An intelligent early warning and response method for slope instability risk

Abstract

This invention provides an intelligent early warning and response method for slope instability risk, belonging to the field of intelligent early warning technology based on data analysis. First, multidimensional environmental geological data is collected and a dual-label mechanism of instability type and spatiotemporal evolution chain is introduced to form a dynamic data set. Subsequently, through slope disturbance sensitivity mapping modeling, key influencing factors and their weights are determined, enabling adaptive optimization of input feature selection and model structure. A deep learning model integrating a multi-scale convolutional neural network with an attention mechanism is used to extract multi-layer spatiotemporal features and dynamically adjust model parameters to accurately predict the probability of slope instability and its evolution path. Finally, by combining evolutionary game theory, risk level, response timeliness, and governance cost are incorporated into multi-objective optimization to generate an optimal intelligent response strategy. This invention achieves dynamic perception, sensitive modeling, intelligent prediction, and scientific response to slope instability risk, effectively improving prediction accuracy and emergency response efficiency.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent early warning based on data analysis, and in particular relates to an intelligent early warning and response method for slope instability risks. Background Art

[0002] With the frequent construction of infrastructure in mountainous areas and the occurrence of extreme climate events, the frequency of slope instability disasters is increasing, seriously threatening the safety of people's lives and property. Traditional slope monitoring and early warning methods mostly rely on manual experience interpretation or single indicator threshold judgment, which is difficult to cope with the actual scenario where geological and environmental factors are highly coupled and the risk evolution path is complex and changeable. As a result, early warnings are not timely, accurate, and the response is not precise. At present, the main methods for predicting slope instability risks include the following methods:

[0003] (1) Slope early warning method based on threshold judgment: This type of method sets an empirical threshold for slope monitoring data (such as displacement, rainfall, groundwater level, etc.), and triggers an early warning when the real-time data exceeds the threshold. For example, when the daily rainfall exceeds a certain critical value, or the horizontal displacement of the monitoring point exceeds the set limit, the system considers that the slope is at risk of instability. This type of method has a simple principle and is easy to deploy, and has been applied in some engineering projects. However, its disadvantage is that it relies on static experience and cannot adapt to different slope conditions and dynamic disturbances. It is easy to generate false alarms or missed alarms and it is difficult to reveal the intrinsic relationship between the evolution mechanism of instability and the influencing factors;

[0004] (2) Slope prediction methods based on numerical simulation of physical mechanisms: This type of method usually uses numerical simulation methods such as finite element, discrete element or fluid-solid coupling to establish a numerical model based on the geological structure and physical parameters of the slope, and simulate its instability process under the action of rain, earthquake, etc., so as to achieve slope stability analysis. Although the simulation accuracy is high and can provide support for the analysis of instability mechanisms, it has problems such as complex modeling, strong parameter dependence, and poor real-time performance, making it difficult to widely promote it to online risk monitoring and early warning scenarios for multiple scenarios and multiple slopes;

[0005] (3) Slope early warning methods based on deep learning: Some technologies use deep learning models such as neural networks, LSTM, and CNN to model slope monitoring data, attempting to predict slope status and identify instability. This type of method has certain nonlinear modeling capabilities, can handle complex time series features, and has strong self-learning capabilities. However, existing methods generally have the problem of fixed model structure and inability to dynamically optimize for different scenarios. They also do not fully consider the response differences and factor sensitivity of disturbance events, which affects the robustness and generalization performance of the model and makes it difficult to meet the needs of accurate early warning in high-risk scenarios.

[0006] Therefore, existing methods have obvious shortcomings in real-time, accuracy, adaptability and interpretability, and are unable to perform accurate early warning tasks for slopes in dynamic disturbance and multi-source driving environments. Summary of the Invention

[0007] To address the above issues, the present invention proposes an intelligent early warning and response method for slope instability risk, which includes the following steps:

[0008] S1, based on a data set containing key geological data and environmental data, through slope disturbance sensitivity mapping modeling, introduces parameter perturbation to simulate the potential anomaly evolution process, and calculates the mask vector of geological data, the mask vector of environmental data and the structural path control coefficient;

[0009] S2, construct an adaptive slope instability prediction model, and use the dataset and structure path control coefficient to optimize and train the model;

[0010] S3, real-time acquisition of multiple key geological data and environmental data to construct multi-source heterogeneous time series data, combined with the mask vector calculated in S1, input into the trained adaptive slope instability prediction model, and output the instability type prediction probability vector and the dual-factor evolution chain prediction result. The dual-factor evolution chain prediction result includes the position detection point evolution sequence prediction result and the time evolution node sequence prediction result;

[0011] S4, based on S3, obtain the predicted probability vector of the instability type and select the instability type with the highest probability as the warning type of the current risk of the slope;

[0012] At the same time, based on the calculation results obtained by S3, the number of monitoring points with rapid instability is calculated and used as the length of the evolution chain. Then, the instability speed of each unstable monitoring point is calculated to form the evolutionary situation state vector. Combined with the defined multiple types of response strategies, it is input into the constructed multi-objective game model, and iterative evolutionary reasoning is performed through Monte Carlo evolutionary game to determine the optimal response strategy under the current risk situation.

