanism model corresponding to the flow slide mode at least includes: the Terzaghi effective stress model and the Biot consolidation theory model.

[0019] In a possible implementation, the landslide physical mechanism model corresponding to the rock-soil coupling mode at least includes: the discrete element model, the finite element model, and the boundary element model.

[0020] In a possible implementation, when the target landslide physical mechanism model is the Mohr-Coulomb model, the physical loss function L of the landslide deformation prediction model MC satisfies:

[0021]

[0022] where M represents the number of landslide data points included in a landslide data sample, M = N s × N t , N s is the number of sampling points, and N t is the number of time points; τ j represents the predicted value of the shear stress at the landslide data point x j ; τ j is the intermediate variable generated by the landslide deformation prediction model when the input is the landslide data point x j ; c represents the cohesion at the landslide data point x j ; σ j represents the normal stress at the landslide data point x j , and φ j represents the internal friction angle at the landslide data point x j .

[0023] When the target landslide physical mechanism model is the Drucker-Prager model, the physical loss function L of the landslide deformation prediction model DP satisfies:

[0024]

[0025] where M represents the number of landslide data points included in a landslide data sample; τ j represents the predicted value of the shear stress at the landslide data point x j ; σ j represents the normal stress at the landslide data point x j , and φ j represents the internal friction angle at the landslide data point x j .

[0026] When the physical mechanism model of the target landslide is the Cambridge model, the physical loss function $L$ of the landslide deformation prediction model NCM satisfies:

[0027]

[0028] where $M$ represents the number of landslide data points included in a landslide data sample; $\varepsilon$ vj represents the predicted value of the volumetric strain at the landslide data point $x$ j , and $\varepsilon$ vj is the intermediate variable generated by the landslide deformation prediction model when the input landslide data point is $x$ j ; $e$ j represents the void ratio at the landslide data point $x$ j .

[0029] When the physical mechanism model of the target landslide is the critical state soil model, the physical loss function $L$ of the above-mentioned landslide deformation prediction model CSM satisfies:

[0030]

[0031] where $M$ represents the number of landslide data points included in a landslide data sample; $\tau$ j represents the predicted value of the shear stress at the landslide data point $x$ j ; $\sigma$ j represents the normal stress at the landslide data point $x$ j , and $\varphi$ j represents the internal friction angle at the landslide data point $x$ j ; $\varepsilon$ vj represents the predicted value of the volumetric strain at the landslide data point $x$ j ; $e$ j represents the void ratio at the landslide data point $x$ j .

[0032] When the physical mechanism model of the target landslide is the Terzaghi effective stress model, the physical loss function $L$ of the landslide deformation prediction model Terzaghi satisfies:

[0033]

[0034] where $M$ represents the number of landslide data points included in a landslide data sample; $\sigma'$ represents the predicted value of the effective stress at the landslide data point $x$ j , and $\sigma'$ is the intermediate variable generated by the above-mentioned landslide deformation prediction model when the input landslide data point is $x$ j ; $c'$ represents the effective cohesion at the landslide data point $x$ j ; $\sigma$ j represents the effective normal stress at the landslide data point $x$ j ; $\varphi$ jDenote the effective internal friction angle at the landslide data point x j ;

[0035] When the physical mechanism model of the target landslide is the Biot consolidation theory model, the physical loss function L of the landslide deformation prediction model Biot satisfies:

[0036]

[0037] where M represents the number of landslide data points included in a landslide data sample; σ vj denotes the predicted value of the vertical stress at the landslide data point x j ; σ vj is an intermediate variable generated by the landslide deformation prediction model when the input is the landslide data point x j ; H i denotes the depth at the landslide data point x j ; C c denotes the compression coefficient at the landslide data point x j ; e0 denotes the void ratio at the landslide data point x j ;

[0038] When the physical mechanism model of the target landslide is the discrete element model, the physical loss function L of the landslide deformation prediction model DEM satisfies:

[0039]

[0040] where M represents the number of landslide data points included in a landslide data sample; F j denotes the predicted value of the contact force at the landslide data point x j ; F j is an intermediate variable generated by the landslide deformation prediction model when the input is the landslide data point x j ; denotes the critical contact force at the landslide data point x j ;

[0041] When the physical mechanism model of the target landslide is the finite element model, the physical loss function L of the landslide deformation prediction model FEM satisfies:

[0042]

[0043] where denotes the predicted value of the stress tensor at the landslide data point x j ; is an intermediate variable generated by the landslide deformation prediction model when the input is the landslide data point x j ; denotes the depth at the landslide data point x jThe critical stress tensor at; the integration region Ω is the monitoring region; the integration variable v is the volume element;

[0044] When the physical mechanism model of the target landslide is the boundary element model, the physical loss function L of the landslide deformation prediction model BEM Satisfy:

[0045]

[0046] Among them, u j Represents the predicted landslide displacement value at the landslide data point x j ; Represents the critical landslide displacement at the landslide data point x j ; the integration region Z is the boundary of the monitoring region; the integration variable s is the area element.

[0047] In a possible implementation, the landslide deformation prediction model is the differential polynomial neural network D-PNN.

[0048] In a second aspect, an embodiment of the present application provides a landslide warning system, including: a landslide deformation prediction module and a landslide warning module; wherein, the landslide deformation prediction module is used to execute the method described in the first aspect or its possible implementation to predict the landslide deformation of the area to be measured and generate landslide deformation data; the landslide warning module is used to generate a landslide warning message according to the landslide deformation data of the area to be measured; the landslide warning message is used to indicate the landslide deformation stage of the area to be measured, and the landslide deformation stage includes: the initial deformation stage, the constant velocity deformation stage, the accelerating deformation stage, and the impending slide stage.

[0049] The landslide warning system provided by the embodiment of the present application can predict the landslide deformation based on data and physical driving, obtain more accurate landslide deformation that conforms to physical laws, and then accurately realize automatic landslide warning based on the landslide deformation, providing data support for subsequent disaster analysis and prevention work.

[0050] In a possible implementation, the landslide warning module is specifically used to process the landslide deformation data of the area to be measured by using a landslide warning model to generate a landslide warning message; the landslide warning model is a random forest.

[0051] In a possible implementation, the landslide deformation data of the area to be measured includes the landslide displacements of multiple measurement points in the area to be measured at multiple time points; the landslide warning module is specifically used for:

[0052] Based on the landslide displacements of multiple measurement points in the area to be measured at multiple time points, calculate the landslide displacement rate and the landslide displacement rate ratio of the area to be measured;

[0053] When the landslide displacement rate is less than or equal to a preset value, the area to be measured is in the initial deformation stage;

[0054] When the landslide displacement rate is greater than a preset value, determine the landslide deformation stage of the area to be measured according to the landslide displacement rate ratio. Specifically, when the landslide displacement rate ratio is greater than the first value and less than or equal to the second value, the landslide deformation stage is the constant velocity deformation stage; when the landslide displacement rate ratio is greater than the second value and less than or equal to the third value, the landslide deformation stage is the accelerating deformation stage; when the landslide displacement rate ratio is greater than the third value, the landslide deformation stage is the pre-sliding stage.

[0055] In a third aspect, the present application provides a landslide deformation prediction device, which includes corresponding functional units or modules, and each functional unit or module interacts with each other to implement the method described in the first aspect or any one of its implementation manners.

[0056] In a fourth aspect, the present application provides a landslide deformation prediction device, including at least one processor; and a memory communicatively connected to the at least one processor; wherein, the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to cause the landslide deformation prediction device to execute the method described in the first aspect or any one of its implementation manners as above.

