Landslide disaster probability prediction and dynamic response method

Abstract

The invention relates to the technical field of landslide prediction and optimization, and provides a landslide disaster probability prediction and dynamic response method, which converts multi-source heterogeneous monitoring data into a comprehensive risk feature vector through a dynamic fusion mechanism based on data quality and environmental perception. And based on a dynamic threshold decision-making mechanism in which the stability state of the landslide mass is linked with the external environment state, self-adaptive accurate adjustment of the early warning standard is realized. And automatically mapping the continuous prediction probability value to a discrete early warning level, and instantly triggering a standardized emergency response process. And analyzing actual response data of the landslide mass after early warning, performing quantitative evaluation on a decision-making effect, and performing closed-loop feedback and self-updating on a prediction model and a threshold rule according to the decision-making effect. According to the method, the problems of low prediction precision, threshold stiffness, response lag and lack of adaptive ability of a traditional method are effectively solved, and intelligent early warning of landslide disasters and continuous optimization of system performance are realized.

Description

Technical Field

[0001] This invention relates to the field of landslide prediction and optimization technology, and in particular to a method for predicting the probability of landslide disasters and dynamic response. Background Technology

[0002] Landslide disasters are typically closely related to multiple factors, including geology, environment, and meteorology. If left uncontrolled, they can cause severe casualties and property damage, destroy infrastructure such as transportation and water conservancy facilities, and potentially trigger secondary disasters like mudslides, resulting in significant negative impacts on society. Therefore, timely monitoring and early warning of such disasters, and accurate assessment of their impact range, are crucial for emergency response resource allocation, dynamic adjustment of mitigation plans, and operational monitoring. However, efficiently and accurately acquiring and processing monitoring data remains a pressing technical challenge. Traditional prediction methods primarily rely on single data sources or static threshold models, which have the following limitations: (1) Data silo problem: Data collected by different monitoring methods (such as sensors, remote sensing, and weather stations) are usually stored and analyzed independently, which makes it impossible to achieve deep fusion of cross-source data, resulting in limited prediction accuracy.

[0003] (2) Static threshold rigidity: Existing models mostly use fixed thresholds to judge the probability of disaster occurrence. The thresholds cannot adapt to the dynamic environmental changes in different regions and seasons, and are prone to false alarms or missed alarms.

[0004] (3) Insufficient timeliness: Traditional methods cannot process massive amounts of monitoring data in real time, resulting in prediction results lagging behind the development trend of disasters and making it difficult to meet the time requirements of emergency response.

[0005] (4) Lack of adaptability: Existing models are poorly adapted to complex geographical environments and extreme climate conditions, and perform poorly when facing new disaster patterns. Once the existing system is deployed, its core model parameters and decision rules become fixed, and it is unable to obtain feedback from the actual effects of each early warning action and perform self-optimization, resulting in the system performance not improving over time and with the accumulation of cases, and low long-term operation and maintenance efficiency.

[0006] Therefore, it is necessary to propose a method for predicting the probability of landslide disasters and for dynamic response to address or at least mitigate the aforementioned shortcomings. Summary of the Invention

[0007] The main objective of this invention is to provide a method for predicting the probability of landslide disasters and for dynamic response, in order to solve the technical problems of traditional prediction methods that mainly rely on a single data source or static threshold model, resulting in limited prediction accuracy, rigid static thresholds, and a lack of adaptive capabilities.

[0008] To achieve the above objectives, the present invention provides a method for predicting the probability of landslide disasters and for dynamic response, comprising the following steps: S1, acquire real-time monitoring data of the landslide body and current environmental status information, and preprocess the real-time monitoring data; S2, generate real-time quality assessment results based on preprocessed real-time monitoring data, calculate the fusion weight of each type of monitoring data according to the real-time quality assessment results and the current environmental status information, and then perform weighted fusion of the preprocessed real-time monitoring data according to the fusion weight to generate a comprehensive risk feature vector; S3, input the comprehensive risk feature vector into the pre-trained landslide probability prediction model to obtain the landslide occurrence probability value; S4. Based on the recent change trend of the comprehensive risk feature vector, the stability state of the landslide body is estimated, and based on the current environmental state information and the stability state of the landslide body, a threshold rule base is queried to obtain a dynamic probability threshold; wherein, the threshold rule base contains the mapping relationship between environmental state information, stability state, and probability threshold. S5, compare the landslide occurrence probability value with the dynamic probability threshold, determine the warning level and generate warning information; S6, execute the preset response action corresponding to the warning level, then collect feedback monitoring data of the landslide body, analyze the effectiveness of the warning decision based on the feedback monitoring data, and trigger the update of the landslide probability prediction model and / or the threshold rule base.

[0009] Preferably, step S1 includes the following steps: S11, acquire real-time monitoring data from various monitoring devices deployed on the landslide body, and simultaneously acquire current environmental status information including current rainfall; wherein, the data types collected by the various monitoring devices include surface and deep displacement data, soil and rock stress data, and rainfall data; S12, preprocessing the real-time monitoring data; wherein: For each type of monitoring data, the mean μ and standard deviation σ of each type of monitoring data within a preset historical period are calculated. Data points in the monitoring data that exceed the interval [μ-3σ, μ+3σ] are corrected to the boundary values ​​corresponding to the interval. For the data sequence after outlier correction, identify the missing data locations; for each missing location, calculate using the nearest valid monitoring data in time; if valid data exists before and after the missing location, use the arithmetic mean of the two to fill the gap; if valid data exists only on one side, use that valid data to fill the gap.

