A method for predicting multi-factor landslide settlement using gradient boosting regression optimization model

Abstract

The present invention relates to the technical field of landslide monitoring and early warning, and in particular to a method for predicting multi-factor landslide settlement by a gradient boosting regression optimization model. The method comprises: training a gradient boosting regression optimization model according to multi-factor sample data to predict landslide settlement; wherein the gradient boosting regression optimization model is composed of multiple gradient boosting regression trees in parallel, and the model improves the model classification strength by training multiple weak classifiers, thereby improving the model's prediction ability; in order to make the parameters of each weak learner different, the training set is updated during the model training process, and multiple weak learners are learned to improve the model's prediction ability; predicting landslide settlement according to current multi-factor sample data to obtain the current landslide settlement prediction value; and determining the threshold of the landslide hazard level based on landslide historical data, geological conditions, soil type and other data, and accurately evaluating the landslide hazard level by comparing the threshold and the prediction value.

Description

Technical Field

[0001] The invention relates to the technical field of landslide monitoring and early warning, and in particular to a method for predicting multi-factor landslide settlement by using a gradient boosting regression optimization model. Background Art

[0002] Landslide is one of the most serious types of geological disasters in my country. It is mainly caused by the large-scale downward sliding of soil caused by the instability of soil or rock. This phenomenon usually occurs in steep slopes, river bank edges and foothills.

[0003] The causes of landslide disasters are very complex. In addition to the internal geological conditions of the slope rock and soil (such as its own gravity and changes in rock properties, etc.), it is also affected by many external factors (such as rainfall, geological movements and human engineering activities, etc.). Therefore, in addition to improving the means of landslide displacement monitoring, the study of reliable prediction models is also one of the major tasks required to deal with landslide disasters.

[0004] In recent years, many machine learning methods have been applied to the research of landslide displacement prediction, including support vector machine regression, neural network regression and Bayesian regression. However, these single methods have limitations and are inefficient and inaccurate, and cannot cope with the accurate analysis of landslide displacement with large amounts of data. Compared with the machine learning methods mentioned above, ensemble learning (EL) has great application potential in landslide prediction. The EL method is a machine learning method that trains multiple learners in parallel; specifically, the EL method first gives each training set the same probability, then uses multiple weak learners to repeatedly train and iterate, and resamples samples based on the iterative results, and finally combines the prediction results of multiple weak learners to obtain a strong learner. In the process of studying ensemble learning, the ensemble strategy is divided into: serial ensemble method and parallel ensemble method; among them, the serial method uses the dependence between the basic models to give a larger weight to the misclassified samples to improve the performance of the model; the parallel method uses the independence of the basic models, and then can greatly reduce the error through averaging. As one of the representatives of serial method improvement algorithms in ensemble learning, the Gradient Boosting Regression (GBR) algorithm can effectively reduce the overall error rate and improve model performance. However, there are few studies on the spatiotemporal evolution of landslide displacement using gradient boosting regression, especially on the prediction of landslide displacement using multiple factors. The landslide displacement prediction accuracy of the algorithm needs to be optimized and improved, which is not conducive to the control of landslides in the later stage.

[0005] To this end, the present invention proposes a method for predicting multi-factor landslide settlement by using a gradient boosting regression optimization model. Summary of the invention

[0006] The invention proposes a method for predicting multi-factor landslide settlement by using a gradient boosting regression optimization model. Firstly, multi-factor sample data of historical predicted landslide settlement is collected and divided into a training set and a test set according to a certain ratio; secondly, a gradient boosting regression optimization model is constructed to predict landslide settlement; wherein the gradient boosting regression optimization model is composed of a plurality of gradient boosting regression trees in parallel, and the model improves the classification strength of the model by training a plurality of weak classifiers, thereby improving the prediction ability of the model; in order to make the parameters of each weak learner different, the training set is updated during the training process of the model, so that m weak learners are trained with m training sets; finally, the landslide settlement is predicted according to the current multi-factor sample data to obtain the current landslide settlement prediction value; and the threshold of the landslide hazard level is determined based on the landslide historical data, geological conditions, soil type and other data, and the landslide hazard level is determined by comparing the threshold and the prediction value.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] The method of predicting multi-factor landslide settlement by gradient boosting regression optimization model includes:

[0009] Collect multi-factor sample data of historical predicted landslide settlement; wherein the multi-factor sample data is calculated through Beidou monitoring station and meteorological station data; including: Beidou monitoring landslide displacement, Beidou atmospheric precipitable water and meteorological station rainfall data;

[0010] The Beidou atmospheric precipitable water is inverted through Beidou monitoring data, and its calculation formula is expressed as:

[0011] ;