[0057] In a fifth aspect, the present application provides a computer-readable storage medium, including computer program instructions, which when executed by a computer, cause the computer to execute the method described in the first aspect or any one of its implementation manners as above.

[0058] In a sixth aspect, the present application provides a computer program product, including computer program instructions, which when run on a computer, cause the computer to execute the method described in the first aspect or any one of its implementation manners as above.

[0059] For the technical effects corresponding to the second to sixth aspects and their possible implementation manners above, reference can be made to the descriptions of the technical effects of the first aspect, the second aspect and their possible implementation manners above, and details are not repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 is one of the schematic diagrams of the landslide deformation prediction method provided by an embodiment of the present application;

[0061] Figure 2 is a schematic diagram of the preliminary matching strategy in the landslide physical mechanism model provided by an embodiment of the present application;

[0062] Figure 3 is another schematic diagram of the landslide deformation prediction method provided by an embodiment of the present application;

[0063] Figure 4It is a schematic structural diagram of the landslide early warning system provided by the embodiments of the present application;

[0064] Figure 5 It is a schematic structural diagram of the system for landslide deformation prediction and landslide early warning provided by the embodiments of the present invention. Detailed implementation manners

[0065] If there are terms such as "first" and "second" in the description and claims of the present application, they are used to distinguish different objects rather than to describe a specific order of the objects.

[0066] "And / or" in the embodiments of the present application represents the relationship between objects. For example, A and / or B may represent the following three situations: A exists alone, B exists alone, and both A and B exist simultaneously.

[0067] In the embodiments of the present application, words such as "exemplary" or "for example" are used to represent examples, illustrations or explanations. Any embodiment or design solution described as "exemplary" or "for example" in the embodiments of the present application should not be construed as being more preferred or having more advantages than other embodiments or design solutions. Rather, the use of words such as "exemplary" or "for example" is intended to present relevant concepts in a specific manner.

[0068] In the description of the embodiments of the present application, unless otherwise specified, the meaning of "a plurality of" refers to two or more.

[0069] First, some technical terms related to the embodiments of the present application are explained.

[0070] 1. Physics Informed Neural Network (PINN)

[0071] PINN is a machine learning model that integrates physical laws and neural networks. The physical laws are incorporated as additional constraints into the training process of the neural network model. Specifically, a physical loss function containing physical information (the physical information follows physical laws) is introduced into the loss function of the neural network model to ensure that the prediction results of the neural network model conform to physical laws. It can be seen that PINN is a machine learning model driven by both data and physics.

[0072] The PINN in the embodiments of the present application integrates the landslide physical mechanism and the neural network model to train a landslide deformation prediction model that follows the landslide physical laws.

[0073] 2. Landslide physical mechanism

[0074] The landslide physical mechanism refers to the physical processes and laws of the occurrence and development of landslides. The landslide physical mechanism includes multiple different modes, and each landslide physical mechanism mode includes multiple different landslide physical mechanism models.

[0075] The physical mechanism models of landslides include at least one of the following: shear failure mode, plastic deformation mode, flow slide mode, or rock-soil coupling mode.

[0076] Among them, the landslide deformation corresponding to the shear failure mode is the deformation that occurs when rock and soil are under shear stress. For example, the landslides corresponding to the shear slide mode include translational landslides, tensile fracture landslides, etc.

[0077] The landslide deformation corresponding to the plastic deformation mode is the deformation that occurs when rock and soil are under plastic deformation. For example, the landslides corresponding to the plastic deformation mode include landslides in soft soil areas, debris flow landslides, etc.

[0078] The landslide deformation corresponding to the flow slide mode is the flow deformation that occurs when rock and soil (such as saturated soil or loose accumulation body) are under pore water pressure. For example, the landslides corresponding to the flow slide mode include debris flow, flow slide, etc.

[0079] The landslide deformation corresponding to the rock-soil coupling mode is the deformation that occurs when rock and soil interact with each other. For example, the landslides corresponding to the rock-soil coupling mode include rock slope landslides, soil-rock mixture landslides, etc.

[0080] The physical mechanism models of landslides corresponding to the above shear failure mode at least include: Mohr-Coulomb model and Drucker-Prager model. The Mohr-Coulomb model is the most classic landslide model and is applicable to the shear failure analysis of most soils and rocks. The Drucker-Prager model is applicable to high-pressure conditions (such as when the pressure is greater than the yield strength), especially for the analysis of the yield behavior of rock and soil.

[0081] The physical mechanism models of landslides corresponding to the above plastic deformation mode at least include: Cambridge model and critical state soil model. The Cambridge model is a model based on the critical state theory and is applicable to describe the plastic deformation of soil under shear and compression. The critical state soil model is applicable to various soils and can describe the behavior of soil reaching the critical state during shear.

[0082] The physical mechanism models of landslides corresponding to the above flow slide mode at least include: Terzaghi effective stress model and Biot consolidation theory model. The Terzaghi effective stress model is applicable to saturated soil landslides and takes into account the influence of pore water pressure. The Biot consolidation theory is applicable to landslides in soft soil areas and takes into account the consolidation process of soil.

[0083] The landslide physical mechanism models corresponding to the above rock-soil coupling models at least include: discrete element models, finite element models, and boundary element models. The discrete element model discretizes the rock-soil mass into elements and is suitable for simulating the fracture, movement, and interaction of the rock-soil mass. The finite element model divides the rock-soil mass into elements and is suitable for simulating the deformation, stress, strain, and temperature field of the rock-soil mass. The boundary element model discretizes the boundary of the rock-soil mass into elements and is suitable for simulating the displacement and stress of the rock-soil mass.

[0084] For more detailed content about the above various physical mechanism models, reference can be made to the existing technical materials, and the embodiments of the present application will not elaborate.

[0085] It can be understood that landslide deformation prediction refers to predicting the landslide displacement of a monitoring area (such as a certain mountain body) based on the landslide data of the monitoring area. Currently, there are numerous methods for landslide deformation prediction based on machine learning. For example, the landslide displacement is predicted based on machine learning models (such as BP neural networks, CNNs, etc.). Specifically, a large number of landslide data samples are used to train the machine learning model, the data loss function of the model is calculated, and the parameters of the model are continuously optimized by minimizing the data loss function of the model to improve the prediction performance of the model. However, since the performance of the machine learning model depends on data-driven, that is, a large amount of landslide data is required for model training, the computational amount is large and the time cost is high, and a large amount of computing resources and time are consumed. More importantly, if the acquisition of landslide data is difficult and the landslide data is limited, the performance of the trained machine learning model is poor, and thus, the accuracy of the predicted landslide displacement is not high.

[0086] In addition, in the above method, the machine learning model usually cannot intuitively explain how to obtain the prediction result from the input data, that is, the machine learning model lacks interpretability. Moreover, due to the influence of different geological conditions, climate conditions, and human activities, the differences in landslide data in different regions are relatively large, and the machine learning model does not have universality. For different landslide data, the model needs to be retrained or adjusted. It can be seen that in the existing methods, the generalization ability of the machine learning model is weak.

