Classification and early warning method of reservoir water type landslide based on multi-source monitoring data

Abstract

This invention discloses a graded early warning method for reservoir-type landslides based on multi-source monitoring data. Belonging to the field of geological disaster prevention and monitoring early warning technology, it addresses the shortcomings in reliability and accuracy of existing methods. The method includes: collecting data on the main controlling environmental factors of the reservoir-type slope under study; inputting this data into a trained network model; determining the output result of the discriminant function based on the model's output reflecting the confidence level of accelerated slope deformation and the early warning discriminant function; determining the optimal response window and its weight for the reservoir water level; obtaining the historical reservoir water levels for the current day and the previous day; and determining the improved effective reservoir water level based on the weights; calculating the landslide displacement rate and improved tangent angle; and determining the risk level at which the improved effective reservoir water level, displacement rate, and improved tangent angle are located; and determining the early warning level for landslides on the reservoir-type slope under study based on the discriminant function output result and the risk level. Using the above scheme for early warning grading can improve the reliability of the early warning results.

Description

Technical Field

[0001] This invention belongs to the field of geological disaster prevention, monitoring and early warning technology, and specifically relates to a graded early warning method for reservoir-type landslides based on multi-source monitoring data. Background Technology

[0002] During the operation of hydropower stations, reservoir water levels often fluctuate periodically. Under these cyclical water level changes, loose deposits on the reservoir banks are prone to collapse and landslides, posing a serious threat to the safe operation of the hydropower station and the safety of life and property of people in the reservoir area. Therefore, conducting research on early warning systems for large deposits in reservoir areas is of great significance.

[0003] Currently, research on early warning models for reservoir landslides mainly falls into three categories: First, methods based on physical model experiments and numerical simulations, which analyze the stress field and deformation response of landslides under hydraulic coupling to reveal the instability mechanism; second, deformation prediction methods based on machine learning algorithms, utilizing models such as long short-term memory networks and convolutional neural networks to predict landslide deformation trends; and third, methods based on historical monitoring data and expert experience to establish empirical thresholds between landslide deformation and reservoir water level changes. While these early warning research methods have been applied in reservoir landslide early warning work, some problems remain: ① Most early warning models focus on the evolution of displacement over time, neglecting the influence of hydrodynamics on landslide deformation under varying reservoir water levels, making it difficult to accurately reflect the deformation mechanism of reservoir-type accumulation landslides; ② Single empirical thresholds are limited by the form of reservoir water level changes, slope structure, and regional factors, resulting in insufficient accuracy in threshold selection and a tendency to misjudge under the combined effects of rainfall and reservoir water level fluctuations. Summary of the Invention

[0004] In view of the above-mentioned shortcomings in the prior art, the reservoir-type landslide classification and early warning method based on multi-source monitoring data provided by the present invention solves the problem of insufficient reliability and accuracy of existing slope early warning methods.

[0005] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: A method for graded early warning of reservoir-type landslides based on multi-source monitoring data is provided, which includes the following steps: S1. Collect the main environmental factor data of the reservoir-type slope to be studied, input them into the trained network model, and determine the output result of the early warning discrimination function of the slope based on the confidence degree of accelerated deformation and the early warning discrimination function of the model output. S2. Determine the optimal response window for reservoir water level using historical reservoir water levels. and its weight, to obtain the current day and previous days The historical reservoir water level of the day was used as a basis to determine the effective improved reservoir water level. S3. Obtain surface displacement monitoring data of the reservoir-type slope to be studied, calculate the displacement rate and improved tangent angle of the landslide, and then determine the risk level of the improved effective reservoir water level, displacement rate and improved tangent angle. S4. Based on the output results of the early warning discrimination function and the risk level of the improved effective reservoir water level, displacement rate, and improved tangent angle, determine the early warning level for landslides on the reservoir-type slope under study.

[0006] Furthermore, the expression for the early warning discrimination function is: ,

[0007] in, For early warning discrimination function, Indicates the sensitive area. Indicates non-sensitive areas; The output of the network model for the controlling environmental factor data z; The confidence threshold; For Youden index; , , and ROC curves were constructed using the output probabilities from the network model validation phase as continuous discriminant values, and different candidate probability thresholds were applied. The statistical results show the number of true positives, false positives, true negatives, and false negatives.

[0008] The beneficial effects of the above technical solution are as follows: Firstly, the discriminant function can transform the accelerated deformation probability output by the network model into two distinct categories: sensitive area and non-sensitive area, facilitating direct application in engineering early warning systems. Secondly, the confidence threshold P... th Determined by the principle of maximizing the ROC curve and Youden index, rather than by manual experience, this method comprehensively considers both true positive and false positive rates, improving the ability to identify accelerated deformation while controlling the risk of false alarms. This approach avoids the problem of fixed probability thresholds being insufficiently adaptable to different bank slopes, making the early warning discrimination boundary more consistent with the characteristics of historical monitoring data of the target bank slope, and providing stable sensitive area identification results for subsequent joint graded early warning systems.

[0009] Furthermore, the training method for the network model includes: S11. Obtain historical surface displacement monitoring data of the reservoir-type slope under study in the monitoring area, and preprocess it. Label the preprocessed data as accelerated deformation, non-accelerated deformation, and transition. S12. Construct a set of environmental factors, remove environmental factors whose correlation with the deformation of the reservoir-type slope under study is lower than the set correlation or whose redundancy is higher than the set redundancy, and then select the main controlling environmental factors by constructing the ROC curve of the environmental factors. S13. Select K main environmental factors, extract their data for several consecutive time steps as model input, and use the sample state label at the end of several consecutive time steps as model output label to construct a dataset. S14. Divide the dataset into a training set and a validation set, and use the training set to train the convolutional neural network; S15. Use a validator to validate the convolutional neural network and determine whether the training accuracy of the convolutional neural network has reached the preset accuracy. If yes, obtain the trained network model; otherwise, let K=K+1 and return to step S13.

