Landslide adaptive early warning system and method based on dynamic model and reinforcement learning

Abstract

The application relates to the technical field of geological disaster monitoring and early warning, and provides a landslide adaptive early warning system and method based on a dynamic model and reinforcement learning. The landslide adaptive early warning system based on the dynamic model and the reinforcement learning comprises a data sensing and preprocessing module, a physical constraint agent model module, a multi-source data real-time assimilation engine module, a deep reinforcement learning early warning decision module and a strategy execution and feedback module. The system can replace a traditional numerical model with the physical constraint agent model to improve the calculation speed, update the state and uncertainty in real time through data assimilation, and realize adaptive decision based on a physical mechanism and risk through deep reinforcement learning, so that the real-time early warning demand is met, and the accuracy and engineering robustness of early warning are improved. In addition, the application also provides a method applied to the landslide adaptive early warning system based on the dynamic model and the reinforcement learning.

Description

Technical Field

[0001] This application relates to the technical field of geological disaster monitoring and early warning, and in particular to an adaptive landslide early warning system and method based on dynamic models and reinforcement learning. Background Technology

[0002] Landslide early warning is an important research direction in the field of geological disaster prevention and control. Among existing technologies, one approach employs data assimilation methods based on high-precision physical numerical models. For example, it uses the finite element method to establish a slope seepage and stress coupling model, and then uses algorithms such as ensemble Kalman filtering to assimilate monitoring data such as displacement and pore water pressure to update the model's state variables and material parameters. This approach can well reflect the physical mechanism of landslide evolution, and the prediction results have high reliability. Another approach uses purely data-driven models or surrogate models with introduced physical constraints for rapid prediction. For example, it uses long short-term memory networks or deep neural networks to establish a mapping relationship from monitoring data to landslide state, thereby improving computational speed and meeting the needs of real-time monitoring.

[0003] However, each of the aforementioned existing technologies has its limitations. While schemes based on high-precision physical numerical models have clear physical mechanisms, single-scale simulations are time-consuming, typically requiring tens of minutes or even longer, making it difficult to meet the high-frequency assimilation requirements for real-time landslide early warning and creating a bottleneck in computational efficiency. Purely data-driven or surrogate-model-based schemes, although advantageous in computational speed, have prediction results highly dependent on the distribution of training data. When monitoring data is sparse, sensors malfunction, or new conditions such as unprecedented extreme rainfall occur, the prediction results may deviate from the fundamental physical laws of seepage and stress coupling, leading to decreased reliability. Furthermore, the early warning decisions of these schemes often employ fixed thresholds or static classifiers, failing to effectively utilize the uncertainties generated during data assimilation and lacking the ability to make sequential decisions in response to the continuous evolution of landslides.

[0004] Therefore, the main challenge facing existing technologies is how to improve real-time computational efficiency while ensuring that the prediction results conform to physical mechanisms, and on this basis, achieve adaptive early warning decisions. Specifically, there is a lack of a technical solution that can organically combine rapid positive prediction, multi-source observation data fusion, and sequential decision-making that considers uncertainties, making it difficult to simultaneously satisfy the real-time performance, physical reliability, and decision-making adaptability of landslide early warning. Summary of the Invention

[0005] In view of this, this application aims to propose a landslide adaptive early warning system and method based on dynamic models and reinforcement learning to solve the above-mentioned technical problems.

[0006] To achieve the above objectives, the technical solution of this application is implemented as follows: In the first aspect, this application proposes a landslide adaptive early warning system based on dynamic models and reinforcement learning, including a data sensing and preprocessing module for receiving and processing multi-source monitoring data, which includes ground sensor data and remote sensing data. The Physical Constraint Proxy Model module is used to construct a neural network proxy model that is deeply embedded in the seepage-stress coupling physical constraints, serving as a fast forward predictor. The multi-source data real-time assimilation engine module uses the physical constraint proxy model as the core predictor and adopts an ensemble Kalman filter framework to fuse observation data and model predictions in real time, and outputs the best estimate of the posterior state estimate and its uncertainty measure. The deep reinforcement learning early warning decision module models the early warning decision as a Markov decision process and trains a deep reinforcement learning agent. The state space of the agent includes the best estimate and uncertainty measure of the posterior state estimate, the action space is a combination of hierarchical early warning levels and disposal measures, and the reward function includes an uncertainty penalty term. The strategy execution and feedback module is used to execute the early warning actions output by the deep reinforcement learning agent, and send the effect feedback and new observation data after the action execution back to the system to form a closed loop.

[0007] Furthermore, the physical constraint proxy model module adopts a physical information neural network or variational physical information neural network architecture, takes key geotechnical parameters as explicit inputs to the network, and embeds the residuals of the seepage-stress coupling control equations into the loss function; the control equations include pore water pressure-seepage equations and displacement-stress relationship equations.

[0008] Furthermore, the physical constraint proxy model module supports batch input of different parameter sets and boundary conditions, and uses GPU parallel computing to obtain multiple state prediction sets in a single forward propagation, thereby accelerating the set Kalman filter prediction step.

[0009] Furthermore, the multi-source data real-time assimilation engine module adopts an adaptive ensemble Kalman filter variant, which can estimate and dynamically adjust the observation noise covariance matrix online, and introduce a mask matrix to handle sensor missing or failure situations.

