Dynamic prediction method of landslide instability time based on machine learning and medium

Abstract

The present invention provides a method for dynamic prediction of landslide instability time based on machine learning, including: preparing original monitoring data of the original landslide, including landslide surface displacement monitoring and landslide crack meter monitoring; using the moving average method to preprocess the data; using the improved tangent angle method to determine the landslide evolution stage of the preprocessed monitoring data; setting a dynamic prediction calculation interval according to the target landslide monitoring period; when it is determined that the landslide has entered the accelerated motion stage, according to the set dynamic interval, predicting the landslide instability time through the basic prediction model; taking the basic prediction model as input and the actual landslide instability time as the target output, using machine learning to integrate the basic prediction model, establishing a comprehensive prediction model for dynamic prediction of landslide instability time, and dynamically predicting the landslide instability time. The accuracy and stability of the instability time prediction are significantly improved, providing a more scientific, intuitive and accurate reference for landslide early warning work.

Description

Technical Field

[0001] The present invention relates to the field of landslide early warning and forecasting, and in particular to a method and medium for dynamic landslide instability time forecasting based on machine learning. Background Art

[0002] Landslide is a common geological disaster that causes a large number of casualties and property losses every year. Therefore, landslide disaster prevention and control is particularly important. Accurate and efficient landslide warning can effectively prevent and control landslide geological disasters. Landslide forecasting is an important part of landslide warning, which mainly includes forecasts on spatial scales and forecasts on time scales. The forecast on time scale mainly refers to the forecast of landslide instability time, which can provide a more direct and reliable reference for landslide warning.

[0003] Generally speaking, the kinematic parameters of landslides, such as the accumulated displacement, velocity (first-order difference), and acceleration (second-order difference), can directly reflect the evolution characteristics of landslides. Therefore, the landslide instability time prediction model mainly uses the kinematic parameters of landslides. At present, a variety of landslide instability time prediction models have been developed at home and abroad, but the current landslide instability time prediction models have the problem of non-uniform calculation intervals. Some models perform well in a specific calculation interval, but landslide evolution is a continuous dynamic process. For the actual monitoring situation on site, the non-uniform calculation interval or the specific calculation interval is not friendly and intuitive for landslide early warning work. In addition, for the landslide instability time prediction model, the prediction rules of different models at different stages of landslide acceleration movement are different. It is necessary to adopt certain means to make full use of the prediction advantages of the basic model at different stages of landslide acceleration movement, and establish a comprehensive prediction model to improve the accuracy and stability of instability time prediction. Summary of the invention

[0004] The purpose of the present invention is to provide a dynamic calculation interval for landslide instability time prediction that meets on-site monitoring conditions and is friendly and intuitive for landslide early warning work, in response to the deficiencies in the above-mentioned background technology, and to integrate the basic landslide instability time prediction model using a machine learning method, which has certain theoretical innovation significance and practical engineering value.

[0005] In order to achieve the above object, the present invention provides a method for dynamic prediction of landslide instability time based on machine learning, comprising the following steps:

[0006] S1, prepare the original monitoring data of the original landslide, including landslide surface displacement monitoring and landslide crack meter monitoring;

[0007] S2, preprocessing the original monitoring data using the moving average method;

[0008] S3, using the tangent angle method to determine the landslide evolution stage based on the pre-processed monitoring data;

[0009] S4, setting the dynamic forecast calculation interval according to the target landslide monitoring period;

[0010] S5, when it is determined that the landslide has entered the accelerated movement stage, the landslide instability time is predicted through the basic prediction model according to the set dynamic range;

[0011] S6, taking the basic forecast model as input and the actual instability time of the landslide as the target output, uses machine learning to integrate the basic forecast model and establish a comprehensive forecast model for dynamic forecast of the landslide instability time to dynamically forecast the landslide instability time.

[0012] Furthermore, the original landslide whose original monitoring data is obtained in S1 is a landslide that has already become unstable and has the same properties as the target landslide.

[0013] Furthermore, the moving average method adopted in S2 selects the filter order according to the monitoring period of the original landslide for obtaining the original monitoring data.

[0014] Furthermore, the dynamic calculation interval in S4 takes the landslide entering the accelerated movement stage as the starting point, and a forecast is made after each landslide monitoring cycle to obtain a dynamic forecast result of the landslide instability time. The forecast result will be updated as the monitoring data is updated.

