Landslide early warning classification method and terminal equipment based on slope deformation rate
By constructing a seepage-stress coupled physical benchmark model and a deep learning neural network, combined with physical constraints, we have achieved accurate classification and determination of landslide early warning, solved the problem of data and mechanical calculation being disconnected in existing technologies, and provided an efficient and reliable landslide early warning method.
Patent Information
- Application Number
- CN202611311623.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-27
- Publication Date
- 2026-11-10
AI Technical Summary
Existing landslide monitoring and early warning methods suffer from a disconnect between data and mechanical calculations, non-standard early warning indicators, poor engineering practicality, and low level of intelligence, failing to meet the needs of real-time dynamic early warning.
The landslide early warning classification method based on slope deformation rate collects historical deformation rate and reservoir water level time series data of the landslide area, constructs a seepage-stress coupled physical benchmark model, and combines deep learning neural network and physical constraints to construct a physical information neural network inversion model to realize the calculation of real-time stability coefficient and early warning level determination.
It enables precise classification and judgment of landslide early warning, and provides scientific, efficient and reliable early warning technology support by combining data-driven and physical and mechanical models, which is adapted to the real-time prevention and control needs of engineering sites.
Smart Images

Figure CN123072174A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a landslide early warning classification method and terminal equipment based on slope deformation rate, belonging to the field of geological disaster early warning technology. Background Technology
[0002] Due to the high proportion of mountainous and hilly areas, the complex geological structure of reservoir banks, and the weathering and fragmentation of soil and rock, coupled with the coupling of multiple factors such as periodic rises and falls in reservoir water levels, infiltration of heavy rainfall, disturbance of the seepage field, and human engineering activities, landslides in reservoir areas have a long-standing and frequent occurrence. These landslides are characterized by their high degree of concealment, long evolution cycle, high suddenness, great destructive power, and long chain of secondary disasters. Overall landslide instability not only damages transportation, houses, water conservancy facilities, and power transmission lines, but may also trigger surge waves impacting the dam, leading to dam overflow or collapse, threatening the lives and property of downstream residents, damaging the watershed's ecological environment, and hindering reservoir operation and regional development. Therefore, conducting accurate, real-time, and reliable monitoring and early warning of landslides in reservoir areas has become a core issue in geological disaster prevention and control.
[0003] Currently, landslide monitoring and early warning in reservoir areas mainly fall into two technical categories. The first is pure data-driven early warning, which relies on algorithms such as machine learning, deep learning, and time series analysis to construct a fitting prediction model based on multi-source monitoring data such as displacement, rainfall, and reservoir water level, and issues warnings through empirical thresholds. Its advantages are fast processing speed and low modeling threshold, but it is essentially a "black box model," lacking the integration of physical and mechanical laws, making it prone to overfitting or underfitting. In complex working conditions, it suffers from severe false alarms and missed alarms, and the early warning indicators lack mechanical basis, making it difficult to guide emergency response. The second is pure numerical mechanics model early warning, which uses methods such as finite element method, discrete element method, and limit equilibrium to construct a seepage-stress coupling model, calculating mechanical parameters such as stress field, deformation field, and stability coefficient for early warning. Its physical meaning is clear, reflecting the internal mechanical state of the slope, but modeling is cumbersome, parameter acquisition is costly and time-consuming, long-term series calculation efficiency is extremely low, it cannot meet the needs of real-time dynamic early warning, and it is severely disconnected from monitoring data, resulting in poor model reusability. Summary of the Invention
[0004] This invention provides a landslide early warning classification method and terminal equipment based on slope deformation rate, which can solve the problems of existing methods such as disconnect between monitoring data and mechanical calculation, non-standard early warning indicators, poor engineering practicality, and low level of intelligence.
[0005] On the one hand, the present invention provides a method for determining the early warning classification of landslides based on slope deformation rate, the method comprising:
[0006] S1. Determine the standard deformation rate data based on the historical deformation rate data of the landslide area in the reservoir area, and determine the historical reservoir water level change rate based on the historical reservoir water level time series change data of the landslide area.
[0007] S2. Based on the seepage data, stress data and geological survey parameters of the landslide area, construct a seepage-stress coupled physical benchmark model, and use the seepage-stress coupled physical benchmark model to perform multi-condition numerical simulation to obtain the historical stability coefficient and historical residual sliding force of the landslide area.
[0008] S3. Construct a standardized sample library with the standard deformation rate data and the historical reservoir water level change rate as inputs and the historical stability coefficient and the historical residual sliding force as outputs, and construct a physical information neural network inversion model based on the standardized sample library;
[0009] S4. Determine the real-time reservoir water level change rate based on the real-time reservoir water level time-series change data of the landslide area, and input the real-time deformation rate data of the landslide area and the real-time reservoir water level change rate into the physical information neural network inversion model to obtain the real-time stability coefficient of the landslide area.
[0010] S5. Determine the landslide warning level of the landslide area at the current moment based on the real-time stability coefficient.
[0011] Optionally, S1 specifically includes:
[0012] Outlier removal, missing value completion, and smoothing preprocessing were performed on the historical deformation rate data and historical reservoir water level time series change data of the landslide area in the reservoir area to obtain the corresponding standard deformation rate data and standard reservoir water level time series change data.
[0013] The historical reservoir water level change rate is determined based on the time-series change data of the standard reservoir water level.
[0014] Optionally, the step S2, which involves constructing a seepage-stress coupled physical benchmark model based on seepage data, stress data, and geological survey parameters of the landslide area, specifically includes:
[0015] Based on the seepage data, stress data, and geological survey parameters of the landslide area, and combined with the saturated-unsaturated seepage control equation and stress balance equation, a seepage-stress coupled physical benchmark model is constructed.
[0016] Optionally, in step S2, multi-condition numerical simulation is performed using the seepage-stress coupled physical benchmark model to obtain the historical stability coefficient and historical residual sliding force of the landslide area, specifically including:
[0017] Set up multiple combinations of different working conditions, including reservoir water level conditions and geotechnical mechanics;
[0018] Based on multiple combinations of different working conditions, and using the seepage-stress coupled physical benchmark model to perform multi-condition numerical simulations, the historical stability coefficient and historical residual sliding force of the deepest sliding surface of the landslide area under the corresponding combinations of different working conditions are obtained.
[0019] Optionally, the step S3, which involves constructing a physical information neural network inversion model based on the standardized sample library, specifically includes:
[0020] A deep learning neural network is trained using the standardized sample library to obtain a deep learning agent model;
[0021] Based on the deep learning agent model and the constraints of landslide physics and mechanics, a physical information neural network is constructed.
