Intelligent debris flow early warning method and system based on combination of physical model and algorithm

Abstract

The invention provides an intelligent debris flow early warning method and system based on a physical model and an algorithm, and belongs to the technical field of geological disaster monitoring, and the method comprises the steps: S1, carrying out an indoor soil physical mechanics test, building a physical model, and carrying out the parameter calibration of the physical model through combining a stability criterion equation of a rainfall infiltration-matrix suction coupling effect; s2, simulating different conditions, and generating and outputting debris flow occurrence process data under different working conditions by the physical model; s3, taking the generation process data and the real-time monitoring data as input, constructing a two-channel input network, and training a correction model to obtain corrected response data; s4, on the basis of an SHAP interpretability framework, constructing a Bayesian probability early warning model driven by multi-source heterogeneous data; and S5, triggering real-time early warning of the early warning platform according to the debris flow real-time occurrence probability of the monitoring point, and constructing a mapping relation between soil deformation and critical parameters through dynamic simulation of a physical model on a soil mechanical behavior.

Description

Technical Field

[0001] This specification relates to the technical field of geological disaster monitoring, and particularly relates to a debris flow intelligent early warning method and system combining physical models with algorithms. Background Art

[0002] Debris flow is a highly destructive geological disaster, and its early warning is of great significance for reducing disaster losses. The core pain points of current debris flow early warning technologies are as follows: 1. Single data dimension: Traditional methods rely on single-point monitoring data such as rain gauges and displacement sensors, and it is difficult to capture the dynamic changes of soil mechanical properties and the coupling effects of multiple physical fields; 2. Insufficient model generalization ability: Empirical models are limited by regional geological characteristics, and it is costly to relocate to a new monitoring area and re-collect historical disaster data; 3. Difficulty and high cost in obtaining high-fidelity data: Complex terrain makes it difficult to deploy monitoring equipment, and the data quality and coverage are limited, resulting in difficulty and high cost in obtaining high-fidelity monitoring data.

[0003] The existing Chinese patent with the publication number CN117829031A and the name of "Dynamics simulation method considering the interaction between runoff and debris flow" includes constructing a rainfall spatial distribution model; constructing a vegetation interception model; constructing a soil infiltration coupling model; determining rainfall data, vegetation intercepted rainwater data, and soil infiltrated rainwater data; determining the remaining rainwater data; obtaining the characteristic data of runoff and debris flow; constructing a two-layer depth-averaged model for the propagation of runoff and debris flow based on the water absorption rate parameter of debris flow on runoff; spatially discretizing the two-layer depth-averaged model to obtain a dynamics model; and simulating the behavior and interaction process of runoff and debris flow. The present invention considers the processes of rainfall, vegetation interception, soil infiltration, runoff generation, and debris flow propagation, proposes a two-layer depth-averaged model for describing the dynamics of runoff and debris flow, introduces a water absorption rate parameter, and uses the dynamics model to accurately and effectively simulate the behavior and interaction process of runoff and debris flow.

[0004] Although the above technology aims to improve the classification ability, it does not consider the physical mechanism constraints, resulting in a deviation between the algorithm output and the actual situation. Therefore, there is an urgent need for a new early warning model that integrates physical laws and data-driven algorithms.

[0005] Therefore, there is a need to provide a debris flow intelligent early warning method and system combining physical models with algorithms to achieve low-cost modeling driven by multi-source data. Summary of the Invention

[0006] The embodiments of this specification provide a debris flow intelligent early warning method and system combining physical models with algorithms. By dynamically simulating the mechanical behavior of soil through a physical model, a mapping relationship between soil deformation and critical parameters is constructed, significantly reducing the dependence on high-precision monitoring data.

[0007] In some embodiments, a debris flow intelligent early warning method combining physical models and algorithms includes: S1: Conduct indoor soil physical and mechanical tests, build a physical model, and calibrate the parameters of the physical model by combining the stability criterion equation of rainfall infiltration-matrix suction coupling effect. S2: Simulate different conditions, and the physical model generates and outputs the occurrence process data of debris flow under different working conditions. S3: Use the occurrence process data and real-time monitoring data as inputs, construct a dual-channel input network, train and correct the model to obtain the corrected response data. S4: Based on the SHAP interpretability framework, quantify the contribution degree of each index to the occurrence probability of debris flow, and construct a Bayesian probability early warning model driven by multi-source heterogeneous data. S5: Trigger the real-time early warning of the early warning platform according to the real-time occurrence probability of debris flow at the monitoring point.