[0015] Furthermore, the basic prediction models for landslide instability time in S5 include the Verhulst model, the GM (1,1) model, and the Saito Toshitaka model.

[0016] Furthermore, when the basic model for predicting the instability time adopts the Verhulst model or the GM (1,1) model, if the original monitoring data is the cumulative displacement, the data is differentiated to obtain the matrix of the coefficients of the white differential equation solution, and the original data is also differentiated when judging the landslide evolution.

[0017] Furthermore, when the basic model for predicting the instability time adopts the Saito Toshitaka model, the instability time calculation process of this model needs to obtain three points with the same displacement spacing, which are obtained by interpolation method.

[0018] Furthermore, the machine learning in S6 is one of artificial neural network, support vector machine regression or decision tree regression.

[0019] Furthermore, the decision tree regression with the best forecasting effect is selected to integrate the basic forecasting models to establish a comprehensive forecasting model to predict the instability time of landslides.

[0020] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for dynamic landslide instability time prediction based on machine learning as described above.

[0021] The above scheme of the present invention has the following beneficial effects:

[0022] The present invention is based on the calculation interval of the dynamic prediction of the landslide instability time under the on-site monitoring situation, and can make an instability time prediction result after each monitoring cycle, fully showing the dynamic prediction law of the basic prediction model of the landslide instability time in the landslide acceleration movement stage, and integrates the prediction results of the basic prediction model by machine learning, making full use of the prediction advantages of different basic prediction models in the landslide acceleration movement stage, significantly improving the accuracy and stability of the instability time prediction, and providing a more scientific, intuitive and accurate reference for landslide early warning work;

[0023] Other beneficial effects of the present invention will be described in detail in the subsequent specific implementation section. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 is a flow chart of the steps of the present invention;

[0025] Figure 2 A diagram for setting the calculation interval of the present invention;

[0026] Figure 3 The original landslide monitoring point distribution map of Case 1 in the present invention;

[0027] Figure 4 It is the original landslide monitoring data diagram of Case 1 in the present invention;

[0028] Figure 5 The data preprocessing of the original landslide and the determination results of the landslide evolution stage in Case 1 of the present invention (two monitoring points are selected as examples);

[0029] Figure 6 The instability time prediction result of the original landslide foundation prediction model of Case 1 in the present invention (two monitoring points are selected as examples);

[0030] Figure 7 The instability time prediction result of the original landslide comprehensive prediction model in Case 1 of the present invention (two monitoring points in the test set are selected as an example);

[0031] Figure 8 The program calculation results of the same type of target landslide in the adjacent area of ​​Case 2 in the present invention. DETAILED DESCRIPTION

[0032] In order to make the technical problems, technical solutions and advantages to be solved by the present invention clearer, the following will be described in detail with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0033] In the description of the present invention, for simplicity of explanation, the method or rule is depicted or described as a series of operations, and its purpose is neither to exhaustively list the experimental operations nor to limit the order of the experimental operations. For example, the experimental operations can be performed in various orders and / or simultaneously, and include other experimental operations that are not described again. In addition, the steps described are not all necessary for the methods and algorithms described herein. Those skilled in the art will recognize and understand that these methods and algorithms can be represented as a series of unrelated states through state diagrams or projects.

[0034] See also Figure 1 The embodiment of the present invention provides a method for dynamic prediction of landslide instability time based on machine learning, which specifically includes the following steps:

[0035] S1, prepare the original monitoring data of the original landslide, including landslide surface displacement monitoring and landslide crack meter monitoring.

[0036] It should be noted that the original landslide from which the original monitoring data is obtained must be an unstable landslide with the same properties as the target landslide, which can improve the comprehensive forecasting effect.

[0037] It should also be noted that the selection of the original data monitoring points in this embodiment has little impact on the final forecast result.

[0038] S2, the original monitoring data is preprocessed using the moving average method.

[0039] It should be noted that the selection of the filter order of the moving average method used in this embodiment is mainly based on the monitoring period of the original landslide from which the original monitoring data is obtained. Generally, the time scale included in the order is about three to seven days.

[0040] S3, the improved tangent angle method is used to determine the landslide evolution stage based on the preprocessed monitoring data.