[0022] Based on the deep learning agent model and the physical information neural network, a physical information neural network inversion model is constructed.
[0023] Optionally, based on the deep learning agent model and the physical information neural network, a physical information neural network inversion model is constructed, specifically including:
[0024] Calculate the data fitting error of the deep learning agent model and the physical constraint residual of the physical information neural network;
[0025] The total loss function is calculated based on the data fitting error and the physical constraint residual.
[0026] The physical information neural network is trained using the total loss function to obtain the physical information neural network inversion model.
[0027] Optionally, calculating the data fitting error of the deep learning agent model specifically includes:
[0028] The training samples in the standardized sample library are input into the deep learning proxy model to obtain the predicted stability coefficient and predicted residual sliding force of the landslide area.
[0029] Calculate the first mean square error between the predicted stability coefficient and the historical stability coefficient, and calculate the second mean square error between the predicted residual glide force and the historical residual glide force;
[0030] The first mean square error and the second mean square error are used as the data fitting error of the deep learning proxy model.
[0031] Optionally, calculating the physical constraint residuals of the physical information neural network specifically includes:
[0032] Based on the predicted stability coefficient and the predicted residual glide force, calculate the limit equilibrium definition residual and the Mohr-Coulomb criterion residual, respectively.
[0033] The residuals defined by the limit equilibrium and the residuals defined by the Mohr-Coulomb criterion are used as the physical constraint residuals of the physical information neural network.
[0034] Optionally, the total loss function is calculated based on the data fitting error and the physical constraint residuals, specifically including:
[0035] The total loss function is obtained by weighted summation of the data fitting error and the physical constraint residual.
[0036] Optionally, the construction of a standardized sample library in S3, using the standard deformation rate data and the historical reservoir water level change rate as inputs and the historical stability coefficient and the historical residual sliding force as outputs, specifically includes:
[0037] Using the standard deformation rate data and the historical reservoir water level change rate under the same combination of different chemical conditions as inputs, and the corresponding historical stability coefficient and the historical residual sliding force as outputs, multiple valid samples are obtained by matching them one by one.
[0038] Integrate all valid samples to form a standardized sample library.
[0039] On the other hand, the present invention provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the landslide early warning classification determination method based on slope deformation rate as described above.
[0040] The beneficial effects that this invention can produce include:
[0041] The landslide early warning and classification method based on slope deformation rate provided by this invention collects historical deformation rate data and historical reservoir water level time-series change data of the landslide area. After outlier removal using the 3σ criterion, missing value completion using linear interpolation, and smoothing preprocessing using moving average, the historical reservoir water level change rate is calculated. Then, combined with seepage, stress data, and geological survey parameters, a seepage-stress coupled physical benchmark model containing saturated-unsaturated seepage control equations and stress balance equations is established. Multiple sets of working conditions and geotechnical parameter combinations are generated through Latin hypercube sampling. Historical stability coefficients and historical residual sliding forces are obtained through numerical simulation, constructing a standard... A sample database was created; then, a deep learning neural network was built based on a multilayer perceptron and trained to obtain a surrogate model. Physical and mechanical constraints were introduced to construct a physical information neural network. The total loss function was constructed by combining the data fitting error and the physical constraint residual, and the PINN inversion model was trained. Real-time deformation rate and reservoir water level data were collected and the real-time reservoir water level change rate was calculated. The data were input into the PINN inversion model to output the real-time stability coefficient. Finally, based on the warning thresholds of 0.98, 1.05, 1.15, and 1.25, the first to fourth level warnings were determined according to the stability coefficient range and the trend of change, so as to achieve accurate graded warning.
[0042] The present invention provides a landslide early warning classification method based on slope deformation rate. This method deeply integrates data-driven approaches with physical and mechanical models. It ensures data quality and sample coverage through standardized preprocessing and multi-condition numerical simulation, while leveraging the inversion model to combine rapid reasoning capabilities with physical rigor. This avoids the black-box limitations of purely data-driven models and the computational lag problems of purely mechanical models. The output real-time stability coefficient serves as the core early warning indicator, and combined with clear thresholds, it achieves classification. The physical meaning is clear, the judgment criteria are unified, and it can accurately capture changes in slope stability. This provides scientific, efficient, and reliable technical support for landslide early warning in reservoir areas and is suitable for real-time prevention and control needs in engineering sites. Attached Figure Description
[0043] Figure 1 A flowchart of a landslide early warning classification method based on slope deformation rate provided in an embodiment of the present invention;
[0044] Figure 2 The structural block diagram of the landslide early warning and classification system based on slope deformation rate provided in the embodiments of the present invention is shown. Detailed Implementation
[0045] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.
[0046] This invention provides a method for determining the early warning classification of landslides based on slope deformation rate, such as... Figure 1 As shown, the method includes:
[0047] S1. Determine the standard deformation rate data based on the historical deformation rate data of the landslide area in the reservoir area, and determine the historical reservoir water level change rate based on the historical reservoir water level time series change data of the landslide area.
[0048] Specifically, this includes: firstly, performing outlier removal, missing value completion, and smoothing preprocessing on the historical deformation rate data and historical reservoir water level time series change data of the landslide area in the reservoir, to obtain the corresponding standard deformation rate data and standard reservoir water level time series change data; then, determining the historical reservoir water level change rate based on the standard reservoir water level time series change data.
[0049] In practical applications, the 3σ criterion can be used to remove outliers from historical deformation rate data and historical reservoir water level time series change data; linear interpolation can be used to fill in missing values from historical deformation rate data and historical reservoir water level time series change data; and moving average method can be used to perform smoothing preprocessing on historical deformation rate data and historical reservoir water level time series change data.
[0050] This invention uses the 3σ criterion to remove outliers caused by equipment failures and signal interference from the collected raw data of deformation rate and reservoir water level in the landslide area of the reservoir. It also uses linear interpolation to fill in missing data caused by brief power outages and signal interruptions, and uses the moving average method to smooth high-frequency noise fluctuations. Through this series of systematic preprocessing operations, high-quality and high-reliability standard deformation rate data and standard reservoir water level time series change data are finally obtained.