[0010] Furthermore, the state space of the deep reinforcement learning early warning decision module is composed of risk features, trend features, and uncertainty features; The risk features include stability safety factor, instability risk, displacement velocity, and pore pressure ratio; the trend features are extracted by the slope within a sliding time window; the uncertainty features are extracted from the posterior covariance matrix, including displacement field variance, pore pressure field variance, parameter field variance, maximum eigenvalue, and cross covariance norm.

[0011] Furthermore, the reward function includes a correct warning reward, a missed warning penalty, a false alarm penalty, an over-escalation penalty, an uncertainty penalty, and an action cost penalty, wherein the missed warning penalty has a higher weight than the false alarm penalty.

[0012] Furthermore, the data perception and preprocessing module employs a hybrid model combining a spatiotemporal attention mechanism and a temporal convolutional network to perform temporal interpolation and spatial downscaling on low-frequency remote sensing areal displacement data, generating virtual observation displacement values ​​and their error covariance that match the ground monitoring points.

[0013] Secondly, this application proposes a method for applying the aforementioned landslide adaptive early warning system based on dynamic models and reinforcement learning, comprising the following steps: Step 1, Multi-source data synchronization and preprocessing: Receive ground sensor data and remote sensing data, perform spatiotemporal interpolation and downscaling on the remote sensing data, generate virtual observation point data, and align it with the ground sensor data in terms of timestamps; Step 2, Prediction of the surrogate model set: Input the posterior state set of the previous time step into the physical constraint surrogate model, perform a forward prediction step, and obtain the prior state set of the current time step; Step 3, Real-time data assimilation update: When new observation data arrives, an ensemble Kalman filter update step is executed, outputting the ensemble mean of the posterior state estimate as the best estimate, and the posterior covariance as the uncertainty measure; Step 4, Deep reinforcement learning agent decision-making: Construct the best estimate and uncertainty measure into a state vector, input it into the trained deep reinforcement learning agent, and output the current optimal warning action; Step 5, Strategy Execution and Feedback: Execute the warning action, collect feedback on the effect of the action and a new round of observation data, and send them back to the system to form a closed loop.

[0014] Furthermore, in step 2, a physical information neural network or a variational physical information neural network is used as a proxy model. Key geotechnical parameters are used as explicit inputs, and multiple parameter sets are input in batches. Multiple state prediction sets are obtained by using GPU parallel computing for forward propagation in one go.

[0015] Furthermore, in step 3, an adaptive ensemble Kalman filter variant is used to estimate and dynamically adjust the observation noise covariance matrix online, and a mask matrix is ​​introduced to handle sensor missing or failure cases. In step 4, the deep reinforcement learning agent is trained offline using a proximal policy optimization algorithm. Its state space consists of risk features, trend features, and uncertainty features, and its reward function includes an uncertainty penalty term.

[0016] Compared with existing technologies, the landslide adaptive early warning system and method based on dynamic models and reinforcement learning proposed in this application have the following advantages: (1) This application uses a system consisting of a data perception and preprocessing module, a physical constraint proxy model module, a multi-source data real-time assimilation engine module, a deep reinforcement learning early warning decision module, and a strategy execution and feedback module to realize the complete process from multi-source data acquisition to early warning action execution and feedback, thereby improving the real-time performance and adaptability of early warning; by adopting a physical information neural network or variational physical information neural network architecture, key geotechnical parameters are used as explicit inputs to the network, and the residuals of the seepage-stress coupling control equation are embedded in the loss function, so that the prediction results follow the basic physical laws of seepage and stress coupling, thereby maintaining high reliability when monitoring data is sparse or new working conditions are encountered; by supporting the input of different parameter sets in batches and using GPU parallel computing to forward propagate to obtain multiple state prediction sets at once, the prediction step of the set Kalman filter is accelerated, thereby improving the real-time performance of high-frequency assimilation.

[0017] (2) This application uses an adaptive ensemble Kalman filter variant to estimate and dynamically adjust the observation noise covariance matrix online, and introduces a mask matrix to handle sensor missing or failure situations, thereby achieving robust handling of data loss and sensor anomalies and reducing estimation bias caused by observation failures; by designing the state space of the deep reinforcement learning early warning decision module to include a combination of risk features, trend features and uncertainty features, the agent can make decisions by comprehensively considering physical risks, changing trends and estimation uncertainties, thus achieving more reasonable adaptive early warning; by including correct early warning rewards, missed warning penalties, false alarm penalties, over-escalation penalties, uncertainty penalties and action cost penalties in the reward function, and setting the weight of the missed warning penalty to be higher than that of the false alarm penalty, the agent is guided to make a dynamic trade-off between false alarm costs and missed warning risks, avoiding taking aggressive actions in a state of high uncertainty.

[0018] (3) The method provided in this application realizes an end-to-end closed-loop early warning process through the steps of multi-source data synchronization and preprocessing, surrogate model set prediction, real-time data assimilation and update, deep reinforcement learning agent decision-making and policy execution and feedback, so that the early warning decision can be dynamically adjusted according to the real-time updated state estimate and its uncertainty; by using physical information neural network or variational physical information neural network as surrogate model, key geotechnical parameters are used as explicit inputs, and multiple parameter sets are input in batches and GPU parallel computing is used to realize the unity of fast prediction and physical credibility; by using adaptive ensemble Kalman filter variant to estimate and dynamically adjust the observation noise covariance matrix online, and by using mask matrix to handle sensor missing or failure cases, and by using the near-end policy optimization algorithm in the deep reinforcement learning agent for offline training, stable decision-making and adaptive early warning under the condition of incomplete data are realized, improving the engineering practicality of the system. Attached Figure Description

[0019] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating the adaptive landslide early warning method based on dynamic models and reinforcement learning described in this application embodiment. Detailed Implementation

[0020] To make the technical solution and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0021] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0022] Furthermore, it should be noted that in the description of this application, if terms such as "upper," "lower," "inner," or "outer" appear, indicating orientation or positional relationship, these are based on the orientation or positional relationship shown in the accompanying drawings and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation on this application. In addition, if terms such as "first" or "second" appear, they are also used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0023] Furthermore, in the description of this application, unless otherwise expressly defined, the terms "installation," "connection," "joining," and "connector" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this application in light of the specific circumstances.