[0012] Wherein, PWV represents the calculated value of the Beidou atmospheric precipitable water; It is expressed as a conversion factor; ZWD represents the tropospheric zenith wet delay;

[0013] The calculation formula of the conversion factor is: ;in, Expressed as water density; Expressed as water vapor density; , Expressed as the constant coefficient of atmospheric refraction; Expressed as weighted mean temperature;

[0014] Furthermore, the multi-factor sample data is preprocessed and a multi-factor sample data set is constructed; wherein the preprocessing mainly includes: performing outlier processing on the multi-factor sample data to obtain first multi-factor sample data; filling missing values ​​of the first multi-factor sample data to obtain second multi-factor sample data; performing smoothing processing on the second multi-factor sample data to obtain third multi-factor sample data; performing standardization processing on the third multi-factor sample data to obtain fourth multi-factor sample data; converting the fourth multi-factor sample data into a specified format to obtain standard multi-factor sample data; and constructing the multi-factor sample data set based on the standard multi-factor sample data.

[0015] Further, the multi-factor sample data set is divided according to a certain ratio to obtain a training set and a test set;

[0016] Further, a gradient boosting regression optimization model is constructed, comprising: m weak learners, and m of the learners are combined in parallel;

[0017] The main body of the gradient boosting regression optimization model uses a gradient boosting regression model to allow multiple weak learners generated in parallel to fit the negative gradient of the loss function of the previously accumulated model, so that the cumulative model loss after adding multiple weak learners decreases in the direction of the negative gradient; wherein the gradient boosting regression tree is used to initialize the weak learners;

[0018] In establishing the gradient boosting regression optimization model, the function of the initial learner can be set to:

[0019] ;

[0020] in, Represented as the initialization model function of the weak learner; represents the training set; argmin() represents the minimum parameter value function; n represents the number of samples in the training set; L() represents the loss function; It is represented as the predicted value of landslide settlement of sample i in the training sample set; Represented as the optimal output value of the fitted leaf node in the gradient boosting regression optimization model;

[0021] Furthermore, the loss function of the initial learner can be expressed as: ;

[0022] In the process of establishing the above model, the negative gradient can be calculated according to the loss function; the negative gradient can be expressed as:

[0023] ;

[0024] in, It is represented as the negative gradient value of training sample i in the mth round of training, and is the number of training rounds; It is represented as the training function of the gradient boosting regression optimization model in the m-1th round;

[0025] The negative gradient of the loss function is used to fit a gradient regression tree in the mth iteration. ; wherein said The calculation formula of the weight coefficient is:

[0026] ;

[0027] in, It is represented as the weight coefficient of the jth leaf node after the mth round of model training; the value range of j is ; Indicates the key areas for establishing landslide settlement prediction;

[0028] Furthermore, the training set is used to train the gradient boosting regression optimization model; wherein the training process includes:

[0029] Further, the training set is input into a first weak learner to obtain a first landslide settlement prediction value;

[0030] Further, calculating the error between the first landslide settlement prediction value and the actual landslide settlement, and optimizing the first weak learner;

[0031] Further, performing data transformation on the training set to obtain a first training set;

[0032] Among them, the data transformation includes: extracting sample features of the training set; performing feature distribution analysis on the sample features to obtain a feature distribution random number sequence; combining the feature distribution random number sequence with the training set to generate an initialization training sample; performing data enhancement on the initialization training sample and performing random changes to obtain an enhanced training sample; re-labeling the enhanced training sample according to the sample label of the training set; verifying whether the enhanced training sample conforms to the feature distribution of the training set, and adjusting it to obtain the first training set.

[0033] Further, the first training set is input into a second weak learner to obtain a second landslide settlement prediction value;

[0034] Further, calculating the error between the second landslide settlement prediction value and the actual landslide settlement, and optimizing the second weak learner;

[0035] Further, repeating the weak learner training process until the m weak learners are optimized;

[0036] Further, according to the prediction results of the M weak learners, a predicted value of landslide settlement is comprehensively output;

[0037] Further, the error between the predicted landslide settlement and the actual landslide settlement is calculated to optimize the global parameters of the gradient boosting regression optimization model;

[0038] Further, according to the preset conditions, the gradient boosting regression optimization model is trained to obtain the optimal gradient boosting regression optimization model; wherein, after m rounds of iterative training, the weak learner can update the weak learner to:

[0039] Among them, record for ;

[0040] The final settlement prediction function model can be expressed as: .

[0041] Further, the optimal gradient boosting regression optimization model is tested using the test set to obtain a landslide settlement test value;

[0042] Specifically, the optimization process of the gradient boosting regression optimization model mainly includes:

[0043] (1) Maximum number of iterations of weak learners: As the number of iterations increases, the model gradually converges. Therefore, in order to avoid underfitting and overfitting, the optimal number of iterations is found to be 41 based on the training results.