[0051] Based on the landslide's topography, geological conditions, deformation characteristics, and potential sliding surface distribution, and following the principles of "full coverage, key density monitoring, combination of shallow and deep monitoring, and dynamic real-time monitoring," a three-dimensional, high-precision monitoring scheme was designed. Multiple types of automated monitoring equipment were deployed to achieve comprehensive, all-around monitoring of slope deformation and reservoir water levels. Specifically, 20 surface GNSS high-precision deformation monitoring points were deployed along the main sliding direction, both sides, the rear edge, and the front edge of the landslide. These points evenly cover all key parts of the landslide. A BeiDou + GPS dual-system GNSS receiver was used, with a positioning accuracy of ±0.1mm. The data acquisition frequency was set to once per 10 minutes, collecting three-dimensional displacement data from each monitoring point in real time. The surface deformation rate was calculated through displacement time-series differential analysis to accurately capture the dynamics of surface deformation. Twelve deep displacement monitoring boreholes were deployed inside the landslide body, randomly positioned at depths of 20m, 25m, and 30m, covering the entire slope. The entire thickness of the landslide body and potential slip surface are measured using a fixed digital inclinometer installed inside the borehole. The measurement accuracy is ±0.01 mm / m, and the data acquisition frequency is set to once every 30 minutes. This allows for real-time acquisition of displacement and deformation rate data at different depths of the slope, accurately capturing deep deformation characteristics and the location of the slip surface. An automatic reservoir water level monitoring station is deployed in the open water area at the landslide front. This station uses a submersible level transmitter with a measurement accuracy of ±0.5 mm and a data acquisition frequency of once every 5 minutes. It collects real-time reservoir water level elevation data, forming a continuous and accurate dataset of reservoir water level changes over time, comprehensively reflecting the dynamic changes in the reservoir water level. All monitoring equipment is equipped with solar power modules and 5G wireless transmission modules, enabling 24-hour uninterrupted power supply and real-time data transmission, ensuring continuous monitoring.
[0052] Complete monitoring data was continuously collected for 12 months (365 days), covering the entire cycle of reservoir water level conditions, including dry season, flood season, slow rise, rapid rise, slow fall, rapid fall, and extreme drop. The data comprehensively recorded the characteristics of slope deformation and reservoir water level changes under various conditions. By the end of the collection, over 1.051 million data points of surface GNSS deformation rate, over 420,000 data points of deep displacement rate, and over 105,000 data points of reservoir water level time series data were acquired, totaling over 1.576 million data points. The data coverage rate reached 99.4%, with only a small amount of missing data due to brief power outages, signal interference, and severe weather. The overall data integrity, continuity, and accuracy are excellent, laying a solid foundation for subsequent data preprocessing and model building.
[0053] This invention addresses the massive amounts of raw monitoring data collected by conducting a three-step refined and standardized preprocessing operation. This thoroughly removes invalid data, fills in missing information, smooths noise fluctuations, improves data quality, and eliminates data interference. The specific operations are as follows: First, precise outlier removal. The 3σ (three standard deviations) criterion is used to determine outliers. Statistical analysis is performed on surface deformation rate, deep displacement rate, and reservoir water level data. The mean and standard deviation of each data sequence are calculated. Data exceeding the mean ± 3σ range are identified as outliers. Such data are often caused by equipment failure, signal interference, or extreme weather disturbances and have no practical reference value; therefore, they are completely removed. This preprocessing removed over 2130 outliers, with an outlier rate of only 0.14%, resulting in excellent overall data quality. Second... The second step involved precise completion of missing values. Addressing the issue of minor data loss due to brief power outages, signal interruptions, or equipment maintenance, linear interpolation was used for precise completion. Based on continuous valid monitoring data before and after the missing period, the values of the missing points were fitted according to a linear variation pattern, ensuring data continuity, no breaks, and no bias. A total of over 9,400 missing data points were completed, achieving a 100% data integrity rate. The third step involved data smoothing and noise reduction. A 5-point moving average method was used to smooth the deformation rate and reservoir water level data, eliminating high-frequency random noise interference, preserving the true trend of data changes, and preventing noise from affecting the accuracy of subsequent model training. After smoothing, the data curves were stable, the trends were clear, and the fluctuations were reasonable, fully meeting the stringent requirements for deep learning model training.
[0054] Based on preprocessed high-precision historical reservoir water level time-series change data, the historical reservoir water level change rate is calculated on a daily scale to accurately reflect the daily rise and fall of the historical reservoir water level and its speed, and to quantitatively characterize the dynamic change characteristics of the reservoir water level. The calculation formula is the difference between the daily final value and the initial value of the reservoir water level divided by the time interval (1 day).
[0055] ;
[0056] in, The rate of change of reservoir water level (m / d) This represents the daily change in reservoir water level (m). The time interval is represented by days (d). A positive rate indicates a rise in reservoir water level, while a negative rate indicates a fall. The larger the absolute value of the rate, the faster the water level rises or falls. This calculation yielded 366 sets of precise historical reservoir water level change rate data for the entire year of 2024, comprehensively covering various change states such as slow rises and falls, sudden rises, sudden falls, and extreme sudden falls. The maximum sudden fall rate reached -0.88 m / d, and the maximum sudden rise rate reached 0.65 m / d, providing crucial and accurate hydrological parameters for subsequent model input.
[0057] The process is clear and standardized. The 3σ criterion can accurately filter out invalid and abnormal data, the linear interpolation method can efficiently fill in missing information and ensure the continuity of data time sequence, and the moving average method can effectively preserve the true trend of data change. The synergistic effect of the three methods not only eliminates noise and interference in historical monitoring data, but also ensures the integrity and accuracy of the data. This lays a solid data foundation for subsequent calculation of reservoir water level change rate, sample library construction and model training, and improves the reliability and accuracy of the entire early warning method.
[0058] S2. Based on the seepage data, stress data and geological survey parameters of the landslide area, a seepage-stress coupled physical benchmark model is constructed, and multi-condition numerical simulation is carried out using the seepage-stress coupled physical benchmark model to obtain the historical stability coefficient and historical residual sliding force of the landslide area.
[0059] Specifically, it includes:
[0060] (1) Based on the seepage data, stress data and geological survey parameters of the landslide area, and combined with the saturated-unsaturated seepage control equation and stress balance equation, a seepage-stress coupled physical benchmark model is constructed.
[0061] This seepage-stress coupling physical benchmark model is used to simulate the stress redistribution, deformation evolution, and stability changes of the slope under the coupling effect of reservoir water level rise and fall and groundwater seepage.
[0062] Based on the collected seepage data, stress data and geological survey parameters of the landslide area, this invention constructs a seepage-stress coupled physical benchmark model by integrating saturated-unsaturated seepage control equations and stress balance equations. This model can accurately simulate the stress redistribution, deformation evolution process and overall stability change law of the slope under the coupled action of reservoir water level rise and fall and groundwater seepage, providing reliable physical and mechanical modeling support for subsequent multi-condition numerical simulation.
