A landslide displacement double-layer fusion prediction method and model

Abstract

The application discloses a landslide displacement double-layer fusion prediction method and model, uses an ICEEMDAN algorithm to decompose an original displacement time sequence, obtains a plurality of IMF components, carries out feature engineering on the IMF components, adopts a trend slope and a window mean value to represent a displacement trend, adopts kurtosis and spectral entropy to represent a mutation early warning, adopts a main frequency and a zero-crossing rate to represent a periodical law, adopts sample entropy and a standard deviation to represent system stability, constructs a three-dimensional feature space fusing time domain and frequency domain, carries out data standardization on the extracted features, eliminates the interference effect of dimensions on the model, and ensures that all feature dimensions are in a unified calculation scale range, constructs a CNN-BiLSTM model for each IMF component, and uses a CPO algorithm to optimize the CNN-BiLSTM model, so that the data acquisition difficulty during model training and use can be reduced, the usability of the model in actual deployment can be enhanced, the prediction precision is improved, and the accuracy of landslide displacement prediction is improved.

Description

Technical Field

[0001] This invention belongs to the field of landslide monitoring technology, specifically relating to a two-layer fusion prediction method and model for landslide displacement. Background Technology

[0002] Landslides are a major hazard in open-pit mines, seriously threatening the safety of personnel and equipment and hindering mining progress. In recent years, with the continuous increase in the depth of open-pit mining, the risk of landslides has become increasingly prominent. To effectively prevent and control disasters and mitigate threats and losses, it is urgent to construct an effective landslide displacement prediction system.

[0003] Currently, landslide displacement prediction models based on machine learning are mainly divided into two types. The first type of prediction model is a multi-factor index prediction method based on the influencing factors of landslide displacement prediction. This method uses environmental factors that may affect landslide displacement as input data to predict landslide displacement. The above method has achieved certain results in specific situations, but it is unsatisfactory in practical applications. This is because it is difficult to obtain effective data, and both training and using the model face the problem of difficulty in obtaining environmental factor data. In addition, there are many factors affecting landslide displacement prediction, and the basic conditions of each coal mine vary greatly. Moreover, existing research has not reached a consensus on the selection of environmental factors, which makes the constructed prediction model lack universality. The second type of prediction model is a prediction method based on landslide displacement time series analysis. Many scholars decompose the landslide displacement time series into a trend term displacement controlled by the time-dependent deformation of the soil and rock mass and a fluctuation term displacement induced by environmental factors, and model and predict them separately. However, they have not made more detailed divisions of the fluctuation term and have not fully considered the nonlinear relationship between the predicted values ​​of the trend term displacement, the predicted values ​​of the fluctuation term displacement and the predicted values ​​of the landslide displacement. Some scholars have considered environmental factors when predicting trend or periodic terms, but like the first method, they face the problem of insufficient data for model training and use. Therefore, it is particularly important to construct a two-layer fusion prediction method for landslide displacement that does not depend on environmental factors. Summary of the Invention

[0004] This invention addresses the problems of existing landslide displacement prediction models, such as over-reliance on environmental influence factors and insufficient consideration of the nonlinear relationship between trend term displacement prediction values, fluctuation term displacement prediction values, and landslide displacement prediction values. It overcomes the shortcomings of existing technologies by providing a two-layer fusion prediction method and model for landslide displacement.

[0005] To achieve the above objectives, the present invention adopts the following technical solution.

[0006] A two-layer fusion prediction method for landslide displacement, the prediction method specifically includes the following steps:

[0007] S1: The original displacement time series is decomposed using the ICEEMDAN algorithm to obtain multiple IMF components with frequencies ranging from high to low.

[0008] S2: Perform feature engineering on each IMF component, use trend slope and window mean to represent displacement trend, kurtosis and spectral entropy to represent abrupt change warning, dominant frequency and zero-crossing rate to represent periodicity, sample entropy and standard deviation to represent system stability, and construct a three-dimensional feature space that integrates time domain and frequency domain.

[0009] S3: Standardize the extracted features to eliminate the interference effect of dimensions on the model and ensure that all feature dimensions are within a uniform computational scale.