[0044] (2) Maximum tree depth: A suitable maximum tree depth can effectively prevent overfitting. Through training, it is found that when the maximum tree depth is set to 6, the training model begins to converge.

[0045] (3) Learning rate: When the learning rate is too small, the convergence speed is very slow. When the learning rate is too large, the model calculation will become unstable in the later stage. According to the training results, the optimal learning rate is 0.5.

[0046] (4) Subsampling in regularization: Its value range is (0,1]. When this parameter is set to 1, it indicates that subsampling is not used in the model calculation process; however, if the value is too small, underfitting will occur. To avoid the above phenomenon, the optimal subsampling set in this application according to the training results is 0.4.

[0047] Furthermore, the landslide hazard level is evaluated based on the current predicted landslide settlement value.

[0048] Furthermore, the landslide hazard level is evaluated according to the current predicted landslide settlement value; wherein the landslide hazard level evaluation process includes:

[0049] Obtain current multi-factor sample data for predicting landslide settlement;

[0050] Performing the preprocessing on the current multi-factor sample data to obtain the current multi-factor prediction sample data;

[0051] Inputting the current multi-factor prediction sample data into the optimal gradient boosting regression optimization model to obtain the current landslide settlement prediction value;

[0052] Determine the threshold value of landslide hazard level based on historical landslide data, geological conditions and soil types;

[0053] The current predicted landslide settlement value is compared with the threshold value to obtain a comparison result.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] 1. In order to accurately predict the landslide settlement, the present invention trains the gradient boosting regression optimization model by collecting multi-factor sample data of historical predicted landslide settlement (Beidou monitoring landslide displacement, Beidou atmospheric precipitable water and meteorological station rainfall data). The model is trained by a gradient boosting regression tree composed of multiple weak learners, and the parameters of the weak learners are optimized during the training process, and finally the prediction results of each learner are combined and output; the gradient boosting regression optimization model can accurately predict the landslide settlement, which is conducive to the effective control of landslides in the later stage.

[0056] 2. In order to improve the learning performance of each weak learner of the gradient boosting regression optimization model, a data conversion method is used to convert the training set for the training of the weak learner during the learning process; wherein, the data conversion method obtains the feature distribution random sequence by analyzing the features of the training samples, initializes the feature distribution random sequence and the original training samples, and performs enhancement and verification to adjust and obtain multiple training samples; this method can expand the training set according to the training samples to improve the prediction performance of the weak learner, so that the gradient boosting regression optimization model is improved as a whole, which is conducive to the effective control of landslides in the later stage.

[0057] 3. The present invention proposes a method for evaluating the landslide grade hazard level. The method uses a gradient boosting regression optimization model to predict the current multi-factor sample data to obtain the current landslide settlement prediction value; and determines the threshold of the landslide hazard level based on the historical landslide data, geological conditions and soil type, and judges the landslide hazard level by comparing the predicted value with the threshold; the method can make an accurate assessment of the landslide hazard level for subsequent landslide control. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1A flow chart of a method for predicting multi-factor landslide settlement using a gradient boosting regression optimization model provided by an embodiment of the present invention;

[0059] Figure 2 A flow chart of a gradient boosting regression optimization model provided by an embodiment of the present invention;

[0060] Figure 3 An algorithm structure diagram of ensemble learning improvement provided by an embodiment of the present invention;

[0061] Figure 4 A schematic diagram of the training convergence of the gradient boosting regression optimization model provided by an embodiment of the present invention;

[0062] Figure 5 A schematic diagram of the deep training convergence of the gradient boosting regression optimization model provided by an embodiment of the present invention;

[0063] Figure 6 A schematic diagram of the learning rate convergence of the gradient boosting regression optimization model provided in an embodiment of the present invention.

[0064] Figure 7 A schematic diagram of the downsampling rate convergence of the gradient boosting regression optimization model provided in an embodiment of the present invention.

[0065] Figure 8 A schematic diagram of the actual deformation amount of landslide settlement provided in an embodiment of the present invention.

[0066] Fig. 9 A schematic diagram of the results of predicting landslide settlement using a gradient boosting regression model provided in an embodiment of the present invention.

[0067] Fig.10 A schematic diagram of the results of predicting landslide settlement using a gradient boosting regression optimization model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0068] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0069] Landslide settlement refers to the phenomenon of lowering the ground surface due to the movement or collapse of the surface soil. This phenomenon usually occurs in areas with loose soil, loose rock or rising groundwater levels, especially after heavy rainfall or natural disasters such as earthquakes. Landslide settlement may have a serious impact on nearby buildings, transportation facilities and personnel safety, so it is very important to accurately predict and prevent it in a timely manner.