[0087] In view of the problems existing in the prior art described above, the embodiments of the present application provide a landslide deformation prediction method and a landslide early warning system based on data and physical driving. In the process of training a landslide deformation prediction model for predicting landslide displacement, a landslide physical mechanism model that can reflect the physical processes and laws of landslide occurrence and development is introduced to construct a physical loss function, so that the prediction results of the model conform to the landslide physical mechanism, and the model training is driven from both the data and physical perspectives. In this way, the prediction effect of the landslide deformation prediction model can be improved, thereby improving the accuracy of the landslide deformation prediction results, and further enabling accurate landslide early warning based on the landslide deformation prediction results. Moreover, since the landslide physical mechanism is used as prior knowledge to train the model, the limitation of insufficient landslide data is overcome, and even in the case of less data, the landslide deformation prediction model can have good performance.

[0088] Furthermore, since the physical laws that the landslide data conforms to are used as constraints in the training process of the landslide deformation prediction model, the prediction process of the landslide deformation prediction model is more intuitive and transparent, facilitating understanding of how the model uses physical laws for prediction, and improving the interpretability of the landslide deformation prediction model.

[0089] In addition, since the physical laws that the landslide data conforms to are used as constraints in the training process of the landslide deformation prediction model, and physical theorems usually have generalization ability, the landslide deformation prediction model has good adaptability to different landslide data under different geological conditions, climate conditions, and human activities, improving the generalization ability of the landslide deformation prediction model, and thus promoting the prediction reliability of the model.

[0090] The method for landslide deformation prediction based on data and physical driving provided by the embodiments of the present application can be executed by an electronic device with processing functions. For example, the electronic device can be a computer, a server, etc. Taking the electronic device as a computer as an example, the hardware part of the computer can include: a processor, a memory, a network interface, a user interface, a communication bus, etc.

[0091] Among them, the processor is used to control the electronic device to execute relevant processing and calculation tasks. The processor can include a central processing unit (CPU), a graphics processing unit (GPU), or other processors (such as a neural processing unit (NPU)). The processor can be single-core or multi-core. For example, the CPU of the electronic device is a multi-core CPU.

[0092] The memory is used to store computer instructions and related data. For example, the memory is used to store computer instructions for executing the landslide prediction method, and store data such as landslide data sets, physical mechanism matching models, landslide deformation prediction models, and landslide prediction results.

[0093] The memory can be a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM), a flash memory, or an optical memory, a magnetic disk storage medium, or any other magnetic storage device, or any other medium capable of storing program code or data that can be accessed by a computer. Optionally, the memory can be integrated within the processor, and the memory can also be independent of the processor.

[0094] The network interface is used for the computer to communicate with other devices or communication networks. The network interface can be a transceiver with sending and receiving functions. Optionally, the network interface can include standard wired interfaces, wireless interfaces (such as WI-FI interfaces, Bluetooth interfaces, 5G interfaces). For example, landslide warning information can be sent to user devices through the network interface.

[0095] The communication bus is used to implement connection and communication between different components. For example, the above-mentioned processor, memory, network interface, and user interface can be interconnected through the communication bus.

[0096] The user interface can include a display screen and an input unit (such as a keyboard). Optionally, the user interface can also include standard wired interfaces and wireless interfaces. For example, the display screen is used to visualize relevant result data of landslide deformation prediction, such as landslide prediction results (displacement fields) and landslide warning information, etc.

[0097] Those skilled in the art can understand that the above computer can also include more or fewer components, or combine certain components, or have different component arrangements, and the embodiments of the present application do not limit this.

[0098] Based on the above content, the technical solutions provided by the embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0099] As Figure 1 shown, the landslide deformation prediction method based on data and physical drive provided by the embodiments of the present application includes S100 - S400.

[0100] S100. Construct a landslide data set, where the landslide data set includes multiple landslide data samples and the true landslide deformation data corresponding to the multiple landslide data samples.

[0101] Among them, a landslide data sample includes landslide data of multiple sampling points in a monitoring area at multiple time points. Specifically, the landslide data includes geological parameters, terrain and geological structure data, and environmental factor data, and the landslide deformation data includes the landslide displacements of multiple sampling points at multiple time points.

[0102] The above geological parameters include one or more of the following: cohesion, internal friction angle, normal stress, yield strength, plasticity index, compression coefficient, porosity ratio, volume strain, or pore water pressure.

[0103] Cohesion is the force formed by the mutual connection between soil particles, and cohesion enables the rock and soil mass to have the ability to resist external force deformation and damage.

[0104] Normal stress is the stress acting perpendicular to a certain cross-section of the rock and soil mass. Normal stress affects the deformation and strength of the rock and soil mass. A larger normal stress will cause the rock and soil mass to compress and settle.

[0105] The internal friction angle is the angle between the shear strength line of the rock and soil mass and the normal stress axis. When the rock and soil mass is under the action of external forces, relative movement and dislocation occur between particles, generating frictional resistance. The internal friction angle is an index of the relationship between frictional resistance and normal stress. The larger the internal friction angle, the tighter the mutual biting between particles, and the more stable the rock and soil mass.

[0106] Yield strength is the stress value that the rock and soil mass bears when it changes from elastic deformation to plastic deformation under the action of load. Yield strength is used to reflect the ability of the rock and soil mass to resist plastic deformation (referring to the deformation generated under the action of external forces and cannot be restored after the external force is removed). The larger the yield strength, the higher the density of the rock and soil mass, and the less likely it is to undergo plastic deformation.

[0107] The plasticity index is the change range of water content when the rock and soil mass is in a plastic state. The larger the plasticity index, the finer the particles, the larger the specific surface area, the more clay minerals, and the stronger the plasticity.

[0108] The porosity ratio is the ratio of the pore volume in the soil to the volume of soil particles. The smaller the porosity ratio, the denser the rock and soil mass.

[0109] Volume strain is the volume change of rock and soil per unit volume during the stress process. The larger the volume strain, the easier the rock and soil mass is to deform.

[0110] Pore water pressure is the pressure of free water existing in the soil pores. When the pore water pressure increases, the effective stress of the rock and soil mass decreases, and the shear strength decreases, triggering landslides.

[0111] The compression coefficient is the ratio of the reduction in the porosity ratio of the rock and soil mass to the increase in effective stress. The compression coefficient is used to reflect the compressibility of the rock and soil mass. The larger the compression coefficient, the stronger the compressibility of the rock and soil mass.

[0112] The above topographic and geological structure data includes one or more of the following: geometric shape data of the landslide body, position and shape data of the sliding surface, and joint fracture data. The geometric shape data of the landslide body can be data describing the contour of the landslide body (such as the coordinates of points on the contour), the position and shape data of the landslide surface (such as the coordinates of points on the landslide surface), and the joint fracture data can be data describing the geometric characteristics, morphological characteristics, mechanical property data, etc. of the joint fractures.

[0113] The above environmental factor data includes rainfall or geological activity data. The geological activity data includes earthquake information, such as the magnitude information of the earthquake. Optionally, the environmental factor data can also include temperature, wind force data, etc.

[0114] Multiple landslide data samples in the above landslide data set (which can also be called the landslide data training set) and the corresponding true landslide deformation data can be obtained through laboratory tests, geological surveys, or existing geological databases, and the embodiments of the present application do not make limitations.

[0115] S200. Use the trained physical mechanism matching model to determine the target landslide physical mechanism model that matches each landslide data sample.

[0116] The input of the above physical mechanism matching model is each landslide data sample in the landslide data set, and the output of the physical mechanism matching model is used to indicate the target landslide physical mechanism model, and the target landslide physical mechanism model is used to describe the physical laws followed by the landslide data in a monitoring area.