[0024] In this application, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to a specific feature, structure, material, or characteristic described in connection with that embodiment or example, which is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0025] The technical solution of this application will be described in detail below with reference to specific embodiments. This embodiment provides a landslide adaptive early warning system and method based on dynamic models and reinforcement learning. The landslide adaptive early warning system based on dynamic models and reinforcement learning in this embodiment includes a data perception and preprocessing module, a physical constraint surrogate model module, a multi-source data real-time assimilation engine module, a deep reinforcement learning early warning decision module, and a strategy execution and feedback module. The data perception and preprocessing module is used to receive and process multi-source monitoring data, including ground sensor data and remote sensing data. The physical constraint surrogate model module is used to construct a neural network surrogate model deeply embedded with seepage stress coupled with physical constraints, as a fast forward predictor. The multi-source data real-time assimilation engine module uses the physical constraint surrogate model as the core predictor, adopts an ensemble Kalman filter framework to fuse observation data and model predictions in real time, and outputs the best estimate of the posterior state estimate and its uncertainty measure. The deep reinforcement learning early warning decision module models the early warning decision as a Markov decision process and trains a deep reinforcement learning agent. The state space of the agent includes the best estimate of the posterior state estimate and the uncertainty measure, the action space is a combination of hierarchical early warning levels and disposal measures, and the reward function includes an uncertainty penalty term. The policy execution and feedback module is used to execute the early warning actions output by the deep reinforcement learning agent, and send the feedback on the effect of the action and new observation data back to the system to form a closed loop.

[0026] To achieve integrated operation from perception to decision-making, the five core modules in this embodiment are connected sequentially according to the data flow order. The output of the data perception and preprocessing module is connected to the input of the physical constraint surrogate model module; the output of the physical constraint surrogate model module is connected to the input of the multi-source data real-time assimilation engine module; the output of the multi-source data real-time assimilation engine module is connected to the input of the deep reinforcement learning early warning decision-making module; the output of the deep reinforcement learning early warning decision-making module is connected to the input of the policy execution and feedback module; and the feedback output of the policy execution and feedback module is connected to the inputs of both the physical constraint surrogate model module and the multi-source data real-time assimilation engine module. Based on the inventive objective of solving the problem of module fragmentation in existing technologies, this embodiment forms an end-to-end topology of a physical surrogate prediction set, assimilation uncertainty perception, deep reinforcement learning decision-making, and decision feedback system optimization, enabling the effective superposition of the technical advantages of each module. Through this integrated architecture, the system can achieve integrated coupling of high-frequency state prediction, real-time multi-source data fusion, and sequential early warning decision-making. In simulation experiments, the early warning accuracy is improved to 94.7% compared to existing technologies and by 18.5% compared to pure data-driven schemes.

[0027] To address the issue of inconsistent spatiotemporal scales in multi-source monitoring data, this embodiment employs a hybrid model combining a spatiotemporal attention mechanism and a temporal convolutional network in its data perception and preprocessing module. This model performs temporal interpolation and spatial downscaling on low-frequency remote sensing areal displacement data, generating virtual observation displacement values ​​and their error covariance that match ground monitoring points. Based on the objective of effectively fusing multi-source data, this embodiment uses a stacked dilated causal convolutional network. Through multi-layer stacking, an exponentially increased receptive field is obtained, effectively modeling the long-term temporal evolution pattern of the displacement field and performing high-precision interpolation of the remote sensing displacement field for missing time points, generating a continuous time series. The spatiotemporal attention mechanism is used to learn the spatial correlation between remote sensing areal data and discrete ground point data, calculating the attention weight between each ground monitoring point location and the remote sensing pixel. Weighted aggregation generates virtual observation displacement values ​​with uniform temporal resolution for each ground monitoring point. The error covariance of the virtual observation is inversely proportional to the entropy of the attention weights, objectively reflecting the reliability of the virtual observation. In the simulation experiment, synthetic aperture radar interferometry data with a 6-day interval and ground sensor data with a 1-minute interval were used. This preprocessing module can generate virtual observations that are time-synchronized with the ground data, providing accurate input for multi-source data assimilation.

[0028] Specifically, in this embodiment, the ground sensor data for the data sensing and preprocessing module comes from Global Navigation Satellite System (GNSS) devices such as displacement gauges, pore water pressure gauges, and rain gauges, while the remote sensing data comes from Synthetic Aperture Radar Interferometry (InSAR). To unify the spatiotemporal scale of data from different sources, this module employs a hybrid model combining a spatiotemporal attention mechanism and a temporal convolutional network. For low-frequency InSAR displacement field data, temporal interpolation is performed using a temporal convolutional network. This network uses dilated causal convolution stacking, and the output of its l-th layer at position t is given by the formula:

[0029] in Indicates the expansion factor as Causal convolution operation, and These are learnable parameters. Through multi-layer stacking, the network has an exponentially increased receptive field, enabling it to effectively model long-term dependencies.