[0063] This invention comprehensively collects detailed engineering geological survey reports, in-situ test data, indoor geotechnical test results, long-term hydrogeological observation data, and reservoir operation regulations for the landslide, and meticulously analyzes the core parameters of the slope: physical and mechanical parameters such as soil layer thickness, natural density, water content, cohesion, internal friction angle, permeability coefficient, elastic modulus, Poisson's ratio, and tensile strength; hydrogeological parameters such as groundwater depth, seepage path, recharge and discharge conditions, and pore water pressure distribution; and model construction parameters such as topographic boundaries, load conditions, initial stress conditions, and seismic motion parameters. Simultaneously, by combining historical reservoir water level data, the invention clarifies the range of reservoir water level changes, rise and fall patterns, extreme water levels, and frequency of sudden rises and falls, providing accurate and comprehensive basis for the construction of the physical benchmark model and the design of working conditions.
[0064] This invention utilizes refined engineering geological data to construct a three-dimensional seepage-stress coupled physical benchmark model of the landslide. The model dimensions are identical to the actual landslide, measuring 220m in length, 160m in width, and 70m in height, covering the slope and surrounding stable bedrock strata. The model mesh is divided according to the principle of "densifying the mesh in key areas and sparser in secondary areas." Key areas such as the potential sliding surface, the water-prone zone at the leading edge, the deep deformation zone, and the crack development zone are meshed with a density of 0.5m-1.0m. The mesh size in the stable bedrock area is appropriately enlarged to 2m-3m. A total of 185,000 structured and unstructured mixed meshes with 132,000 nodes are created, ensuring a balance between computational accuracy and efficiency. The model integrates the saturated-unsaturated seepage control equations and stress balance equations to achieve bidirectional real-time coupling of the seepage field and stress field. The Mohr-Coulomb strength criterion is selected as the failure criterion for slope soil and rock. It accurately simulates the changes in pore water pressure, effective stress redistribution, deformation evolution, plastic zone development, and stability changes of the slope under the influence of reservoir water level fluctuations, groundwater seepage, and rainfall infiltration. The model boundary conditions are set as follows: the bottom and sides are fixed displacement boundaries, the leading edge is a reservoir water seepage boundary, and the trailing edge and surface are free boundaries. Initial conditions are set according to the measured in-situ ground stress, groundwater level, and initial deformation to ensure that the model closely matches actual working conditions and restores the true mechanical state of the slope.
[0065] This method integrates core mechanics and seepage equations to accurately characterize the water-force coupling effect, realistically reproducing the dynamic response of the slope under complex hydrological conditions in the reservoir area. It overcomes the shortcomings of single-equation modeling, which cannot take into account the mutual influence of seepage and stress. The model can directly serve multi-condition simulation and sample library construction, providing basic data that conforms to the actual engineering laws for subsequent model training, and significantly improving the scientificity and physical rigor of the entire early warning method.
[0066] (2) Set up multiple combinations of different working conditions including reservoir water level and geotechnical mechanics; then, based on multiple combinations of different working conditions, and using the seepage-stress coupling physical benchmark model, perform multi-condition numerical simulation to obtain the historical stability coefficient and historical residual sliding force of the deepest sliding surface of the slope in the corresponding combination of different working conditions.
[0067] This invention employs the Latin hypercube sampling method to sample multiple sets of reservoir water level conditions and geotechnical parameters to generate multiple differentiated combinations of conditions covering the entire scenario and multiple dimensions.
[0068] This invention combines actual reservoir scheduling rules, weathering and deterioration patterns of slope soil and rock parameters, and the impact of rainfall infiltration. It selects five core variables: reservoir water level, reservoir water level fluctuation rate, cohesion, internal friction angle, and permeability coefficient. A full range of differentiated working conditions is designed, and the Latin hypercube sampling method is used to generate 2000 representative working condition combinations, comprehensively covering various common, extreme, and special working conditions in the reservoir area, avoiding redundancy and omissions. Specifically, the reservoir water level variable ranges from 270m (extreme low water level) to 320m (normal storage level), and the reservoir water level fluctuation rate ranges from -0.9m / d (extreme sudden drop) to 0.7m / d (extreme sudden rise). The soil and rock mechanical parameters consider the deterioration effects caused by rainfall softening, long-term weathering, and wet-dry cycles. The cohesion value ranges from 6kPa to 24kPa, the internal friction angle ranges from 8° to 22°, and the permeability coefficient fluctuates within a reasonable range according to the characteristics of the soil and rock mass, taking into account extreme cases of parameter deterioration. Numerical simulations were conducted for each of the 2000 working conditions. Each working condition was calculated until the slope deformation stabilized and the stress field and seepage field reached equilibrium. The convergence accuracy was controlled within 10⁻⁵ to ensure calculation accuracy. The stability coefficient was calculated using the strength reduction method, and the deepest potential sliding surface (most dangerous sliding surface) was determined based on the distribution of the plastic zone / maximum shear strain rate. Deformation rates of 20 surface monitoring points, deformation rates of 12 deep displacement monitoring points, reservoir water level change rates, stability coefficient of the deepest sliding surface, and residual sliding force data were extracted for the corresponding working conditions to form complete data for each working condition, ensuring that the data is comprehensive, accurate, and consistent with reality.
[0069] S3. Construct a standardized sample library with standard deformation rate data and historical reservoir water level change rate as inputs and historical stability coefficient and historical residual sliding force as outputs, and construct a physical information neural network inversion model based on the standardized sample library.
[0070] Specifically, it includes:
[0071] (1) The standard deformation rate data and historical reservoir water level change rate under the same combination of different chemical conditions are used as inputs, and the corresponding historical stability coefficient and historical residual sliding force are used as outputs. Multiple valid samples are obtained by matching them one by one. Then, all valid samples are integrated to form a standardized sample library.
[0072] The historical stability coefficients and historical residual sliding forces obtained above are matched and screened one by one with the standard deformation rate data and historical reservoir water level change rate under the corresponding working conditions to obtain multiple valid samples. All valid samples are integrated to construct a standardized sample library with standard deformation rate data and historical reservoir water level change rate as input and historical stability coefficients and historical residual sliding forces of the deepest sliding surface of the slope as output.
[0073] This invention standardizes and filters 2000 sets of data extracted from numerical simulations, removing invalid samples due to simulation convergence failure, missing data, abnormal parameters, or deviations from established patterns, ultimately yielding 1912 valid samples to construct a standardized, structured sample library. The sample library adopts an "input-output" design, with 32 input parameters including deformation rates at 20 surface monitoring points, deformation rates at 12 deep displacement monitoring points, and reservoir water level fluctuation rates; and 2 output parameters including the historical stability coefficient of the deepest slip surface and the historical residual sliding force. The sample library comprehensively covers various working conditions such as reservoir water level fluctuations, soil and rock parameter deterioration, extreme hydrology, and rainfall infiltration. The data is evenly distributed, has reasonable dispersion, closely reflects the actual mechanical laws of landslides, and has a sufficient number of samples to fully meet the needs of the entire process of deep learning model training, validation, and testing, providing high-quality and reliable data support for subsequent surrogate model training.