[0070] In recent years, many machine learning methods have been applied to landslide displacement prediction research, including support vector machine regression, neural network regression, and Bayesian regression. However, these single methods have limitations and are inefficient and inaccurate, and cannot cope with the precise analysis of landslide displacement with large amounts of data.

[0071] Compared with the above mentioned methods, ensemble learning has great application potential in landslide prediction. In the process of ensemble learning research, the ensemble strategy is divided into: serial ensemble method and parallel ensemble method; among them, the serial method gives a larger weight to the misclassified samples through the dependence between the basic models to improve the performance of the model; the parallel method uses the independence of the basic models and then can greatly reduce the error through averaging. The gradient boosting regression algorithm, as one of the representatives of the serial method boosting algorithm in ensemble learning, can effectively reduce the overall error rate and improve the model performance. However, there are few studies on the use of gradient boosting regression to analyze the spatiotemporal evolution of landslide displacement, especially the use of multiple factors to predict landslide displacement. The landslide displacement prediction accuracy of the algorithm needs to be optimized and improved, which is not conducive to the control of landslides in the later stage.

[0072] To this end, the present invention proposes a method for predicting multi-factor landslide settlement by using a gradient boosting regression optimization model and describes it in detail through an embodiment.

[0073] Embodiment 1

[0074] In the embodiment of the present application, a method for predicting multi-factor landslide settlement by using a gradient boosting regression optimization model is used to predict landslide settlement on the collected multi-factor data samples; wherein, Figure 1 The specific steps of the method of the present invention are given in the text, including: S10. Collecting multi-factor sample data of historical predicted landslide settlement; S20. Preprocessing the multi-factor sample data and constructing a multi-factor sample data set; S30. Dividing the multi-factor sample data set into a training set and a test set; S40. Constructing a gradient boosting regression optimization model; S50. Training the gradient boosting regression optimization model to obtain the optimal gradient boosting regression optimization model; S60. Testing the optimal gradient boosting regression optimization model; S70. Evaluating the landslide hazard level.

[0075] The method is described in detail according to the steps S10 to S70 given above;

[0076] Further, according to Figure 1 In step S10, multi-factor sample data of historical predicted landslide settlement are collected; wherein the multi-factor sample data are calculated through Beidou monitoring station and meteorological station data; including: Beidou monitoring landslide displacement, Beidou atmospheric precipitable water and meteorological station rainfall data;

[0077] In the embodiment of the present application, relevant multi-factor data is collected to train the gradient boosting regression optimization model; in the prediction of landslide settlement, the landslide displacement, Beidou atmospheric precipitable water and meteorological station rainfall data are considered as external influencing factors; among them, the landslide displacement can reflect the degree and rate of landslide activity; Beidou atmospheric precipitable water includes rain, snow, etc. It is one of the important meteorological factors for landslides and has a direct impact on the saturation and stability of the soil; the rainfall data of the meteorological station is used to provide more accurate and real-time rainfall information, and the size and duration of rainfall are one of the main factors leading to landslides.

[0078] The Beidou atmospheric precipitable water is inverted through Beidou monitoring data, and its calculation formula is expressed as:

[0079] ;

[0080] Wherein, PWV represents the calculated value of the Beidou atmospheric precipitable water; It is expressed as a conversion factor; ZWD represents the tropospheric zenith wet delay;

[0081] The calculation formula of the conversion factor is: ;in, Expressed as water density; Expressed as water vapor density; , Expressed as the constant coefficient of atmospheric refraction; set , ; Expressed as weighted mean temperature;

[0082] The tropospheric zenith wet delay is calculated by obtaining meteorological elements through meteorological observation, mainly including air pressure, air temperature, humidity and dew point temperature; in order to obtain the authenticity and scientificity of the Beidou atmospheric precipitable water, the embodiment of the present application detects relevant data through multiple sounding stations; refer to Table 1, which gives the detection data of some sounding stations at different altitudes.

[0083] Table 1. Detection data of some sounding points

[0084] Altitude / m Air pressure / hPa Temperature / ℃ Dew point temperature / ℃ Relative humidity / % 16 1010 -3.1 -5.0 87 67 1005 -2.9 -6.4 76 90 1000 -4.2 -7.1 73 218 982 -6.8 -8.0 80 506 939 -8.4 -8.2 89

[0085] In the embodiment of the present application, the Beidou atmospheric precipitable water is obtained by an inversion calculation method. The calculation method obtains accurate convective zenith delay based on the detection data in Table 1, and then performs scientific calculations in combination with the conversion factor. Such a calculation method can accurately obtain the Beidou atmospheric precipitable water, ensure the accuracy of the data, and is conducive to the training of the gradient boosting regression optimization model.