[0010] Preferably, step S2 includes the following steps: S21. For each type of monitoring data, the abnormal data ratio and missing data rate are calculated based on the number of data points marked as outliers during preprocessing and the number of missing data points imputed, respectively, as the real-time quality assessment result of that type of monitoring data. S22, based on the real-time quality assessment results and the current rainfall in the current environmental status information, a dynamic fusion weight is calculated for each type of monitoring data; wherein, the calculation rule of the dynamic fusion weight is as follows: first, a basic weight is assigned to each type of monitoring data according to the data anomaly ratio and data missing rate, and the value of the basic weight decreases as the data anomaly ratio and data missing rate of the monitoring data type increase; if the current rainfall exceeds a preset rainfall threshold, the basic weight of the monitoring data category corresponding to the rainfall data is increased. S23, the preprocessed monitoring data of various types are weighted and fused according to the dynamic fusion weight to generate the comprehensive risk feature vector; wherein, the comprehensive risk feature vector includes feature values ​​corresponding to displacement monitoring data, feature values ​​corresponding to stress monitoring data and feature values ​​corresponding to rainfall data.

[0011] Preferably, the landslide probability prediction model in step S3 is obtained through the following steps: S31, Obtain historical monitoring data of the landslide body and historical landslide event tags that correspond to the historical monitoring data in time; S32, with the optimization objective of maximizing the model performance quantification index, the hyperparameters of the XGBoost model are iteratively tuned using a Bayesian optimization algorithm. The termination condition for the iterative tuning is reaching a preset maximum number of evaluations. In each iteration, for a selected set of hyperparameters to be evaluated, the following sub-steps S321~S322 are executed: S321, using this set of hyperparameters, perform K different training and validation set partitions on the historical monitoring data, and train K XGBoost models accordingly. Calculate the model performance quantification index on the validation set to obtain K index values. S322, Calculate the average value of K index values, and use the average value as the comprehensive performance target value characterizing the set of hyperparameters; S33, after reaching the preset maximum number of evaluations, select the group with the highest comprehensive performance target value from all evaluated hyperparameter combinations as the optimal hyperparameter combination; S34. Using the optimal hyperparameter combination, the landslide probability prediction model is trained based on the historical monitoring data and the historical landslide event labels.

[0012] Preferably, step S4 includes the following steps: S41, For the characteristic components of the displacement and stress monitoring data corresponding to the comprehensive risk characteristic vector, calculate the average rate of change and average acceleration of the characteristic components within a preset time window. S42, based on whether the average rate of change and the average acceleration of change exceed the preset rate threshold and acceleration threshold, the current stability state of the landslide body is determined to be a stable period, a deformation period, or an acceleration period. S43, using the stability state of the landslide body and the seasonal information extracted from the current environmental state information as a composite query key; based on the composite query key, query the threshold rule base to obtain the corresponding dynamic probability threshold; wherein, the threshold rule base stores the mapping relationship between environmental state information, stability state, and probability threshold.

[0013] Preferably, step S5 includes the following steps: The numerical range into which the landslide occurrence probability value falls is matched with multiple preset probability threshold ranges, and each probability threshold range corresponds to a warning level; wherein, the warning level includes four levels: low risk, medium risk, high risk and ultra-high risk, and the higher the landslide occurrence probability value, the higher the risk level represented by the matched warning level.

[0014] Preferably, step S6 includes the following steps: S61, execute the preset response action corresponding to the warning level; S62, within the preset observation time window after the preset response action is executed, collect feedback monitoring data of the landslide body; S63, based on the changing trend of the feedback monitoring data relative to the monitoring data before the warning level was determined, analyze the effectiveness of the warning decision; wherein, calculate the acceleration value of the deformation data in the feedback monitoring data within a preset observation time window; if the acceleration value is less than or equal to zero, it is determined to be a valid warning; if the acceleration value calculated at each sampling moment within the preset observation time window is greater than a preset positive acceleration threshold, it is determined to be an invalid warning; S64, Based on the validity analysis results, selectively trigger update operations; where: If the warning is determined to be valid or invalid, incremental learning and updating of the landslide probability prediction model is triggered; wherein the training samples used for incremental learning include the comprehensive risk feature vector that triggered the warning, the landslide occurrence probability value, the warning level, and the validity analysis results; If the warning is deemed invalid, a calibration update of the threshold rule base is triggered simultaneously, and the probability threshold associated with the current environmental state information and the stability state of the landslide is lowered according to a preset adjustment range.

[0015] Preferably, determining the warning level in step S5 includes the following steps: The landslide occurrence probability value is compared with three preset probability thresholds of 0.5, 0.8 and 0.95; Based on the comparison results, the warning level is determined as follows: When 0 ≤ P < 0.5, the warning level is low risk; when 0.5 ≤ P < 0.8, the warning level is medium risk; when 0.8 ≤ P < 0.95, the warning level is high risk; when 0.95 ≤ P ≤ 1, the warning level is extremely high risk.

[0016] Preferably, the warning level is associated with a reference range of the slope stability safety factor Fs; wherein, the ultra-high risk level corresponds to Fs<1.0; the high risk level corresponds to 1.0≤Fs<1.05; the medium risk level corresponds to 1.05≤Fs<1.15; and the low risk level corresponds to Fs≥1.15.

[0017] Preferably, step S42 includes the following steps: The average rate of change is compared with a preset rate threshold, and the average acceleration is compared with a preset acceleration threshold. If the average rate of change is less than the rate threshold and the average acceleration is less than the acceleration threshold, then it is determined to be a stable period. If the average rate of change is greater than or equal to the rate threshold, and the average acceleration is less than the acceleration threshold, then it is determined to be the deformation period; If the average change in acceleration is greater than or equal to the acceleration threshold, it is determined to be an acceleration period.