[0074] This invention generates differentiated working condition combinations through Latin hypercube sampling, ensuring both comprehensive and representative coverage of working conditions while avoiding redundant calculations, thus efficiently adapting to complex reservoir scenarios. The multi-working-condition numerical simulation combined with a parameter-one-to-one matching sample construction method ensures that the sample data closely matches actual mechanical laws, with clear input-output correspondences. The standardized sample library provides high-quality, highly reliable data support for subsequent deep learning model training, effectively improving the model's fitting accuracy and generalization ability, laying a solid foundation for the accuracy of the entire early warning method.
[0075] (2) Use a standardized sample library to train a deep learning neural network to obtain a deep learning agent model.
[0076] Specifically, the deep learning neural network is first constructed, including: building the deep learning neural network based on a multilayer perceptron; ensuring the number of neurons in the input layer of the deep learning neural network matches the sum of the number of neurons in the standard deformation rate data and the historical water level change rate; ensuring the number of neurons in the output layer matches the sum of the number of neurons in the historical stability coefficient and the historical residual sliding force; using the ReLU activation function in the hidden layers and the linear activation function in the output layer. The use of the ReLU activation function in the hidden layers avoids gradient vanishing and improves training stability, while the use of the linear activation function in the output layer ensures the unbounded and accurate output of continuous mechanical parameters such as stability coefficient and residual sliding force, forming a network architecture adapted to the data mapping requirements of the sample database.
[0077] This invention utilizes the numerical characteristics of 32 input parameters and 2 output parameters from a sample database to construct a fully connected deep neural network (DNN). It employs a classic optimized architecture of "input layer - multiple hidden layers - output layer," balancing model fitting accuracy, inference speed, and generalization ability. The input layer has 32 neurons, corresponding to the 32 input parameters. Five hidden layers are used, with 256, 128, 64, 32, and 16 neurons respectively. The ReLU activation function is used in the hidden layers to effectively avoid gradient vanishing and exploding problems, improving model training stability and convergence speed. The output layer has 2 neurons, corresponding to the historical stability coefficient and historical residual sliding force. A linear activation function is used in the output layer to ensure the continuity, accuracy, and lack of range limitations of the output parameters. The network weights are initialized using the Xavier uniform initialization method to ensure stable gradient propagation, accelerate early model convergence, and improve training efficiency.
[0078] This invention aligns with the core requirements of landslide mechanical parameter mapping. The multilayer perceptron architecture ensures the ability to fit the nonlinear relationship between input and output. The precise matching of the number of neurons in the input and output layers avoids data dimensional redundancy or missing data. The reasonable combination of ReLU activation function and linear activation function balances training efficiency and output accuracy. The overall architecture is simple, efficient, and highly adaptable, facilitating rapid iterative training and providing a stable foundation for subsequently introducing physical constraints to build the PINN (Physical Information Neural Network) inversion model, effectively improving the accuracy and reliability of data mapping.
[0079] After constructing the deep learning neural network, this invention uses a standardized sample library as training data to iteratively train the deep learning neural network to obtain a deep learning surrogate model. Specifically, the standardized sample library is randomly divided into training, validation, and test sets according to a preset dataset ratio; the training set is used as the training data, and a learning rate decay strategy and an early stopping strategy are employed to optimize the iterative training process; the parameters of the deep learning neural network are continuously updated iteratively through backpropagation until the deep learning neural network converges to obtain the deep learning surrogate model.
[0080] This invention first pre-determines a reasonable dataset partitioning ratio, randomly dividing a standardized sample library into a training set, a validation set, and a test set according to this ratio. Then, iterative training of a deep learning neural network is carried out using the training set as the core training data. During the training process, a learning rate decay strategy and an early stopping strategy are adopted to optimize the training process. The backpropagation algorithm is used to continuously update the network's weights, biases, and other parameters, and continuously monitor the model's convergence status until the neural network reaches the convergence standard, ultimately obtaining a deep learning surrogate model.
[0081] In one specific embodiment, 1912 standardized samples are randomly divided into a training set, a validation set, and a test set in a golden ratio of 8:1:1. This ensures the randomness, representativeness, and balance of the sample division, avoiding data bias from affecting the model's accuracy and generalization ability. The training set comprises 1530 sets (80%), used for iterative model training to learn data mapping patterns and intrinsic relationships; the validation set comprises 191 sets (10%), used for real-time verification of model accuracy, dynamic adjustment of hyperparameters, and overfitting monitoring during training, allowing for timely optimization of the training strategy; and the test set comprises 191 sets (10%), used for testing the model's generalization ability, inversion accuracy, and stability after training, objectively evaluating model performance and ensuring the model's adaptability to unknown conditions.
[0082] The core hyperparameters for model training were set, and the Adam adaptive moment estimator was selected as the optimizer. This optimizer has fast convergence speed, high accuracy, and strong stability, making it suitable for the complex training requirements of this model. The initial learning rate was set to 0.001, and a cosine annealing learning rate decay strategy was adopted. The learning rate was dynamically adjusted every 50 iterations to improve the convergence accuracy of the model in the later stages. The batch size was set to 64 to balance training efficiency and gradient stability. The maximum number of iterations was set to 2000, and the training stopped when the validation set loss did not decrease for 80 consecutive iterations, i.e., an early stopping strategy to avoid overfitting and improve generalization ability. The mean squared error (MSE) between the model's output stability coefficient, the predicted residual glide force, and the numerical simulation ground truth was used as the basic loss function. Iterative training was carried out, and the changes in the training set loss and validation set loss curves were monitored in real time during training. Hyperparameters were dynamically adjusted to optimize network weights and biases. After 1260 iterations of training, the model converged, the validation set loss tended to stabilize, and training was terminated. At this point, the model's data fitting accuracy reached its optimal level, with no overfitting or underfitting.
[0083] After training, a comprehensive and rigorous test was conducted on the model's performance using an independent test set to evaluate its inversion accuracy, generalization ability, and stability. The test results show that the deep learning surrogate model's average absolute error in predicting the stability coefficient of the deepest slip surface is less than 0.022, with a maximum error not exceeding 0.035; the average absolute error in predicting the residual sliding force is less than 2.0 kN / m, with a maximum error not exceeding 3.0 kN / m. The model exhibits extremely high data mapping accuracy and computational efficiency improvements of over 99%, fully meeting the engineering accuracy and real-time requirements for landslide stability inversion in the reservoir area. This yields a deep learning surrogate model capable of rapid inference, laying a solid foundation for subsequent PINN coupled modeling.