[0013] Furthermore, the geological data includes the horizontal displacement of the ground surface detected by GNSS fixed point detection. ; Underwater level data collected using an automatic water level gauge ; Width of landslide cracks collected using linear displacement meters ; and the internal seepage water pressure of the slope collected by the piezometer ;

[0014] The environmental data includes precipitation collected by automatic rain gauges and duration of precipitation ; Temperature collected by weather stations and air humidity ; Earthquake acceleration collected by seismograph .

[0015] Furthermore, the dataset for training the adaptive slope instability prediction model is constructed in the following ways:

[0016] For the slope to be tested, uniform selection of monitoring points, and collect geological data of each monitoring point at the same time; directly collect geological data for the slope to be tested. Environmental data at each point in time;

[0017] The corresponding instability type prediction probability vector labels are divided into no instability, tension instability, shear instability, fluid-plastic instability and combined instability;

[0018] The corresponding two-factor evolution chain prediction label is set as a sequence of position detection points and a sequence of time evolution points;

[0019] The evolution sequence of the position detection points is set by analyzing the abnormal response degree of the geological parameters of each monitoring point before instability, screening out the monitoring point sequence with significant anomalies and triggering instability signs one after another, and forming the position instability propagation path. ,in, Indicates the location of the first monitoring point that triggers the instability sign. Similarly, Indicates the The location of the monitoring point that triggers the instability signs, and ;

[0020] The time evolution node sequence is set based on the position instability propagation path , determine the time point when each monitoring point triggers the instability sign as the important stage node of the instability evolution, and construct the evolution path of the time dimension :in, Indicates the first monitoring time point that triggers the instability sign. Similarly, Indicates the A monitoring time point that triggers signs of instability, and .

[0021] Furthermore, the specific process of S1 is as follows:

[0022] S11, for geological data and input environmental data , for each geological parameter and environmental parameter, the size is Add a small disturbance and repeat the operation times, and obtain the geological data set after adding disturbance and environmental datasets ;in, is the disturbance amplitude coefficient, is the standard deviation of the corresponding factor in the original data;

[0023] S12, based on the constructed multi-source dynamic dataset of slope risk, the geological data and environmental data are spliced ​​together and clustered using the K-means algorithm. The number of clusters is preset to Q, and Q initial cluster centers are obtained. ; Afterwards, the disturbed geological dataset will be added and environmental datasets Add to the data sample to be clustered and get Q new cluster centers ; At the same time, we get the The classification result set of data samples ;

[0024] S13, based on the degree of category drift before and after the perturbation sample clustering, factor sensitivity analysis is performed and calculated one by one After the perturbation is added, the cluster label of the perturbation sample changes. The proportion of cluster label changes in the perturbation sample corresponding to different factors is counted and defined as the sensitivity score of the factor. , Indicates the The sensitivity of the factors; reconstruct the first p factors into a mask vector of geological data , reconstruct other factors into mask vectors of environmental data ;

[0025] S14, based on the initial cluster center And the cluster centers after adding disturbance , the overall offset of the cluster center is calculated using Euclidean distance: ; Then, combine the obtained geological mask vector and the environment mask vector , calculate the structural control coefficients of the geological data pathway and the environmental data pathway respectively and .

[0026] Furthermore, the adaptive slope instability prediction model in S2 includes a geological structure evolution extraction module, an environmental driving feature modeling module and a slope stability evolution discrimination module;

[0027] The mass structure evolution extraction module will Vector with geological mask Perform inner product to obtain geological mask features ; Then, through multiple chained TCN multi-scale convolution blocks The geological feature evolution is extracted to obtain the deep feature representation of the geological structure evolving over time; the output of each TCN convolution block is used as the input of the next TCN convolution block, thus forming a progressive feature extraction path; finally, The outputs of the TCN multi-scale convolutional blocks are fused through a multi-head attention layer to obtain stable and discriminative geological evolution features. ;

[0028] The environment-driven feature modeling module uses environmental data As input, with the environment mask vector Perform inner product to obtain geological mask features ; Then through multiple chained ESN network layers Extract environmental features, capture the dynamic changes of environmental driving factors in the time dimension, and obtain The output features of each ESN layer are used as the input of the next layer, and finally, the multi-head attention fusion The output features of the ESN layer are used to obtain a stable representation of the dynamic characteristics of the environment. ;

[0029] The slope stability evolution discrimination module will output the characteristics and output features Splicing is performed to obtain the comprehensive feature vector ; Then, the feature dimension is reduced through two fully connected layers, and the ReLU activation function is used to enhance the nonlinear expression ability to obtain the discriminant feature ;

[0030] Finally, Input BatchNorm layer and Softmax activation function to get the instability type prediction probability vector ;Will Input the Dropout layer and Sigmoid activation function to obtain the two-factor evolution chain prediction result .

[0031] Furthermore, the TCN multi-scale convolution block sets parallel convolution kernels with receptive fields of 1×1, 3×3, and 5×5, respectively, to extract characteristic response patterns at local and global spatial scales, thereby coping with the heterogeneity and hierarchy of geological data at different spatial scales. Subsequently, through the ReLU activation function and channel splicing operations, a feature embedding representation with rich spatial perception capabilities is finally generated.

[0032] Furthermore, the Through the structural control coefficient Adaptive determination: ,in, To preset the minimum number of TCN multi-scale convolution blocks, The number of buffers for the preset adaptive TCN multi-scale convolution blocks;

[0033] described Through the structural control coefficient Calculation yields: ,in, To preset the minimum ESN layer, The number of preset adaptive ESN layer buffers.