[0008] Furthermore, in S1, establish a soil dynamic stability model based on the modified Mohr-Coulomb criterion, introduce the Bishop effective stress formula to characterize the influence of matrix suction on shear strength, combine the Richards equation to describe the rainfall infiltration process, calibrate the soil hydraulic parameters through triaxial tests, and construct a physical model.

[0009] Furthermore, in S1, the stability criterion equation of rainfall infiltration-matrix suction coupling effect includes: Where: is the effective cohesion of the soil; is the shear strength of the soil under effective stress; is the friction angle related to matrix suction; is the pore air pressure; is the pore water pressure; is the normal stress; is the soil unit weight; is the thickness of the sliding mass; is the inclination angle of the slip surface.

[0010] Furthermore, in S2, different conditions include different rainfall intensities, rainfall durations, and initial water contents, and the occurrence process data includes soil displacement changes, pore water pressure changes, and soil stability changes.

[0011] Further, in S3, the dual-channel input network includes a GAN adversarial generation network and an improved ConvLSTM network. The original monitoring data is sent as input to the improved ConvLSTM network for distribution. The physical model generates simulated data through the GAN adversarial generation network. By using the simulated data and the original monitoring data as input, and taking whether the slope in the original monitoring data is unstable as the training target, the correction model is trained. The response data output by the physical model is corrected by the trained correction model.

[0012] Further, in S5, after embedding the physical model into the early warning platform, by using the sensor data stream as the input of the Bayesian probability early warning model, the probability calculation result is output to trigger the corresponding level of early warning.

[0013] It also includes a debris flow intelligent early warning system combining a physical model and an algorithm. Further, it includes an early warning platform, a monitoring unit, and a physical model set on the platform. The multi-source real-time monitoring data is obtained through the monitoring unit. The multi-source real-time monitoring data is preprocessed by the physical model. By comparing the existing simulated data with the preprocessed data, the response data is output. The early warning platform includes a multi-index coupling early warning model. By quantifying the decision contribution degree of the multi-index coupling early warning model, the real-time occurrence probability of debris flow of the response data is calculated.

[0014] Further, the monitoring unit includes a GNSS ground displacement monitoring station, a MEMS micro-displacement sensor, and a fiber Bragg grating pore water pressure gauge.

[0015] Further, the preprocessing includes using the wavelet transform-Kalman filter joint algorithm to perform time-frequency domain denoising on the multi-source real-time monitoring data, and extracting key indicators through feature engineering. The key indicators include displacement acceleration and hydraulic gradient mutation points.

[0016] Further, the quantification of the decision contribution degree includes critical rainfall intensity, displacement acceleration inflection point, and safety factor decay rate.

[0017] The beneficial effects of the present invention are: 1. Achieve low-cost modeling driven by multi-source data: Through the dynamic simulation of the mechanical behavior of the soil body by the physical model, the mapping relationship between soil body deformation and critical parameters is constructed, significantly reducing the dependence on high-precision monitoring data. Combining with transfer learning technology, the model can achieve cross-regional generalization application based on limited samples, greatly reducing the data acquisition cost compared with traditional methods; 2. Establish an intelligent early warning system for the fusion mechanism: Adopt a dual-engine architecture of physical model and deep learning, where the physical engine is responsible for the numerical simulation of the soil stress-strain relationship, and the intelligent algorithm performs non-linear modeling of the multi-parameter coupling relationship. Through the dynamic safety factor correction module and displacement trend prediction unit, the accuracy of early warning of this architecture is much higher than that of traditional early warning models; 3. Form an adaptive risk prediction ability: The risk quantification module based on the improved Monte Carlo simulation can real-time analyze the joint probability distribution of parameters such as rainfall intensity, displacement acceleration, and pore water pressure, and early warn of potential instability risks through the failure mode recognition algorithm. The model supports multi-scale parameter migration and can still maintain a high prediction reliability in areas where monitoring data is missing. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] This specification will be further described in the form of exemplary embodiments, and these exemplary embodiments will be described in detail through the drawings. These embodiments are not restrictive. In these embodiments, the same numbers represent the same structures, where: Figure 1 is a schematic diagram of the working principle shown in some embodiments of this specification. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] To more clearly illustrate the technical solutions of the embodiments of this specification, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some examples or embodiments of this specification. For those of ordinary skill in the art, without creative efforts, this specification can also be applied to other similar scenarios based on these drawings. Unless obvious from the language context or otherwise stated, the same reference numerals in the drawings represent the same structures or operations.