[0084] This invention effectively avoids data bias affecting model accuracy by randomly dividing the dataset into multiple classes, ensuring the rationality of model training and the objectivity of generalization ability evaluation. The combination of learning rate decay strategy and early stopping strategy can improve the convergence accuracy of the model in the later stage and effectively prevent overfitting, thereby improving training efficiency. The application of the backpropagation algorithm realizes the precise iterative optimization of network parameters, ensuring that the trained deep learning proxy model has efficient and accurate mapping ability from input data to mechanical parameters, laying a high-quality foundation for the subsequent construction of the PINN inversion model.
[0085] (3) Construct a physical information neural network based on the deep learning agent model and the constraints of landslide physics.
[0086] This invention overcomes the limitations of pure data fitting by introducing constraints from landslide physics and mechanics onto a trained deep learning proxy model. This achieves deep coupling between data-driven approaches and physical laws, eliminating the physical biases inherent in pure data fitting. The introduced physical constraints mainly fall into two categories: First, the Mohr-Coulomb criterion residual constrains the balance between shear stress and shear strength in the slope's soil and rock mass, clarifying the critical conditions for slope failure and ensuring that the stress state output by the model fully conforms to the shear failure laws of the soil and rock mass, guaranteeing mechanical rationality. Second, the limit equilibrium mechanical definition of the stability coefficient of the deepest sliding surface constrains the ratio of the total resisting force to the total sliding force, ensuring that the stability coefficient obtained through inversion strictly adheres to the landslide limit equilibrium principle. This corrects the biases of pure data fitting from a physical perspective, ensuring that the model output conforms to the intrinsic mechanism of slope instability.
[0087] (4) Construct a physical information neural network inversion model based on the deep learning agent model and the physical information neural network.
[0088] include:
[0089] 1) Calculate the data fitting error of the deep learning agent model and the physical constraint residual of the physical information neural network.
[0090] Specifically, the data fitting error of the deep learning proxy model is calculated as follows: First, the training samples in the standardized sample library are input into the deep learning proxy model to obtain the predicted stability coefficient and the predicted residual sliding force of the landslide area; then, the first mean square error between the predicted stability coefficient and the historical stability coefficient is calculated, and the second mean square error between the predicted residual sliding force and the historical residual sliding force is calculated; finally, the first mean square error and the second mean square error are used as the data fitting error of the deep learning proxy model.
[0091] Specifically, the physical constraint residuals of the physical information neural network are calculated as follows: First, based on the predicted stability coefficient and the predicted residual sliding force, the limit equilibrium definition residuals of the deepest potential sliding surface stability coefficient and the Mohr-Coulomb criterion residuals are calculated respectively; then, the limit equilibrium definition residuals of the deepest potential sliding surface stability coefficient and the Mohr-Coulomb criterion residuals are used as the physical constraint residuals of the physical information neural network.
[0092] 2) Calculate the total loss function based on the data fitting error and physical constraint residuals.
[0093] Specifically, the total loss function is obtained by weighted summation of the data fitting error and the physical constraint residual.
[0094] 3) Train the physical information neural network using the total loss function to obtain the physical information neural network inversion model.
[0095] This invention weighted couples the data fitting error of the training samples with the physical constraint residuals to construct the total loss function of the PINN framework. The formula for calculating the total loss function is as follows:
[0096] ;
[0097] in, For the total loss function, The data fitting error (using the mean squared error loss from the deep learning surrogate model) is used. This refers to the physical constraint residuals (i.e., the mean square error of the physical rule residuals). This is the physical loss weighting coefficient, used to balance the weights of data fitting and physical constraints. This embodiment, through multiple debugging sessions, comparative experiments, and accuracy verification, has finally set this value. =0.40, which ensures both the accuracy of model data fitting and strengthens the physical constraints, so that the model output satisfies both data accuracy and physical rigor, taking into account both efficiency and mechanism.
[0098] This invention, based on a constructed total loss function, employs the Adam optimizer for secondary optimization training of the PINN framework. Hyperparameters such as batch size and learning rate decay strategy remain constant, with the initial learning rate adjusted to 0.0005 and the maximum number of iterations set to 1500. During training, changes in data loss, physical loss, and total loss are monitored synchronously, and network parameters are optimized in real time until the total loss converges to a minimum, indicating model stability. After 920 iterations, the model's total loss stabilizes, demonstrating excellent convergence. At this point, the mechanical parameters output by the model not only meet the data fitting accuracy requirements but also strictly conform to the physical mechanics of landslides, completely eliminating the physical bias of pure data fitting. This yields the final PINN inversion model, which combines millisecond-level fast inference capabilities with high-level physical rigor, perfectly meeting the needs for real-time inversion and accurate early warning of landslides in reservoir areas.
[0099] This invention accurately quantifies the deviation between the model's predicted values and the actual values through mean square error, ensuring the accuracy and rationality of the calculation of data fitting error. Physical constraint residuals are integrated into the core mechanics and seepage equations of landslides, allowing model training to be constrained by physical laws and avoiding the black box defects of pure data fitting. By constructing a total loss function through weighted summation, the accuracy of data fitting and the fit with physical laws can be flexibly balanced, so that the PINN model trained subsequently has both accurate data mapping capabilities and strictly conforms to the actual mechanical evolution laws of landslides, greatly improving the reliability and generalization ability of the model.
[0100] S4. Determine the real-time reservoir water level change rate based on the real-time reservoir water level change data of the landslide area, and input the real-time deformation rate data and the real-time reservoir water level change rate of the landslide area into the physical information neural network inversion model to obtain the real-time stability coefficient of the landslide area.
[0101] This invention first collects real-time deformation rate data and real-time reservoir water level time series change data in the landslide area. Then, it performs real-time preprocessing on the real-time reservoir water level time series change data to obtain standard continuous water level data. Finally, it calculates the real-time reservoir water level change rate according to a fixed time scale, with a positive rate representing a rise in reservoir water level and a negative rate representing a fall in reservoir water level. Simultaneously, it acquires real-time deformation rate data and real-time reservoir water level change rate data that can be directly input into the model.