[0018] Compared with the prior art, the present invention has the following beneficial effects: This application transforms multi-source heterogeneous data into a unified, high signal-to-noise ratio comprehensive risk feature vector by employing a dynamic weight fusion mechanism based on real-time data quality and environmental perception. This constructs a high-quality feature vector that better characterizes the comprehensive risk state of landslides from the source. This application creatively employs a dynamic threshold decision-making mechanism that links the landslide's own stability state (e.g., stable, deformed, and accelerating phase) with external environmental conditions (e.g., dry season, rainy season). This makes the early warning standard no longer a fixed value, but rather adaptively adjustable according to internal and external risk conditions. It maintains a high threshold in low-risk scenarios to reduce false alarms and lowers the threshold in high-risk scenarios to avoid missed alarms. This application automatically and quickly maps continuous predicted probability values ​​to discrete early warning levels and directly triggers associated standardized response processes, achieving automation and immediacy of early warning response. This application can use actual landslide response data after each early warning to quantitatively evaluate the decision-making effect and update the prediction model or threshold rules accordingly. This not only adapts to unknown disaster patterns but also achieves steady long-term performance improvement and significant reduction in maintenance costs, solving the industry problem of traditional static system performance degradation with environmental changes. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0020] Figure 1 This is a schematic flowchart of one embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the trend of days, landslide probability values, and early warning levels in one embodiment of the present invention. Figure 3 This is a schematic diagram of the landslide occurrence probability distribution in a test set according to one embodiment of the present invention.

[0021] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0022] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0024] In this invention, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature. Furthermore, the technical solutions of the various embodiments can be combined with each other, but only on the basis of being achievable by those skilled in the art. If the combination of technical solutions is contradictory or impossible to implement, such a combination of technical solutions should be considered non-existent and not within the scope of protection claimed by this invention.

[0025] Please refer to Figures 1 to 3 The present invention provides a method for predicting the probability of landslide disasters and for dynamic response, comprising the following steps: S1, acquire real-time monitoring data of the landslide body and current environmental status information, and preprocess the real-time monitoring data; S2, generate real-time quality assessment results based on preprocessed real-time monitoring data, calculate the fusion weight of each type of monitoring data according to the real-time quality assessment results and the current environmental status information, and then perform weighted fusion of the preprocessed real-time monitoring data according to the fusion weight to generate a comprehensive risk feature vector; S3, input the comprehensive risk feature vector into the pre-trained landslide probability prediction model to obtain the landslide occurrence probability value; S4. Based on the recent change trend of the comprehensive risk feature vector, the stability state of the landslide body is estimated, and based on the current environmental state information and the stability state of the landslide body, a threshold rule base is queried to obtain a dynamic probability threshold; wherein, the threshold rule base contains the mapping relationship between environmental state information, stability state, and probability threshold. S5, compare the landslide occurrence probability value with the dynamic probability threshold, determine the warning level and generate warning information; S6, execute the preset response action corresponding to the warning level, then collect feedback monitoring data of the landslide body, analyze the effectiveness of the warning decision based on the feedback monitoring data, and trigger the update of the landslide probability prediction model and / or the threshold rule base.

[0026] This application utilizes a dynamic weighted fusion mechanism based on real-time data quality and environmental perception to transform multi-source heterogeneous data into a unified, high signal-to-noise ratio comprehensive risk feature vector, thus constructing a high-quality feature vector that better characterizes the comprehensive risk state of landslides from the source. Furthermore, this application innovatively employs a dynamic threshold decision-making mechanism that links the landslide's own stability state (e.g., stable, deformed, and accelerating phases) with external environmental conditions (e.g., dry season, rainy season). This ensures that the early warning standard is no longer a fixed value but can adaptively adjust according to internal and external risk conditions, maintaining a high threshold in low-risk scenarios to reduce false alarms and decreasing it in high-risk scenarios. The low threshold avoids missed warnings; this application automatically and quickly maps continuous predicted probability values ​​to discrete warning levels and directly triggers the associated standardized response process, realizing the automation and immediacy of warning response; this application can use the actual landslide response data after each warning to quantitatively evaluate the decision-making effect, and accordingly, differentiate and accurately update the prediction model or threshold rules. It can not only adapt to unknown disaster patterns, but also achieve steady long-term performance improvement and significant reduction in maintenance costs, solving the industry problem of traditional static system performance degradation with environmental changes.

[0027] In a preferred embodiment, step S1 includes the following steps: S11, acquire real-time monitoring data from various monitoring devices deployed on the landslide body, and simultaneously acquire current environmental status information including current rainfall; wherein, the data types collected by the various monitoring devices include surface and deep displacement data, soil and rock stress data, and rainfall data; As a preferred example, various types of monitoring equipment are deployed on the landslide body, including: (1) Deformation monitoring equipment: such as GNSS receivers for acquiring surface displacement D1, and deep displacement gauges buried in boreholes (e.g., displacement gauges at depths of 15 meters and 25 meters) for acquiring deep displacement D2 and deep displacement D3 respectively. (2) Stress and strain monitoring equipment: such as stress and strain gauges installed on anchor cables for synchronously monitoring anchor cable stress and anchor cable strain. (3) Environmental monitoring equipment: such as rain gauges for real-time acquisition of rainfall data. At the same time, the current rainfall is extracted in real time from the rain gauge data stream and integrated with independent magnitude monitoring information (which can be obtained from the regional seismic network) to jointly constitute the current environmental status information.