[0010] S4: Construct a CNN-BiLSTM model for each IMF component;

[0011] S5: The CPO (Crowned Porcupine) optimization algorithm is used to optimize the CNN-BiLSTM model. The number of convolutional kernels, the number of LSTM units, the learning rate, and the loss rate are selected as hyperparameters to be optimized. Multiple initial values ​​are set for each hyperparameter to form a hyperparameter space. The CPO optimization algorithm randomly generates initial solutions based on the hyperparameter combinations in the search space and substitutes them into the CNN-BiLSTM model for training. The validation set loss function is used as the evaluation metric. The parameters are updated globally through visual and auditory intimidation methods and locally through odor and physical attacks methods to find the parameter combination with the minimum displacement prediction error. In each iteration, the patience value set by the early stopping strategy is used to determine whether to enter the next iteration, and the program is stopped based on the number of iterations. The final optimization result is set as the model hyperparameters.

[0012] S6: For each IMF component, construct a CNN-BiLSTM model using the optimal hyperparameter combination determined in S5, train the model and save the model to obtain multiple IMF component learners.

[0013] S7: Use the prediction results of each IMF component learner as the input of the XGBoost regression model, divide the training set, validation set and test set according to the ratio of 8:1:1 to train the model and obtain the final slope displacement prediction result.

[0014] Furthermore, step S1 includes the following specific steps:

[0015] S101: Let the original displacement time series be... x ( t The goal of decomposition is to obtain k IMF components and final residuals r ( t Define the modality extraction operator. This indicates the extraction of the first signal from the signal. k The decomposition process of the first-order IMF components is described as follows:

[0016] When calculating each IMF component, it is necessary to perform... N Round iteration, generating N Grouped noisy residual sequences:

[0017] ;

[0018] In the formula: Indicates that during the first k The first IMF component i The noisy residual sequence generated during round iteration, where ; r k-1 ( t ) indicates the first k The residuals of each IMF component, and r 0 ( t )= x ( t ); β k-1 It is the first k Noise amplitude coefficient of each IMF component; It is the first i Gaussian white noise is added during round iteration; This indicates the extraction of the first noise from the added noise. k- First-order IMF components;

[0019] S102: Process the generated N sets of noisy residual sequences to obtain the first... k One IMF component:

[0020] ;

[0021] In the formula: IMF k ( t ) indicates the first k One IMF component; N This represents the number of iterations. This indicates the extraction of the first [item] from the noisy residual sequence. k IMF components of order;

[0022] S103: Update residuals:

[0023] ;

[0024] In the formula: r k ( t ) indicates the first k+ The residual of one IMF component;

[0025] S104: Repeat S101-S103, when the residual signal... r k ( t If the number of extreme points is less than 3 or the number of IMF components decomposed reaches the preset upper limit, the decomposition process will be terminated immediately.

[0026] Furthermore, step S2 includes the following specific steps:

[0027] S201: Using a dynamic window mechanism, the window size is dynamically configured according to the frequency of IMF components, and features are extracted from the data in the sliding window to construct nine features: window mean, standard deviation, skewness, kurtosis, dominant frequency, spectral entropy, approximate entropy, trend slope, and zero-crossing rate.

[0028] S202: Use a random forest regressor to perform feature importance analysis and remove features whose contribution is close to zero in all components.

[0029] Furthermore, step S3 includes the following specific steps:

[0030] S301: Divide all the feature-engineered data into training, validation, and test sets in a 7:1:2 ratio; first, calculate the mean of each feature. m and standard deviation s :

[0031] ,

[0032] ;

[0033] Then, for each feature x i Standardize using the following formula:

[0034] ;

[0035] In the formula: m The characteristic average; n The number of samples; x i For training set x The Middle i The original sequence of each feature, ; s Standard deviation; z i The new sequence after transformation, ;

[0036] S302: Use the mean and standard deviation calculated from the training set to perform the same standardization process on the validation and test sets.