[0030] By stacking multiple layers, an exponentially increased receptive field is obtained, effectively modeling long-term dependencies. Based on this, a spatiotemporal attention mechanism is used to calculate the attention weights between each ground monitoring point location and InSAR pixels. :

[0031] in, Ground point The feature query vector is obtained by encoding the historical observation data of that point through a gated recurrent unit (GRU), that is... ; InSAR pixels The feature key vector is extracted from the displacement values ​​within the spatiotemporal neighborhood of that pixel using a lightweight convolutional network (CNN). , Represents a pixel Spatial neighborhood; The similarity function uses a scaled dot product. By using weighted aggregation, a matching "virtual observation" displacement value with a uniform time resolution is generated for each ground monitoring point.

[0032] Simultaneously estimate the error covariance matrix of the virtual observation. Its diagonal elements are inversely proportional to the entropy of the attention weights, that is... This provides accurate observational information for subsequent assimilation.

[0033] To ensure the physical reliability of the surrogate model predictions, the physical constraint surrogate model module in this embodiment adopts a physical information neural network or variational physical information neural network architecture. Key geotechnical parameters are used as explicit inputs to the network, and the residuals of the seepage-stress coupling control equations are embedded in the loss function. The control equations include the pore water pressure seepage equation and the displacement-stress relationship equation.

[0034] To enhance the generalization ability of the model, this embodiment employs an encoder-solver-decoder paradigm for constructing the physical constraint surrogate model. The encoder consists of multiple convolutional and fully connected layers. For spatial field input, a UNet-style encoder extracts multi-scale features; for temporal scalar input, the features are projected through fully connected layers and concatenated with the spatial features along the channel dimension. The solver uses the residuals of the seepage stress coupling control equations as the core component of the loss function. The total loss function comprises data fitting loss, physical constraint loss, boundary condition loss, initial condition loss, observation consistency loss, and a weighted regularization term. The physical constraint loss corresponds to the residuals of the mechanical equilibrium equation, effective stress equation, displacement-stress constitutive relation, and pore water pressure evolution equation, respectively.

[0035] To adapt to different seepage stages, the Richards equation is used to describe water transport during the unsaturated rainfall infiltration stage, while the Darcy-Biot simplified form is used to describe pore pressure evolution during the near-saturated or saturated consolidation stage. Key soil and rock parameters are used as explicit input channels to the network, enabling the model to directly reflect the impact of parameter changes on displacement and pore pressure state, thus supporting joint estimation of parameters and state during assimilation. Through deep embedding of physical constraints, the surrogate model can still follow basic physical laws under new working conditions not covered by the training data. In the generalization test of extreme rainfall patterns, the mean square error of displacement prediction by the surrogate model is stable at 1.2 × 10⁻⁶. -4 m 2 Compared to a purely data-driven long short-term memory network model, it reduces performance by 86.2%.

[0036] To improve the computational efficiency of the ensemble Kalman filter prediction step, the physical constraint surrogate model module in this embodiment supports batch input of different parameter sets and boundary conditions. It utilizes the graphics processing unit (GPU) for parallel computation to obtain multiple state prediction sets in a single forward propagation, thereby accelerating the ensemble Kalman filter prediction step. Based on the invention's objective of overcoming real-time bottlenecks, the batch processing capability of the surrogate model in this embodiment is highly compatible with the ensemble propagation requirements of the ensemble Kalman filter. The ensemble Kalman filter requires the simultaneous propagation of multiple state set members to estimate the statistical properties of the states. Traditional numerical models require calculating each set member individually, with computation time increasing linearly with the set size. The surrogate model in this embodiment leverages the parallel computing capabilities of the GPU, enabling it to process tens to hundreds of set members simultaneously in a single forward propagation. In simulation experiments, using an ensemble Kalman filter framework with 100 set members, the surrogate model completes one ensemble prediction in 1.8 seconds, a 99.9% reduction compared to the 1520 seconds required by the high-precision finite element model, laying the foundation for achieving a high-frequency assimilation loop once per minute.

[0037] Based on the invention's objective of achieving stable and reliable multi-source data fusion, in this embodiment, the adaptive ensemble Kalman filter can update the elements of the observation noise covariance matrix in real time according to the residual statistical characteristics of the observed data and model predictions, adapting to changes in noise levels and data quality fluctuations from different sensors. When there are missing or failed sensors, the system introduces a mask matrix to filter valid observations, constructing observation vectors and operators containing only valid observations, thereby ensuring that assimilation updates are performed only within the valid observation subspace. In a fault scenario simulating 12 consecutive hours of missing data from three key pore water pressure gauges in the middle of a slope, the average root mean square error of the system state estimation in this embodiment remained at 0.015 MPa, while the root mean square error of the purely data-driven long short-term memory network model rose to 0.082 MPa. Simultaneously, the parameter uncertainty output by the assimilation engine increased by approximately 300%, and this uncertainty information can be effectively utilized by subsequent decision-making modules.