[0117] Optionally, the above physical mechanism matching model is a graph neural network (GNN). The matching rules of the physical mechanism matching model (such as GNN) include a preliminary matching strategy and a fine matching strategy. The preliminary matching strategy is used to match the landslide physical mechanism pattern, and the fine matching strategy is used to match the landslide physical mechanism model.

[0118] For more descriptions of the landslide physical mechanism pattern and the corresponding landslide physical mechanism models for various patterns, reference can be made to the technical term introduction part in the above embodiments.

[0119] Optionally, the above preliminary matching strategy and fine matching strategy can be trained during the process of training the physical mechanism matching model.

[0120] In the embodiments of the present application, the training process of the GNN model used to match the landslide physical mechanism model can include the following steps:

[0121] Step 1. Initialize the GNN model.

[0122] First, define the node types and edge types of the GNN model. The node types include landslide physical mechanism modes, geological parameters, terrain and geological structure data, and the edge types include the relationships between geological parameters and landslide physical mechanisms, the relationships between terrain and geological structure data and landslide physical mechanisms, etc.

[0123] For example, define 4 nodes to represent shear failure mode, plastic deformation mode, flow slide mode, and rock-soil coupling mode respectively; define 9 nodes to represent Mohr-Coulomb model, Drucker-Prager model, Cambridge model, critical state soil model, Terzaghi effective stress model, Biot consolidation theory model, discrete element model, finite element model, and boundary element model respectively; define 9 nodes to represent cohesion, internal friction angle, normal stress, yield strength, plastic index, compression coefficient, void ratio, volume strain, or pore water pressure respectively; define 3 nodes to represent the geometric shape data of the landslide body, the position and shape data of the slip surface, and joint fracture data respectively; define 2 nodes to represent rainfall and geological activity data respectively.

[0124] Secondly, initialize the parameters of the GNN model, including initializing network structure parameters and weight parameters, etc. It should be noted that the above-mentioned preliminary matching strategy and fine matching strategy are also used as learning parameters of the GNN model and need to be initialized.

[0125] Reference Figure 2 , as an embodiment, the initialized preliminary matching strategy is: According to on-site investigation or geological exploration data, determine the failure mode of the landslide body, and judge whether the failure mode of the landslide body is shear failure. If so, determine the landslide physical mechanism mode as the shear failure mode; otherwise, determine whether the failure mode of the landslide body is plastic deformation. If so, determine the landslide physical mechanism mode as the plastic deformation mode; otherwise, judge whether the failure mode of the landslide body is flow failure. If so, determine the landslide physical mechanism mode as the flow slide mode; otherwise, if the landslide body is composed of multiple rock and soil masses and the failure mode of the landslide is related to the interaction between the rock and soil masses, determine the landslide physical mechanism mode as the rock-soil coupling mode.

[0126] As an embodiment, the above-mentioned initialized fine matching strategy is:

[0127] When the landslide physical mechanism mode is the shear failure mode, if the geological parameters in the landslide data mainly include cohesion and internal friction angle, select the Mohr-Coulomb model as the landslide physical mechanism model; if the geological parameters in the landslide data include not only cohesion and internal friction angle but also yield strength, select the Drucker-Prager model as the landslide physical mechanism model.

[0128] When the physical mechanism model of the landslide is the plastic deformation model, if the geological parameters in the landslide data mainly include the void ratio and volumetric strain, the Cambridge model should be selected as the landslide physical mechanism model. If the geological parameters in the landslide data are very rich, including the void ratio, volumetric strain, cohesion, and internal friction angle, the critical state soil model is selected as the landslide physical mechanism model.

[0129] When the physical mechanism model of the landslide is the flow slide model, if the geological parameters in the landslide data mainly include the compression coefficient of the soil mass and the void ratio of the soil mass, the Biot consolidation theory is selected as the landslide physical mechanism model; if the geological parameters in the landslide data mainly include the pore water pressure, the Terzaghi effective stress model is selected as the landslide physical mechanism model.

[0130] When the physical mechanism model of the landslide is the rock-soil coupling model, if the terrain and geological structure data in the landslide data contain a large amount of joint fracture data, the discrete element model is selected as the landslide physical mechanism model; if the boundary conditions of the landslide are relatively vague or complex, the boundary element model is selected as the landslide physical mechanism model. Among them, the geometric shape of the landslide body can be determined according to the collected terrain and geological structure data. If the boundary is discontinuous or the slope surface is complex, it is determined that the boundary conditions of the landslide are vague or complex; in other cases that do not meet the above conditions, the finite element model is selected as the landslide physical mechanism model.

[0131] Step 2: Data preparation.

[0132] Data preparation includes collecting multiple landslide data samples and the landslide physical mechanism models that each landslide data sample conforms to. For example, a landslide knowledge graph containing landslide data and the corresponding landslide physical mechanism models can be constructed using a graph database (Neo4j).

[0133] Step 3: Optimize the model parameters to obtain a trained GNN.

[0134] Based on the above landslide knowledge graph and the initialized model matching rules, train the GNN model so that it learns the relationship between different landslide data and physical mechanism models. For example, the backpropagation algorithm and the gradient descent method can be used to optimize the parameters of the GNN model.

[0135] In one implementation, combine Figure 1 , such as Figure 3 shown, the above S200 specifically includes S201 - S202.

[0136] S201: Input each landslide data sample in the landslide dataset into the trained physical mechanism matching model, perform a preliminary match based on the preliminary match strategy of the physical mechanism matching model, and determine the landslide physical mechanism mode corresponding to each landslide data sample. Each landslide physical mechanism mode corresponds to multiple landslide physical mechanism models.

[0137] S202. Perform fine matching using the fine matching strategy based on the physical mechanism matching model, and determine the target landslide physical mechanism model that matches each landslide data sample from multiple landslide physical mechanism models corresponding to the landslide physical mechanism pattern.

[0138] The target landslide physical mechanism model is one of the above-mentioned multiple landslide physical mechanism models.

[0139] In summary, the GNN model can learn the relationship between landslide data and physical mechanism models. Thus, based on the graph neural network model, the landslide physical mechanism model that conforms to the landslide data can be efficiently and accurately matched.

[0140] S300. Use each landslide data sample in the landslide dataset as the input value, and use the true landslide deformation data corresponding to each landslide data sample as the observed value, and train the landslide deformation prediction model using a machine learning algorithm; among them, the loss function for optimizing the landslide deformation prediction model includes a data loss function and a physical loss function, and the physical loss function is generated based on the target landslide physical mechanism model.

[0141] Due to the complex non-linear relationship in the landslide deformation prediction problem, optionally, the landslide deformation prediction model in the embodiments of the present application is a differential polynomial neural network (D-PNN) model. The D-PNN model captures the dynamic relationship between input variables (landslide data), such as complex non-linear relationships, by constructing differential equations. In the D-PNN model, a polynomial function is used to represent the relationship between landslide data and landslide displacement.

[0142] It should be understood that the structure of the D-PNN model includes an input layer, a hidden layer, and an output layer. The input layer receives landslide data (including geological parameters, terrain and geological structure data, environmental factor data). The hidden layer consists of multiple neurons, and each neuron represents a polynomial function. The output layer provides the landslide prediction result (i.e., landslide deformation data), and the landslide deformation data includes the landslide displacements of multiple sampling points in the monitoring area at multiple time points.