[0037] Further, step S4 includes:

[0038] S401: The CNN-BiLSTM network adopts a layered and progressive architecture, including an input layer, a convolutional layer, a pooling layer, two bidirectional LSTM layers, two Dropout layers, a self-attention layer, and a fully connected layer. The network training uses a batch size of 32, an early stopping tolerance value of 30, and uses the validation set loss as the monitoring metric. The initial training epochs are set to 200 epochs, and the EarlyStopping callback function of the TensorFlow framework is integrated to realize dynamic termination of the training process and saving of the optimal model.

[0039] Furthermore, step S5 includes the following specific steps:

[0040] S501: The core parameters of the CPO algorithm are set as follows: initial population size 100, early stopping patience value 30, maximum number of iterations 35, defense factor 0.7, and attack factor 0.5;

[0041] S502: The search space of CPO-CNN-BiLSTM consists of four dimensions: the number of convolutional kernels (16, 32, 64), the number of LSTM units (32, 64, 128), the learning rate (0.0001-0.01), and the dropout rate (0.1-0.5). First, the population is initialized based on the search space parameter combinations, using the following formula:

[0042] ;

[0043] In the formula: It is the first i The initial position of each individual; It is the lower bound vector of the search space; It is the upper bound vector of the search space; It is a random vector in the range [0,1], which controls the random distribution of the initial solution; It is the Hadamard product symbol, representing a matrix operation based on element-wise multiplication;

[0044] S503: Parameter updates are performed through visual intimidation, auditory intimidation, olfactory attacks, and physical attacks;

[0045] Visual scare tactics represent the process by which an individual moves closer to the current optimal solution, thus expanding the search range.

[0046] ,

[0047] ;

[0048] In the formula: It is the first t+1 iterations i The location of each individual; It is the first t In the first iteration i The location of each individual; Indicates the position of the current best individual; This indicates the predator's location; , and These are three randomly selected individual locations; t 1, t 2, t 3 is a random number within [0,1], controlling the exploration direction and step size;

[0049] Voice intimidation represents an information exchange between individuals, enhancing group diversity:

[0050] ;

[0051] In the formula: It is a binary vector, where each element takes the value 0 or 1, which determines whether to avoid the predator; , and These are three randomly selected individual locations; t 4 Use a random number within [0,1] to adjust the movement range;

[0052] Odor attack represents the process by which an individual converges to a local optimum:

[0053] ;

[0054] In the formula: and These are two randomly selected individual locations; t 5 Use random numbers within [0,1] to adjust the local search range;

[0055] Physical attacks represent the process by which an individual rapidly approaches the global optimum:

[0056] ;

[0057] In the formula: β It is the attenuation coefficient, which controls the speed of convergence towards the optimal solution; T max Represents the maximum number of iterations;

[0058] S504: By updating the solution set through group collaboration, the solution is gradually approached to the global optimum until the predetermined number of iterations is reached or the stopping condition is met, thus obtaining the optimal combination of hyperparameters.

[0059] Furthermore, step S7 includes the following specific steps:

[0060] S701: The XGBoost regression model is an ensemble learning algorithm based on gradient boosting. It treats each IMF component learner as a regression tree, and progressively stacks multiple regression trees through an additive model. The output of each tree is used as a residual correction term, and the final prediction is the weighted sum of the outputs of all trees.

[0061] ;

[0062] In the formula: Represents the final predicted value; K m It represents the number of regression trees; or k Indicates the first k The learning rate of each regression tree; f k ( x i ) is the first k The output of the regression tree;

[0063] S702: The prediction model is evaluated using MAE (mean absolute error), RMSE (root mean square error), and R² (coefficient of determination). MAE represents the mean absolute deviation between the predicted and actual values, reflecting the accuracy of the prediction. The closer the value is to 0, the higher the consistency between the predicted and actual values. RMSE reflects the dispersion of the predicted values ​​relative to the actual values. The closer the value is to 0, the smaller the average deviation of the predicted values, and the better the model performance. R² describes the model's ability to explain data fluctuations. The closer the value is to 1, the stronger the model's fitting ability.