[0038] Specifically, in this embodiment, the physical constraint proxy model module is used to construct a neural network proxy model deeply embedded with seepage-stress coupled physical constraints, serving as a fast forward predictor. To achieve joint parameter-state estimation, the proxy model employs a Physical Information Neural Network (PINN) or Variational Physical Information Neural Network (VPINN) architecture, using key geotechnical parameters as explicit inputs to the network and embedding the residuals of the seepage-stress coupled control equations into the loss function. The control equations include the pore water pressure-seepage equation and the displacement-stress relationship equation. Specifically, the network inputs include the current boundary conditions such as rainfall intensity P(t) and reservoir water level H(t), material parameters such as permeability field k and elastic modulus field E, and the state at the previous time step. The network uses an encoder-solver-decoder structure. The total loss function... Defined as:

[0039] in, Represents the displacement field predicted by the network. , pore pressure field The fitting loss between the high-fidelity numerical simulation data and the data itself. This represents the residual loss obtained by substituting the network output into the control equation. , These represent boundary conditions and initial conditions, respectively. This indicates consistency constraints with on-site multi-source monitoring and observation. This represents the regularization term for network parameters.

[0040] Richards' equation is used for the unsaturated rainfall infiltration stage, and Darcy-Biot's form is used for the near-saturated or saturated consolidation stage. The formula is as follows:

[0041]

[0042] in, For pressure head, Water content It is an unsaturated hydraulic conductivity function. For the specific water storage coefficient, The dynamic viscosity of water. For the body to adapt to strain, , These are rainfall infiltration source term and pore pressure source term, respectively.

[0043] The mechanical equilibrium equation, the effective stress equation, and the displacement-stress relationship can be expressed as:

[0044]

[0045]

[0046] The effective stress equation described above clearly demonstrates the weakening effect of pore water pressure on the stress of the soil skeleton. Based on this, the Mohr-Coulomb strength criterion can be further used to calculate the shear strength of the potential sliding surface:

[0047] in, For effective cohesion, For the effective internal friction angle, This is the effective normal stress. For the pre-defined potential sliding surface... Integrating the anti-slip force and the sliding force on the surface yields the stability safety factor:

[0048] in, To actually mobilize shear stress along the potential sliding surface, To prevent extremely small positive numbers with a denominator of zero, the instability risk can be further defined as a monotonic mapping of the safety factor:

[0049] in, For risk sensitivity coefficient; when hour, The rapid increase provides continuous quantitative risk indicators for subsequent early warning decisions.

[0050] In the strong-form PINN implementation, the residuals of the governing equations can be written as:

[0051]

[0052] When the Richards equation is used to describe the unsaturated infiltration process, the corresponding seepage residual can be replaced by:

[0053] Accordingly, each loss item can be expanded as follows:

[0054]

[0055]

[0056]

[0057]

[0058]

[0059] in, For the observation operator, To observe the noise covariance, These are the learnable parameters for the neural network. If VPINN is used, the strong-form residuals described above are replaced with trial functions. , The weak integral constraint, that is, the requirement and This improves stability in complex boundary and sparse sampling scenarios.

[0060] This proxy model was specifically designed to serve real-time data assimilation, including: Explicit input: Input key geotechnical parameters (such as...) As an explicit input channel to the network, the model can directly reflect the influence of parameter changes on the state (displacement, pore pressure), thus supporting joint estimation of parameters and state during assimilation.

[0061] Batch set prediction: The network supports inputting different sets of parameters in batches. Given the boundary conditions, the results are obtained through a single forward propagation using the parallel computing capabilities of the GPU. A set of state predictions This greatly improves the efficiency of the prediction step in the ensemble Kalman filter.

[0062] Fast gradient calculation: Utilizing the automatic differentiation (AutoDiff) feature of modern deep learning frameworks, the Jacobian matrix of the network output with respect to the input parameters can be calculated efficiently and accurately. This provides support for certain assimilation algorithm variants that require gradient information.

[0063] To improve the robustness of the data assimilation process to abnormal data and sensor failures, the multi-source data real-time assimilation engine module in this embodiment adopts an adaptive ensemble Kalman filter variant, which can estimate and dynamically adjust the observation noise covariance matrix online, and introduce a mask matrix to handle sensor missing or failure situations.

[0064] Specifically, in this embodiment, the multi-source data real-time assimilation engine module uses a physical constraint proxy model as the core predictor and employs an ensemble Kalman filter (EnKF) framework to achieve real-time fusion of multi-source observation data and model predictions. The state vector is defined as follows: ,in These represent the key material parameters to be inverted.

[0065] At any moment It has a set of states. For each set member Using the proxy model Perform a forward prediction step:

[0066] in Given the known boundary conditions (rainfall, water level). To address process noise, small perturbations are added to the input of the surrogate model. To achieve this, in which This is the preset process noise covariance matrix.

[0067] When new observation data Upon arrival, calculate the mean of the set of samples. Covariance Then, through Kalman gain... Incorporate observational information into each set member:

[0068] in For the observation operator, To address observation noise, this module employs an adaptive EnKF (AEnKF) variant, enabling online estimation and dynamic adjustment of the observation noise covariance matrix. It can robustly handle missing data, asynchronous arrivals, and outliers.

[0069] To facilitate the unified inclusion of observations from different sources into the assimilation framework, the observation vector is defined as:

[0070] in:

[0071] For the There are GNSS displacement observation points.

[0072] For the One pore water pressure gauge, with

[0073] For the There are 10 InSAR virtual observation points.

[0074] in, , Let represent the sampling matrices for extracting displacement and pore pressure components from the state field, respectively. This represents the InSAR line-of-sight direction vector. This represents the interpolation / downscaling weights determined by the spatiotemporal attention mechanism.