[0034] Furthermore, the function of the multi-objective game model in S4 is:

[0035]

[0036] in, Indicates the adoption of a strategy The profit value at To adopt a strategy The decrease in the probability of instability after To adopt a strategy The number of instability monitoring points decreased after To adopt a strategy The acceleration drop value of the instability monitoring point after and Respectively indicate the adoption of strategies the subsequent costs and time to effect; is the preset weight; the ultimate goal of the multi-objective game is to maximize .

[0037] Furthermore, four optional response strategies are predefined: early warning strategy , Engineering disposal strategy , Traffic Strategy , Engineering disposal strategy ; then the response policy set , and each response strategy corresponds to a system behavior, and the game goal is to choose the optimal strategy To guide the system evolution towards a low-risk state.

[0038] Furthermore, the dynamic adjustment of system state and strategy selection is achieved through iterative evolutionary reasoning using Monte Carlo evolutionary games, including:

[0039] Calculate the multi-objective benefit function value of each alternative response strategy based on the current risk situation state vector;

[0040] Gradually adjust the strategy distribution through the evolutionary selection mechanism to strengthen the strategy tendency of guiding the system evolution to a low-risk state;

[0041] Dynamically adjust weight parameters and response strategy space to ensure that the multi-objective optimization process converges to the global optimal solution;

[0042] Ultimately determine the optimal response strategy under the current risk situation .

[0043] Compared with the prior art, the innovative features of the present invention have the following beneficial effects:

[0044] (1) Slope instability data modeling method integrating “double-factor evolution chain”: Incorporating “location detection point sequence” and “time evolution point sequence” into the labeling mechanism, the spatial propagation path and temporal evolution process of slope instability are fully characterized, significantly improving the structured and dynamic modeling capabilities of the dataset;

[0045] (2) Perturbation-driven factor sensitivity mapping and mask generation mechanism: Without relying on the prediction model, perturbation injection and response clustering techniques are introduced, and a generation mechanism of mask vectors and structural control coefficients is proposed for adaptive optimization of model structure, thereby improving factor interpretability and scientific input selection;

[0046] (3) Adaptive prediction model integrating multi-scale convolution and attention mechanism: The temporal convolutional network (TCN), echo state network (ESN) and attention mechanism are integrated to achieve hierarchical modeling based on the spatiotemporal heterogeneity of slope risk, and the structure is dynamically adjusted in combination with factor sensitivity to improve prediction accuracy and model robustness;

[0047] (4) Response strategy generation framework based on evolutionary game: By introducing a dynamic game reasoning module, risk prediction results and governance costs are incorporated into multi-objective optimization to achieve automated decision-making and optimal selection of slope risk response plans. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a flow chart of the overall technical route of the present invention.

[0049] Figure 2 This is a network architecture diagram of the adaptive slope instability prediction model of the present invention.

[0050] Figure 3 This is the network structure diagram of the TCN multi-scale convolutional block of the present invention.

[0051] Figure 4 This is the probability distribution thermal map of slope instability types of the present invention.

[0052] Figure 5 The figure shows the comparison results of the three models in the prediction accuracy of instability types.

[0053] Figure 6 The figure shows the comparison of the accuracy of the three models in predicting spatial evolution chains.

[0054] Figure 7 The figure shows the comparison of the accuracy of the three models in time evolution chain prediction. DETAILED DESCRIPTION

[0055] This invention proposes an intelligent early warning method for slope instability risk. The overall technical route is as follows: Figure 1 As shown:

[0056] Construction of a multi-source dynamic dataset for slope risk: This involves collecting and integrating multi-source heterogeneous time series data, including geological monitoring data and environmental monitoring data, and adding data labels, including slope instability types and a two-factor evolution chain containing detection points and time points, to establish a multi-source dynamic dataset for slope risk.

[0057] Slope disturbance sensitivity mapping modeling: Based on key geological and environmental parameters collected in the constructed dataset, parameter perturbations are introduced to simulate the evolution of potential anomalies, quantitatively assessing the impact of various factors on the slope instability response. The resulting output is a mask vector and structural control coefficient, which are used to guide the input variable masking and structural path adjustment of the adaptive slope instability prediction model.

[0058] Adaptive Slope Instability Prediction Model Construction: This model integrates multi-scale convolution and attention mechanisms. The model's input feature selection and network structure are optimized based on the output sensitivity factors (output mask vectors and structural control coefficients), thereby improving the model's prediction accuracy and response robustness in highly sensitive areas. The model is trained using a constructed dataset, ultimately outputting multidimensional prediction results, including instability probability and evolution chains.

[0059] Risk level game reasoning and response strategy generation: Based on the prediction results, a multi-objective game reasoning module is constructed that integrates factors such as instability risk and governance cost. The optimal risk response strategy is comprehensively calculated through the dynamic game evolution process, and ultimately the generation of slope instability risk response plans based on intelligent reasoning is achieved.

[0060] The invention will be further described below with reference to specific embodiments.