[0064] ,

[0065] ,

[0066] ;

[0067] In the formula: n The number of samples; y i For the first i The true value of each sample; It is the first i Predicted values ​​for each sample; This is the mean of the true values.

[0068] In addition, the present invention also provides a landslide displacement dual-layer fusion prediction model, which is constructed by the aforementioned landslide displacement dual-layer fusion prediction method.

[0069] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0070] 1. This invention uses only landslide displacement time series as the sole input data, eliminating the need for environmental factors such as temperature, humidity, wind force, cloud cover, slope, and slope height. This significantly reduces the difficulty of data acquisition during model training and use, and enhances the ease of use of the model in actual deployment.

[0071] 2. This invention performs feature engineering on the IMF components obtained by the ICEEMDAN method and injects domain knowledge into the model through manual feature extraction. This enhances the interpretability of the model and achieves efficient learning with small samples. Combined with the spatiotemporal feature extraction capability of the CNN-BiLSTM model, it ultimately achieves accurate and persistent prediction of displacement components at different frequencies.

[0072] 3. This invention effectively solves the problem of difficult manual and precise parameter setting in CNN-BiLSTM models through the CPO algorithm, corrects the initialization deviation of its temporal dependencies, reduces trial and error costs, and improves the convergence speed of the model.

[0073] 4. This invention successfully constructed a two-layer prediction model for landslide displacement monitoring. Comparative experiments using eight combined models revealed that the CNN-BiLSTM-XGBoost algorithm exhibited the best prediction performance, with a mean absolute error of 0.072 mm, a root mean square error of 0.097 mm, and a coefficient of determination of 0.9981. Compared to single-layer prediction models, which directly superimpose trend and fluctuation displacement predictions to obtain landslide displacement predictions, this significantly improves prediction accuracy and demonstrates the superiority of the CNN-BiLSTM-XGBoost model in landslide displacement prediction. Attached Figure Description

[0074] Figure 1 This is a schematic diagram of the overall structure of the landslide displacement dual-layer fusion prediction method of the present invention.

[0075] Figure 2 This is a diagram of the CNN-BiLSTM network architecture of the landslide displacement dual-layer fusion prediction method of the present invention.

[0076] Figure 3 This is the original displacement time series diagram of the landslide displacement dual-layer fusion prediction method of the present invention.

[0077] Figure 4 This is a flowchart of the CPO-CNN-BiLSTM model establishment process for the landslide displacement dual-layer fusion prediction method of this invention.

[0078] Figure 5This is a comparison chart of the MAE and RMSE indices of eight combined models of the landslide displacement dual-layer fusion prediction method of this invention.

[0079] Figure 6 This is a comparison chart of the R² index of eight combined models of the landslide displacement dual-layer fusion prediction method of this invention. Detailed Implementation

[0080] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0081] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0082] Reference Figure 1 to Figure 6 The landslide displacement two-layer fusion prediction method provided in this embodiment of the invention specifically includes the following steps:

[0083] S1: The cumulative displacement in the X direction of point S28 on the east side of an open-pit mine was selected as the research object. NaN and Null values ​​in the original data were deleted, and the missing values ​​were filled using the fillna method in the pandas library, resulting in 16,950 sets of valid monitoring sequence data as the original displacement time series for model training. The ICEEMDAN algorithm was used to decompose the original displacement time series to obtain multiple IMF components with frequencies from high to low.

[0084] S2: Perform feature engineering on each IMF component, use trend slope and window mean to represent displacement trend, kurtosis and spectral entropy to represent abrupt change warning, dominant frequency and zero-crossing rate to represent periodicity, sample entropy and standard deviation to represent system stability, and construct a three-dimensional feature space that integrates time and frequency domains.

[0085] S3: Standardize the extracted features to eliminate the interference effect of dimensions on the model and ensure that all feature dimensions are within a uniform computational scale.

[0086] S4: Build a CNN-BiLSTM model for each IMF component.