[0010] Furthermore, methods for labeling preprocessed data as accelerated deformation or non-accelerated deformation include: S111. Calculate the displacement rate at the monitoring time based on the preprocessed data. ; S112. Based on historically known stable deformation stages and accelerated deformation stages, the maximum displacement rate of the stable deformation stage is taken as the low-velocity boundary. The minimum displacement rate during the accelerated deformation stage is used as the high-speed boundary. ; S113, when When, the corresponding preprocessed data is labeled as non-accelerated deformation; when At that time, the corresponding labels of the preprocessed data are marked as accelerated deformation; S114, when At that time, calculate the improved tangent angle at the corresponding monitoring time. The maximum improved tangent angle of the known stable deformation stage is taken as the tangent angle boundary of the stable stage. The minimum improved tangent angle of the known accelerated deformation stage is taken as the boundary of the tangent angle of the accelerated stage. ; S115, when When, the corresponding labels of the preprocessed data are marked as accelerated deformation; when When, the corresponding preprocessed data is labeled as non-accelerated deformation; when At that time, the corresponding label of the preprocessed data is marked as transition.

[0011] Furthermore, methods for selecting the controlling environmental factors by constructing ROC curves of environmental factors include: S121. Obtain continuous historical values ​​of environmental factors. When a value is greater than or equal to a preset value, use it as a discrimination value. When a value is less than a preset value, use the opposite of the value as a discrimination value. Sort all discrimination values ​​of the same environmental factor in descending order. S122. Select different values ​​sequentially as discrimination thresholds, and determine the predicted labels of the discrimination values ​​of environmental factors based on the discrimination thresholds:

[0012] in, For environmental factors at the discrimination threshold The i-th discriminant value s i Predicted labels; S123, the discriminant value s i Predicted labels With real labels Compare and statistically analyze the results at the discrimination threshold. True positive number False positives True negative numbers and false negatives ; S124. Calculate the discrimination threshold. True positive rate and false positive rate: , Wherein, TPR(θ) and FPR(θ) are the true positive rate and false positive rate at the discrimination threshold θ, respectively; S125. Repeat steps S122 to S124 until all the discriminant values ​​of the same environmental factor have been used as the discriminant threshold. Use all (FPR(θ), TPR(θ)) points of the same environmental factor to plot its ROC curve. S126. Calculate the area under the ROC curve using numerical integration to obtain the AUC value of the environmental factor; use the area under each discrimination threshold... and Calculate sensitivity, and Specificity was calculated, and the Youden index was calculated using sensitivity and specificity. S127. Among all environmental factors, select those that meet the preset conditions in terms of AUC value, sensitivity, specificity, and Youden index as key environmental factors.

[0013] The beneficial effects of the above technical solution are as follows: This solution uses the method of constructing ROC curves of environmental factors to determine the controlling environmental factors, which can quantitatively evaluate the ability of each candidate environmental factor to distinguish between accelerated deformation samples and non-accelerated deformation samples. Through indicators such as AUC, sensitivity, specificity, and Youden index, the overall discrimination ability, hazard state identification ability, and false alarm control ability of candidate environmental factors can be comprehensively judged, thereby avoiding the subjectivity caused by simply relying on experience to select environmental factors. This method can prioritize the screening of environmental factors with strong discrimination ability for accelerated landslide deformation states, and eliminate factors with weak correlation or insufficient discrimination ability, providing more effective input features for subsequent convolutional neural network early warning discrimination models, improving model training efficiency and early warning discrimination accuracy.

[0014] Furthermore, the reservoir-type landslide classification and early warning method also includes optimizing key environmental factors: S128. According to different water storage stages, the continuous historical values ​​of each key environmental factor are divided into multiple stages. For each key environmental factor, steps S121 to S126 are executed in each stage to obtain the AUC value, Youden index, sensitivity and specificity. S129. Based on the Youden index and AUC value of the same key environmental factor at each water storage stage, calculate the mean Youden index and mean AUC value of the key environmental factor, as well as the coefficient of variation of AUC and Youden index at each water storage stage. When the coefficient of variation is less than the preset coefficient of variation, the key environmental factor is marked as having good stage stability. S1210. Compare the Youden index and AUC value of the same key environmental factor in all stages, select the largest Youden index, and if the AUC value of the stage in which the largest Youden index is located is greater than the set AUC value, then mark the key environmental factor as having the strongest applicability. S1211. Select key environmental factors whose AUC value, Youden index, sensitivity, and specificity all meet the overall discrimination conditions. From the remaining key environmental factors, select the key environmental factors with good stage stability and the strongest applicability. Use the selected key environmental factors as the final key environmental factors.

[0015] The beneficial effects of the above technical solution are as follows: By calculating the AUC, Youden index, sensitivity, and specificity of key environmental factors according to different water storage stages, this solution can further evaluate the discriminative ability of the factors at different water storage stages. This step can not only identify environmental factors that maintain stable discriminative ability across multiple water storage stages, but also discover environmental factors that have significant discriminative effects only at specific water storage stages, thereby avoiding misjudgments caused by simply selecting factors based on the overall ROC index of the entire sample. By comprehensively considering overall discriminative ability, stage stability, and stage applicability, the reliability of the screening results of the main control environmental factors can be further improved, providing more stable and effective input features for subsequent early warning discrimination models, and improving the accuracy and applicability of reservoir-type landslide graded early warning.

[0016] Furthermore, the method for determining the optimal response window and its weights for reservoir water levels using historical reservoir water levels further includes: Obtain historical reservoir water level monitoring sequences, set multiple candidate response windows, and for any candidate response window... Construct a cumulative index of reservoir water level changes within this time window. ; Calculated based on historical displacement monitoring data. The displacement rate corresponding to the day According to each candidate response window and Calculate the Pearson correlation coefficient between the two. The candidate response window corresponding to the maximum Pearson correlation coefficient is selected as the optimal response window. And calculate the current day and the previous day. The daily change in reservoir water level is then randomly assigned several weights to the change in reservoir water level. Calculate the improved effective reservoir water level for each set of weights. and according to corresponding and each group of weights Calculate the Pearson correlation coefficient between the two. The optimal response window is selected by choosing a set of weights corresponding to the maximum Pearson correlation coefficient. The weight within is used as the current day and previous days. The weight of daily changes in reservoir water level.