[0102] This invention establishes a 5G wireless real-time data transmission interface, configuring data parsing and format conversion modules to seamlessly integrate collected and preprocessed real-time deformation rate data and real-time reservoir water level change rate data with the PINN inversion model. This enables automatic synchronous transmission and format adaptation of monitoring data, with the data input frequency matching the on-site monitoring frequency: surface deformation rate is input every 10 minutes, deep displacement rate every 30 minutes, and reservoir water level change rate updated daily, ensuring the real-time, accurate, and consistent nature of the input data. Upon receiving the real-time monitoring data, the PINN inversion model, relying on a lightweight and fast inference engine, completes the entire calculation process within 0.5 seconds without requiring time-consuming numerical simulations. It outputs the real-time stability coefficient of the deepest slip surface of the slope at the current moment and automatically generates a stability coefficient time-series variation curve and deformation rate field cloud map, dynamically displaying the slope stability change trend and deformation distribution characteristics in real time. This allows for timely capture of subtle fluctuations and abrupt changes in stability, providing accurate basis for early warning judgments.
[0103] This embodiment selects the extreme drop in reservoir water level during the flood season (July-September) for real-time inversion verification. During this period, the reservoir water level rapidly drops from 319m to 273m, with a water level change rate of -0.88m / d, which is an extreme drop condition. The deformation rate of both the surface and deep parts of the slope increases significantly, and the stability fluctuates violently. The inversion results show that the stability coefficient of the deepest slip surface of the slope gradually decreases from 1.24 to 1.00 during this period. The inversion process is lag-free, the data is accurate, and the trend is clear. It is highly consistent with the actual deformation state of the slope and the results of on-site inspections. It can timely and accurately reflect the dynamic changes in slope stability, fully verifying the real-time performance, accuracy, and reliability of the PINN inversion model of this invention.
[0104] This invention ensures the continuity and standardization of real-time reservoir water level data through targeted preprocessing, providing an accurate foundation for calculating the rate of water level change. The fixed time scale calculation method unifies the quantitative standard of water level change rate and makes the results comparable. The setting of positive and negative values can also intuitively reflect the trend of water level rise and fall. The overall process is simple and efficient, adapting to the timeliness requirements of real-time monitoring. It can quickly obtain two types of real-time data that meet the model input standards, laying the data foundation for the rapid inversion of stability coefficients and ensuring the real-time nature of landslide early warning.
[0105] S5. Determine the landslide warning level of the landslide area at the current moment based on the real-time stability coefficient.
[0106] In practical applications, landslide warning thresholds can be preset to 0.98, 1.05, 1.15, and 1.25. When the real-time stability coefficient is in the range of 1.15 to 1.25 and shows a continuous downward trend, it is determined to be a Level 1 warning; when the real-time stability coefficient is in the range of 1.05 to 1.15, it is determined to be a Level 2 warning; when the real-time stability coefficient is in the range of 0.98 to 1.05, it is determined to be a Level 3 warning; and when the real-time stability coefficient is less than 0.98, it is determined to be a Level 4 warning.
[0107] This invention determines the warning level of the current landslide area in the reservoir by pre-setting landslide warning thresholds of 0.98, 1.05, 1.15 and 1.25 based on the slope deformation rate, and by classifying the real-time stability coefficient of the slope output by the PINN inversion model in combination with the numerical range and the trend of change.
[0108] The Level 1 warning is determined by a real-time stability coefficient between 1.15 and 1.25 that shows a continuous downward trend. This indicates that the slope is in a basically stable state, but its stability is gradually decreasing, showing a tendency towards instability. Therefore, monitoring frequency needs to be increased, and attention should be paid to slope deformation dynamics. As the reservoir water level begins to slowly decline, the stability coefficient retrieved from the PINN model drops from 1.24 to 1.17, remaining within the Level 1 warning range and showing a significant downward trend. The system automatically issues a blue warning, and the warning information is simultaneously pushed to on-site management and technical personnel via SMS and the app. This requires increasing the frequency of surface GNSS monitoring to once every 5 minutes, conducting one on-site foot patrol daily, focusing on observing the expansion of surface cracks, seepage at the leading edge, and rockfalls. Monitoring data and patrol results must be recorded, and abnormal situations must be reported in real time.
[0109] Level 2 Warning: The judgment rule is that the real-time stability coefficient is in the range of 1.05 to 1.15, and the deformation rate field shows the characteristics of shear band penetration. This indicates that the slope is in an unstable state, local deformation is intensifying, potential slip surface shear bands are gradually penetrating, and the risk of instability is gradually increasing. It is necessary to increase the frequency of inspections and comprehensively investigate potential deformation hazards. If the reservoir water level drops rapidly and the stability coefficient drops to the range of 1.06 to 1.12, and at the same time the deformation rate at the monitoring points in the middle and leading edge of the slope increases significantly, and the deformation rate field shows a clear trend of shear band penetration, the system will automatically issue a yellow warning, and on-site management personnel will increase the frequency of inspections.
[0110] Level 3 warning: The judgment rule is that the real-time stability coefficient is in the range of 0.98 to 1.05.
[0111] Level 4 warning: The judgment rule is that the real-time stability coefficient is less than 0.98.
[0112] This invention uses a stability coefficient with a clear physical meaning as the core indicator. The preset warning threshold is consistent with the actual instability evolution law of landslides in the reservoir area. The judgment rules are clear, quantifiable, and standardized. When classifying, both the numerical range and the trend of the coefficient are taken into account, which can accurately capture the dynamic changes in slope stability. The gradient design of the four-level warning makes the classification of the degree of danger more detailed. The warning results are intuitive and easy to understand. They can directly guide the differentiated emergency prevention and control, patrol monitoring and personnel evacuation work on site, which greatly improves the practicality and pertinence of landslide warning in the reservoir area.
[0113] Figure 2 This embodiment provides a landslide early warning and classification system based on slope deformation rate. The system includes:
[0114] The data acquisition and preprocessing module is used to collect historical deformation rate data and historical reservoir water level time series change data in the landslide area of the reservoir area. It determines the standard deformation rate data based on the historical deformation rate data of the landslide area and determines the historical reservoir water level change rate based on the historical reservoir water level time series change data of the landslide area.
[0115] The multi-condition simulation calculation module is connected to the data acquisition and preprocessing module. It is used to collect seepage data, stress data and geological exploration parameters of the landslide area. Based on the seepage data, stress data and geological exploration parameters, a seepage-stress coupled physical benchmark model is constructed. The seepage-stress coupled physical benchmark model is used to perform multi-condition numerical simulation to obtain the historical stability coefficient and historical residual sliding force of the landslide area.
[0116] The sample library construction module, connected to the multi-condition simulation calculation module, is used to construct a standardized sample library with standard deformation rate data and historical reservoir water level change rate as inputs and historical stability coefficient and historical residual sliding force as outputs.