[0087] S5: The Crested Porcupine Optimization (CPO) algorithm is used to optimize the CNN-BiLSTM model. The number of convolutional kernels, the number of LSTM units, the learning rate, and the loss rate are selected as hyperparameters to be optimized. Multiple initial values ​​are set for each hyperparameter to form a hyperparameter space. CPO randomly generates initial solutions based on the hyperparameter combinations in the search space and substitutes them into the CNN-BiLSTM model for training. The validation set loss function is used as the evaluation metric. Global parameter updates are performed through visual and auditory intimidation, and local parameter updates are performed through odor and physical attacks to find the parameter combination with the minimum displacement prediction error. In each iteration, the patience value set by the early stopping strategy determines whether to enter the next iteration, and the program stops based on the number of iterations. The final optimization result is set as the model hyperparameters.

[0088] S6: For each IMF component, construct a CNN-BiLSTM model using the optimal hyperparameter combination determined in S5, train the model, and save the model to obtain multiple IMF component learners.

[0089] S7: Use the prediction results of each IMF component learner as the input of the XGBoost regression model, divide the training set, validation set and test set according to the ratio of 8:1:1 to train the model and obtain the final slope displacement prediction result.

[0090] Specifically, step S1 includes the following specific steps:

[0091] S101: Let the original displacement time series be... x ( t The goal of decomposition is to obtain k IMF components and final residuals r ( t Define the modality extraction operator. This indicates the extraction of the first signal from the signal. k The decomposition process of the first-order IMF components is described as follows:

[0092] When calculating each IMF component, it is necessary to perform... N Round iteration, generating N Grouped noisy residual sequences:

[0093] ;

[0094] In the formula: Indicates that during the first k The first IMF component i The noisy residual sequence generated during round iteration, where ; r k-1 ( t ) indicates the first kThe residuals of each IMF component, and r 0 ( t )= x ( t ); β k-1 It is the first k Noise amplitude coefficient of each IMF component; It is the first i Gaussian white noise is added during round iteration; This indicates the extraction of the first noise from the added noise. k- First-order IMF components;

[0095] S102: Process the generated N sets of noisy residual sequences to obtain the first... k One IMF component:

[0096] ;

[0097] In the formula: IMF k ( t ) indicates the first k One IMF component; N This represents the number of iterations. This indicates the extraction of the first [item] from the noisy residual sequence. k IMF components of order;

[0098] S103: Update residuals:

[0099] ;

[0100] In the formula: r k ( t ) indicates the first k+ The residual of one IMF component;

[0101] S104: Repeat S101-S103, when the residual signal... r k ( t If the number of extreme points is less than 3 or the number of IMF components decomposed reaches the preset upper limit, the decomposition process will be terminated immediately.

[0102] Specifically, step S2 includes the following specific steps:

[0103] S201: A dynamic window mechanism is used to dynamically configure the window size according to the frequency of the IMF components. For example, high-frequency components are set to 10-20 time steps, mid-frequency components to 25-35 time steps, and low-frequency components to 40-60 time steps. Features are extracted from the data in the sliding window to construct nine features: window mean, standard deviation, skewness, kurtosis, dominant frequency, spectral entropy, approximate entropy, trend slope, and zero-crossing rate.

[0104] S202: Use the Random Forest Regressor (RandomForestRegressor in the sklearn.ensemble library) to perform feature importance analysis and remove features whose contribution is close to zero in all components.

[0105] Specifically, step S3 includes the following specific steps:

[0106] S301: Divide all the feature-engineered data into training, validation, and test sets in a 7:1:2 ratio. First, calculate the mean of each feature. m and standard deviation s :

[0107] ,

[0108] ;

[0109] Then, for each feature x i Standardize using the following formula:

[0110] ;

[0111] In the formula: m The characteristic average; n The number of samples; x i For training set x The Middle i The original sequence of each feature, ; s Standard deviation; z i The new sequence after transformation, .

[0112] Table 1 shows some of the data after IMF1 component normalization:

[0113] Table 1. Partial data after IMF1 component normalization.

[0114]

[0115] S302: Use the mean and standard deviation calculated from the training set to perform the same standardization process on the validation and test sets.