[0017] Furthermore, the cumulative index of reservoir water level changes The expression is: , in, For the first On the day Cumulative indicators of reservoir water level changes under the time window of the response; For the first Daily changes in reservoir water level; and The first Hiyori The reservoir water level; Improve effective reservoir water level The calculation formula is:

[0018] in, for The weight.

[0019] Furthermore, the risk levels include stable state, initial response state, significant acceleration state, and high-risk state, and the output of the early warning discrimination function includes two types: sensitive area and non-sensitive area. When the improved effective reservoir water level, displacement rate, and improved tangent angle of the reservoir-type slope under study are all in a stable state, and the output result of the early warning discrimination function is a non-sensitive area, the early warning level is Attention Level I; when at least one of the improved effective reservoir water level, displacement rate, and improved tangent angle of the reservoir-type slope under study is in an initial response state, and the output result of the early warning discrimination function is a non-sensitive area, the early warning level is Warning Level II. When at least two of the improved effective reservoir water level, displacement rate, and improved tangent angle of the reservoir-type slope under study are in a significantly accelerated state, and the output result of the early warning discrimination function is a sensitive area, the early warning level is Alert Level III; when at least two of the improved effective reservoir water level, displacement rate, and improved tangent angle of the reservoir-type slope under study are in a high-risk state, and the output result of the early warning discrimination function is a sensitive area, the early warning level is Alert Level IV.

[0020] Furthermore, the method for determining the risk level of the displacement rate, improved tangent angle, and improved effective reservoir water level includes: A1. When the displacement rate is lower than the first rate threshold and its rate of change is less than the first change threshold, it is set to a steady state; when the rate of change of the displacement rate is greater than the second change threshold, it is set to an initial response state; when the rate of change of the displacement rate is greater than the third change threshold, it is set to a significant acceleration state; when the rate of change of the displacement rate is greater than the fourth change threshold, it is set to a high-risk state. A2. When the improved tangent angle is lower than the first tangent angle threshold and its rate of change is less than the first tangent angle change threshold, it is set to a steady state; when the rate of change of the improved tangent angle is greater than the second tangent angle change threshold, it is set to an initial response state; when the rate of change of the improved tangent angle is greater than the third tangent angle change threshold, it is set to a significant acceleration state; when the rate of change of the improved tangent angle is greater than the fourth tangent angle change threshold, it is set to a high-risk state. A3. Obtain historical surface displacement monitoring data of the reservoir-type slope under study in the monitoring area, and classify the displacement rate and improved tangent angle corresponding to the surface displacement monitoring data to their corresponding risk levels; determine the boundary between the data distribution of two adjacent risk levels based on the data distribution of the four risk levels, and use it as the boundary threshold. A4. Classify the displacement rate and improved tangent angle according to the boundary threshold, and compare their true levels. When the proportion of high-level data being classified into low-level data is greater than the first set proportion, the corresponding boundary threshold is reduced by a preset proportion; when the proportion of low-level data being classified into high-level data is greater than the second set proportion, the corresponding boundary threshold is increased by a preset proportion. A5. Return to step A4 based on the adjusted boundary thresholds until the relationship between the divided levels and the true levels meets the first set ratio or the second set ratio. Then obtain the boundary thresholds for displacement rate and improved tangent angle, which are used to determine the risk level of displacement rate and improved tangent angle. A6. The risk level of historical samples is calibrated, and then the improved effective reservoir water level corresponding to the historical samples is calculated. The improved effective reservoir water level samples are divided into four groups according to the stable state, initial response state, obvious acceleration state and high-risk state to which the historical samples belong. A7. Statistically analyze the distribution characteristics of the improved effective reservoir water level in the four groups of samples, and determine the boundary threshold of the improved effective reservoir water level based on the mean or distribution boundary of the adjacent risk level sample groups; based on the boundary threshold, classify the improved effective reservoir water level into stable state, initial response state, obvious acceleration state and high risk state.

[0021] The beneficial effects of this invention are as follows: This solution can preliminarily predict the confidence level reflecting the accelerated deformation of the slope by acquiring data of the main environmental factors of the slope and combining them with a network model. By inputting the early warning discrimination function, it can preliminarily determine whether the slope has entered the deformation sensitive zone. Then, by combining the improved effective reservoir water level, displacement rate and the risk level of the improved tangent angle, the early warning level of landslides on reservoir-type slopes can be determined. This effectively reduces the uncertainty of single threshold indicators and the discrete interference of monitoring data, and significantly improves the accuracy and reliability of early warning of reservoir-type landslides, thereby improving the value of engineering applications.

[0022] This invention identifies the main controlling environmental factors of reservoir-type landslides under study by using multi-source monitoring data, avoiding the problems of fixed early warning input parameters and insufficient applicability in existing methods. Instead of directly using all environmental factors as model input, this approach first constructs ROC curves based on the historical values ​​and sample state labels of candidate environmental factors. Then, it evaluates the discriminative ability of each environmental factor on the accelerated deformation state of the slope using indicators such as AUC, sensitivity, specificity, and Youden index, thereby screening out the environmental factors that have a major controlling effect on landslide deformation. This method reduces the interference of weakly correlated or redundant factors on subsequent model training, making the input features more targeted and improving the adaptability of the early warning model to different reservoir-type slopes.

[0023] This scheme further optimizes the analysis of key environmental factors according to different water storage stages, which can improve the stability and engineering applicability of the screening results of the main control environmental factors. Since the deformation response characteristics of reservoir-type slopes may differ in the initial water storage, high water level stabilization, precipitation receding, and periodic fluctuations stages, screening environmental factors solely based on the overall discrimination ability of the entire sample easily overlooks the stage-specific differences in the effects of environmental factors. This scheme calculates the AUC, Youden index, sensitivity, and specificity of key environmental factors at different water storage stages, and further evaluates their stage stability and stage applicability. This allows for the screening of main control environmental factors that possess both overall discrimination ability and can maintain effective effects in specific or multiple water storage stages, thereby improving the reliability of the input features of subsequent early warning discrimination models.