[0020] It should be understood that the "system", "device", "unit" and / or "module" used herein is a method for distinguishing different components, elements, parts, portions or assemblies at different levels. However, if other words can achieve the same purpose, the said words can be replaced by other expressions.

[0021] As shown in this specification and the claims, unless the context clearly indicates an exception, words such as "a", "an", "one" and / or "the" are not specifically singular and may also include the plural. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of the clearly identified steps and elements, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements.

[0022] Flowcharts are used in this specification to illustrate the operations performed by the systems according to the embodiments of this specification. It should be understood that the preceding or subsequent operations are not necessarily executed precisely in sequence. On the contrary, the steps can be processed in reverse order or simultaneously. Also, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0023] Embodiment: Please refer to Figure 1 , for the target monitoring area, collect representative undisturbed soil samples and prepare them according to relevant specifications to ensure that the soil samples can truly reflect the physical and mechanical properties of the in-situ soil mass; conduct a series of indoor physical and mechanical tests, including but not limited to: basic physical property tests such as the density, water content, and particle size distribution of the soil mass, soil strength tests such as direct shear tests and triaxial compression tests, and permeability characteristic tests to analyze the impact of rainfall infiltration on soil mass stability.

[0024] Based on unsaturated soil mechanics theory, establish a stability criterion equation considering the coupling effect of rainfall infiltration - matrix suction, In the formula, is the effective cohesion of the soil mass, is the shear strength of the soil mass under effective stress, is the friction angle related to matrix suction, is the pore air pressure, is the pore water pressure, is the normal stress, is the unit weight of the soil mass, is the thickness of the sliding mass, is the inclination angle of the sliding surface.

[0025] It should be noted that using the above physical model, simulate the occurrence process of debris flow under the influence of different conditions (different rainfall intensities, rainfall durations, initial water contents, etc.), obtain the safety factor, displacement field, stress field, etc. of the soil mass, and output response data including soil mass displacement changes, pore water pressure changes, soil mass stability changes, etc.

[0026] Construct a dual-channel input network: The output data (displacement field, pore pressure field) of the physical model and the real-time monitoring data are input into the improved ConvLSTM network after standardization, and the feature weights are dynamically allocated through the attention mechanism; an adversarial generative network (GAN) is introduced to generate realistic soil instability sequence data to solve the problem of model overfitting in small-sample scenarios. The simulated data generated by the physical model and the actual monitoring data are used as inputs, and the result of whether the slope is unstable in the monitoring data is used as the training target to train and correct the model. The trained correction model is used to correct the response data generated by the physical model, so that the corrected data has the same accuracy and reliability as the measured data.

[0027] Based on the SHAP (Shapley Additive Explanations) interpretability framework, the contribution degrees of key indicators such as critical rainfall intensity (≥50mm / h), cumulative displacement (>10cm / 24h), and dynamic safety factor (threshold <1.05) to the probability of debris flow occurrence are systematically quantified, and the non-linear coupling mechanism of each indicator is revealed. Through the calculation of Shapley values, the marginal effects of different features in model prediction are clarified. For example, the contribution weight of rainfall intensity to sudden debris flow is higher than that of cumulative displacement, and the dynamic adjustment of the safety factor can reflect the critical state of soil instability. Combining the historical disaster case database and the geological mechanics expert knowledge base, a Bayesian probability early warning model driven by multi-source heterogeneous data is constructed, and its output result is the probability value of debris flow occurrence in the range of 0-1.