[0061] 1. Construction of a multi-source dynamic dataset of slope risk

[0062] This paper aims to achieve slope instability risk early warning based on multi-source heterogeneous slope data. Therefore, a multi-source dynamic dataset of slope risk is first constructed. The input of this dataset includes geological and environmental monitoring data. The output includes the slope instability type and a two-factor evolution chain consisting of detection points and time points. The specific contents include:

[0063] 1. Input parameter selection: Since the slope geological data and environmental data directly reflect the slope instability, the present invention selects geological data and environmental data as input parameters respectively;

[0064] Specifically, geological data include: surface horizontal displacement detected by GNSS fixed points ; Underwater level data collected using an automatic water level gauge ; Width of landslide cracks collected using linear displacement meters ; and the internal seepage water pressure of the slope collected by the piezometer ; Therefore, the geological data parameter combination is .

[0065] 2. Environmental data include: precipitation collected by automatic rain gauges and duration of precipitation ; Temperature collected by weather stations and air humidity ; Earthquake acceleration collected by seismograph ; Therefore, the environmental data parameter combination is .

[0066] Multi-source heterogeneous input data collection: Since the geological data at different locations on the slope are different, and the corresponding geological parameters will change dynamically when the slope is unstable; while the environmental data at different locations are relatively stable; therefore, the present invention collects geological data from multiple monitoring points and overall environmental data; specifically including:

[0067] (1) Input geological data collection: uniformly select the common Monitoring points, collect geological data of each monitoring point at the same time, and obtain the Geological data collected from multiple monitoring points , Indicates in The geological data collected at each monitoring point, and ; Afterwards, in order to capture the dynamic changes of geological data, a total of The geological data of multiple monitoring points at each time point are obtained, and the complete input geological data is finally obtained. ,in ;

[0068] (2) Input environmental data collection: Since environmental data is relatively stable, it is directly collected for the slope to be tested. Environmental data at each time point, to obtain complete input environmental data ,in, Indicates at the time point The collected slope environmental data, and .

[0069] 3. Slope instability type label setting: Slope instability type is an important label parameter for slope risk assessment, which can effectively characterize the failure mode and evolution trend of the slope driven by different mechanisms. Based on existing geological engineering practice experience and historical instability case data, this paper divides slope instability labels into five categories: (1) no instability; (2) tensile instability; (3) shear instability; (4) plastic instability; and (5) composite instability. The corresponding instability type is determined based on the subsequent slope monitoring results. Finally, an instability type label corresponding to the input sample is generated. .

[0070] 4. Dual-factor evolution chain label setting: To comprehensively reflect the evolution path and risk accumulation mechanism during the slope instability process, this paper further proposes a dual-factor evolution chain label mechanism based on "position detection point sequence" and "time evolution point sequence", specifically including:

[0071] (1) Setting the evolution sequence of position detection points: Professional technicians analyze the abnormal response degree of geological parameters at each monitoring point before instability, select the monitoring point sequence with significant anomalies and triggering instability signs one after another, and form the position instability propagation path. ,in, Indicates the location of the first monitoring point that triggers the instability sign. Similarly, Indicates the The location of the last monitoring point that triggers the instability sign, and .

[0072] (2) Time evolution node sequence setting: based on the position instability propagation path , determine the time point when each monitoring point triggers the instability sign as the important stage node of the instability evolution, and construct the evolution path of the time dimension :in, Indicates the first monitoring time point that triggers the instability sign. Similarly, Indicates the The last monitoring time point that triggers the instability sign, and ;

[0073] Finally, the two-factor evolution chain label is set to .

[0074] 5. Complete data set construction: collect the input geological data and input environmental data As the input of the entire dataset; and the constructed instability type label And the constructed dual-factor evolution chain label As output; based on this principle, we collect and build The two sets of data sets together constitute a multi-source dynamic data set of slope risk.

[0075] 2. Modeling of Slope Disturbance Sensitivity Mapping

[0076] Without building a complete risk prediction model, this paper proposes a model-free sensitivity modeling method based on disturbance injection and response clustering to obtain mask vectors and structural control coefficients, and to adaptively adjust the model architecture. The method mainly includes the following steps:

[0077] Perturbation sample construction: for the data to be tested, including input geological data and input environmental data , for each geological parameter and environmental parameter, the size is Add a small disturbance and repeat the operation times, and obtain the geological data set after adding disturbance and environmental datasets ;in, is the disturbance amplitude coefficient, is the standard deviation of the corresponding factor in the original data.

[0078] Cluster analysis: Based on the constructed multi-source dynamic dataset of slope risk, the geological data and environmental data were spliced ​​together and clustered using the K-means algorithm. The number of clusters was preset to 5, and 5 initial cluster centers were obtained. ; Afterwards, the disturbed geological dataset will be added and environmental datasets Add to the data sample to be clustered and get 5 new cluster centers ; At the same time, we get the The classification result set of data samples .

[0079] Geological-environmental factor sensitivity modeling: Factor sensitivity analysis is performed based on the degree of category drift before and after the disturbance sample clustering. Specifically, the factor sensitivity is calculated one by one. After the perturbation is added, the cluster label of the perturbation sample changes. The proportion of cluster label changes in the perturbation sample corresponding to different factors is counted and defined as the sensitivity score of the factor. , Indicates the The sensitivity of the factors, and .