[0143] Optionally, since the displacement field of the monitoring area (the displacement field can be understood as the set of landslide displacements of each data point) and the landslide displacement-time curve (a curve used to describe the relationship between landslide displacement and time) can be obtained based on the landslide displacements of multiple sampling points in the monitoring area at multiple time points, the content of the landslide deformation data can also be the displacement field of the monitoring area and the landslide displacement-time curve.

[0144] Specifically, the basic form of the D-PNN model can be expressed as:

[0145]

[0146] where N(X) is the output of the D-PNN model (i.e., landslide deformation data), m represents the number of neurons in the D-PNN model (since one neuron can represent a polynomial function, m also represents the number of polynomial functions in the D-PNN model), and P i (X) represents the i-th polynomial function, and w i is the weight of the i-th polynomial function.

[0147] Optionally, the above polynomial function P i (X) can be expressed as:

[0148]

[0149] where x1, x2, ……, x n represent n variables, represents the coefficient, and i1, i2, …, i n represent the powers of each variable. i1, i2, …, i n are all non-negative integers.

[0150] During the process of training a landslide deformation prediction model (such as the D-PNN model), the parameters of the model are optimized based on the loss function of the landslide deformation prediction model. The loss function is used to measure the difference between the model prediction value and the actual observed value.

[0151] Optionally, the backpropagation algorithm and the gradient descent method can be used to optimize the parameters of the D-PNN model until the value of the loss function reaches the minimum or a predetermined number of training times.

[0152] Specifically, first, the input landslide data undergoes forward propagation in the D-PNN model, and the model prediction value is calculated through the polynomial function, and the error between the model prediction value and the actual observed value is calculated; the backpropagation algorithm is executed, that is, the gradient of the loss function with respect to the model parameters is calculated, and starting from the output layer, the error is propagated backward to each layer of the network, and the gradient of each layer is updated; second, the gradient descent algorithm is executed, for example, the gradient descent is implemented through the Adam optimization algorithm, so that the prediction value is as close as possible to the actual observed value to optimize the model parameters, and then the trained D-PNN model is obtained.

[0153] After completing the training of the D-PNN model, an independent dataset (i.e., the validation set) can also be used to validate the model to evaluate the generalization ability of the model. For the selection of the validation set, it should be ensured that it is consistent with the dataset matching the landslide physical mechanism model determined during the above training model process. The validation process includes: calculating the statistical error between the model prediction value and the actual observation value. For example, the statistical error can be the mean square error (MSE) or the mean absolute error (MAE). MSE is the average of the squares of the differences between the prediction value and the actual observation value, and MAE is the average of the absolute values of the differences between the prediction value and the actual observation value.

[0154] Furthermore, according to the validation results of the model (i.e., the statistical error), the structure of the D-PNN model can be adjusted, such as increasing or decreasing the number of layers and adjusting the number of neurons, to improve the prediction performance of the model. Exemplarily, if the value of MSE is large, there may be overfitting or underfitting, and one can try to increase the number of layers and / or increase the number of neurons.

[0155] Optionally, the loss function of the landslide deformation prediction model is the weighted sum of the data loss function and the physical loss function. The loss function L is specifically expressed as the following formula:

[0156] L = λ data L data + λ PDE L PDE

[0157] where L data is the data loss function, L PDE is the physical loss function, λ data and λ PDE are the corresponding weight coefficients, and the sum of λ data and λ PDE is 1. The weight coefficients are used to balance the importance of different loss function terms in the loss function.

[0158] Optionally, the weight coefficients of the loss function can be set according to experience. For example, first set the weight coefficients based on experience, and then adjust the weight coefficients according to the results of multiple experiments to make the validation set and the prediction results close.

[0159] Optionally, the mean square error is used as the data loss function, and the data loss function L data is expressed as the following formula:

[0160]

[0161] where N s represents the number of sampling points, N t represents the number of time points, x i represents the i-th sampling point, and t jDenote the j-th time point, u pred (x i , t j ; θ) represents the predicted value of the landslide deformation data corresponding to the landslide data at the sampling point x i at time point t j . u obs (x i , t j ) represents the true value of the landslide deformation data corresponding to the landslide data at the sampling point x i at time point t j .

[0162] The physical loss function of the above landslide deformation prediction model is constructed based on the target landslide physical mechanism model that the landslide data conforms to. The physical loss functions corresponding to various landslide physical mechanism models are introduced below respectively.

[0163] When the target landslide physical mechanism model is the Mohr-Coulomb model, the physical loss function L MC of the landslide deformation prediction model satisfies:

[0164]

[0165] where M represents the number of landslide data points included in a landslide data sample, M = N s × N t , N s is the number of sampling points, and N t is the number of time points; τ j represents the predicted value of the shear stress at the landslide data point x j ; τ j is the intermediate variable generated by the landslide deformation prediction model when the input landslide data point is x j ; c represents the cohesion at the landslide data point x j ; σ j represents the normal stress at the landslide data point x j , and φ j represents the internal friction angle at the landslide data point x j .

[0166] When the target landslide physical mechanism model is the Drucker-Prager model, the physical loss function L DP of the landslide deformation prediction model satisfies:

[0167]

[0168] where M represents the number of landslide data points included in a landslide data sample; τ j represents the predicted value of the shear stress at the landslide data point x j ; σ jDenote the normal stress at the landslide data point x j , φ j Denote the landslide data point x j The internal friction angle at.

[0169] When the physical mechanism model of the target landslide is the Cambridge model, the physical loss function L of the landslide deformation prediction model NCM Satisfies:

[0170]

[0171] Among them, M represents the number of landslide data points included in a landslide data sample; ε vj Denote the predicted value of the volumetric strain at the landslide data point x j , ε vj Is the intermediate variable generated by the landslide deformation prediction model when the input landslide data point is x j ; e j Denote the void ratio at the landslide data point x j .

[0172] When the physical mechanism model of the target landslide is the critical state soil model, the physical loss function L of the landslide deformation prediction model CSM Satisfies:

[0173]

[0174] Among them, M represents the number of landslide data points included in a landslide data sample; τ j Denote the predicted value of the shear stress at the landslide data point x j ; σ j Denote the normal stress at the landslide data point x j , φ j Denote the landslide data point x j The internal friction angle at; ε vj Denote the predicted value of the volumetric strain at the landslide data point x j ; e j Denote the void ratio at the landslide data point x j .

[0175] When the physical mechanism model of the target landslide is the Terzaghi effective stress model, the physical loss function L of the landslide deformation prediction model Terzaghi Satisfies:

[0176]

[0177] Among them, M represents the number of landslide data points included in a landslide data sample; σ' represents the predicted value of the effective stress at the landslide data point x j , σ' is the input landslide data point x jWhen, the intermediate variable generated by the landslide deformation prediction model; c' represents the effective cohesion at the landslide data point x j ; σ j represents the effective normal stress at the landslide data point x j ; φ j represents the effective internal friction angle at the landslide data point x j .

[0178] It should be noted that each of the above physical quantities (such as effective stress, effective cohesion, effective normal stress, effective internal friction angle) represents the actual value of the physical quantity for the rock and soil mass. For example, the effective stress represents the actual bearing capacity of the rock and soil mass. The pore water pressure and total stress are measured by a pore water pressure sensor and a stress sensor respectively, and the difference between the total stress and the pore water pressure is the effective stress.