[0117] The PINN inversion modeling module, connected to the sample library construction module, is used to build a deep learning neural network. It uses a standardized sample library as training data to iteratively train the deep learning neural network to obtain a deep learning surrogate model. Based on the deep learning surrogate model, constraints of landslide physics and mechanics are introduced to construct a physical information neural network. The data fitting error of the deep learning surrogate model and the physical constraint residual of the physical information neural network are calculated. The total loss function is calculated based on the data fitting error and the physical constraint residual. The physical information neural network is trained using the total loss function to obtain the PINN inversion model.
[0118] The real-time inversion calculation module is connected to the PINN inversion modeling module. It is used to collect real-time deformation rate data and real-time reservoir water level time series change data in the landslide area, calculate the real-time reservoir water level change rate based on the real-time reservoir water level time series change data, input the real-time deformation rate data and the real-time reservoir water level change rate into the PINN inversion model, and output the corresponding real-time stability coefficient.
[0119] The graded early warning determination module is connected to the real-time inversion calculation module. It is used to preset the landslide early warning threshold and determine the landslide early warning level of the landslide area in the reservoir area at the current time based on the real-time stability coefficient obtained by inversion and the landslide early warning threshold.
[0120] The system adopts a modular hierarchical design, with each module having independent functions while maintaining data interoperability and close integration. It achieves closed-loop automation of the entire process from data acquisition and preprocessing to model training, real-time inversion, and tiered early warning, requiring minimal manual intervention and significantly improving the efficiency of early warning work. The interconnected design between modules ensures efficient and smooth data flow. The PINN inversion modeling module integrates data-driven approaches with physical and mechanical laws, the real-time inversion calculation module ensures the timeliness of early warnings, and the tiered early warning judgment module enables quantitative and accurate judgments. The overall system balances scientific rigor, real-time performance, and practicality, adapting to the routine and intelligent monitoring and early warning needs of landslides in various reservoir areas. It can also be flexibly adapted to different geological conditions in different reservoir areas, making it highly scalable.
[0121] Another embodiment of the present invention provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the landslide early warning classification method based on slope deformation rate as described above.
[0122] The above descriptions are merely a few embodiments of the present invention and are not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any modifications or alterations made by those skilled in the art without departing from the scope of the technical solution of the present invention using the disclosed technical content are equivalent to equivalent implementation cases and fall within the scope of the technical solution.
Claims
1. A method for determining the early warning and classification of landslides based on slope deformation rate, characterized in that, The method includes: S1. Determine the standard deformation rate data based on the historical deformation rate data of the landslide area in the reservoir area, and determine the historical reservoir water level change rate based on the historical reservoir water level time series change data of the landslide area. S2. Based on the seepage data, stress data and geological survey parameters of the landslide area, construct a seepage-stress coupled physical benchmark model, and use the seepage-stress coupled physical benchmark model to perform multi-condition numerical simulation to obtain the historical stability coefficient and historical residual sliding force of the landslide area. S3. Construct a standardized sample library with the standard deformation rate data and the historical reservoir water level change rate as inputs and the historical stability coefficient and the historical residual sliding force as outputs, and construct a physical information neural network inversion model based on the standardized sample library; S4. Determine the real-time reservoir water level change rate based on the real-time reservoir water level time-series change data of the landslide area, and input the real-time deformation rate data of the landslide area and the real-time reservoir water level change rate into the physical information neural network inversion model to obtain the real-time stability coefficient of the landslide area. S5. Determine the landslide warning level of the landslide area at the current moment based on the real-time stability coefficient.
2. The method according to claim 1, characterized in that, S1 specifically includes: Outlier removal, missing value completion, and smoothing preprocessing were performed on the historical deformation rate data and historical reservoir water level time series change data of the landslide area in the reservoir area to obtain the corresponding standard deformation rate data and standard reservoir water level time series change data. The historical reservoir water level change rate is determined based on the time-series change data of the standard reservoir water level.
3. The method according to claim 1, characterized in that, The seepage-stress coupled physical benchmark model constructed in S2 based on the seepage data, stress data, and geological survey parameters of the landslide area specifically includes: Based on the seepage data, stress data, and geological survey parameters of the landslide area, and combined with the saturated-unsaturated seepage control equation and stress balance equation, a seepage-stress coupled physical benchmark model is constructed.
4. The method according to claim 1, characterized in that, In step S2, multi-condition numerical simulations are performed using the seepage-stress coupled physical benchmark model to obtain the historical stability coefficient and historical residual sliding force of the landslide area, specifically including: Set up multiple combinations of different working conditions, including reservoir water level conditions and geotechnical mechanics; Based on multiple combinations of different working conditions, and using the seepage-stress coupled physical benchmark model to perform multi-condition numerical simulations, the historical stability coefficient and historical residual sliding force of the deepest sliding surface of the landslide area under the corresponding combinations of different working conditions are obtained.
5. The method according to claim 1, characterized in that, The step S3, which involves constructing a physical information neural network inversion model based on the standardized sample library, specifically includes: A deep learning neural network is trained using the standardized sample library to obtain a deep learning agent model; Based on the deep learning agent model and the constraints of landslide physics and mechanics, a physical information neural network is constructed. Based on the deep learning agent model and the physical information neural network, a physical information neural network inversion model is constructed.
6. The method according to claim 5, characterized in that, Based on the deep learning proxy model and the physical information neural network, a physical information neural network inversion model is constructed, specifically including: Calculate the data fitting error of the deep learning agent model and the physical constraint residual of the physical information neural network; The total loss function is calculated based on the data fitting error and the physical constraint residual. The physical information neural network is trained using the total loss function to obtain the physical information neural network inversion model.
7. The method according to claim 6, characterized in that, Calculating the data fitting error of the deep learning agent model specifically includes: The training samples in the standardized sample library are input into the deep learning proxy model to obtain the predicted stability coefficient and predicted residual sliding force of the landslide area. Calculate the first mean square error between the predicted stability coefficient and the historical stability coefficient, and calculate the second mean square error between the predicted residual glide force and the historical residual glide force; The first mean square error and the second mean square error are used as the data fitting error of the deep learning proxy model.
8. The method according to claim 7, characterized in that, Calculating the physical constraint residuals of the physical information neural network specifically includes: Based on the predicted stability coefficient and the predicted residual glide force, calculate the limit equilibrium definition residual and the Mohr-Coulomb criterion residual, respectively. The residuals defined by the limit equilibrium and the residuals defined by the Mohr-Coulomb criterion are used as the physical constraint residuals of the physical information neural network.
9. The method according to claim 6, characterized in that, The total loss function is calculated based on the data fitting error and the physical constraint residuals, specifically including: The total loss function is obtained by weighted summation of the data fitting error and the physical constraint residual.
10. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the landslide early warning classification method based on slope deformation rate as described in any one of claims 1 to 9.