[0024] This scheme utilizes convolutional neural networks to learn from continuous time-step data of key environmental factors, enabling a better characterization of the nonlinear relationship between temporal changes in environmental factors and accelerated slope deformation. Compared to methods that rely solely on environmental parameters or empirical thresholds at a single moment, this scheme extracts key environmental factor data from several consecutive time steps as model input. This allows the model to learn the temporal variation patterns of environmental factors such as reservoir water level, reservoir water level change rate, and rainfall, and their correspondence with accelerated slope deformation, thereby improving the ability to identify slope deformation-sensitive states.

[0025] This scheme constructs an improved effective reservoir water level index, which can characterize the cumulative effect of recent reservoir water level changes on current slope deformation. The deformation of reservoir-type slopes is often influenced not only by the current day's reservoir water level but also by the sustained and lagging effects of reservoir water level changes over several previous days. This scheme first determines the optimal response window by analyzing the correlation between the cumulative reservoir water level change index and the displacement rate. Then, within this response window, weights are assigned to the current day's and previous days' reservoir water level changes to form the improved effective reservoir water level. This quantifies the cumulative effect of recent reservoir water level changes into an index that can be used for early warning judgment, reflecting the actual response mechanism of reservoir-type landslides more effectively than simply using the current reservoir water level or a single day's reservoir water level change.

[0026] This scheme jointly assesses the slope risk status from multiple perspectives, including the output of the early warning discriminant function, improved effective reservoir water level, displacement rate, and improved tangent angle. Specifically, the early warning discriminant function determines whether the current combination of controlling environmental factors has entered a sensitive zone of accelerated deformation; the improved effective reservoir water level reflects the historical cumulative effect of reservoir water changes; the displacement rate characterizes the current deformation intensity of the slope; and the improved tangent angle characterizes the deformation stage of the slope. This multi-indicator joint assessment effectively reduces the risk of misjudgment caused by single threshold indicators being susceptible to monitoring noise, local anomalies, or short-term disturbances, thus improving the accuracy and reliability of early warning results for reservoir-type landslides.

[0027] This scheme categorizes displacement rate, improved tangent angle, and improved effective reservoir water level into stable, initial response, significant acceleration, and high-risk states, respectively. Combined with the sensitive area identification results of the early warning discrimination function, it outputs attention, warning, alert, and alarm levels, achieving a complete technical closed loop from monitoring data processing, main control factor screening, model discrimination, threshold determination to graded early warning output. This grading method provides clear judgment criteria and engineering implications for early warning results, facilitating managers to take corresponding monitoring intensification, on-site verification, risk management, or emergency prevention and control measures based on different early warning levels, thereby enhancing the engineering application value of reservoir-type landslide early warning methods. Attached Figure Description

[0028] Figure 1 This is a flowchart of a graded early warning method for reservoir-type landslides based on multi-source monitoring data.

[0029] Figure 2 This is a schematic diagram of a convolutional neural network (CNN) structure.

[0030] Figure 3 This is a graph showing the change in the accuracy of the model on the test set during the training of a convolutional neural network (CNN). Detailed Implementation

[0031] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0032] refer to Figure 1 , Figure 1 A flowchart of a graded early warning method for reservoir-type landslides based on multi-source monitoring data is shown; for example... Figure 1 As shown, the method S includes steps S1 to S4.

[0033] In step S1, the main environmental factor data of the reservoir-type slope to be studied are collected, input into the trained network model, and the output result of the early warning discrimination function of the slope is determined based on the confidence degree of accelerated deformation of the slope and the early warning discrimination function output by the model.

[0034] In implementation, the preferred expression for the early warning discrimination function in this scheme is: ,

[0035] in, For early warning discrimination function, Indicates the sensitive area. Indicates non-sensitive areas; The output of the network model for the controlling environmental factor data z; The confidence threshold; For Youden index; , , and ROC curves were constructed using the output probabilities from the network model validation phase as continuous discriminant values, and different candidate probability thresholds were applied. The statistical results show the number of true positives, false positives, true negatives, and false negatives.

[0036] In step S2, the optimal response window for determining the reservoir water level is determined using historical reservoir water levels. and its weight, to obtain the current day and previous days The historical reservoir water level of the day was used as a basis to determine the effective improved reservoir water level. In landslide early warning, numerical target variables form the basis for trend prediction, while categorical target variables are more suitable for determining the threshold of early warning indicators, especially when identifying critical indicators of key control factors. Therefore, after completing the joint calibration of historical sample states, constructing a candidate environmental factor set is crucial. Candidate environmental factors should include at least: the current reservoir water level, the rate of change of reservoir water level, the cumulative change of reservoir water level over multiple consecutive days, the cumulative rainfall, deformation monitoring parameters, and other environmental monitoring parameters that may affect the deformation of the target slope.

[0037] In one embodiment of the present invention, the method for determining the optimal response window and its weights for reservoir water levels using historical reservoir water levels further includes: Obtain historical reservoir water level monitoring sequences, set multiple candidate response windows, and for any candidate response window... Construct a cumulative index of reservoir water level changes within this time window. : The cumulative index of reservoir water level change The expression is: , in, For the first On the day Cumulative indicators of reservoir water level changes under the time window of the response; For the first Daily changes in reservoir water level; and The first Hiyori The reservoir water level; Calculated based on historical displacement monitoring data. The displacement rate corresponding to the day According to each candidate response window and Calculate the Pearson correlation coefficient between the two:

[0038] in, This represents the total number of candidate response windows used in the calculation. Candidate response window The mean of the cumulative index series of changes in the lower reservoir water level; The mean of the displacement rate sequence; Indicates candidate response window The correlation coefficient between the cumulative index of water level change in the lower reservoir and the displacement rate.

[0039] The candidate response window corresponding to the maximum Pearson correlation coefficient is selected as the optimal response window. And calculate the current day and the previous day. The daily change in reservoir water level is recorded, and then several weights are randomly assigned to the change in reservoir water level; the sum of all weights in each weight group is equal to 1, and each weight is greater than 0.

[0040] Calculate the improved effective reservoir water level for each set of weights. :

[0041] in, for The weight.

[0042] according to corresponding and each group of weights Calculate the Pearson correlation coefficient between the two:

[0043] in, Candidate response window Improve the mean of the cumulative index series of effective reservoir water level changes.

[0044] The optimal response window is selected by choosing a set of weights corresponding to the maximum Pearson correlation coefficient. The weight within is used as the current day and previous days. The weight of daily changes in reservoir water level.