[0086] Further, the collected data is preprocessed according to the content of step S20; wherein the preprocessing mainly includes: performing outlier processing on the multi-factor sample data to obtain first multi-factor sample data; filling missing values ​​of the first multi-factor sample data to obtain second multi-factor sample data; performing smoothing processing on the second multi-factor sample data to obtain third multi-factor sample data; performing standardization processing on the third multi-factor sample data to obtain fourth multi-factor sample data; converting the fourth multi-factor sample data into a prescribed format to obtain standard multi-factor sample data; and constructing the multi-factor sample data set according to the standard multi-factor sample data.

[0087] In the embodiments of the present application, a series of preprocessing steps are performed to ensure the consistency and integrity of the data, which is beneficial to improving the training effect and prediction accuracy of the model. These steps help to eliminate noise and inconsistency in the data, reduce the risk of overfitting of the model, and extract the most informative features, thereby better capturing the relationship between landslide settlement and influencing factors.

[0088] Further, as shown in step S30, the multi-factor sample data set is divided according to a certain ratio to obtain a training set and a test set; in the embodiment of the present application, Python programming is used to automatically divide the preprocessed data set, and the division ratio is set to 7:3 in the program to randomly divide it into a training set and a test set;

[0089] Further, as in the above step S40, refer to Figure 2 , construct a gradient boosting regression optimization model; wherein the main body of the gradient boosting regression optimization model uses a gradient boosting regression model, allowing multiple weak learners generated in parallel to fit the negative gradient of the loss function of the previously accumulated model, so that the cumulative model loss after adding multiple weak learners decreases in the direction of the negative gradient; wherein the gradient boosting regression tree is used to initialize the weak learners;

[0090] In the embodiment of the present application, the gradient boosting regression optimization model uses multiple gradient regression trees as the main body of the model, and a more powerful model is constructed by integrating multiple weak learners (gradient regression trees). Each gradient regression tree is trained based on the residual of the previous tree. Through continuous iterative optimization, the prediction results of multiple trees are finally weighted and combined to obtain the final prediction result; this integrated learning method can effectively improve the generalization ability and prediction accuracy of the model.

[0091] like Figure 3 As shown, the gradient boosting regression optimization model includes m weak learners, and the m weak learners are combined in parallel; wherein the initial model function of the weak learner is expressed as:

[0092] ;

[0093] in, Represented as the initialization model function of the weak learner; represents the training set; argmin() represents the minimum parameter value function; n represents the number of samples in the training set; L() represents the loss function; It is represented as the predicted value of landslide settlement of sample i in the training sample set; Represented as the optimal output value of the fitted leaf node in the gradient boosting regression optimization model;

[0094] Among them, the mean square error is used as the loss function in the model design process, and its expression is: ;

[0095] In the embodiment of the present application, the model function of the weak learner is initialized by the collected multi-factor data; the initialization of the weak learner is based on the error between the leaf node value output in the gradient regression tree and the actual value. Such initialization setting can make the gradient regression tree have a certain accuracy and fitting ability when it starts to iterate, thereby accelerating the convergence process of the model and improving the training efficiency and prediction performance of the model.

[0096] Further, the gradient boosting regression optimization model is trained according to step S50; the specific training process includes:

[0097] Inputting the training set into the first weak learner to obtain a first landslide settlement prediction value;

[0098] Further, calculating the error between the first landslide settlement prediction value and the actual landslide settlement, and optimizing the first weak learner;

[0099] Further, performing data transformation on the training set to obtain a first training set;

[0100] Among them, the data transformation includes: extracting sample features of the training set; performing feature distribution analysis on the sample features to obtain a feature distribution random number sequence; combining the feature distribution random number sequence with the training set to generate an initialization training sample; performing data enhancement on the initialization training sample and performing random changes to obtain an enhanced training sample; re-labeling the enhanced training sample according to the sample label of the training set; verifying whether the enhanced training sample conforms to the feature distribution of the training set, and adjusting it to obtain the first training set.

[0101] In the embodiment of the present application, in order to enable multiple weak learners of the gradient boosting regression optimization model to learn different parameters, the original training set is transformed; wherein, the data transformation method obtains a feature distribution random sequence by analyzing the features of the training samples, initializes the feature distribution random sequence and the original training samples, and performs enhancement and verification to adjust and obtain multiple training samples; this method can expand the training set according to the training samples to improve the prediction performance of the weak learner, so that the gradient boosting regression optimization model is improved as a whole, which is conducive to the effective control of landslides in the later stage.