[0080] The higher the sensitivity, the higher the impact of the factor on slope risk warning; Mask vector reconstructed into geological data ,Will Reconstructed mask vector of environment data , which is used to select important geological and environmental variables in advance and access the model in subsequent models, thereby achieving shallow input optimization.

[0081] Structural path parameter coefficient acquisition: According to the initial cluster center And the cluster centers after adding disturbance , the overall offset of the cluster center is calculated using Euclidean distance: ; Then, combine the obtained geological mask vector and the environment mask vector , calculate the structural control coefficients of the geological data pathway and the environmental data pathway respectively:

[0082]

[0083]

[0084] in, and represent the structural control coefficients of the geological data pathway and the environmental data pathway, The first one represents the geological mask vector mask parameters, represents the environment mask vector mask parameters; represents the regulatory weight of factor sensitivity to pathway structure, represents the regulatory weight of cluster structure deviation on pathway structure; is the normalization function;

[0085] final, and It will be used as the control coefficient of the geological channel and environmental channel network structure in the control network model, and will be used for the subsequent configuration of the geological and environmental input network structure, thereby achieving adaptive optimization at the structural level.

[0086] 3. Construction of adaptive slope instability prediction model

[0087] In order to achieve rapid and stable detection of slope instability, the present invention designs an adaptive slope instability prediction model for complex scenarios; the input of the model is geological data and environmental data , the output is the predicted probability of instability type And the prediction results of the two-factor evolution chain .

[0088] Furthermore, the geological mask vector obtained by and the environment mask vector It can effectively reflect the sensitivity of different input factors and use the structural control coefficient and Dynamically adjust the network structure of the model to match the complexity of the actual scene; Adaptive slope instability prediction model such as Figure 2 As shown in the figure, it specifically includes a geological structure evolution extraction module, an environmental driving feature modeling module and a slope stability evolution discrimination module.

[0089] Design of geological structure evolution extraction module: This module is based on geological data As input, the input factors are firstly weighted and screened through the geological mask layer to obtain the significant feature subset; specifically, Vector with geological mask Perform inner product to obtain geological mask features ; Then, through multiple chained TCN multi-scale convolution blocks Perform geological feature evolution extraction to obtain a deep feature representation of the geological structure evolving over time; i.e. The output features of a TCN multi-scale convolution block are ; Moreover, the output of each TCN convolution block is used as the input of the next TCN convolution block, thus forming a progressive feature extraction path; finally, since the multi-head attention layer can effectively deal with variable input features (i.e. Dynamically changing) scenes, so The outputs of the TCN multi-scale convolutional blocks are fused through a multi-head attention layer to obtain stable and discriminative geological evolution features. ;

[0090] at the same time, Through the structural control coefficient Adaptive determination: ,in, To preset the minimum number of TCN multi-scale convolution blocks, this embodiment sets , The number of adaptive TCN multi-scale convolution block buffers is preset so that the model can adaptively control the modeling depth to match the complexity of geological disturbance characteristics.

[0091] In addition, the network structure of TCN multi-scale convolutional block is as follows Figure 3 As shown in the figure, the module sets parallel convolution kernels with receptive fields of 1×1, 3×3, and 5×5 respectively to extract characteristic response patterns at local and global spatial scales, thereby coping with the heterogeneity and hierarchy of geological data at spatial scales; then, through the ReLU activation function and channel splicing operations, it finally generates a feature embedding representation with rich spatial perception capabilities.

[0092] Design of environment-driven feature modeling module: This module uses environmental data As input, firstly, the sensitive environmental features are extracted through the environmental mask layer; the specific method is similar to the above, with the environment mask vector Perform inner product to obtain geological mask features ; Then through multiple chained ESN network layers Extract environmental features, capture the dynamic changes of environmental driving factors in the time dimension, and obtain The output features of the ESN layer are , the output of each layer of ESN is used as the input of the next layer, and finally; through multi-head attention fusion The output features of the ESN layer are used to obtain a stable representation of the dynamic characteristics of the environment. ;

[0093] at the same time, Through the structural control coefficient Calculation yields: ,in, To preset the minimum ESN level, also set , The number of adaptive ESN layer buffers is preset to adaptively adjust the network complexity according to the sensitivity of environmental disturbances, so that the model has good generalization ability when dealing with both sudden and gradual environmental changes.

[0094] Slope stability evolution discrimination module design: Afterwards, the output features and output features Splicing is performed to obtain the comprehensive feature vector ; Then, the feature dimension is reduced through two fully connected layers, and the ReLU activation function is used to enhance the nonlinear expression ability to obtain the discriminant feature ;

[0095] Finally, Input BatchNorm layer and Softmax activation function to get the instability type prediction probability vector ;Will Input the Dropout layer and Sigmoid activation function to obtain the two-factor evolution chain prediction result ,and Including position detection point evolution sequence prediction results , and the time evolution node sequence prediction results .

[0096] Model training: Based on the constructed multi-source dynamic dataset of slope risk, the cross entropy loss and mean absolute error function are used to predict the instability type. and the evolution chain prediction results The optimization is performed; and the training process adopts the stochastic gradient descent method. The training is terminated after reaching the preset maximum number of iterations, and a trained adaptive slope instability prediction model is obtained.