[0045] In step S3, surface displacement monitoring data of the reservoir-type slope under study is obtained, the landslide displacement rate and improved tangent angle are calculated, and then the risk level of the improved effective reservoir water level, displacement rate, and improved tangent angle is determined; the formula for calculating the improved tangent angle is: , in, and Cumulative displacement , With average deformation rate The ratio; α i For improved tangent angle; and These are the i-th and i-1-th monitoring times, respectively.

[0046] In step S4, based on the output of the early warning discrimination function and the risk level of the improved effective reservoir water level, displacement rate, and improved tangent angle, the early warning level for landslides on the reservoir-type slope under study is determined.

[0047] In this scheme, the risk levels include stable state, initial response state, significant acceleration state, and high-risk state. The early warning discrimination function outputs two types of results: sensitive area and non-sensitive area.

[0048] Methods for determining the early warning level of landslides on reservoir-type slopes under study include: When the improved effective reservoir water level, displacement rate, and improved tangent angle of the reservoir-type slope under study are all in a stable state, and the output result of the early warning discrimination function is a non-sensitive area, the early warning level is Attention Level I; when at least one of the improved effective reservoir water level, displacement rate, and improved tangent angle of the reservoir-type slope under study is in an initial response state, and the output result of the early warning discrimination function is a non-sensitive area, the early warning level is Warning Level II. When at least two of the improved effective reservoir water level, displacement rate, and improved tangent angle of the reservoir-type slope under study are in a significantly accelerated state, and the output result of the early warning discrimination function is a sensitive area, the early warning level is Alert Level III; when at least two of the improved effective reservoir water level, displacement rate, and improved tangent angle of the reservoir-type slope under study are in a high-risk state, and the output result of the early warning discrimination function is a sensitive area, the early warning level is Alert Level IV.

[0049] In one embodiment of the present invention, the training method of the network model includes: S11. Obtain historical surface displacement monitoring data of the reservoir-type slope under study in the monitoring area, and preprocess the data, labeling the preprocessed data as accelerated deformation, non-accelerated deformation, and transition; in practice, this scheme preferably uses the following methods to label the preprocessed data as accelerated deformation and non-accelerated deformation: S111. Calculate the displacement rate at the monitoring time based on the preprocessed data. Detailed monitoring data for the monitored area can be found in Table 1.

[0050] Table 1. Detailed Monitoring Data of Deformation Zone

[0051] S112. Based on historically known stable deformation stages and accelerated deformation stages, the maximum displacement rate of the stable deformation stage is taken as the low-velocity boundary. The minimum displacement rate during the accelerated deformation stage is used as the high-speed boundary. ; S113, when When, the corresponding preprocessed data is labeled as non-accelerated deformation; when At that time, the corresponding labels of the preprocessed data are marked as accelerated deformation; S114, when At that time, calculate the improved tangent angle at the corresponding monitoring time. The maximum improved tangent angle of the known stable deformation stage is taken as the tangent angle boundary of the stable stage. The minimum improved tangent angle of the known accelerated deformation stage is taken as the boundary of the tangent angle of the accelerated stage. ; S115, when When, the corresponding labels of the preprocessed data are marked as accelerated deformation; when When, the corresponding preprocessed data is labeled as non-accelerated deformation; when At that time, the corresponding label of the preprocessed data is marked as transition.

[0052] Determining data labels can provide a more robust state label foundation for subsequent identification of master control environmental factors and training of convolutional neural networks.

[0053] S12. Construct a set of environmental factors, and remove environmental factors whose correlation with the deformation of the reservoir-type slope under study is lower than the set correlation or whose redundancy is higher than the set redundancy. It is preferable to use correlation coefficient, mutual information, variance contribution rate or collinearity analysis methods. Then, select the main controlling environmental factors by constructing ROC curves of environmental factors.

[0054] In one embodiment of the present invention, the method for selecting the controlling environmental factor by constructing the ROC curve of the environmental factor includes: S121. Obtain continuous historical values ​​of environmental factors. When a value is greater than or equal to a preset value, use it as a discrimination value. When a value is less than a preset value, use the opposite of the value as a discrimination value. Sort all discrimination values ​​of the same environmental factor in descending order. S122. Select different values ​​sequentially as discrimination thresholds, and determine the predicted labels of the discrimination values ​​of environmental factors based on the discrimination thresholds:

[0055] in, For environmental factors at the discrimination threshold The i-th discriminant value s i Predicted labels; S123, the discriminant value s i Predicted labels With real labels Compare and statistically analyze the results at the discrimination threshold. True positive number False positives True negative numbers and false negatives ; S124. Calculate the discrimination threshold. True positive rate and false positive rate: , Wherein, TPR(θ) and FPR(θ) are the true positive rate and false positive rate at the discrimination threshold θ, respectively; S125. Repeat steps S122 to S124 until all the discriminant values ​​of the same environmental factor have been used as the discriminant threshold. Use all (FPR(θ), TPR(θ)) points of the same environmental factor to plot its ROC curve. S126. Calculate the area under the ROC curve using numerical integration to obtain the AUC value of the environmental factor; use the area under each discrimination threshold... and Calculate sensitivity, and Specificity was calculated, and the Youden index was calculated using sensitivity and specificity. Among them, the sensitivity is Specificity is The Youden index is .

[0056] S127. Among all environmental factors, select those that meet the preset conditions in terms of AUC value, sensitivity, specificity, and Youden index as key environmental factors.

[0057] The preset conditions are as follows: firstly, candidate environmental factors with low AUC values, small Youden indexes, and poor sensitivity and specificity are eliminated, because such factors have a weak ability to distinguish between accelerated and non-accelerated deformation states; candidate environmental factors with high AUC values, large Youden indexes, and relatively balanced sensitivity and specificity are retained, because such factors have a strong overall discrimination ability.

[0058] To further ensure the accuracy of the selected key environmental factors, the reservoir-type landslide classification and early warning method also includes optimizing key environmental factors: S128. According to different water storage stages, the continuous historical values ​​of each key environmental factor are divided into multiple stages. For each key environmental factor, steps S121 to S126 are performed in each stage to obtain the AUC value, Youden index, sensitivity and specificity. This scheme preferably includes four stages, namely the initial water storage stage, the high water level stabilization stage, the precipitation receding stage and the periodic fluctuation stage.