[0028] After the model is embedded in the real-time monitoring platform, the sensor data stream is dynamically parsed through the edge computing module, and the index status and probability calculation results are updated every 5 minutes. A three-level early warning mechanism is adopted: when the probability value P≥0.7, a red early warning (evacuate immediately) is triggered; when 0.5≤P<0.7, an orange early warning (control of key areas) is launched; when 0.3≤P<0.5, a yellow early warning (strengthen patrol) is issued. The model is built with a feedback learning mechanism that can optimize the feature weight parameters using the actual occurrence of each early warning event, achieving adaptive iteration of the early warning threshold.

[0029] This design emphasizes the collaborative optimization of physical mechanisms and data-driven approaches. The SHAP interpretability module not only verifies the scientificity of the index system but also generates a visual attribution report, providing a traceable quantitative basis for emergency decision-making.

[0030] It should be noted that under the framework of unsaturated soil mechanics, a dynamic stability model of soil mass based on the modified Mohr-Coulomb criterion is established. The Bishop effective stress formula is introduced to characterize the influence of matrix suction on shear strength, and the Richards equation is combined to describe the rainfall infiltration process. The hydraulic parameters of the soil (such as soil-water characteristic curve, permeability coefficient function) are calibrated through triaxial tests, and a multi-physics field coupling dynamic analysis model of soil stability including pore water pressure gradient and displacement field evolution is constructed. By coupling the pore water pressure evolution equation and the stress-strain relationship of the soil mass, the instability thresholds of the soil mass under different rainfall intensities and infiltration paths are simulated, and a multi-dimensional data set including rainfall spatio-temporal distribution, displacement field evolution, pore water pressure gradient and safety factor matrix is generated. At the same time, a transfer learning framework (adversarial neural network DANN) is introduced, and cross-domain mapping is carried out between the data generated by the model and the field monitoring data to improve the generalization ability of the model.

[0031] A multi-source sensing network (GNSS surface displacement monitoring station, MEMS micro-displacement sensor, fiber Bragg grating pore water pressure gauge) is used to achieve centimeter-level deformation monitoring and minute-level data transmission, and the InSAR satellite remote sensing is combined to invert the regional surface deformation field. The wavelet transform-Kalman filter joint algorithm is used to denoise the monitoring data in the time-frequency domain, and key indicators such as displacement acceleration and hydraulic gradient mutation points are extracted through feature engineering. A spatial data alignment model based on weighted regression (GWR) is established to unify the spatio-temporal benchmarks of physical simulation data and multi-source monitoring data.

[0032] A physics-guided neural network architecture (PINN) is constructed, and the soil constitutive equation is embedded into the LSTM network as a regularization term, and the simulation data and the monitoring data stream are synchronously processed through a dual-channel input mechanism. The Bayesian deep learning framework is used to quantify the epistemic uncertainty in the data correction process, and the Monte Carlo Dropout strategy is used to dynamically evaluate the correction confidence. A hybrid loss function is designed, comprehensively considering the mean square error (MSE), dynamic time warping (DTW) and physical conservation constraints, to ensure that the correction results satisfy both data matching and mechanical laws.

[0033] A multi-index coupling early warning model is constructed based on the SHAP-XGBoost framework to quantify the decision contribution degrees of indicators such as critical rainfall intensity (≥50mm / h), displacement acceleration inflection point (second derivative > 0.05cm / h²), safety factor decay rate (ΔSF / Δt < -0.1h⁻¹). The federated learning mechanism is used to realize the collaborative update of the parameters of the regional sub-model and the central model, and support the local inference of distributed edge nodes (such as on-site RTU controllers). A lightweight early warning engine is deployed, integrating an online learning module, which automatically triggers the incremental update of the model when the monitoring data deviates from the training distribution, ensuring the continuous reliability of the system in a non-stationary environment.

[0034] In summary, an intelligent early warning system integrating mechanisms is established: adopting a dual-engine architecture of physical models and deep learning, where the physical engine is responsible for the numerical simulation of the soil stress-strain relationship, and the intelligent algorithm conducts non-linear modeling of the multi-parameter coupling relationship. Through the dynamic safety factor correction module and the displacement trend prediction unit, the accuracy of this architecture's early warning is much higher than that of traditional early warning models.