[0116] Specifically, step S4 includes:

[0117] S401: The CNN-BiLSTM network adopts a layered and progressive architecture, including an input layer, a convolutional layer (with a kernel size of 3 and the ReLU activation function), a pooling layer, two bidirectional LSTM layers, two Dropout layers, a self-attention layer, and a fully connected layer. The network training uses a batch size of 32, an early stopping tolerance value of 30, and uses the validation set loss (val_loss) as the monitoring metric. The initial training epochs are set to 200 epochs, and the EarlyStopping callback function of the TensorFlow framework (tf.keras.callbacks.EarlyStopping) is integrated to realize the dynamic termination of the training process and the saving of the optimal model.

[0118] Specifically, step S5 includes the following specific steps:

[0119] S501: The core parameters of the CPO algorithm are set as follows: initial population size 100, early stopping patience value 30, maximum number of iterations 35, defense factor 0.7, and attack factor 0.5.

[0120] S502: The search space of CPO-CNN-BiLSTM consists of four dimensions: the number of convolutional kernels (16, 32, 64), the number of LSTM units (32, 64, 128), the learning rate (0.0001-0.01), and the dropout rate (0.1-0.5). First, the population is initialized based on the search space parameter combinations, using the following formula:

[0121] ;

[0122] In the formula: It is the first i The initial position of each individual; It is the lower bound vector of the search space; It is the upper bound vector of the search space; It is a random vector in the range [0,1], which controls the random distribution of the initial solution; It is the Hadamard product symbol, representing a matrix operation based on element-wise multiplication.

[0123] S503: Parameter updates are performed through visual intimidation, auditory intimidation, olfactory attacks, and physical attacks;

[0124] Visual scare tactics represent the process by which an individual moves closer to the current optimal solution, thus expanding the search range.

[0125] ,

[0126] ;

[0127] In the formula: It is the firstt +1 iterations i The location of each individual; It is the first t In the first iteration i The location of each individual; Indicates the position of the current best individual; This indicates the predator's location; , and These are three randomly selected individual locations; t 1, t 2, t 3 is a random number within [0,1], which controls the exploration direction and step size.

[0128] Voice intimidation represents an information exchange between individuals, enhancing group diversity:

[0129] ;

[0130] In the formula: It is a binary vector, where each element takes the value 0 or 1, which determines whether to avoid the predator; , and These are three randomly selected individual locations; t 4 Use a random number within [0,1] to adjust the movement range.

[0131] Odor attack represents the process by which an individual converges to a local optimum:

[0132] ;

[0133] In the formula: and These are two randomly selected individual locations; t 5 Use random numbers within [0,1] to adjust the local search range.

[0134] Physical attacks represent the process by which an individual rapidly approaches the global optimum:

[0135] ;

[0136] In the formula: β It is the attenuation coefficient, which controls the speed of convergence towards the optimal solution; T max This represents the maximum number of iterations.

[0137] S504: By updating the solution set through group collaboration, the solution is gradually approached to the global optimum until the predetermined number of iterations is reached or the stopping condition is met, thus obtaining the optimal combination of hyperparameters.

[0138] Specifically, step S7 includes the following specific steps:

[0139] S701: The XGBoost regression model is an ensemble learning algorithm based on gradient boosting. It treats each IMF component learner as a regression tree, and progressively stacks multiple regression trees through an additive model. The output of each tree is used as a residual correction term, and the final prediction is the weighted sum of the outputs of all trees.

[0140] ;

[0141] In the formula: Represents the final predicted value; K m It represents the number of regression trees; or k Indicates the first k The learning rate of each regression tree; f k ( x i ) is the first k The output of the regression tree.

[0142] S702: The prediction model is evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Coefficient of Determination (R²). MAE represents the average absolute deviation between the predicted and actual values, reflecting the accuracy of the prediction; the closer its value is to 0, the higher the consistency between the predicted and actual values. RMSE reflects the dispersion of the predicted values ​​relative to the actual values; the closer its value is to 0, the smaller the average deviation of the predicted values, and the better the model performance. R² describes the model's ability to explain data fluctuations; the closer its value is to 1, the stronger the model's fitting ability.