[0041] For a certain landslide displacement S-time t monitoring curve, the rate of the uniform motion stage is relatively constant. The S axis can be transformed into the same time dimension as the t axis by dividing the cumulative displacement S by the rate v of the uniform motion stage:

[0042]

[0043] Where: T(i) is the ordinate value after dimension transformation; S(i) is the cumulative displacement of the landslide at time i, and v is the velocity of the landslide in the uniform motion stage. After the time dimension transformation is completed, the tangent angle of the improved T(i)-t curve can be obtained:

[0044]

[0045] According to the above calculation results, since the S-axis of the St displacement time curve is the same as the t-axis after dimension transformation, when the landslide is in the uniform motion stage, its α i The calculated result is equal to 45°; when the landslide is in the initial movement stage, α i The calculated result is less than 45°; when the landslide is in the acceleration stage, α i The calculated result is greater than 45°.

[0046] S4, setting the dynamic forecast calculation interval according to the target landslide monitoring period.

[0047] It should be noted that the dynamic calculation interval in this embodiment is mainly based on the time when the landslide enters the accelerated motion stage. A forecast is performed after each landslide monitoring cycle to obtain a dynamic forecast result of the landslide instability time. The forecast result will be updated as the monitoring data is updated. Figure 2 shown.

[0048] The prediction law of the basic prediction model is analyzed at different stages of landslide acceleration movement (mainly including the initial acceleration stage, the middle acceleration stage and the accelerated stage). The prediction law can be used to reflect the current stage of landslide acceleration movement. It can also be integrated according to the idea of ​​model stacking to establish a comprehensive prediction model, give full play to the prediction advantages of different types of models in different movement stages, and realize accurate and stable dynamic prediction of instability time throughout the acceleration movement, thereby providing a more reliable early warning reference.

[0049] S5, when it is determined that the landslide has entered the accelerated movement stage, the landslide instability time is predicted through the basic prediction model according to the set dynamic range.

[0050] In this embodiment, the basic prediction models for instability time include the Verhulst model, the GM (1,1) model, and the Saito Tokusaka model.

[0051] It should be noted that when the basic model for instability time prediction adopts the Verhulst model or the GM (1,1) model, if the original monitoring data is the cumulative displacement, it is necessary to differentiate the data to obtain the matrix of the coefficients for solving the white differential equation, and the original data also need to be differentiated when judging the landslide evolution. Generally, the prediction conditions are met when the monitoring period of the acceleration movement stage of the original monitoring data of the original landslide is greater than three periods.

[0052] When the basic model for instability time prediction adopts the Saito Toshitaka model, the instability time calculation process of this model needs to obtain three points with the same displacement spacing, which can be obtained by interpolation method, and the instability time prediction effect is very good in actual verification.

[0053] S6, taking the basic forecast model as input and the actual instability time of the landslide as the target output, machine learning is used to integrate the basic forecast model, and a comprehensive forecast model for dynamic forecast of the landslide instability time is established to dynamically forecast the landslide instability time.

[0054] Among them, the machine learning adopted includes support vector machine regression, artificial neural network and decision tree regression, etc. After calculating the relevant evaluation indicators of time series analysis, the machine learning method with the best forecasting effect is selected after comparison to integrate the basic forecasting model. The machine learning method of decision tree regression with the best forecasting effect is shown here. The principle of decision tree regression is to make decisions based on the tree structure. A decision regression tree contains a root node (input of the data set), several leaf nodes (target output of the data set) and several internal nodes. The present invention adopts Decisiontree regression based on the skikit-learn framework. During the training process, the input data of the root node is the input of the sample set data, and the leaf node is the target output of the sample set data. Each node divides the input feature plane into several units according to the judgment of yes and no. The forecast result of each input data is equal to the specific output of the corresponding divided area. The division rule is to find the best division point in the input feature so that the error between the predicted instability time of the training set and the actual instability is minimized. Its mathematical form is as follows:

[0055]

[0056] Wherein, j, s represent the division point s of the jth variable, R1(j, s) represents the area on the left side of the division, R2(j, s) represents the area on the right side of the division, c1 and c2 represent the mean values ​​of the dependent variables in each area. The dependent variables of the regression tree training process of the present invention are 3, which are the instability time prediction results of the 3 basic models, y i is the actual instability time result at each time node.