[0059] S129. Based on the Youden index and AUC value of the same key environmental factor at each water storage stage, calculate the mean Youden index and mean AUC value of the key environmental factor, as well as the coefficient of variation of AUC and Youden index at each water storage stage. When the coefficient of variation is less than the preset coefficient of variation, the key environmental factor is marked as having good stage stability. The expressions for the mean Youden index and mean AUC of key environmental factors are: , in, and These are the mean AUC index and the mean Youden value of the key environmental factors, respectively. and These represent the AUC value and Youden index of the key environmental factors at the k-th impoundment stage, respectively. This represents the total number of water storage stages.

[0060] Coefficients of variation of AUC and Youden index at each stage and The expressions are as follows: , in, and These represent the standard deviations of the AUC and Youden index for this key environmental factor at each water storage stage, respectively.

[0061] S1210. Compare the Youden index and AUC value of the same key environmental factor in all stages, select the largest Youden index, and if the AUC value of the stage in which the largest Youden index is located is greater than the set AUC value, then mark the key environmental factor as having the strongest applicability. S1211. Select key environmental factors whose AUC value, Youden index, sensitivity, and specificity all meet the overall discrimination criteria. From the remaining key environmental factors, select those with good stage stability and the strongest applicability. Use these selected key environmental factors as the final key environmental factors. The overall discrimination criteria are: high AUC across all control environmental factors, a large Youden index, and relatively balanced sensitivity and specificity.

[0062] Based on the above analysis and judgment, the set of main environmental factors controlling the target slope was determined and sorted according to their comprehensive discrimination ability. In the subsequent training of the network model, the key environmental factors ranked higher were given priority for model training.

[0063] S13. Select K main environmental factors, extract their data for several consecutive time steps as model input, and use the sample state label at the end of several consecutive time steps as model output label to construct a dataset. Suppose that two main environmental factors are selected for model training: reservoir water level and reservoir water level change rate. During training, when the model judges whether "accelerated deformation will occur today", it does not look at today's data, but rather the data from the previous few days. Suppose that the data from the previous 5 days is used. The reservoir water level and reservoir water level change rate are shown in Table 2 below.

[0064] Table 2 Reservoir water level and rate of change of reservoir water level

[0065] Then, organize the data from the previous few days into a table / array that the model can understand, according to a fixed format. For example, the data from the previous 5 days and 2 factors can be organized into a 5×2 data block.

[0066] If day 5 is identified as "accelerated deformation," then after inputting the data from the previous 4 days, the goal is to obtain the result for day 5 as "accelerated deformation" as much as possible. Therefore, the data of the master control environmental factors for 5 consecutive days are used as input, and the monitoring data and labels of day 5 are used as output. The convolutional neural network learns the pattern through a large number of such "input + output" data. The trained model outputs the probability of entering the accelerated deformation stage, i.e., the confidence level.

[0067] S14. Divide the dataset into a training set and a validation set, and use the training set to train the convolutional neural network; for example... Figure 2 As shown, the convolutional neural network includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The convolutional layer is used to extract local features from the temporal changes of the driving factors of the master environment. The pooling layer is used to reduce noise interference and compress feature dimensions. The fully connected layer is used to synthesize the extracted feature information. The output layer is used to provide the confidence distribution of the target slope entering the accelerated deformation stage under the current combination of master environmental factors.

[0068] During convolutional neural network training, the cross-entropy loss function is preferred for measuring classification error. The Adam optimizer is used for parameter updates, and a Bayesian optimization algorithm is combined to fine-tune the kernel size, learning rate, dropout rate, and regularization parameters. L2 regularization and early stopping mechanisms are also introduced to suppress overfitting and improve the model's generalization ability. Through this process, an early warning discrimination model for target reservoir-type slopes under different combinations of driving factors in the main control environment can be established, providing a foundation for subsequent probability threshold determination and sensitive area identification. The accuracy changes of the model on the test set during the overall convolutional neural network training process can be referenced... Figure 3 .

[0069] S15. Use a validator to validate the convolutional neural network and determine whether the training accuracy of the convolutional neural network has reached the preset accuracy. If yes, obtain the trained network model; otherwise, let K=K+1 and return to step S13.

[0070] During implementation, the preferred methods for determining the risk level of displacement rate, improved tangent angle, and improved effective reservoir water level in this scheme include: A1. When the displacement rate is lower than the first rate threshold and its rate of change is less than the first change threshold, it is set to a steady state; when the rate of change of the displacement rate is greater than the second change threshold, it is set to an initial response state; when the rate of change of the displacement rate is greater than the third change threshold, it is set to a significant acceleration state; when the rate of change of the displacement rate is greater than the fourth change threshold, it is set to a high-risk state. A2. When the improved tangent angle is lower than the first tangent angle threshold and its rate of change is less than the first tangent angle change threshold, it is set to a steady state; when the rate of change of the improved tangent angle is greater than the second tangent angle change threshold, it is set to an initial response state; when the rate of change of the improved tangent angle is greater than the third tangent angle change threshold, it is set to a significant acceleration state; when the rate of change of the improved tangent angle is greater than the fourth tangent angle change threshold, it is set to a high-risk state. A3. Obtain historical surface displacement monitoring data of the reservoir-type slope under study in the monitoring area, and classify the displacement rate and improved tangent angle corresponding to the surface displacement monitoring data to their corresponding risk levels; determine the boundary between the data distribution of two adjacent risk levels based on the data distribution of the four risk levels, and use it as the boundary threshold. A4. Classify the displacement rate and improved tangent angle according to the boundary threshold, and compare their true levels. When the proportion of high-level data being classified into low-level data is greater than the first set proportion, the corresponding boundary threshold is reduced by a preset proportion; when the proportion of low-level data being classified into high-level data is greater than the second set proportion, the corresponding boundary threshold is increased by a preset proportion. A5. Return to step A4 based on the adjusted boundary thresholds until the relationship between the divided levels and the true levels meets the first set ratio or the second set ratio. Then obtain the boundary thresholds for displacement rate and improved tangent angle, which are used to determine the risk level of displacement rate and improved tangent angle. A6. The risk level of historical samples is calibrated, and then the improved effective reservoir water level corresponding to the historical samples is calculated. The improved effective reservoir water level samples are divided into four groups according to the stable state, initial response state, obvious acceleration state and high-risk state to which the historical samples belong. A7. Statistically analyze the distribution characteristics of the improved effective reservoir water level in the four groups of samples, and determine the boundary threshold of the improved effective reservoir water level based on the mean or distribution boundary of the adjacent risk level sample groups; based on the boundary threshold, classify the improved effective reservoir water level into stable state, initial response state, obvious acceleration state and high risk state.