[0028] S12, preprocessing the real-time monitoring data; wherein: For each type of monitoring data, the mean μ and standard deviation σ of each type of monitoring data within a preset historical period are calculated. Data points in the monitoring data that exceed the interval [μ-3σ, μ+3σ] are corrected to the boundary values ​​corresponding to the interval. For each type of monitoring data, calculate the data for that type within a preset historical time period (e.g., the past 24 hours). For each new data point arriving, perform the following operations: First, calculate the mean μ and standard deviation σ of all valid data within the preset historical time period. Next, determine if the new data point falls outside the interval [μ-3σ, μ+3σ]. Based on the characteristics of the normal distribution, the probability of falling outside this interval is low, and therefore it can be reasonably considered an anomaly. If the new data point > μ+3σ, correct it to μ+3σ; if the new data point < μ-3σ, correct it to μ-3σ, thereby filtering out spikes caused by momentary equipment interference or communication errors.

[0029] For the data sequence after outlier correction, identify the missing data locations; for each missing location, calculate using the nearest valid monitoring data in time; if valid data exists before and after the missing location, use the arithmetic mean of the two to fill the gap; if valid data exists only on one side, use that valid data to fill the gap.

[0030] After outlier correction, the system scans the data sequence to identify any missing data points (e.g., due to transmission delays or brief device downtime). For each missing location, the system searches for the nearest valid data point in time before and after it. If both a preceding and following valid data point can be found, the arithmetic mean of the two is used to fill the missing area. If the missing data occurs at the beginning or end of the sequence, and only one valid neighboring data point can be found, that valid data point is used directly for filling. This embodiment achieves automatic outlier identification and correction, ensuring the spatiotemporal continuity of the monitored data stream.

[0031] In a preferred embodiment, step S2 includes the following steps: S21: For each type of monitoring data, the outlier ratio and missing data rate are calculated based on the number of data points marked as outliers during preprocessing and the number of missing data points imputed, respectively, as the real-time quality assessment result for that type of monitoring data; that is, the outlier ratio and missing data rate are calculated based on the number of outlier data points marked and corrected by the 3σ rule in S12 and the number of missing data points filled by the imputation method in S12, respectively, to constitute the real-time quality assessment result for that type of data, with lower values ​​indicating higher data quality.

[0032] S22, based on the real-time quality assessment results and the current rainfall in the current environmental status information, a dynamic fusion weight is calculated for each type of monitoring data. The calculation rule for the dynamic fusion weight is as follows: first, a basic weight is assigned to each type of monitoring data based on the data anomaly ratio and data missing rate. The value of the basic weight decreases as the data anomaly ratio and data missing rate of that type of monitoring data increase. If the current rainfall exceeds a preset rainfall threshold, the basic weight of the monitoring data category corresponding to the rainfall data is increased. For example, an initial basic weight can be assigned to each type of monitoring data (e.g., initialized to 1). The value of the basic weight is inversely proportional to the data anomaly ratio and data missing rate; that is, the higher the anomaly ratio or missing rate, the lower the assigned basic weight. Then, adjustments are made based on the corresponding quality assessment results: it is determined whether the current rainfall exceeds a preset rainfall threshold (e.g., whether the 24-hour cumulative rainfall exceeds 50 mm, which can be set according to the regional climate characteristics). If it exceeds, it is determined that rainfall has become the dominant disaster-causing factor, and its information importance needs to be increased. Therefore, the basic weight of the rainfall data category should be increased, for example, by an increase factor of 1.5.

[0033] S23, the preprocessed monitoring data of various types are weighted and fused according to the dynamic fusion weights to generate the comprehensive risk feature vector; wherein, the comprehensive risk feature vector includes feature values ​​corresponding to displacement monitoring data, feature values ​​corresponding to stress monitoring data, and feature values ​​corresponding to rainfall data. The latest feature values ​​(such as displacement value, stress value, and rainfall value) corresponding to various types of monitoring data after preprocessing are extracted to form a set of basic feature values. Then, the final dynamic fusion weights corresponding to various types of data calculated in S22 are used as coefficients to weight and fuse this set of basic feature values ​​to form a multi-dimensional vector, which is the comprehensive risk feature vector.

[0034] As a preferred embodiment, the landslide probability prediction model in step S3 is obtained through the following steps: S31, acquire historical monitoring data of the landslide body and historical landslide event labels corresponding to the historical monitoring data in time; label the historical monitoring data (i.e. a feature vector containing multi-source features) at each time point with the corresponding historical landslide event label (for example, use "1" to mark data within a certain period before the landslide occurs, and use "0" to mark data during the stable period).

[0035] S32, with maximizing the model performance quantification index as the optimization objective, the hyperparameters of the XGBoost model are iteratively tuned using a Bayesian optimization algorithm. The termination condition for the iterative tuning is reaching a preset maximum number of evaluations. In each iteration, for a selected set of hyperparameters to be evaluated, the following sub-steps S321-S322 are executed: This implementation uses a Bayesian optimization framework to automatically search for the optimal hyperparameter combination of the XGBoost model. Its core innovation lies in the evaluation method for each set of parameters. A model performance quantification index (e.g., for a binary classification problem of whether a landslide has occurred, preferably the area under the curve (AUC)) is selected as the optimization objective, with the goal of maximizing this index. A preset maximum number of evaluations is set (e.g., 50 times). In each iteration, after the Bayesian optimization framework recommends a set of hyperparameter combinations to be evaluated, the system executes the following noise resistance evaluation process: S321, using this set of hyperparameters, perform K different training and validation set partitions on the historical monitoring data, and train K XGBoost models accordingly. Calculate the model performance quantification index on the validation set to obtain K index values. The value of K can be set according to actual needs, for example, 5. After each partition, train an XGBoost model using the training set data and calculate the AUC value on the corresponding validation set to obtain K AUC values.