[0179] When the physical mechanism model of the target landslide is the Biot consolidation theory model, the physical loss function L of the landslide deformation prediction model Biot satisfies:

[0180]

[0181] where M represents the number of landslide data points included in a landslide data sample; σ vj represents the predicted value of the vertical stress at the landslide data point x j , σ vj is the intermediate variable generated by the landslide deformation prediction model when the input landslide data point is x j ; H i represents the depth at the landslide data point x j ; C c represents the compression coefficient at the landslide data point x j ; e0 represents the void ratio at the landslide data point x j .

[0182] When the physical mechanism model of the target landslide is the discrete element model, the physical loss function L of the landslide deformation prediction model DEM satisfies:

[0183]

[0184] where M represents the number of landslide data points included in a landslide data sample; F j represents the predicted value of the contact force at the landslide data point x j , F j is the intermediate variable generated by the landslide deformation prediction model when the input landslide data point is x j ; represents the critical contact force at the landslide data point x j . The critical contact force It can be obtained by numerical simulation using a historical dataset.

[0185] When the physical mechanism model of the target landslide is a finite element model, the physical loss function L of the landslide deformation prediction model FEM satisfies:

[0186]

[0187] where represents the predicted value of the stress tensor at the landslide data point x j ; is the intermediate variable generated by the landslide deformation prediction model when inputting the landslide data point x j ; represents the critical stress tensor at the landslide data point x j ; the integration region Ω is the monitoring region; the integration variable v is the volume element. The critical stress tensor It can be obtained by numerical simulation using a historical dataset.

[0188] When the physical mechanism model of the target landslide is a boundary element model, the physical loss function L of the landslide deformation prediction model BEM satisfies:

[0189]

[0190] where u j represents the predicted value of the landslide displacement at the landslide data point x j ; represents the critical landslide displacement at the landslide data point x j ; the integration region Z is the boundary of the monitoring region; the integration variable s is the area element. The critical landslide displacement It can be obtained by numerical simulation using a historical dataset.

[0191] Combining the above content, the embodiments of the present application adopt PINN (Physics-Informed Neural Network), which is driven bilayerly from the perspectives of data and physics, and trains the landslide deformation prediction model in combination with the physical mechanism of the landslide. It does not need to rely on a large number of data samples like traditional machine learning models. In this way, the cost and time of data collection can be reduced, and because the physical mechanism model has generalization ability, the applicability of the landslide deformation prediction model in different regions can be improved.

[0192] S400. Input the landslide data of the area to be measured into the landslide deformation prediction model to predict the landslide deformation data of the area to be measured.

[0193] In the embodiments of the present application, the data of multiple sampling points in the area to be measured at multiple time points are respectively input into the landslide deformation prediction model, and the landslide displacements of multiple sampling points in the area to be measured at multiple time points can be obtained. Based on the landslide displacements of multiple sampling points at multiple time points, the displacement field and the landslide displacement-time curve of the area to be measured can be obtained.

[0194] In summary, the core of the landslide deformation prediction method based on data and physical drive provided by the embodiments of the present application lies in that, in the process of training the landslide deformation prediction model for predicting landslide displacement, a landslide physical mechanism model that can reflect the physical processes and laws of landslide occurrence and development is introduced to construct a physical loss function, so that the prediction results of the model conform to the landslide physical mechanism. In this way, the prediction effect of the landslide deformation prediction model can be improved, thereby improving the accuracy of the landslide deformation prediction results. Furthermore, based on the landslide deformation prediction results, accurate landslide early warning can be realized. And because the landslide physical mechanism is used as prior knowledge to train the model, the limitation of insufficient landslide data is overcome. Even in the case of less data, the landslide deformation prediction model can still have good performance.

[0195] Furthermore, since the physical laws that the landslide data conforms to are used as constraints in the training process of the landslide deformation prediction model, the prediction process of the landslide deformation prediction model is more intuitive and transparent, which is convenient to understand how the model uses physical laws for prediction, and improves the interpretability of the landslide deformation prediction model.

[0196] In addition, since the physical laws that the landslide data conforms to are used as constraints in the training process of the landslide deformation prediction model, and physical theorems usually have generalization ability, the landslide deformation prediction model has good adaptability to different landslide data under different geological conditions, climate conditions and human activities, which improves the generalization ability of the landslide deformation prediction model, thereby promoting the prediction reliability of the model.

[0197] In some embodiments, landslide early warning can be carried out by analyzing the change of the landslide displacement predicted by the above landslide deformation prediction method. Based on this, as Figure 4 shown, the embodiments of the present application also provide a landslide early warning system, which includes a landslide prediction module 401 and a landslide early warning module 402. Among them, the landslide prediction module 401 is used to execute the steps S100-S400 and related steps in the above method embodiments to predict the landslide deformation of the area to be measured and generate the landslide deformation data of the area to be measured; the landslide early warning module 402 is used to generate a landslide early warning message according to the landslide deformation data of the area to be measured, and the landslide early warning message is used to indicate the landslide deformation stage of the area to be measured, and the landslide deformation stage includes: the initial deformation stage, the constant velocity deformation stage, the accelerated deformation stage and the impending slide stage.

[0198] Optionally, the above landslide warning information further includes: the location information of the area to be measured, the generation time of the warning information, future landslide prediction information, and recommended measures to be taken, etc. The future landslide prediction information may include, for example, the estimated time for the area to be measured to enter the acceleration or impending sliding stage, and the confidence interval of the landslide occurrence in the area to be measured.

[0199] In one implementation, the above landslide warning module 402 is specifically configured to, according to the Saito creep theory, calculate the landslide displacement rate and the landslide displacement rate ratio of the area to be measured based on the landslide displacements of multiple measurement points in the area to be measured at multiple time points; and then determine the landslide deformation stage of the area to be measured according to the landslide displacement rate and the landslide displacement rate ratio of the area to be measured.

[0200] The landslide displacement rate refers to the displacement amount of the landslide body per unit time. The calculation formula for the landslide displacement rate v is:

[0201]

[0202] Δs is the displacement change amount within a preset time period, and Δt is the duration of the time period.

[0203] The displacement rate ratio refers to the ratio of the displacement rates in two consecutive time periods. The calculation formula for the displacement rate ratio R is:

[0204]

[0205] v n is the displacement rate in the current time period, and v n-1 is the displacement rate in the previous time period.

[0206] Specifically, when the landslide displacement rate is less than or equal to the preset value, the area to be measured is in the initial deformation stage. Optionally, the preset value of the displacement rate can be 10 mm / year, and the preset value of the displacement rate can also be changed according to the historical data set and the actual situation of the area to be measured.

[0207] When the landslide displacement rate is greater than the preset value, determine the landslide deformation stage of the area to be measured according to the landslide displacement rate ratio. Specifically, when the landslide displacement rate ratio is greater than the first value and less than or equal to the second value, the landslide deformation stage is the constant velocity deformation stage; when the landslide displacement rate ratio is greater than the second value and less than or equal to the third value, the landslide deformation stage is the acceleration deformation stage; when the landslide displacement rate ratio is greater than the third value, the landslide deformation stage is the impending sliding stage.

[0208] Exemplarily, the above first value is 0, the second value is 2, and the third value is 8. If 0 < R ≤ 2, the area to be measured is in the constant velocity deformation stage; if 2 < R ≤ 8, the area to be measured is in the acceleration deformation stage; if R > 8, the area to be measured is in the impending sliding stage.

[0209] Optionally, when the area to be measured is in the stage of accelerated deformation, it can be further determined whether it is in the initial accelerated deformation stage or the mid-accelerated deformation stage. For example, if 2 < R ≤ 6, the area to be measured is in the initial accelerated deformation stage, and if 6 < R ≤ 8, the area to be measured is in the mid-accelerated deformation stage.