[0057] After the training of the Decisiontree regression model based on the skikit-learn framework is completed, the Joblib toolkit is called for encapsulation and the computer program stored in the computer-readable storage medium of the present invention is called. The data preparation and training process are as follows:

[0058] (1) Selecting original landslide monitoring data, wherein the data of each monitoring point is a vector, after data preprocessing, determining the landslide evolution stage, importing the landslide acceleration movement stage into the selected basic prediction model to obtain the basic model instability time prediction result, and storing the prediction result of each basic model as a vector. The present invention selects three basic instability time prediction models, with a total of three vectors, which constitute the input of the sample set data;

[0059] (2) Based on the actual occurrence time of the original landslide, the time from each time node in the acceleration phase of the landslide to the actual occurrence time of the landslide is reversed and stored as a vector to constitute the target output of the sample set data.

[0060] (3) Divide the sample data into a certain proportion of training set and test set, retrieve the training set data, train it according to the selected machine learning method and test it in the test set, calculate the time series analysis related evaluation indicators (such as RMSE) for comparison, and select the optimal model.

[0061] The method is based on the dynamic prediction calculation interval of the landslide instability time of the on-site monitoring situation, and can make an instability time prediction result after each monitoring cycle, fully demonstrating the dynamic prediction law of the basic prediction model of the landslide instability time in the landslide acceleration movement stage, that is, each basic prediction model has different prediction advantages in different stages of the landslide acceleration movement. The Verhulst model and GM (1,1) model used in the present invention have better prediction results in the early and middle stages of the landslide acceleration movement, while the Saito Tokusaka model has better prediction results in the late stage of the landslide acceleration movement. The prediction results of the basic prediction models are integrated by machine learning, that is, the stacking integration idea in the machine learning algorithm, also known as the expert hybrid algorithm, can establish a model that provides a minimum error result (according to the principle of machine learning training) in the whole process from the landslide entering the acceleration movement to the instability, fully utilizing the prediction advantages of different basic prediction models in the landslide acceleration movement stage, significantly improving the accuracy and stability of the instability time prediction, and providing a more scientific, intuitive and accurate reference for landslide early warning work.

[0062] In the specific case 1 provided by the present invention, the original landslide is a landslide distributed in a mountainous area. The overall terrain of the area where it is located is high in the north and low in the south. The upper and lower parts are steeper and the middle part is relatively flat. According to on-site observations, the landslide is about 500 meters long and about 300 meters wide. The landslide volume is about 3 million cubic meters, which is a large soil landslide. The landslide uses a real-time slope surface displacement meter (GNSS) based on the Beidou satellite system. According to the development and deformation characteristics of the landslide, a total of 8 surface displacement monitoring stations are deployed, such as Figure 3 As shown in the figure, the monitoring mainly started on April 7, 2020. The landslide creep was relatively stable in the initial monitoring stage. With the increase of rainfall from July 4 to July 6, the displacement change rate continued to increase until the landslide became unstable. Figure 4 Shown

[0063] Since the movement scales of monitoring points at different locations vary greatly, the moving average method is used to preprocess the original monitoring data (according to the monitoring period of the landslide, the filter order is selected as 7), and the improved tangent angle method is used to determine the evolution stage of the landslide, such as Figure 5 It can be seen that the data preprocessing and the determination of the landslide evolution stage are consistent with the actual situation of Case 1.

[0064] After obtaining the data of the accelerated movement stage of the landslide in Case 1, the instability time prediction of the basic prediction model is carried out, such as Figure 6 As shown, according to Figure 6 The dynamic prediction results show that the basic prediction models have certain prediction rules in different stages of landslide acceleration. The Verhulst model and the GM (1,1) model have better prediction effects in the early stage of landslide acceleration, and the Saito Tokusaka model has better prediction effects in the late stage of landslide acceleration. In order to make full use of the prediction advantages of different basic prediction models in the stage of landslide acceleration to improve the accuracy and stability of instability time prediction, different machine learning methods are used to integrate the basic prediction models to establish a comprehensive prediction model. The selected machine learning methods include artificial neural network, support vector machine regression and decision tree regression. After comparing the evaluation indicators of time series analysis, the comprehensive prediction model based on decision tree regression training has the best prediction effect. The comparison results are shown in Figure 7 shown.

[0065] Based on the same inventive concept, this embodiment also provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the aforementioned method for dynamic prediction of landslide instability time based on machine learning is implemented.