[0036] S322, calculate the average value of K index values, and use the average value as the comprehensive performance target value of the hyperparameter group; for example, calculate the arithmetic mean of these K AUC values, and use this average value as the comprehensive performance target value of the hyperparameter group in the current iteration.

[0037] S33, after reaching the preset maximum number of evaluations, select the group with the highest comprehensive performance target value from all evaluated hyperparameter combinations as the optimal hyperparameter combination; when the preset maximum number of evaluations is reached, the system selects the combination with the highest comprehensive performance target value (i.e., the highest average AUC after K verifications) from all evaluated hyperparameter combinations and determines it as the global optimal hyperparameter combination.

[0038] S34, using the optimal hyperparameter combination, the landslide probability prediction model is trained based on the historical monitoring data and the historical landslide event labels. Using the optimal hyperparameter combination determined in step S33, the XGBoost model is finally trained based on all historical monitoring data and corresponding landslide event labels from S31 to obtain a landslide probability prediction model that can be used for real-time prediction.

[0039] This embodiment aims to address common challenges in existing machine learning applications: model performance heavily relies on hyperparameter settings, while manual or grid search for hyperparameter tuning is inefficient and prone to getting trapped in local optima; simultaneously, model evaluation is greatly affected by the randomness of data partitioning, resulting in poor robustness of the found "optimal" parameters and unstable model performance in actual deployment. This embodiment utilizes an automated model training process centered on noise resistance evaluation, deeply integrating XGBoost with landslide early warning. By finding a set of robust parameters that can withstand the interference of data randomness and ensure stable model output, the prediction accuracy of the final generated landslide probability prediction model is guaranteed.

[0040] In a preferred embodiment, step S4 includes the following steps: S41, for the characteristic components of the displacement and stress monitoring data corresponding to the comprehensive risk feature vector, calculate the average rate of change and average acceleration of change of the characteristic components within a preset time window; extract the core components that directly reflect the mechanical behavior of the landslide from the comprehensive risk feature vector: the characteristic values ​​of the corresponding displacement monitoring data (such as surface displacement values) and the characteristic values ​​of the corresponding stress monitoring data (such as anchor cable stress values). For the time-series data of these characteristic values ​​within a preset time window (e.g., the most recent 6 hours), calculate their average rate of change and average acceleration of change respectively.

[0041] S42, based on whether the average rate of change and the average acceleration of change exceed the preset rate threshold and acceleration threshold, the current stability state of the landslide body is determined to be a stable period, a deformation period, or an acceleration period. The obtained average rate of change and average acceleration are compared with preset rate thresholds and acceleration thresholds. These thresholds can be set based on historical data, geomechanical models, or regional experience; for example, the rate threshold could be set to 2.0 mm / h and the acceleration threshold to 0.1 mm / h. 2 .

[0042] Preferably, step S42 includes the following steps: The average rate of change is compared with a preset rate threshold, and the average acceleration is compared with a preset acceleration threshold. If the average rate of change is less than the rate threshold and the average acceleration is less than the acceleration threshold, then it is determined to be a stable period. If the average rate of change is greater than or equal to the rate threshold, and the average acceleration is less than the acceleration threshold, then it is determined to be the deformation period; If the average change in acceleration is greater than or equal to the acceleration threshold, it is determined to be an acceleration period.

[0043] S43, using the stability state of the landslide body and the seasonal information extracted from the current environmental state information as a composite query key; based on the composite query key, query the threshold rule base to obtain the corresponding dynamic probability threshold; wherein, the threshold rule base stores the mapping relationship between environmental state information, stability state, and probability threshold.

[0044] Based on the stability state of the landslide and the seasonal information extracted from the current environmental state information, seasonal information is extracted (e.g., divided into dry season and rainy season according to system date or rainfall climate characteristics). A pre-set threshold rule base is queried using the composite key (landslide stability state, season). This threshold rule base stores recommended warning probability thresholds for different scenario combinations in the form of a two-dimensional table or equivalent data structure. For example, the rule base sets: for (stable period, dry season), the dynamic probability threshold is 0.95 (extremely high threshold, very few alarms); for (deformation period, rainy season), the threshold may be set to 0.75; and for (acceleration period, rainy season), the threshold may drop sharply to 0.60.

[0045] This embodiment addresses the problem of rigid static thresholds explicitly mentioned in the background section. Existing early warning models use fixed thresholds, which cannot adapt to dynamic environmental changes in different regions and seasons, leading to false alarms during dry seasons and missed alarms during rainy seasons. This embodiment achieves intelligent dynamic adjustment of the early warning threshold by introducing a dual dimension of landslide body self-state perception and external environment state perception.

[0046] Preferably, step S5 includes the following steps: The numerical range into which the landslide occurrence probability value falls is matched with multiple preset probability threshold ranges, and each probability threshold range corresponds to a warning level; wherein, the warning level includes four levels: low risk, medium risk, high risk and ultra-high risk, and the higher the landslide occurrence probability value, the higher the risk level represented by the matched warning level.