[0210] In another implementation, the landslide warning module 402 is specifically configured to process the landslide deformation data of the area to be measured by using a landslide warning model to generate landslide warning information. The landslide warning model is a random forest. The random forest can be used to divide the landslide displacement-time curve into different deformation stages, and based on the random forest, rapid response and real-time warning can be achieved.

[0211] Use historical landslide displacement-time curves and real landslide deformation stage data to train the random forest so that it can learn the characteristics of different deformation stages. Then, input the prediction output result of the D-PNN model into the trained random forest model. Once it is detected that the landslide body enters the accelerated deformation stage, the landslide warning system will automatically generate warning information. At the same time, the system will record all monitoring data and warning events, providing data support for subsequent disaster analysis and prevention work.

[0212] For example, for huge topographic height differences and complex tectonic features (such as some plateaus), it is difficult to conduct large-scale geological and geophysical parameter monitoring or some data are missing due to equipment failures during the landslide deformation monitoring process; for another example, for some mountainous areas or reservoirs, considering the landslides with multiple physical mechanisms in the area and the impact of reservoir impoundment on the landslides, by using the landslide warning system provided in the embodiments of the present application and combining the landslide physical mechanism, landslide deformation prediction and automatic warning can be carried out for plateau areas.

[0213] The landslide warning system provided in the embodiments of the present application can predict landslide deformation based on data and physical driving, obtain more accurate landslide deformation that conforms to physical laws, and then accurately achieve automatic landslide warning based on the landslide deformation, providing data support for subsequent disaster analysis and prevention work.

[0214] Combined with the landslide prediction method and landslide warning system based on data and physical driving described in the above embodiments, the following refers to Figure 5 Describe a system architecture for landslide deformation prediction and automatic warning. As Figure 5 shown, the system architecture may include a data collection module 501, a model prediction module 502, and a model warning module 503.

[0215] Among them, the data collection module 501 may include a data acquisition unit 5011 and a model matching unit 5012. The data acquisition unit 5011 is used to collect landslide data, such as geological parameters, terrain and geological structure data, and environmental factor data. The collected landslide data can be used to construct a landslide data set and an application prediction set. The model matching unit 5012 is used to determine the landslide physical mechanism model matched by each landslide data sample after preliminary matching and detailed matching according to the matching strategy based on the graph neural network.

[0216] The model prediction unit 502 includes a model training unit 5021, a model prediction unit 5022, a model optimization unit 5023, and a model verification unit 5024. Among them, the model training unit 5021 is used to train a landslide deformation prediction model (D-PNN model) according to the landslide data set. The loss function of the landslide deformation prediction model includes a loss function generated based on the landslide physical mechanism model. The model optimization unit 5O23 is used to optimize the model parameters obtained by the model training unit 5021 by using the backpropagation algorithm and the gradient descent method. The model verification unit 5024 is used to verify the landslide deformation prediction model optimized by the model optimization unit 5023 based on the verification set, and determine whether the model needs to be adjusted according to MSE or MAE. The landslide prediction unit 5022 is used to process the prediction application set by using the trained landslide deformation prediction model, and predict the landslide displacement field and the landslide displacement-time curve of the area to be measured.

[0217] The model warning module 503 includes a warning model training unit 5031 and a warning model application unit 5032. Among them, the warning model can be a random forest or a warning method based on the Saito creep theory. The warning model training unit 5031 is used to train a random forest with historical landslide displacement-time curves and real landslide deformation stages and other data. The warning model application unit 5032 is used to analyze the landslide deformation based on the warning model to generate landslide warning information to send a warning notice to the user. The warning model application unit 5032 may also include a visualization module, and the visualization module is used to display the displacement field predicted by the landslide deformation prediction model. The warning model application unit 5032 may also perform more warning applications based on the landslide displacement-time curve. For example, determine the estimated time for the area to be measured to enter the acceleration or impending landslide stage, the confidence interval of the landslide occurrence in the area to be measured, etc.

[0218] Correspondingly, the embodiment of the present application provides a landslide deformation prediction device, which includes corresponding functional units or modules. Each functional unit or module interacts with each other to implement each step in the above method embodiment, and all relevant contents involved in the above method embodiment can be cited to the function description of the corresponding functional module, which will not be repeated here.

[0219] The embodiment of the present application further provides a landslide deformation prediction device, including: a processor and a memory coupled to the processor; the memory is used to store computer instructions, and when the landslide deformation prediction device runs, the processor executes the computer instructions stored in the memory to execute the method described in the above embodiment.

[0220] The embodiment of the present application further provides a computer-readable storage medium, which includes a computer program that, when running on a computer, is used to execute the method described in the above embodiment.

[0221] The embodiment of the present application further provides a computer program product, which includes computer program instructions that, when running on a computer, are used to execute the method described in the above embodiment.

[0222] Each embodiment in this specification is described in a progressive manner. For the same or similar parts among the embodiments, reference can be made to each other. The key point of each embodiment is to illustrate the differences from other embodiments.

[0223] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A landslide deformation prediction method based on data and physical driving, characterized in that, Including: Constructing a landslide dataset; The landslide dataset includes a plurality of landslide data samples and the corresponding true landslide deformation data of the plurality of landslide data samples. One landslide data sample includes the landslide data of a plurality of sampling points in a monitoring area at a plurality of time points. Among them, the landslide data includes geological parameters, terrain and geological structure data, and environmental factor data, and the landslide deformation data includes the landslide displacements of the plurality of sampling points at a plurality of time points; Using the trained physical mechanism matching model to determine the target landslide physical mechanism model that matches each landslide data sample. The input of the physical mechanism matching model is each landslide data sample in the landslide dataset, and the output of the physical mechanism matching model is used to indicate the target landslide physical mechanism model. The target landslide physical mechanism model is used to describe the physical law followed by the landslide data in a monitoring area; Taking each landslide data sample in the landslide dataset as an input value and taking the corresponding true landslide deformation data of each landslide data sample as an observed value, and training a landslide deformation prediction model using a machine learning algorithm. Among them, the loss function used to optimize the landslide deformation prediction model includes a data loss function and a physical loss function, and the physical loss function is generated based on the target landslide physical mechanism model; Inputting the landslide data of the area to be measured into the landslide deformation prediction model to predict the landslide deformation data of the area to be measured.

2. The method according to claim 1, wherein The geological parameters include one or more of the following: cohesion, internal friction angle, normal stress, yield strength, plastic index, compression coefficient, porosity ratio, volume strain, or pore water pressure; The terrain and geological structure data includes one or more of the following: geometric shape data of the landslide body, position and shape data of the slip surface, and joint fracture data; The environmental factor data includes rainfall or geological activity data.

3. The method according to claim 1 or 2, characterized in that, The physical mechanism matching model is a graph neural network GNN. The physical mechanism matching model has a preliminary matching strategy and a fine matching strategy, and the preliminary matching strategy and the fine matching strategy are trained during the process of training the physical mechanism matching model; The step of using the trained physical mechanism matching model to determine the target landslide physical mechanism model that matches each landslide data sample includes: Inputting each landslide data sample in the landslide dataset into the trained physical mechanism matching model, and performing preliminary matching based on the preliminary matching strategy of the physical mechanism matching model to determine the landslide physical mechanism mode corresponding to each landslide data sample. Each landslide physical mechanism mode corresponds to a plurality of landslide physical mechanism models; Performing fine matching based on the fine matching strategy of the physical mechanism matching model, and determining the target landslide physical mechanism model that matches each landslide data sample from the plurality of landslide physical mechanism models corresponding to the landslide physical mechanism mode.