[0066] The computer readable medium includes, but is not limited to, any type of disk (including floppy disk, hard disk, optical disk, CD-ROM, and magneto-optical disk), ROM, RAM, EPROM (Erasable Programmable Read-Only Memory), EEPROM, flash memory, magnetic card or optical card. In other words, the readable medium includes any medium that can store or transmit information in a readable form by a device (such as a computer).

[0067] It should be noted that when writing the program, attention should be paid to the distinction between different acceleration stages. Due to the influence of internal and external factors in the landslide evolution process, local abnormal acceleration stages are inevitable. Different acceleration stages need to be screened. Generally, the time scale of the abnormal acceleration stage is small, or less than three cycles, and can be updated and covered in the process of dynamic forecasting.

[0068] It should be noted that attention should be paid to the writing of units when writing the program. The units of the forecast results need to be the same as the monitoring period in order to adapt to landslide monitoring work in different monitoring periods.

[0069] The computer program stored in the computer-readable storage medium is connected to the field real-time monitoring database of the same type of target landslide in the neighboring area of ​​Case 2 for calculation, such as Figure 8 As shown in Figure 2, since Case 2 is a real-time monitoring of landslides, which has not yet occurred, the calculation results are more consistent with the on-site monitoring results, and different abnormal acceleration stages are selected in the program to adapt to the abnormal acceleration that may occur during the actual monitoring process.

[0070] Among them, abnormal acceleration points usually do not meet the minimum number of monitoring period items for instability time forecast and are not forecasted, or are updated and covered during dynamic monitoring and dynamic forecasting. The smaller the time scale of the monitoring period, the more detection period items of abnormal acceleration points will increase, but at the same time, the speed of dynamic update will also be faster. In addition, since the instability time forecast model has certain forecasting rules at different stages of landslide acceleration movement, it can be used to reflect the current landslide evolution stage, significantly reducing the possibility of false alarms.

[0071] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for dynamic prediction of landslide instability time based on machine learning, characterized in that: The steps include: S1, prepare the original monitoring data of the original landslide, including landslide surface displacement monitoring and landslide crack meter monitoring; S2, using the moving average method to pre-process the original monitoring data; the moving average method used selects the filter order according to the monitoring period of the original landslide from which the original monitoring data is obtained; S3, using the tangent angle method to determine the landslide evolution stage based on the pre-processed monitoring data; S4, according to the target landslide monitoring period, set the dynamic forecast calculation interval; the dynamic calculation interval takes the landslide entering the acceleration movement stage as the starting point, and a forecast is made after each landslide monitoring period to obtain the dynamic forecast result of the landslide instability time, and the forecast result will be updated as the monitoring data is updated; S5, when it is determined that the landslide has entered the accelerated movement stage, the landslide instability time is predicted through the basic prediction model according to the set dynamic range; the basic prediction model includes the Verhulst model, the GM (1,1) model, and the Saito Toshitaka model; When the prediction basic model of instability time adopts Verhulst model or GM (1,1) model, if the original monitoring data is cumulative displacement, the data is differentiated to obtain the matrix of coefficients of the white differential equation solution, and the original data is also differentiated when determining the landslide evolution; When the basic model for predicting the instability time adopts the Saito Tokusaka model, three points with the same displacement spacing are obtained in the instability time calculation process of the model, and the interpolation method is used to obtain them; S6, taking the basic forecast model as input and the actual instability time of the landslide as the target output, uses machine learning to integrate the basic forecast model and establish a comprehensive forecast model for dynamic forecast of the landslide instability time to dynamically forecast the landslide instability time.

2. The method for dynamic landslide instability prediction based on machine learning according to claim 1 is characterized in that: The original landslide from which the original monitoring data is obtained in S1 is a landslide that has already become unstable and has the same properties as the target landslide.

3. The method for dynamic landslide instability prediction based on machine learning according to claim 1 is characterized in that: The machine learning in S6 is one of artificial neural network, support vector machine regression or decision tree regression.

4. The method for dynamic prediction of landslide instability time based on machine learning according to claim 3 is characterized in that: Decision tree regression is used to integrate the basic forecasting model to establish a comprehensive forecasting model to predict the instability time of landslides.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method for dynamic landslide instability prediction based on machine learning as described in any one of claims 1 to 4 is implemented.

Images