Claims

1. A debris flow intelligent early warning method combining physical model with algorithm, characterized in that: include, S1: Conduct indoor soil physical and mechanical tests, build a physical model, and calibrate the parameters of the physical model based on the stability criterion equation of the rainfall infiltration-matrix suction coupling effect; S2: Simulate different conditions, and the physical model generates and outputs the data of debris flow occurrence process under different working conditions; S3: Take the process data and real-time monitoring data as input, build a dual-channel input network, train the correction model, and obtain the corrected response data; S4: Based on the SHAP interpretability framework, the contribution of each indicator to the probability of debris flow is quantified, and a Bayesian probabilistic early warning model driven by multi-source heterogeneous data is constructed; S5: Trigger the real-time warning of the early warning platform based on the real-time probability of debris flow at the monitoring point.

2. The intelligent early warning method for debris flow combining physical model and algorithm as claimed in claim 1, characterized in that: In S1, a soil dynamic stability model based on the modified Mohr-Coulomb criterion is established, the Bishop effective stress formula is introduced to characterize the influence of matrix suction on shear strength, the Richards equation is combined to describe the rainfall infiltration process, the soil hydraulic parameters are calibrated through triaxial tests, and a physical model is constructed.

3. The intelligent early warning method for debris flow combining physical model and algorithm as claimed in claim 2, characterized in that: In S1, the stability criterion equation for the rainfall infiltration-matrix suction coupling effect includes: In the formula, is the effective cohesion of soil, is the shear strength of soil under effective stress, is the friction angle related to matrix suction, is the pore gas pressure, is the pore water pressure, is the normal stress, is the soil bulk density, is the thickness of the sliding body, is the sliding surface inclination angle.

4. The intelligent early warning method for debris flow combining physical model and algorithm as claimed in claim 3, characterized in that: The different conditions in S2 include different rainfall intensities, rainfall durations, and initial moisture contents, and the occurrence process data include changes in soil displacement, pore water pressure, and soil stability.

5. The intelligent early warning method for debris flow combining physical model and algorithm as claimed in claim 4, characterized in that: In S3, the dual-channel input network includes a GAN adversarial generation network and a ConvLSTM improved network. The original monitoring data is sent as input to the ConvLSTM improved network for distribution. The physical model generates simulation data through the GAN adversarial generation network. The simulation data and the original monitoring data are used as input, and whether the slope in the original monitoring data is unstable is used as a training target to train the correction model. The response data output by the physical model is corrected by the trained correction model.

6. The intelligent early warning method for debris flow combining physical model and algorithm as claimed in claim 5, characterized in that: In S5, after the physical model is embedded in the early warning platform, the sensor data stream is used as the input of the Bayesian probabilistic early warning model to output the probability calculation result and trigger the corresponding level of early warning.

7. An intelligent debris flow warning system combining physical model and algorithm, characterized in that: It includes an early warning platform, a monitoring unit and a physical model installed on the platform. Acquire multi-source real-time monitoring data through monitoring units; Preprocess multi-source real-time monitoring data through physical models, and output response data by comparing existing simulation data with preprocessed data; The early warning platform includes a multi-indicator coupling early warning model, and calculates the real-time probability of debris flow occurrence of the response data by quantifying the decision contribution of the multi-indicator coupling early warning model.

8. The intelligent debris flow early warning system combining physical model and algorithm as claimed in claim 7, characterized in that: The monitoring unit includes a GNSS surface displacement monitoring station, a MEMS micro-displacement sensor, and a fiber grating pore water pressure gauge.

9. The intelligent debris flow early warning system combining physical model and algorithm as claimed in claim 7, characterized in that: Preprocessing includes using the wavelet transform-Kalman filter joint algorithm to denoise the multi-source real-time monitoring data in the time and frequency domain, and extracting key indicators through feature engineering. The key indicators include displacement acceleration and hydraulic gradient mutation points.

10. The intelligent debris flow early warning system combining physical model and algorithm as claimed in claim 7, characterized in that: The quantitative decision contribution includes critical rainfall intensity, displacement acceleration inflection point, and safety factor attenuation rate.

Images