[0097] 4. Risk Level Game Reasoning and Response Strategy Generation

[0098] After completing the slope instability probability prediction and evolution path output, in order to further achieve efficient response to slope instability, this paper proposes a risk level reasoning and response strategy generation mechanism based on multi-objective evolutionary game, which specifically includes the following steps:

[0099] Constructing risk situation state space: According to the adaptive slope instability prediction model, the instability type prediction probability vector can be obtained , and the prediction results of the two-factor evolution chain ; and through Position detection point evolution sequence prediction results in And the time evolution node sequence prediction results , calculate the number of monitoring points with rapid instability and use it as the length of the evolution chain , and calculate the instability velocity of each instability monitoring point at the same time , which ultimately constitutes the evolutionary state vector .

[0100] Response strategy space definition: predefined 4 types of optional corresponding strategy sets: early warning strategy (such as issuing red, yellow and blue warning information); engineering disposal strategy (such as reinforcement, support, shotcrete); traffic strategy (such as closing roads or tunnels near slopes); engineering disposal strategies (such as encrypted monitoring sampling, extended perception link); then the response strategy set , and each strategy combination corresponds to a system behavior, the game goal is to choose the optimal strategy To guide the system evolution towards a low-risk state.

[0101] Multi-objective game modeling: The multi-objective response decision function of the construction system is:

[0102]

[0103] in, Indicates the adoption of a strategy The profit value at To adopt a strategy The decrease in the probability of instability after To adopt a strategy The number of instability monitoring points decreased after To adopt a strategy The acceleration drop value of the instability monitoring point after and Respectively indicate the adoption of strategies The subsequent cost and time to effect; is the preset weight; therefore, the ultimate goal of the multi-objective game is to maximize .

[0104] Game reasoning: Based on constructed multi-objective game modeling, iterative evolutionary reasoning using Monte Carlo evolutionary games is used to achieve dynamic adjustment of system status and strategy selection; specifically, it includes:

[0105] (1) Calculate the multi-objective benefit function value of each alternative response strategy based on the current risk situation state vector;

[0106] (2) Gradually adjust the strategy distribution through the evolutionary selection mechanism to strengthen the strategy tendency of guiding the system evolution to a low-risk state;

[0107] (3) Dynamically adjust the weight parameters and response strategy space to ensure that the multi-objective optimization process converges to the global optimal solution, thereby ensuring the effectiveness and applicability of the strategy;

[0108] (4) Ultimately determine the optimal response strategy under the current risk situation .

[0109] The warning result output and response strategy are generated synchronously: Based on the instability type prediction probability vector , the instability type with the highest probability is selected as the warning type of the current risk of the slope, including no instability, tension instability, shear instability, flow-plastic instability and combined instability; at the same time, the optimal response strategy obtained based on game reasoning ; and push the early warning results and response strategies to the risk management platform and emergency response units simultaneously, thereby effectively improving the scientificity and efficiency of slope disaster prevention and control.

[0110] 5. Comparative Experimental Design and Result Analysis of Slope Instability Early Warning Methods

[0111] In order to verify the effectiveness and superiority of the adaptive slope instability prediction model proposed in this paper (hereinafter referred to as "this model"), two representative comparison models were used: (1) LSTM model: using the traditional long short-term memory network to perform multi-category probability prediction of slope instability types; (2) traditional machine learning model: using random forest (RF) as the benchmark model to perform predictions on the same task.

[0112] The experiment mainly consists of two parts:

[0113] 1. Comparative experiment on probability prediction of instability type

[0114] This study used time-series monitoring data from 10 historical slope instability cases to predict the probability distribution of each type of slope instability. The model, along with LSTM and random forest models, was used to predict the probability distribution of each type of slope instability. All three models were trained on the same training set and output the probability distribution of each detection point belonging to one of the five instability types in the future.

[0115] In addition, the evaluation index uses multi-category accuracy to quantitatively evaluate the prediction results. The heat map of the probability distribution of instability types is shown as follows: Figure 4 As shown in the figure, it can intuitively reflect the difference in probability prediction of each model for different detection points and instability types. Figure 5 The comparison of the classification accuracy of the three models on 10 slope cases is shown in the form of a line graph. The horizontal axis represents the numbers of the 10 slopes to be monitored, and the vertical axis represents the corresponding classification accuracy.

[0116] From the heat map of the probability distribution ( Figure 4 ) It can be seen that the model can accurately focus on the dominant instability type and the predicted probability distribution of different slopes is clearer, showing strong identification ability. For example, for slopes 4 and 9, the model accurately identified the high probability of shear instability. In addition, through the quantitative analysis of multi-category accuracy ( Figure 5 ), the average accuracy of this model reaches 90%, which is significantly higher than the 82% of the LSTM model and the 68% of the random forest.

[0117] This result shows that the adaptive feature extraction mechanism and spatiotemporal coupling design adopted by this model, the former efficiently mines the key spatiotemporal information in time series data by dynamically adjusting the input feature weights, thereby enhancing the ability to learn complex nonlinear relationships; the latter integrates the location information of spatial detection points with temporal dynamic changes, realizing deep interaction between spatial and temporal information, thereby improving the ability to capture the temporal change characteristics of the slope, and improving the accuracy of instability type identification and the reliability of early warning.

[0118] 2. Comparative Experiment on Evolution Chain Generation Results

[0119] This experiment further tested the performance of the three models in generating spatial evolution chains (sequences of detection points) and temporal evolution chains (sequences of time nodes). The accuracy of the evolution chain reflects the degree of match between the predicted chain and the actual evolution process, and is directly related to the early warning system's ability to dynamically understand and respond to slope instability processes.