[0075] When there is sensor failure or malfunction, a mask matrix is ​​introduced. Valid observations are screened:

[0076] This ensures that assimilation updates are performed only within the valid observation subspace.

[0077] To comprehensively characterize the risk status and decision-making basis of landslides, the state space of the deep reinforcement learning early warning decision-making module in this embodiment consists of risk features, trend features, and uncertainty features. Risk features include the stability safety factor, instability risk, displacement velocity, and pore pressure ratio. Trend features are extracted using the slope within a sliding time window. Uncertainty features are extracted from the posterior covariance matrix, including the displacement field variance, pore pressure field variance, parameter field variance, maximum eigenvalue, and cross-covariance norm.

[0078] Based on the invention's objective of providing sufficient input for intelligent decision-making, in this embodiment, risk features are calculated from the assimilated posterior mean state. First, the displacement, pore pressure, and parameter fields are recovered, and the displacement velocity and acceleration are calculated. Then, the stability safety factor of the potential sliding surface is calculated based on the effective stress and the Mohr-Coulomb strength criterion. Finally, continuous instability risk indicators are obtained through monotonic mapping. Trend features are extracted within a sliding time window of 12 to 24 hours. The slope of each risk indicator is calculated using linear regression to reflect the dynamic trend of landslide evolution. Choosing a sliding time window of 12 to 24 hours effectively filters short-term monitoring noise while retaining key trend information about accelerated landslide instability. Uncertainty features are obtained through block decomposition and statistical calculation of the posterior covariance matrix, representing the estimation uncertainties of the displacement, pore pressure, and parameter fields, as well as the coupling uncertainties between different state variables. This state-space design can simultaneously provide the current risk level evolution trend and estimation reliability of the landslide, providing a comprehensive basis for the agent's sequential decision-making.

[0079] To guide the agent in learning the optimal early warning strategy, the reward function in this embodiment includes a correct early warning reward, a missed warning penalty, a false alarm penalty, an over-escalation penalty, an uncertainty penalty, and an action cost penalty, with the missed warning penalty having a higher weight than the false alarm penalty. Based on the inventive objective of achieving a dynamic trade-off between the cost of false alarms and the risk of missed alarms, in this embodiment, the correct early warning reward is weighted according to the early warning lead time, encouraging earlier but not excessive early warnings. The upper limit of the lead time factor is set to 1, corresponding to early warnings of 3 hours or more. The weight of the missed warning penalty is set to 5 to 10 times the weight of the false alarm penalty to reflect the high risk of loss of life and property caused by landslide instability. The over-escalation penalty is used to constrain unnecessary early warning escalation relative to the actual risk. The uncertainty penalty is positively correlated with the current estimated uncertainty and the early warning level, inhibiting aggressive early warning actions under high uncertainty. The action cost penalty is used to quantify the engineering costs and social impacts corresponding to different early warning actions. Through this reward function design, the agent can learn the optimal decision-making strategy that comprehensively considers the uncertainty of risk trends and costs. In simulation experiments, the agent trained with this reward function achieved a false alarm rate of 5.1% and a false negative rate of 2.4%, which are 76.7% and 71.8% lower than the pure data-driven scheme, respectively. The average early warning time reached 3.5 hours, which is about 66.7% higher than the pure data-driven scheme.

[0080] Specifically, in this embodiment, the deep reinforcement learning-based early warning decision module models the early warning decision as a Markov Decision Process (MDP) and trains a DRL agent to learn the optimal early warning strategy. The environment in which the agent exists is composed of a physically constrained agent model updated by a real-time assimilation engine. This is a dynamic environment whose internal state (the physical state of the landslide) evolves in real time with the assimilation process. The environment changes at each decision step... Provide a state to the agent State transition probability of the environment It is determined by both the prediction step and the update step of the assimilation engine.

[0081] It includes not only the best state estimate after assimilation, such as the displacement field mean , average value of pore pressure field and its changing trends More fundamentally, it includes uncertainty measures, such as the variance field of the displacement field. or posterior covariance matrix The main diagonal element. The state representation can be designed as follows:

[0082] In engineering implementation, the aforementioned high-dimensional field state is further compressed into a combined vector of "risk characteristics + trend characteristics + uncertainty characteristics". The extraction steps are as follows: The first step is to use the posterior mean. Recover the displacement, pore pressure, and parametric field, and calculate the displacement velocity and acceleration:

[0083] The second step is to calculate the risk characteristics based on the formula for effective stress and safety factor:

[0084] in, For the first The empirical critical pore pressure threshold for a region or monitoring point.

[0085] The third step is to [do something] in a length of [length]. Trend features are extracted within a sliding time window. For any scalar sequence... Its trend slope is defined as:

[0086] Based on this, trend characteristics are constructed:

[0087] The fourth step is to extract uncertainty features from the posterior covariance matrix. Decomposed into state blocks as follows:

[0088] Then it can be defined as:

[0089] in, Represents the largest eigenvalue. This represents the Frobenius norm. Ultimately, the DRL state vector can be specifically represented as:

[0090] The design is hierarchical, including warning levels (such as blue, yellow, orange, and red) and specific combinations of response recommendations (such as "strengthen monitoring", "activate Plan A", and "organize evacuation").