[0071] In summary, by combining the early warning discrimination function, displacement rate, improved tangent angle, and improved effective reservoir water level, a graded early warning output for landslides can be achieved, improving the reliability of the early warning results and their engineering application value.

Claims

1. A graded early warning method for reservoir-type landslides based on multi-source monitoring data, characterized in that, Including the following steps: S1. Collect the main environmental factor data of the reservoir-type slope to be studied, input them into the trained network model, and determine the output result of the early warning discrimination function of the slope based on the confidence degree of accelerated deformation and the early warning discrimination function of the model output. S2. Determine the optimal response window for reservoir water level using historical reservoir water levels. and its weight, to obtain the current day and previous days The historical reservoir water level of the day was used as a basis to determine the effective improved reservoir water level. S3. Obtain surface displacement monitoring data of the reservoir-type slope to be studied, calculate the displacement rate and improved tangent angle of the landslide, and then determine the risk level of the improved effective reservoir water level, displacement rate and improved tangent angle. S4. Based on the output results of the early warning discrimination function and the risk level of the improved effective reservoir water level, displacement rate, and improved tangent angle, determine the early warning level for landslides on the reservoir-type slope under study.

2. The graded early warning method for reservoir-type landslides according to claim 1, characterized in that, The expression for the early warning discrimination function is: ， ； ； in, For early warning discrimination function, Indicates the sensitive area. Indicates non-sensitive areas; The output of the network model for the controlling environmental factor data z; The confidence threshold; For Youden index; , , and ROC curves were constructed using the output probabilities from the network model validation phase as continuous discriminant values, and different candidate probability thresholds were applied. The statistical results show the number of true positives, false positives, true negatives, and false negatives.

3. The graded early warning method for reservoir-type landslides according to claim 1, characterized in that, The training methods for the network model include: S11. Obtain historical surface displacement monitoring data of the reservoir-type slope under study in the monitoring area, and preprocess it. Label the preprocessed data as accelerated deformation, non-accelerated deformation, and transition. S12. Construct a set of environmental factors, remove environmental factors whose correlation with the deformation of the reservoir-type slope under study is lower than the set correlation or whose redundancy is higher than the set redundancy, and then select the main controlling environmental factors by constructing the ROC curve of the environmental factors. S13. Select K main environmental factors, extract their data for several consecutive time steps as model input, and use the sample state label at the end of several consecutive time steps as model output label to construct a dataset. S14. Divide the dataset into a training set and a validation set, and use the training set to train the convolutional neural network; S15. Use a validator to validate the convolutional neural network and determine whether the training accuracy of the convolutional neural network has reached the preset accuracy. If yes, obtain the trained network model; otherwise, let K=K+1 and return to step S13.

4. The graded early warning method for reservoir-type landslides according to claim 3, characterized in that, Methods for labeling preprocessed data as accelerated or non-accelerated deformation include: S111. Calculate the displacement rate at the monitoring time based on the preprocessed data. ; S112. Based on historically known stable deformation stages and accelerated deformation stages, the maximum displacement rate of the stable deformation stage is taken as the low-velocity boundary. The minimum displacement rate during the accelerated deformation stage is used as the high-speed boundary. ; S113, when When, the corresponding preprocessed data is labeled as non-accelerated deformation; when At that time, the corresponding labels of the preprocessed data are marked as accelerated deformation; S114, when At that time, calculate the improved tangent angle at the corresponding monitoring time. The maximum improved tangent angle of the known stable deformation stage is taken as the tangent angle boundary of the stable stage. The minimum improved tangent angle of the known accelerated deformation stage is taken as the boundary of the tangent angle of the accelerated stage. ; S115, when When, the corresponding labels of the preprocessed data are marked as accelerated deformation; when When, the corresponding preprocessed data is labeled as non-accelerated deformation; when At that time, the corresponding label of the preprocessed data is marked as transition.

5. The graded early warning method for reservoir-type landslides according to claim 3, characterized in that, Methods for selecting the controlling environmental factors by constructing ROC curves of environmental factors include: S121. Obtain continuous historical values ​​of environmental factors. When a value is greater than or equal to a preset value, use it as a discrimination value. When a value is less than a preset value, use the opposite of the value as a discrimination value. Sort all discrimination values ​​of the same environmental factor in descending order. S122. Select different values ​​sequentially as discrimination thresholds, and determine the predicted labels of the discrimination values ​​of environmental factors based on the discrimination thresholds: ； in, For environmental factors at the discrimination threshold The i-th discriminant value s i Predicted labels; S123, the discriminant value s i Predicted labels With real labels Compare and statistically analyze the results at the discrimination threshold. True positive number False positives True negative numbers and false negatives ; S124. Calculate the discrimination threshold. True positive rate and false positive rate: ， ； Wherein, TPR(θ) and FPR(θ) are the true positive rate and false positive rate at the discrimination threshold θ, respectively; S125. Repeat steps S122 to S124 until all the discriminant values ​​of the same environmental factor have been used as the discriminant threshold. Use all (FPR(θ), TPR(θ)) points of the same environmental factor to plot its ROC curve. S126. Calculate the area under the ROC curve using numerical integration to obtain the AUC value of the environmental factor; use the area under each discrimination threshold... and Calculate sensitivity, and Specificity was calculated, and the Youden index was calculated using sensitivity and specificity. S127. Among all environmental factors, select those that meet the preset conditions in terms of AUC value, sensitivity, specificity, and Youden index as key environmental factors.