[0047] Figure 3 This is a schematic diagram of the probability distribution of landslide occurrence values ​​in the test set. The horizontal axis represents the predicted landslide probability (i.e., the probability value of landslide occurrence), and the vertical axis represents the number of predictions (i.e., the number of samples). The orange pattern represents the number of occurrences in the probability interval, with higher values ​​indicating a higher probability.

[0048] In a preferred embodiment, step S6 includes the following steps: S61, execute the preset response action corresponding to the warning level; the preset response action is a pre-set standardized operation process, such as: (1) Information release: automatically send warning information containing specific location and risk level to the mobile phone of the preset responsible person and the large screen of the monitoring center; (2) Monitoring enhancement: automatically increase the sampling frequency of the sensors in the relevant landslide area (e.g., from 1 time / hour to 1 time / 10 minutes) to obtain more dense deformation data; (3) Contingency plan activation: identify and trigger the corresponding level of emergency response process in the system, prompting relevant personnel to enter the preset handling process.

[0049] S62, within a preset observation time window following the execution of the preset response action, feedback monitoring data of the landslide body is collected; a preset observation time window (e.g., the next 24 hours) is initiated. Within this observation time window, feedback monitoring data of the landslide body is continuously collected using monitoring equipment deployed on the landslide body (the same monitoring equipment involved in S1). This data is consistent in type with the original monitoring data and is used to assess the impact of the preset response action on the stability of the landslide body.

[0050] S63, based on the changing trend of the feedback monitoring data relative to the monitoring data before the warning level was determined, analyze the effectiveness of the warning decision; wherein, calculate the acceleration value of the deformation data in the feedback monitoring data within a preset observation time window; if the acceleration value is less than or equal to zero, it is determined to be a valid warning; if the acceleration value calculated at each sampling moment within the preset observation time window is greater than a preset positive acceleration threshold, it is determined to be an invalid warning; Using the deformation trend reflected in the monitoring data before the warning level was determined as a benchmark, the deformation acceleration value, which best reflects changes in stability, was calculated over the entire observation window (S62). The acceleration values ​​of deformation data (such as deep displacement) in the monitoring data were then compared. If the calculated acceleration value was less than or equal to 0, it indicated that the deformation rate was stable or decreasing, and the landslide was stabilizing or improving, thus determining that the warning decision and subsequent response were effective. If, within the entire observation window, the instantaneous acceleration value calculated at each sampling moment was greater than the preset positive acceleration threshold, it indicated that the deformation had not only failed to be contained but was instead accelerating and deteriorating, thus determining that the warning decision was invalid.

[0051] S64, Based on the validity analysis results, selectively trigger update operations; where: If the warning is determined to be valid or invalid, incremental learning and updating of the landslide probability prediction model is triggered; wherein the training samples used for incremental learning include the comprehensive risk feature vector that triggered the warning, the landslide occurrence probability value, the warning level, and the validity analysis results; If the warning is deemed invalid, a calibration update of the threshold rule base is triggered simultaneously, and the probability threshold associated with the current environmental state information and the stability state of the landslide is lowered according to a preset adjustment range.

[0052] When an early warning decision produces an invalid or valid warning, the event itself constitutes a valuable labeled sample. This triggers incremental learning updates to the landslide probability prediction model. The training samples used for learning contain a complete decision context: the comprehensive risk feature vector that triggered the warning, the landslide probability value given by the model, the final warning level issued, and the validity label for this event. By adding this sample, the model learns which feature patterns lead to valid or invalid results in which situations.

[0053] If an alert is deemed invalid, it indicates that the current alert threshold is too lenient for the given situation and fails to trigger a sufficiently strong response in a timely manner. Therefore, the system simultaneously triggers a calibration update of the threshold rule base. Specifically, the system lowers the probability threshold associated with the current environmental state information and the landslide stability state, which was used in step S4 to generate the current dynamic probability threshold, by a preset adjustment margin (e.g., multiplying the corresponding threshold by 0.9 or subtracting a fixed value). This ensures that in similar high-risk situations in the future, the system will issue higher-level alerts earlier and more sensitively.

[0054] As a preferred example, determining the warning level in step S5 includes the following steps: The landslide occurrence probability value is compared with three preset probability thresholds of 0.5, 0.8, and 0.95. Based on the comparison results, the warning level is determined as follows: When 0 ≤ P < 0.5, the warning level is low risk; when 0.5 ≤ P < 0.8, the warning level is medium risk; when 0.8 ≤ P < 0.95, the warning level is high risk; when 0.95 ≤ P ≤ 1, the warning level is extremely high risk.

[0055] Preferably, the warning level is associated with a reference range of the slope stability safety factor Fs; wherein, the ultra-high risk level corresponds to Fs<1.0; the high risk level corresponds to 1.0≤Fs<1.05; the medium risk level corresponds to 1.05≤Fs<1.15; and the low risk level corresponds to Fs≥1.15.

[0056] In this embodiment, when generating early warning information, not only is the early warning level output, but also the reference range of the slope stability safety factor Fs associated with that level is output or explicitly indicated. This makes the early warning information no longer just a suggestion from the model, but is transformed into information that engineers can directly understand.