[0120] The evaluation index uses the accuracy of the evolution chain sequence (Sequence Accuracy) to evaluate the matching degree between the actual evolution chain and the predicted evolution chain. The experimental process also selected 10 historical unstable slope cases for testing, and used each model to predict the evolution chain and calculate various indicators; the final prediction accuracy comparison results of the two evolution chains are as follows Figure 6 and Figure 7 shown.

[0121] Results show that the proposed model achieved an 88% accuracy rate for spatial evolution chains, significantly outperforming the LSTM (75%) and random forest (63%) models. In terms of temporal evolution chain accuracy, the proposed model also achieved an 87% accuracy rate, surpassing the LSTM (74%) and random forest (62%) models. This advantage demonstrates that the proposed model not only excels in identifying instability types but also more accurately captures the spatiotemporal evolution of slope instability.

[0122] These results demonstrate the model's multi-factor interaction modeling capabilities and adaptive model adjustment. By integrating spatial location information and temporal dynamics, the model deeply explores the complex coupling relationships among multiple factors influencing slope instability and effectively simulates the spatial propagation and temporal evolution mechanisms of the instability process. Compared to traditional single-time series or spatial models, the designed spatiotemporal coupling module significantly enhances the modeling capabilities of complex evolution chains, enabling the early warning system to have stronger dynamic response and predictive accuracy.

[0123] Therefore, the method proposed in the present invention not only performs excellently in distinguishing the instability type, but also can accurately reflect the spatial-temporal evolution law of slope instability, significantly improving the timeliness and accuracy of the early warning system.

[0124] The foregoing description is merely a preferred embodiment of the present application and is not intended to limit the present application. Persons skilled in the art will readily appreciate that various modifications and variations are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present application shall be included within the scope of protection of the present application.

[0125] Although the above describes the specific implementation methods of the present invention, it does not limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. An intelligent early warning and response method for slope instability risk, characterized by: The following processes are included: S1, based on a data set containing key geological data and environmental data, through slope disturbance sensitivity mapping modeling, introduces parameter perturbation to simulate the potential anomaly evolution process, and calculates the mask vector of geological data, the mask vector of environmental data and the structural path control coefficient; S2, construct an adaptive slope instability prediction model, and use the dataset and structure path control coefficient to optimize and train the model; S3, real-time acquisition of multiple key geological data and environmental data to construct multi-source heterogeneous time series data, combined with the mask vector calculated in S1, input into the trained adaptive slope instability prediction model, and output the instability type prediction probability vector and the dual-factor evolution chain prediction result. The dual-factor evolution chain prediction result includes the position detection point evolution sequence prediction result and the time evolution node sequence prediction result; S4, based on S3, obtain the predicted probability vector of the instability type and select the instability type with the highest probability as the warning type of the current risk of the slope; At the same time, based on the calculation results obtained by S3, the number of monitoring points with rapid instability is calculated and used as the length of the evolution chain. Then, the instability speed of each unstable monitoring point is calculated to form the evolutionary situation state vector. Combined with the defined multiple types of response strategies, it is input into the constructed multi-objective game model, and iterative evolutionary reasoning is performed through Monte Carlo evolutionary game to determine the optimal response strategy under the current risk situation.

2. The intelligent early warning and response method for slope instability risk according to claim 1, characterized in that: The geological data includes the horizontal displacement of the ground surface detected by GNSS fixed points ; Underwater level data collected using an automatic water level gauge ; The width of landslide cracks collected using linear displacement meters ; and the internal seepage water pressure of the slope collected by the piezometer ; The environmental data includes precipitation collected by automatic rain gauges and duration of precipitation ; Temperature collected by weather stations and air humidity ; Earthquake acceleration collected by seismograph .

3. The intelligent early warning and response method for slope instability risk according to claim 2, characterized in that: The dataset used to train the adaptive slope instability prediction model is constructed in the following ways: For the slope to be tested, select monitoring points, and collect data from each monitoring point at the same time. Geological data at each time point; directly collect data for the slope to be tested Environmental data at each point in time; The corresponding instability type prediction probability vector labels are divided into no instability, tension instability, shear instability, fluid-plastic instability and combined instability; The corresponding two-factor evolution chain prediction label is set as a sequence of position detection points and a sequence of time evolution points; The evolution sequence of the position detection points is set by analyzing the abnormal response degree of the geological parameters of each monitoring point before instability, screening out the monitoring point sequence with significant anomalies and triggering instability signs one after another, and forming the position instability propagation path. ,in, Indicates the location of the first monitoring point that triggers the instability sign. Similarly, Indicates the The location of the monitoring point that triggers the instability signs, and ; The time evolution node sequence is set based on the position instability propagation path , determine the time point when each monitoring point triggers the instability sign as the important stage node of the instability evolution, and construct the evolution path of the time dimension :in, Indicates the first monitoring time point that triggers the instability sign. Similarly, Indicates the A monitoring time point that triggers signs of instability, and .