[0102] Further, the first training set is input into a second weak learner to obtain a second landslide settlement prediction value;

[0103] Further, calculating the error between the second landslide settlement prediction value and the actual landslide settlement, and optimizing the second weak learner;

[0104] Further, repeating the weak learner training process until the m weak learners are optimized;

[0105] The calculation formula of the negative gradient of the loss function is expressed as:

[0106] ;

[0107] in, It is represented as the negative gradient value of training sample i in the mth round of training, and is the number of training rounds; It is represented as the training function of the gradient boosting regression optimization model in the m-1th round;

[0108] In the embodiment of the present application, the negative gradient of the loss function is calculated and the split direction of each tree can be determined according to the negative gradient, so that each step of model optimization is carried out in the direction of reducing the loss function, which helps the model to be optimized more efficiently, thereby improving the performance and generalization ability of the model. In the training process of the weak learner, the gradient regression tree is fitted by the negative gradient of the loss function, and the weight of each node of the gradient regression tree is obtained; including: fitting a gradient regression tree at the mth iteration ; wherein said The calculation formula of the weight coefficient is:

[0109] ;

[0110] in, It is represented as the weight coefficient of the jth leaf node after the mth round of model training; the value range of j is ; Indicates the key area for establishing landslide settlement prediction.

[0111] In the embodiment of the present application, the interpretability of the model can be enhanced by calculating the weight of the gradient regression tree. By analyzing the weight of each node, it is possible to better understand how the model makes predictions, thereby improving the understanding and trust of the model behavior.

[0112] Further, according to the prediction results of the m weak learners, a predicted value of landslide settlement is comprehensively output;

[0113] Further, the error between the predicted landslide settlement and the actual landslide settlement is calculated to optimize the global parameters of the gradient boosting regression optimization model;

[0114] Further, according to the preset conditions, the gradient boosting regression optimization model is trained to obtain the optimal gradient boosting regression optimization model; wherein, after m rounds of iterative training, the weak learner can update the weak learner to:

[0115] ;

[0116] Among them, for ;

[0117] The final settlement prediction function model can be expressed as: .

[0118] Among them, the main optimization parameters in the optimization process of the gradient boosting regression optimization model are shown in Table 2, in which the optimization values ​​of the key parameters are given;

[0119] Table 2. Gradient boosting regression optimization model parameter optimization

[0120] name Numeric Loss function ls Maximum number of iterations of weak learners (n_estimators) 41 Maximum tree depth (max_depth) 6 Learning rate (learning_rate) 0.5 Subsampling in Regularization 0.4

[0121] In order to illustrate that the gradient boosting regression optimization model parameters given in this embodiment are the optimal parameters; Figure 4 The scores of the gradient boosting regression optimization model after multiple rounds of training are given; Figure 5 The results of training different depths of the gradient boosting regression optimization model; Figure 6 About the changes in learning rate during training of the gradient boosting regression optimization model; Figure 7 It is the performance of the gradient boosting regression optimization model under sampling conditions of different proportions. In the embodiment of the present application, the final landslide settlement prediction model is obtained after m rounds of iterative training, which is also the optimal gradient boosting regression optimization model. The final model after multiple rounds of iterative training can more accurately predict the landslide settlement. Through continuous iteration, the model can gradually optimize and learn the complex patterns and relationships in the data, thereby improving the accuracy of the prediction.

[0122] Further, according to step S60, the trained optimal gradient boosting regression optimization model is tested using the test set to obtain a test value of landslide settlement;

[0123] The gradient boosting regression optimization model is optimized according to the optimization parameters in Table 2 to obtain the optimal gradient boosting regression optimization model; in order to illustrate that the model can improve the prediction ability after optimization, the model before and after optimization is experimentally described in the embodiments of the present application; refer to Table 3, which gives the experimental comparison.

[0124] Table 3. Comparative analysis of prediction accuracy before and after GBR model optimization

[0125] Model MAE / mm RMSE / mm <![CDATA[R 2 ]]> GBR 3.83 5.04 0.84 Optimal GBR 2.66 2.57 0.94

[0126] The table gives a comparison between the gradient boosting regression model and the gradient boosting regression optimization model, including: mean absolute error, mean square error evaluation indicators and goodness of fit; from the data in the table, it can be seen that the performance of the gradient boosting regression optimization model is better than that of the unoptimized model.

[0127] In addition, in order to further illustrate the accuracy of the gradient boosting regression optimization model in predicting landslide settlement; Figure 8 The actual deformation of the landslide settlement from 2021 / 4 / 15 to 2021 / 8 / 30 is given in Fig. 9 and Fig.10 The two models were compared in prediction experiments; Fig. 9 The landslide settlement prediction from 2021 / 4 / 15 to 2021 / 8 / 30 is based on the gradient boosting regression model (GBR); Fig.10 The landslide settlement prediction from 2021 / 4 / 15 to 2021 / 8 / 30 is based on the gradient boosting regression optimization model (i.e. Optimal GBR); according to the comparison Fig.10 The prediction results and Figure 8 The data in are very close.