[0057] To further illustrate the technical solution of this application, the inventors also compared the model of this application with the existing XGBoost model: Standard XGBoost greedily constructs each decision tree, selecting only the currently optimal feature and split point when splitting nodes, pursuing single-step optimization, but potentially getting trapped in local optima. In contrast, the embodiments of this application use a Bayesian optimization algorithm to iteratively fine-tune the hyperparameters of the XGBoost model, ultimately obtaining a landslide probability prediction model: the NRBO-XGBoost model. It should be understood by those skilled in the art that NRBO (Noisy Response Bayesian Optimization) combines NRBO with XGBoost to specifically address hyperparameter optimization problems in scenarios where model performance evaluation involves noise (fluctuations). NRBO-XGBoost is a global optimization strategy: instead of selecting only the currently optimal split point, it uses batch sampling + global search to find a globally better solution among multiple candidate split points; the core improvement is that in the decision tree node splitting stage, NRBO replaces the greedy splitting of XGBoost, balancing global optima and computational efficiency.

[0058] Table 1 Partial Sample Data Table 2 Evaluation Metrics for NRBO-XGBoost Model Table 3 Evaluation metrics for the XGBoost model As shown in Table 2, where: (1) Accuracy = 0.964 Meaning: In multi-class classification scenarios, accuracy refers to the proportion of samples where the model's predicted warning level is completely consistent with the actual level. Calculation: Accuracy = (Total number of samples with the same predicted and actual levels) / (Total number of samples). Note: High accuracy does not necessarily mean the model performs well at high / very high-risk levels, because low-risk samples usually constitute the vast majority, easily inflating the overall accuracy.

[0059] (2) AUC value (Area Under the Curve) = 0.932 Meaning: In multi-level early warning, AUC usually refers to the area under the ROC curve of the binary classification problem of "whether a landslide has occurred", that is, the model's ability to distinguish between "there is a risk of landslide (P≥0.5)" and "there is no risk of landslide (P<0.5)".

[0060] Engineering significance: The AUC of 0.932 indicates that the model performs exceptionally well in distinguishing between the two major categories of "risky" and "no risk," which forms the basis for multi-level early warning systems.

[0061] III. Sensitivity / Recall = 0.955 Meaning: In multi-level early warning systems, sensitivity typically refers to the proportion of landslides that are predicted by the model to be of high / very high risk (P≥0.8) among actual landslide cases. Calculation: Sensitivity = TP / (FN + TP) (TP (True Positive): The number of samples where a landslide actually occurred and the model predicted it as high / very high risk (P≥0.8). FN (False Negative): The number of samples where a landslide actually occurred but the model predicted it as low / medium risk (P<0.8).) Engineering Significance: A sensitivity of 0.955 means that in areas where landslides actually occurred, 95.5% were successfully identified by the model as high or very high risk, which is crucial for disaster prevention and mitigation.

[0062] A comparison of the two models shows that the NRBO-XGBoost model in this application has better accuracy, AUC value, and sensitivity than the existing XGBoost model.

[0063] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for predicting the probability of landslide disasters and for dynamic response, characterized in that, Includes the following steps: S1, acquire real-time monitoring data of the landslide body and current environmental status information, and preprocess the real-time monitoring data; S2, generate real-time quality assessment results based on preprocessed real-time monitoring data, calculate the fusion weight of each type of monitoring data according to the real-time quality assessment results and the current environmental status information, and then perform weighted fusion of the preprocessed real-time monitoring data according to the fusion weight to generate a comprehensive risk feature vector; S3, input the comprehensive risk feature vector into the pre-trained landslide probability prediction model to obtain the landslide occurrence probability value; S4. Based on the recent change trend of the comprehensive risk feature vector, the stability state of the landslide body is estimated, and based on the current environmental state information and the stability state of the landslide body, a threshold rule base is queried to obtain a dynamic probability threshold; wherein, the threshold rule base contains the mapping relationship between environmental state information, stability state, and probability threshold. S5, compare the landslide occurrence probability value with the dynamic probability threshold, determine the warning level and generate warning information; S6, execute the preset response action corresponding to the warning level, then collect feedback monitoring data of the landslide body, analyze the effectiveness of the warning decision based on the feedback monitoring data, and trigger the update of the landslide probability prediction model and / or the threshold rule base.

2. The landslide disaster probability prediction and dynamic response method according to claim 1, characterized in that, Step S1 includes the following steps: S11, acquire real-time monitoring data from various monitoring devices deployed on the landslide body, and simultaneously acquire current environmental status information including current rainfall; wherein, the data types collected by the various monitoring devices include surface and deep displacement data, soil and rock stress data, and rainfall data; S12, preprocessing the real-time monitoring data; wherein: For each type of monitoring data, the mean μ and standard deviation σ of each type of monitoring data within a preset historical period are calculated. Data points in the monitoring data that exceed the interval [μ-3σ, μ+3σ] are corrected to the boundary values ​​corresponding to the interval. For the data sequence after outlier correction, identify the missing data locations; for each missing location, calculate using the nearest valid monitoring data in time; if valid data exists before and after the missing location, use the arithmetic mean of the two to fill the gap; if valid data exists only on one side, use that valid data to fill the gap.

3. The landslide disaster probability prediction and dynamic response method according to claim 2, characterized in that, Step S2 includes the following steps: S21. For each type of monitoring data, the abnormal data ratio and missing data rate are calculated based on the number of data points marked as outliers during preprocessing and the number of missing data points imputed, respectively, as the real-time quality assessment result of that type of monitoring data. S22, based on the real-time quality assessment results and the current rainfall in the current environmental status information, a dynamic fusion weight is calculated for each type of monitoring data; wherein, the calculation rule of the dynamic fusion weight is as follows: first, a basic weight is assigned to each type of monitoring data according to the data anomaly ratio and data missing rate, and the value of the basic weight decreases as the data anomaly ratio and data missing rate of the monitoring data type increase; if the current rainfall exceeds a preset rainfall threshold, the basic weight of the monitoring data category corresponding to the rainfall data is increased. S23, the preprocessed monitoring data of various types are weighted and fused according to the dynamic fusion weight to generate the comprehensive risk feature vector; wherein, the comprehensive risk feature vector includes feature values ​​corresponding to displacement monitoring data, feature values ​​corresponding to stress monitoring data and feature values ​​corresponding to rainfall data.