4. The intelligent early warning and response method for slope instability risk according to claim 1, characterized in that: The specific process of S1 is: S11, for geological data and input environmental data , for each geological parameter and environmental parameter, the size is Add a small disturbance and repeat the operation times, and obtain the geological data set after adding disturbance and environmental datasets ;in, is the disturbance amplitude coefficient, is the standard deviation of the corresponding factor in the original data; S12, based on the constructed multi-source dynamic dataset of slope risk, the geological data and environmental data are spliced ​​together and clustered using the K-means algorithm. The number of clusters is preset to Q, and Q initial cluster centers are obtained. ; Afterwards, the disturbed geological dataset will be added and environmental datasets Add to the data sample to be clustered and get Q new cluster centers ; At the same time, we get the The classification result set of data samples ; S13, based on the degree of category drift before and after the perturbation sample clustering, factor sensitivity analysis is performed and calculated one by one After the perturbation is added, the cluster label of the perturbation sample is changed. The proportion of cluster label changes in the perturbation sample corresponding to different factors is counted and defined as the sensitivity score of the factor. , Indicates the The sensitivity of the factors; reconstruct the first p factors into a mask vector of geological data , reconstruct other factors into mask vectors of environmental data ; S14, based on the initial cluster center And the cluster centers after adding disturbance , the sum of the overall offsets of the cluster centers calculated using Euclidean distance is ; Then, combine the obtained geological mask vector and the environment mask vector , calculate the structural control coefficients of the geological data pathway and the environmental data pathway respectively and .

5. The intelligent early warning and response method for slope instability risk according to claim 1, characterized in that: The adaptive slope instability prediction model in S2 includes a geological structure evolution extraction module, an environmental driving feature modeling module and a slope stability evolution discrimination module; The mass structure evolution extraction module will Vector with geological mask Perform inner product to obtain geological mask features ; Then, through multiple chained TCN multi-scale convolution blocks Extract geological feature evolution to obtain a deep feature representation of geological structure evolution over time; the output of each TCN convolution block serves as the input of the next TCN convolution block, thus forming a progressive feature extraction path; Finally, The outputs of the TCN multi-scale convolutional blocks are fused through a multi-head attention layer to obtain stable and discriminative geological evolution features. ; The environment-driven feature modeling module uses environmental data As input, with the environment mask vector Perform inner product to obtain geological mask features ; Then through multiple chained ESN network layers Extract environmental features, capture the dynamic changes of environmental driving factors in the time dimension, and obtain The output features of each ESN layer are used as the input of the next layer, and finally, the multi-head attention fusion The output features of the ESN layer are used to obtain a stable representation of the dynamic characteristics of the environment. ; The slope stability evolution discrimination module will output the characteristics and output features Splicing is performed to obtain the comprehensive feature vector ; Then, the feature dimension is reduced through two fully connected layers, and the ReLU activation function is used to enhance the nonlinear expression ability to obtain the discriminant feature ; Finally, Input BatchNorm layer and Softmax activation function to get the instability type prediction probability vector ;Will Input the Dropout layer and Sigmoid activation function to obtain the two-factor evolution chain prediction result .

6. The intelligent early warning and response method for slope instability risk according to claim 5, characterized in that: The TCN multi-scale convolution block sets parallel convolution kernels with receptive fields of 1×1, 3×3, and 5×5, respectively, to extract characteristic response patterns at local and global spatial scales, thereby addressing the heterogeneity and hierarchy of geological data in spatial scales. It then uses the ReLU activation function and channel splicing operations to ultimately generate a feature embedding representation with rich spatial perception capabilities.

7. The intelligent early warning and response method for slope instability risk according to claim 5, characterized in that: described Through the structural control coefficient Adaptive determination: ,in, To preset the minimum number of TCN multi-scale convolution blocks, The number of buffers for the preset adaptive TCN multi-scale convolution blocks; described Through the structural control coefficient Calculation yields: ,in, To preset the minimum ESN layer, The number of preset adaptive ESN layer buffers.

8. The intelligent early warning and response method for slope instability risk according to claim 1, characterized in that: The function of the multi-objective game model in S4 is: in, Indicates the adoption of a strategy The profit value at To adopt a strategy The decrease in the probability of instability after To adopt a strategy The number of instability monitoring points decreased after To adopt a strategy The acceleration drop value of the instability monitoring point after and Respectively indicate the adoption of strategies the subsequent costs and time to effect; is the preset weight; the ultimate goal of the multi-objective game is to maximize .

9. The intelligent early warning and response method for slope instability risk according to claim 8, characterized in that: Predefined 4 types of optional response strategies: early warning strategy , Engineering disposal strategy , Traffic Strategy , Engineering disposal strategy ; then the response policy set , and each response strategy corresponds to a system behavior, and the game goal is to choose the optimal strategy To guide the system evolution towards a low-risk state.

10. The intelligent early warning and response method for slope instability risk according to claim 8, characterized in that: Iterative evolutionary reasoning through Monte Carlo evolutionary games enables dynamic adjustment of system status and strategy selection, specifically including: Calculate the multi-objective benefit function value of each alternative response strategy based on the current risk situation state vector; Gradually adjust the strategy distribution through the evolutionary selection mechanism to strengthen the strategy tendency of guiding the system evolution to a low-risk state; Dynamically adjust weight parameters and response strategy space to ensure that the multi-objective optimization process converges to the global optimal solution; Ultimately determine the optimal response strategy under the current risk situation .

Images