[0128] Further, according to step S70, the current landslide hazard level is evaluated using the predicted landslide settlement value; the specific evaluation process includes:

[0129] Obtain current multi-factor sample data for predicting landslide settlement;

[0130] Further, the current multi-factor sample data is preprocessed to obtain current multi-factor prediction sample data;

[0131] Further, the current multi-factor prediction sample data is input into the optimal gradient boosting regression optimization model to obtain the current landslide settlement prediction value;

[0132] Further, a threshold value of the landslide hazard level is determined based on historical landslide data, geological conditions and soil types;

[0133] Further, the current predicted landslide settlement value is compared with the threshold value to obtain a comparison result;

[0134] Furthermore, the hazard level of the landslide is determined according to the comparison result; wherein the hazard level includes: low, medium and high.

[0135] In the embodiment of the present application, the landslide hazard level is evaluated according to the prediction results of the gradient boosting regression optimization model. The method uses the gradient boosting regression optimization model to predict the current multi-factor sample data to obtain the current landslide settlement prediction value; and determines the threshold of the landslide hazard level based on the historical landslide data, geological conditions and soil type, and judges the hazard level of the landslide by comparing the predicted value with the threshold; the method can make an accurate assessment of the degree of landslide hazard for subsequent landslide control.

[0136] In the first embodiment, the gradient boosting regression optimization model is used to predict the landslide settlement, which specifically includes: first, by collecting multi-factor sample data of historical predicted landslide settlement, and dividing the multi-factor sample data into a training set and a test set according to a certain ratio; second, constructing a gradient boosting regression optimization model to predict the landslide settlement; wherein the gradient boosting regression optimization model is composed of multiple gradient boosting regression trees in parallel, and the model improves the model classification strength by training multiple weak classifiers, thereby improving the prediction ability of the model; in order to make the parameters of each weak learner different, the training set is updated during the training process of the model, so that m weak learners are trained with m training sets, thereby improving the prediction ability of the model; finally, the landslide settlement is predicted according to the current multi-factor sample data to obtain the current landslide settlement prediction value; and the threshold of the landslide hazard level is determined based on the landslide historical data, geological conditions, soil type and other data, and the landslide hazard level is accurately evaluated by comparing the threshold and the prediction value.

[0137] Embodiment 2

[0138] In Example 1, the gradient boosting regression optimization model is used to achieve accurate prediction of landslide settlement. In order to further illustrate the feasibility of the method proposed in the present invention, it is further illustrated by the embodiment of the present application. The specific method flow is basically the same as that of Example 1, and includes the following steps:

[0139] Collect multi-factor sample data of historical predicted landslide settlement;

[0140] Further, the multi-factor sample data is preprocessed and a multi-factor sample data set is constructed;

[0141] Further, the multi-factor sample data set is divided according to 7:3 to obtain a training set and a test set;

[0142] Further, a gradient boosting regression optimization model is constructed, comprising: m weak learners, and m of the learners are combined in parallel;

[0143] Further, the gradient boosting regression optimization model is trained using the training set;

[0144] Further, the error between the predicted landslide settlement and the actual landslide settlement is calculated to optimize the global parameters of the gradient boosting regression optimization model;

[0145] Further, according to the preset conditions, the gradient boosting regression optimization model is trained to obtain the optimal gradient boosting regression optimization model; wherein, after m rounds of iterative training, the weak learner can update the weak learner to:

[0146] Among them, record for ;

[0147] The final settlement prediction function model can be expressed as: .

[0148] Furthermore, the landslide hazard level is evaluated according to the current predicted landslide settlement value; wherein the landslide hazard level evaluation process includes:

[0149] Obtain current multi-factor sample data for predicting landslide settlement;

[0150] Performing the preprocessing on the current multi-factor sample data to obtain the current multi-factor prediction sample data;

[0151] Inputting the current multi-factor prediction sample data into the optimal gradient boosting regression optimization model to obtain the current landslide settlement prediction value;

[0152] Determine the threshold value of landslide hazard level based on historical landslide data, geological conditions and soil types;

[0153] The current predicted landslide settlement value is compared with the threshold value to obtain a comparison result.