6. The graded early warning method for reservoir-type landslides according to claim 5, characterized in that, This also includes optimizing key environmental factors: S128. According to different water storage stages, the continuous historical values ​​of each key environmental factor are divided into multiple stages. For each key environmental factor, steps S121 to S126 are executed in each stage to obtain the AUC value, Youden index, sensitivity and specificity. S129. Based on the Youden index and AUC value of the same key environmental factor at each water storage stage, calculate the mean Youden index and mean AUC value of the key environmental factor, as well as the coefficient of variation of AUC and Youden index at each water storage stage. When the coefficient of variation is less than the preset coefficient of variation, the key environmental factor is marked as having good stage stability. S1210. Compare the Youden index and AUC value of the same key environmental factor in all stages, select the largest Youden index, and if the AUC value of the stage in which the largest Youden index is located is greater than the set AUC value, then mark the key environmental factor as having the strongest applicability. S1211. Select key environmental factors whose AUC value, Youden index, sensitivity, and specificity all meet the overall discrimination conditions. From the remaining key environmental factors, select the key environmental factors with good stage stability and the strongest applicability. Use the selected key environmental factors as the final key environmental factors.

7. The method for graded early warning of reservoir-type landslides according to claim 1, characterized in that, The method for determining the optimal response window and its weights for reservoir water levels using historical reservoir water levels further includes: Obtain historical reservoir water level monitoring sequences, set multiple candidate response windows, and for any candidate response window... Construct a cumulative index of reservoir water level changes within this time window. ; Calculated based on historical displacement monitoring data. The displacement rate corresponding to the day According to each candidate response window and Calculate the Pearson correlation coefficient between the two. The candidate response window corresponding to the maximum Pearson correlation coefficient is selected as the optimal response window. And calculate the current day and the previous day. The daily change in reservoir water level is then randomly assigned several weights to the change in reservoir water level. Calculate the improved effective reservoir water level for each set of weights. and according to corresponding and each group of weights Calculate the Pearson correlation coefficient between the two. The optimal response window is selected by choosing a set of weights corresponding to the maximum Pearson correlation coefficient. The weight within is used as the current day and previous days. The weight of daily changes in reservoir water level.

8. The method for graded early warning of reservoir-type landslides according to claim 7, characterized in that, The cumulative index of reservoir water level change The expression is: ， ； in, For the first On the day Cumulative indicators of reservoir water level changes under the time window of the response; For the first Daily changes in reservoir water level; and The first Hiyori The reservoir water level; Improve effective reservoir water level The calculation formula is: ； in, for The weight.

9. The graded early warning method for reservoir-type landslides according to claim 1, characterized in that, The risk levels include stable state, initial response state, obvious acceleration state and high risk state. The output of the early warning discrimination function includes two types: sensitive area and non-sensitive area. When the improved effective reservoir water level, displacement rate, and improved tangent angle of the reservoir-type slope under study are all in a stable state, and the output result of the early warning discrimination function is a non-sensitive area, the early warning level is attention level I. When at least one of the improved effective reservoir water level, displacement rate and improved tangent angle of the reservoir-type slope under study is in the initial response state, and the output result of the early warning discrimination function is a non-sensitive area, the early warning level is warning level II. When at least two of the improved effective reservoir water level, displacement rate and improved tangent angle of the reservoir-type slope under study are in a state of significant acceleration, and the output result of the early warning discrimination function is a sensitive area, the early warning level is Alert Level III. When at least two of the improved effective reservoir water level, displacement rate, and improved tangent angle of the reservoir-type slope under study are in a high-risk state, and the output result of the early warning discrimination function is a sensitive area, the early warning level is alarm level IV.

10. The graded early warning method for reservoir-type landslides according to claim 9, characterized in that, The methods for determining the risk level of displacement rate, improved tangent angle, and improved effective reservoir water level include: A1. When the displacement rate is lower than the first rate threshold and its rate of change is less than the first change threshold, it is set to a steady state; when the rate of change of the displacement rate is greater than the second change threshold, it is set to an initial response state; when the rate of change of the displacement rate is greater than the third change threshold, it is set to a significant acceleration state; when the rate of change of the displacement rate is greater than the fourth change threshold, it is set to a high-risk state. A2. When the improved tangent angle is lower than the first tangent angle threshold and its rate of change is less than the first tangent angle change threshold, it is set to a steady state; when the rate of change of the improved tangent angle is greater than the second tangent angle change threshold, it is set to an initial response state; when the rate of change of the improved tangent angle is greater than the third tangent angle change threshold, it is set to a significant acceleration state; when the rate of change of the improved tangent angle is greater than the fourth tangent angle change threshold, it is set to a high-risk state. A3. Obtain historical surface displacement monitoring data of the reservoir-type slope under study in the monitoring area, and classify the displacement rate and improved tangent angle corresponding to the surface displacement monitoring data to their corresponding risk levels; determine the boundary between the data distribution of two adjacent risk levels based on the data distribution of the four risk levels, and use it as the boundary threshold. A4. Classify the displacement rate and improved tangent angle according to the boundary threshold, and compare their true levels. When the proportion of high-level data being classified into low-level data is greater than the first set proportion, the corresponding boundary threshold is reduced by a preset proportion; when the proportion of low-level data being classified into high-level data is greater than the second set proportion, the corresponding boundary threshold is increased by a preset proportion. A5. Return to step A4 based on the adjusted boundary thresholds until the relationship between the divided levels and the true levels meets the first set ratio or the second set ratio. Then obtain the boundary thresholds for displacement rate and improved tangent angle, which are used to determine the risk level of displacement rate and improved tangent angle. A6. The risk level of historical samples is calibrated, and then the improved effective reservoir water level corresponding to the historical samples is calculated. The improved effective reservoir water level samples are divided into four groups according to the stable state, initial response state, obvious acceleration state and high-risk state to which the historical samples belong. A7. Statistically analyze the distribution characteristics of the improved effective reservoir water level in the four groups of samples, and determine the boundary threshold of the improved effective reservoir water level based on the mean or distribution boundary of the adjacent risk level sample groups; based on the boundary threshold, classify the improved effective reservoir water level into stable state, initial response state, obvious acceleration state and high risk state.

Images