4. The landslide disaster probability prediction and dynamic response method according to claim 3, characterized in that, The landslide probability prediction model in step S3 is obtained through the following steps: S31, Obtain historical monitoring data of the landslide body and historical landslide event tags that correspond to the historical monitoring data in time; S32, with the optimization objective of maximizing the model performance quantification index, the hyperparameters of the XGBoost model are iteratively tuned using a Bayesian optimization algorithm. The termination condition for the iterative tuning is reaching a preset maximum number of evaluations. In each iteration, for a selected set of hyperparameters to be evaluated, the following sub-steps S321~S322 are executed: S321, using this set of hyperparameters, perform K different training and validation set partitions on the historical monitoring data, and train K XGBoost models accordingly. Calculate the model performance quantification index on the validation set to obtain K index values. S322, Calculate the average value of K index values, and use the average value as the comprehensive performance target value characterizing the set of hyperparameters; S33, after reaching the preset maximum number of evaluations, select the group with the highest comprehensive performance target value from all evaluated hyperparameter combinations as the optimal hyperparameter combination; S34. Using the optimal hyperparameter combination, the landslide probability prediction model is trained based on the historical monitoring data and the historical landslide event labels.

5. The landslide disaster probability prediction and dynamic response method according to claim 4, characterized in that, Step S4 includes the following steps: S41, For the characteristic components of the displacement and stress monitoring data corresponding to the comprehensive risk characteristic vector, calculate the average rate of change and average acceleration of the characteristic components within a preset time window. S42, based on whether the average rate of change and the average acceleration of change exceed the preset rate threshold and acceleration threshold, the current stability state of the landslide body is determined to be a stable period, a deformation period, or an acceleration period. S43, using the stability state of the landslide body and the seasonal information extracted from the current environmental state information as a composite query key; based on the composite query key, query the threshold rule base to obtain the corresponding dynamic probability threshold; wherein, the threshold rule base stores the mapping relationship between environmental state information, stability state, and probability threshold.

6. The landslide disaster probability prediction and dynamic response method according to claim 5, characterized in that, Step S5 includes the following steps: The numerical range into which the landslide occurrence probability value falls is matched with multiple preset probability threshold ranges, and each probability threshold range corresponds to a warning level; wherein, the warning level includes four levels: low risk, medium risk, high risk and ultra-high risk, and the higher the landslide occurrence probability value, the higher the risk level represented by the matched warning level.

7. The landslide disaster probability prediction and dynamic response method according to claim 6, characterized in that, Step S6 includes the following steps: S61, execute the preset response action corresponding to the warning level; S62, within the preset observation time window after the preset response action is executed, collect feedback monitoring data of the landslide body; S63, based on the changing trend of the feedback monitoring data relative to the monitoring data before the warning level was determined, analyze the effectiveness of the warning decision; wherein, calculate the acceleration value of the deformation data in the feedback monitoring data within a preset observation time window; if the acceleration value is less than or equal to zero, it is determined to be a valid warning; if the acceleration value calculated at each sampling moment within the preset observation time window is greater than a preset positive acceleration threshold, it is determined to be an invalid warning; S64, Based on the validity analysis results, selectively trigger update operations; where: If the warning is determined to be valid or invalid, incremental learning and updating of the landslide probability prediction model is triggered; wherein the training samples used for incremental learning include the comprehensive risk feature vector that triggered the warning, the landslide occurrence probability value, the warning level, and the validity analysis results; If the warning is deemed invalid, a calibration update of the threshold rule base is triggered simultaneously, and the probability threshold associated with the current environmental state information and the stability state of the landslide is lowered according to a preset adjustment range.

8. The landslide disaster probability prediction and dynamic response method according to claim 5, characterized in that, Determining the warning level in step S5 includes the following steps: The landslide occurrence probability value is compared with three preset probability thresholds of 0.5, 0.8 and 0.95; Based on the comparison results, the warning level is determined as follows: When 0 ≤ P < 0.5, the warning level is low risk; when 0.5 ≤ P < 0.8, the warning level is medium risk; when 0.8 ≤ P < 0.95, the warning level is high risk; when 0.95 ≤ P ≤ 1, the warning level is extremely high risk.

9. The landslide disaster probability prediction and dynamic response method according to claim 8, characterized in that, in, The warning levels are associated with a reference range for the slope stability safety factor Fs; where, the ultra-high risk level corresponds to Fs<1.0; the high risk level corresponds to 1.0≤Fs<1.05; the medium risk level corresponds to 1.05≤Fs<1.15; and the low risk level corresponds to Fs≥1.

15.

10. The landslide disaster probability prediction and dynamic response method according to claim 5, characterized in that, Step S42 includes the following steps: The average rate of change is compared with a preset rate threshold, and the average acceleration is compared with a preset acceleration threshold. If the average rate of change is less than the rate threshold and the average acceleration is less than the acceleration threshold, then it is determined to be a stable period. If the average rate of change is greater than or equal to the rate threshold, and the average acceleration is less than the acceleration threshold, then it is determined to be the deformation period; If the average change in acceleration is greater than or equal to the acceleration threshold, it is determined to be an acceleration period.

Images