[0091] reward function It is key to guiding the agent's learning, and its design is as follows:

[0092] Describe the warning level function These correspond to four warning levels: blue, yellow, orange, and red. This indicates the target warning level obtained from actual instability evolution or expert annotation. If we represent the early warning time relative to the instability event, then:

[0093]

[0094]

[0095]

[0096]

[0097]

[0098] in, Rewards are given for accurate early warnings, and "earlier but not excessive" warnings are encouraged through an advance time factor; The penalty for underreporting is usually... To reflect the high risk of underreporting; Penalties for false alarms; Used to constrain excessive escalation behavior relative to the target level; As an uncertainty penalty term, this term increases significantly when the posterior covariance is large and the action level is high, thereby inhibiting the rash implementation of aggressive early warning actions under high uncertainty. This is used to characterize the engineering costs of actions such as encrypted monitoring, activation of contingency plans, and organization of evacuation.

[0099] Advanced DRL algorithms, such as Proximal Policy Optimization (PPO), are employed for offline training of the agent. The PPO algorithm stabilizes policy updates through importance sampling and pruning mechanisms. (Policy Network) and value network Typically, this is an MLP or GNN. The update objective is to maximize the alternative objective function after pruning.

[0100] in , This is the estimated advantage function. After training, the agent is deployed online, based on the real-time input state. Directly output the optimal early warning action .

[0101] This module is responsible for executing the early warning actions output by the DRL agent, such as issuing warning information and activating emergency plans. Simultaneously, it collects feedback on the effects of these actions and new observation data, sending it back to the system. The system supports fine-tuning the DRL strategy using new online data, enabling lifelong learning.

[0102] Furthermore, this embodiment also provides a landslide adaptive early warning method based on dynamic models and reinforcement learning, applied to the aforementioned landslide adaptive early warning system based on dynamic models and reinforcement learning, with reference to... Figure 1 This includes the following steps: Step 1: Multi-source data synchronization and preprocessing. Receive ground sensor data and remote sensing data, perform spatiotemporal interpolation and downscaling on the remote sensing data to generate virtual observation point data, and align it with the ground sensor data in terms of timestamps.

[0103] Step 2, surrogate model set prediction. Input the posterior state set from the previous time step into the physical constraint surrogate model, perform a forward prediction step, and obtain the prior state set at the current time step.

[0104] Step 3, Real-time Data Assimilation and Update. When new observation data arrives, an ensemble Kalman filter update step is performed, outputting the ensemble mean of the posterior state estimate as the best estimate, and the posterior covariance as an uncertainty measure.

[0105] Step 4: Deep reinforcement learning agent decision-making. The best estimate and uncertainty metric are constructed into a state vector, which is then input into the trained deep reinforcement learning agent to output the current optimal warning action.

[0106] Step 5: Strategy Execution and Feedback. Execute the early warning action, collect feedback on the effect of the action and a new round of observation data, and send them back to the system to form a closed loop.

[0107] To ensure the method's scalability, each step in this embodiment is executed cyclically according to a fixed time period, with the assimilation and decision-making time period set to 1 minute. This 1-minute time period meets the response requirements for real-time early warning while also taking into account reasonable consumption of computational resources. Through this method, the system can continuously track and dynamically warn of landslide conditions, constantly updating state estimates and early warning strategies throughout the entire landslide evolution process.

[0108] To further improve the efficiency of surrogate model ensemble prediction, in step two of this embodiment, a physical information neural network or variational physical information neural network is used as the surrogate model. Key geotechnical parameters are used as explicit inputs, and multiple parameter sets are input in batches. A graphics processing unit (GPU) is used for parallel computation to perform a single forward propagation to obtain multiple state prediction sets. Based on the invention objective of achieving high-frequency assimilation loops, in this embodiment, the batch input size of the surrogate model is set to 50 to 200 set members. Choosing 50 to 200 set members allows for control of computational resource consumption while ensuring the accuracy of ensemble statistical characteristic estimation. In simulation experiments, using a configuration of 100 set members, the surrogate model completes one ensemble prediction in 1.8 seconds, which fully meets the requirement of one assimilation loop per minute.

[0109] To improve the robustness and decision-making performance of the method, in step three of this embodiment, an adaptive ensemble Kalman filter variant is used to estimate and dynamically adjust the observation noise covariance matrix online, and a mask matrix is ​​introduced to handle sensor missing or failure scenarios. In step four, the deep reinforcement learning agent is trained offline using a proximal policy optimization algorithm. Its state space consists of risk features, trend features, and uncertainty features, and its reward function includes an uncertainty penalty term. Based on the invention's objective of achieving stable and reliable adaptive early warning, in this embodiment, the proximal policy optimization algorithm stabilizes policy updates through importance sampling and pruning mechanisms. The policy network and value network adopt a three-layer multilayer perceptron structure, with the number of hidden layer neurons set to 128 to 256. Offline training uses a large amount of landslide evolution simulation data generated based on a high-precision finite element model, and the training process covers different rainfall patterns, reservoir water level fluctuation conditions, and parameter distributions. In a failure scenario where sensor data is missing for 12 consecutive hours, this method can maintain stable state estimation, and the agent can output reasonable early warning actions based on increased uncertainty, avoiding aggressive decisions due to insufficient information.

[0110] The above descriptions are merely some embodiments of this application and are not intended to limit this application. The technical features or structures in the foregoing different embodiments can be arbitrarily combined to form other specific technical solutions as needed. For those skilled in the art, this application can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of the claims of this application.