[0154] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting multi-factor landslide settlement using a gradient boosting regression optimization model, characterized in that: include: Collect multi-factor sample data of historical predicted landslide settlement; The multi-factor sample data is calculated through Beidou monitoring station and meteorological station data; including: Beidou monitoring landslide displacement, Beidou atmospheric precipitable water and meteorological station rainfall data; The Beidou atmospheric precipitable water is inverted through Beidou monitoring data, and its calculation formula is expressed as: PWV = Π × ZWD; Wherein, PWV is the calculated value of the Beidou atmospheric precipitable water; Π is the conversion factor; ZWD is the tropospheric zenith wet delay; The calculation formula of the conversion factor is: Among them, ρ w Expressed as water density; ρ v It is expressed as water vapor density; k'2, k3 are atmospheric refraction constants; T m Expressed as weighted mean temperature; Preprocessing the multi-factor sample data and constructing a multi-factor sample data set; dividing the multi-factor sample data set to obtain a training set and a test set; Constructing a gradient boosting regression optimization model, including multiple weak learners; wherein the main body of the gradient boosting regression optimization model uses a gradient boosting regression model, allowing multiple weak learners generated in parallel to fit the negative gradient of the loss function of the previously accumulated model, so that the cumulative model loss after adding multiple weak learners decreases in the direction of the negative gradient; wherein the weak learner uses a gradient boosting regression tree; Among them, the initialization model function of the weak learner is expressed as: Wherein, F0() represents the initialization model function of the weak learner; x={x i |1≤i≤n} represents the training set; argmin() represents the minimum parameter value function; n represents the number of samples in the training set; L() represents the loss function; y i It represents the predicted value of landslide settlement of sample i in the training set; γ represents the optimal output value of the fitting leaf node in the gradient boosting regression optimization model; Among them, the calculation formula of the negative gradient of the loss function is expressed as: Among them, g mi It is represented as the negative gradient value of training sample i in the mth round of training, and m (m = 1, 2, ..., M) is the number of training rounds; F m-1 () represents the training function of the gradient boosting regression optimization model in the m-1th round; The negative gradient of the loss function is used to fit a gradient regression tree h in the mth iteration. m (x i ); wherein said h m (x i The calculation formula of the weight coefficient is: Among them, γ mj It is represented as the weight coefficient of the jth leaf node after the mth round of model training; the value range of j is [1,...,J]; R mj Indicates the key areas for establishing landslide settlement prediction; Among them, the learner after m rounds of iterations can be updated as follows: Where I is h m (x); The gradient boosting regression optimization model is trained, including: Inputting the training set into the first weak learner to obtain a first landslide settlement prediction value; Calculating the error between the first landslide settlement prediction value and the actual landslide settlement, and optimizing the first weak learner; Performing data transformation on the training set to obtain a first training set; wherein the data transformation includes: extracting sample features of the training set; performing feature distribution analysis on the sample features to obtain a feature distribution random number sequence; combining the feature distribution random number sequence with the training set to generate an initialization training sample; performing data enhancement on the initialization training sample to perform random changes to obtain an enhanced training sample; re-assigning labels to the enhanced training sample according to the sample labels of the training set; verifying whether the enhanced training sample conforms to the feature distribution of the training set, and adjusting it to obtain the first training set; Inputting the first training set into a second weak learner to obtain a second landslide settlement prediction value; Calculating the error between the second landslide settlement prediction value and the actual landslide settlement, and optimizing the second weak learner; Repeat the weak learner training process until multiple weak learners are optimized; Outputting a predicted value of landslide settlement according to the prediction results of the plurality of weak learners; Calculating the error between the predicted landslide settlement and the actual landslide settlement, and optimizing the global parameters of the gradient boosting regression optimization model; According to the preset conditions, the model training is completed and the optimal gradient boosting regression optimization model is obtained; Wherein, in the optimal gradient boosting regression optimization model, the maximum number of iterations of the weak learner is 41, the maximum tree depth is 6, and the learning rate is 0.5; the subsampling in the regularization is 0.4; Using the test set to test the optimal gradient boosting regression optimization model, to obtain a landslide settlement test value; Assess the landslide hazard level based on the current predicted landslide settlement value.

2. The method for predicting multi-factor landslide settlement by gradient boosting regression optimization model according to claim 1, characterized in that: The landslide hazard level assessment process includes: Obtain current multi-factor sample data for predicting landslide settlement; Performing the preprocessing on the current multi-factor sample data to obtain the current multi-factor prediction sample data; Inputting the current multi-factor prediction sample data into the optimal gradient boosting regression optimization model to obtain the current landslide settlement prediction value; Determine the threshold value of landslide hazard level based on historical landslide data, geological conditions and soil types; Comparing the current predicted landslide settlement value with the threshold value to obtain a comparison result; The hazard level of the landslide is determined according to the comparison result; wherein the hazard levels include: low, medium and high.

Images