4. The method according to claim 3, wherein The landslide physical mechanism model includes at least one of the following: shear failure mode, plastic deformation mode, flow slide mode, or rock-soil coupling mode; wherein, the landslide deformation corresponding to the shear failure mode is the deformation of rock and soil under the action of shear stress, the landslide deformation corresponding to the plastic deformation mode is the deformation of rock and soil under the action of plastic deformation, the landslide deformation corresponding to the flow slide mode is the flow deformation of rock and soil under the action of pore water pressure, and the landslide deformation corresponding to the rock-soil coupling mode is the deformation of rock and soil under the interaction of rock and soil masses; The landslide physical mechanism model corresponding to the shear failure mode at least includes: Mohr-Coulomb model and Drucker-Prager model; The landslide physical mechanism model corresponding to the plastic deformation mode at least includes: Cambridge model and critical state soil model; The landslide physical mechanism model corresponding to the flow slide mode at least includes: Terzaghi effective stress model and Biot consolidation theory model; The landslide physical mechanism model corresponding to the rock-soil coupling mode at least includes: discrete element model, finite element model, and boundary element model.

5. The method according to claim 4, wherein, When the physical mechanism model of the target landslide is the Mohr-Coulomb model, the physical loss function L of the landslide deformation prediction model MC satisfies: Wherein, M represents the number of landslide data points included in a landslide data sample, and M = N s × N t , N s is the number of sampling points, and N t is the number of time points; τ j represents the predicted value of shear stress at the landslide data point x j ; τ j is the intermediate variable generated by the landslide deformation prediction model when the input landslide data point is x j ; c represents the cohesion at the landslide data point x j ; σ j represents the normal stress at the landslide data point x j ; φ j represents the internal friction angle at the landslide data point x j . When the physical mechanism model of the target landslide is the Drucker-Prager model, the physical loss function L of the landslide deformation prediction model DP satisfies: Among them, M represents the number of landslide data points included in a landslide data sample; τ j represents the predicted value of shear stress at the landslide data point x j ; σ j represents the normal stress at the landslide data point x j ; φ j represents the internal friction angle at the landslide data point x j ; When the physical mechanism model of the target landslide is the Cambridge model, the physical loss function \(L\) of the landslide deformation prediction model NCM satisfies: Where M represents the number of landslide data points included in a landslide data sample; ε vj represents the predicted value of the volumetric strain at the landslide data point x j , ε vj is the intermediate variable generated by the landslide deformation prediction model when the input landslide data point is x j ; e j represents the void ratio at the landslide data point x j ; When the physical mechanism model of the target landslide is the critical state soil model, the physical loss function L of the landslide deformation prediction model CSM satisfies: Where M represents the number of landslide data points included in a landslide data sample; τ j represents the predicted value of shear stress at the landslide data point x j ; σ j represents the normal stress at the landslide data point x j ; φ j represents the internal friction angle at the landslide data point x j ; ε vj represents the predicted value of volumetric strain at the landslide data point x j ; e j represents the void ratio at the landslide data point x j . When the physical mechanism model of the target landslide is the Terzaghi effective stress model, the physical loss function L of the landslide deformation prediction model Terzaghi satisfies: Where M represents the number of landslide data points included in a landslide data sample; σ' represents the predicted value of the effective stress at the landslide data point x j ; σ' is an intermediate variable generated by the landslide deformation prediction model when the input is the landslide data point x j ; c' represents the effective cohesion at the landslide data point x j ; σ j ′ represents the effective normal stress at the landslide data point x j ; φ j ′ represents the effective internal friction angle at the landslide data point x j . When the physical mechanism model of the target landslide is the Biot consolidation theory model, the physical loss function L of the landslide deformation prediction model Biot satisfies: where M represents the number of landslide data points included in a landslide data sample; σ vj represents the predicted value of the vertical stress at the landslide data point x j , σ vj is an intermediate variable generated by the landslide deformation prediction model when the input landslide data point is x j ; H i represents the depth at the landslide data point x j ; C c represents the compression coefficient at the landslide data point x j ; e0 represents the void ratio at the landslide data point x j ; When the physical mechanism model of the target landslide is a discrete element model, the physical loss function L of the landslide deformation prediction model DEM satisfies: Where M represents the number of landslide data points included in a landslide data sample; F j represents the predicted value of the contact force at the landslide data point x j , and F j is the intermediate variable generated by the landslide deformation prediction model when the input landslide data point is x j ; represents the critical contact force at the landslide data point x j ; When the physical mechanism model of the target landslide is a finite element model, the physical loss function L of the landslide deformation prediction model FEM satisfies: Among them, represents the predicted value of the stress tensor at the landslide data point x j ; is the intermediate variable generated by the landslide deformation prediction model when the landslide data point x j is input; represents the critical stress tensor at the landslide data point x j ; the integration region Ω is the monitoring region; the integration variable v is the volume element. When the physical mechanism model of the target landslide is a boundary element model, the physical loss function L of the landslide deformation prediction model BEM satisfies: Among them, u j represents the predicted landslide displacement value at the landslide data point x j ; represents the critical landslide displacement at the landslide data point x j ; the integration region Z is the boundary of the monitoring area; the integration variable s is the area element.

6. The method according to claim 1, wherein, The landslide deformation prediction model is a differential polynomial neural network D-PNN.

7. A landslide deformation prediction device, characterized in that it includes: At least one processor; And a memory communicatively connected to the at least one processor; Wherein, the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to execute the method according to any one of claims 1 to 6.

8. A landslide warning system, characterized in that, It includes: A landslide deformation prediction module and a landslide warning module; wherein, The landslide deformation prediction module is used to execute the method according to any one of claims 1 to 6 to predict the landslide deformation of the area to be measured and generate landslide deformation data; The landslide warning module is used to generate a landslide warning message according to the landslide deformation data of the area to be measured; the landslide warning message is used to indicate the landslide deformation stage of the area to be measured, and the landslide deformation stage includes: initial deformation stage, constant velocity deformation stage, accelerated deformation stage, and impending slide stage.

9. The landslide warning system according to claim 8, wherein, The landslide warning module is specifically used to process the landslide deformation data of the area to be measured by using a landslide warning model to generate the landslide warning message; the landslide warning model is a random forest.

10. The landslide warning system according to claim 8, characterized in that, The landslide deformation data of the area to be measured includes the landslide displacements of multiple measurement points in the area to be measured at multiple time points; The landslide warning module is specifically used for: Based on the landslide displacements of multiple measurement points in the area to be measured at multiple time points, calculate the landslide displacement rate and the landslide displacement rate ratio of the area to be measured; When the landslide displacement rate is less than or equal to a preset value, the area to be measured is in the initial deformation stage; When the landslide displacement rate is greater than the preset value, determine the landslide deformation stage of the area to be measured according to the landslide displacement rate ratio; Specifically, when the landslide displacement rate ratio is greater than the first value and less than or equal to the second value, the landslide deformation stage is the constant velocity deformation stage; when the landslide displacement rate ratio is greater than the second value and less than or equal to the third value, the landslide deformation stage is the accelerated deformation stage; when the landslide displacement rate ratio is greater than the third value, the landslide deformation stage is the pre-sliding stage.

Images