Claims

1. A landslide adaptive early warning system based on dynamic models and reinforcement learning, characterized in that, include: The data sensing and preprocessing module is used to receive and process multi-source monitoring data, which includes ground sensor data and remote sensing data. The Physical Constraint Proxy Model module is used to construct a neural network proxy model that is deeply embedded in the seepage-stress coupling physical constraints, serving as a fast forward predictor. The multi-source data real-time assimilation engine module uses the physical constraint proxy model as the core predictor and adopts an ensemble Kalman filter framework to fuse observation data and model predictions in real time, and outputs the best estimate of the posterior state estimate and its uncertainty measure. The deep reinforcement learning early warning decision module models the early warning decision as a Markov decision process and trains a deep reinforcement learning agent. The state space of the agent includes the best estimate and uncertainty measure of the posterior state estimate, the action space is a combination of hierarchical early warning levels and disposal measures, and the reward function includes an uncertainty penalty term. The strategy execution and feedback module is used to execute the early warning actions output by the deep reinforcement learning agent, and send the effect feedback and new observation data after the action execution back to the system to form a closed loop.

2. The landslide adaptive early warning system based on dynamic model and reinforcement learning according to claim 1, characterized in that, The physical constraint proxy model module adopts a physical information neural network or variational physical information neural network architecture, takes key geotechnical parameters as explicit inputs to the network, and embeds the residuals of the seepage-stress coupling control equations into the loss function; the control equations include pore water pressure-seepage equations and displacement-stress relationship equations.

3. The landslide adaptive early warning system based on dynamic model and reinforcement learning according to claim 2, characterized in that, The physical constraint proxy model module supports batch input of different parameter sets and boundary conditions, and uses GPU parallel computing to obtain multiple state prediction sets in one forward propagation, thereby accelerating the Kalman filter prediction step of the set.

4. The landslide adaptive early warning system based on dynamic model and reinforcement learning according to claim 1, characterized in that, The multi-source data real-time assimilation engine module adopts an adaptive ensemble Kalman filter variant, which can estimate and dynamically adjust the observation noise covariance matrix online, and introduce a mask matrix to handle sensor missing or failure situations.

5. The landslide adaptive early warning system based on dynamic model and reinforcement learning according to claim 1, characterized in that, The state space of the deep reinforcement learning early warning decision module consists of risk characteristics, trend characteristics, and uncertainty characteristics. The risk features include stability safety factor, instability risk, displacement velocity, and pore pressure ratio; the trend features are extracted by the slope within a sliding time window; the uncertainty features are extracted from the posterior covariance matrix, including displacement field variance, pore pressure field variance, parameter field variance, maximum eigenvalue, and cross covariance norm.

6. The landslide adaptive early warning system based on dynamic model and reinforcement learning according to claim 1, characterized in that, The reward function includes a reward for correct early warning, a penalty for missed warning, a penalty for false alarm, a penalty for over-escalation, a penalty for uncertainty, and a penalty for action cost, with the penalty for missed warning having a higher weight than the penalty for false alarm.

7. The landslide adaptive early warning system based on dynamic model and reinforcement learning according to claim 1, characterized in that, The data perception and preprocessing module employs a hybrid model combining a spatiotemporal attention mechanism and a temporal convolutional network to perform temporal interpolation and spatial downscaling on low-frequency remote sensing areal displacement data, generating virtual observation displacement values ​​and their error covariance that match ground monitoring points.

8. A landslide adaptive early warning method based on dynamic models and reinforcement learning, applied to the landslide adaptive early warning system based on dynamic models and reinforcement learning as described in any one of claims 1 to 7, characterized in that, Includes the following steps: Step 1, Multi-source data synchronization and preprocessing: Receive ground sensor data and remote sensing data, perform spatiotemporal interpolation and downscaling on the remote sensing data, generate virtual observation point data, and align it with the ground sensor data in terms of timestamps; Step 2, Prediction of the surrogate model set: Input the posterior state set of the previous time step into the physical constraint surrogate model, perform a forward prediction step, and obtain the prior state set of the current time step; Step 3, Real-time data assimilation update: When new observation data arrives, an ensemble Kalman filter update step is executed, outputting the ensemble mean of the posterior state estimate as the best estimate, and the posterior covariance as the uncertainty measure; Step 4, Deep reinforcement learning agent decision-making: Construct the best estimate and uncertainty measure into a state vector, input it into the trained deep reinforcement learning agent, and output the current optimal warning action; Step 5, Strategy Execution and Feedback: Execute the warning action, collect feedback on the effect of the action and a new round of observation data, and send them back to the system to form a closed loop.

9. The landslide adaptive early warning method based on dynamic model and reinforcement learning according to claim 8, characterized in that, In step 2, a physical information neural network or variational physical information neural network is used as a proxy model. Key geotechnical parameters are used as explicit inputs, and multiple parameter sets are input in batches. Multiple state prediction sets are obtained by using GPU parallel computing for forward propagation in one go.

10. The landslide adaptive early warning method based on dynamic model and reinforcement learning according to claim 8, characterized in that, In step 3, an adaptive ensemble Kalman filter variant is used to estimate and dynamically adjust the observation noise covariance matrix online, and a mask matrix is ​​introduced to handle sensor missing or failure situations. In step 4, the deep reinforcement learning agent is trained offline using a proximal policy optimization algorithm. Its state space consists of risk features, trend features, and uncertainty features, and its reward function includes an uncertainty penalty term.

Images