[0143] ,

[0144] ,

[0145] ;

[0146] In the formula: n The number of samples; y i For the first i The true value of each sample; It is the first i Predicted values ​​for each sample; This is the mean of the true values.

[0147] S703: Four machine learning algorithms—BP neural network, ELM, LSTM, and CNN-BiLSTM—were used as the first-layer learners, and Ridge regression model and XGBoost regression model were used as the second-layer learners. A total of eight combined models were used for comparative experiments, and the experimental results are shown in Table 2.

[0148] Table 2 Experimental results of 8 combined models

[0149]

[0150] S704: As shown in Table 1 of S703, compared with F-1, F-2, F-3, and F-4, F-5, F-6, F-7, and F-8 have smaller MAE and RMSE indices, indicating that the XGBoost regression model performs better than the Ridge regression model in both MAE and RMSE, demonstrating higher prediction accuracy. Furthermore, the R² indices of F-5, F-6, F-7, and F-8 are significantly higher than those of F-1, F-2, F-3, and F-4, confirming their stronger fitting ability and stability. Therefore, the XGBoost regression model, as the second-layer learner, exhibits superior prediction performance in this research scenario. When the second-layer learner uniformly adopts the XGBoost regression model, the CNN-BiLSTM model achieves the lowest MAE and RMSE values ​​and the highest R² value, verifying its optimal performance as the base learner. Based on the above experimental results, this study ultimately selects the CNN-BiLSTM-XGBoost fusion model as the final architecture for landslide displacement prediction.

[0151] In addition, this embodiment of the invention also provides a landslide displacement dual-layer fusion prediction model, which is constructed by the aforementioned landslide displacement dual-layer fusion prediction method.

[0152] It is understood that, although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A two-layer fusion prediction method for landslide displacement, characterized in that: The prediction method specifically includes the following steps: S1: The original displacement time series is decomposed using the ICEEMDAN algorithm to obtain multiple IMF components with frequencies ranging from high to low. S2: Perform feature engineering on each IMF component, use trend slope and window mean to represent displacement trend, kurtosis and spectral entropy to represent abrupt change warning, dominant frequency and zero-crossing rate to represent periodicity, sample entropy and standard deviation to represent system stability, and construct a three-dimensional feature space that integrates time domain and frequency domain. S3: Standardize the extracted features to eliminate the interference effect of dimensions on the model and ensure that all feature dimensions are within a uniform computational scale. S4: Construct a CNN-BiLSTM model for each IMF component; Step S4 includes: S401: The CNN-BiLSTM network adopts a layered and progressive architecture, including an input layer, a convolutional layer, a pooling layer, two bidirectional LSTM layers, two Dropout layers, a self-attention layer, and a fully connected layer. The network training uses a batch size of 32, an early stopping tolerance value of 30, and uses the validation set loss as the monitoring metric. The initial training epochs are set to 200 epochs, and the EarlyStopping callback function of the TensorFlow framework is integrated to realize dynamic termination of the training process and saving of the optimal model. S5: The CPO (Crowned Porcupine) optimization algorithm is used to optimize the CNN-BiLSTM model. The number of convolutional kernels, the number of LSTM units, the learning rate, and the loss rate are selected as hyperparameters to be optimized. Multiple initial values ​​are set for each hyperparameter to form a hyperparameter space. The CPO optimization algorithm randomly generates initial solutions based on the hyperparameter combinations in the search space and substitutes them into the CNN-BiLSTM model for training. The validation set loss function is used as the evaluation metric. The parameters are updated globally through visual and auditory intimidation methods and locally through odor and physical attacks methods to find the parameter combination with the minimum displacement prediction error. In each iteration, the patience value set by the early stopping strategy is used to determine whether to enter the next iteration, and the program is stopped based on the number of iterations. The final optimization result is set as the model hyperparameters. S6: For each IMF component, construct a CNN-BiLSTM model using the optimal hyperparameter combination determined in S5, train the model and save the model to obtain multiple IMF component learners. S7: Use the prediction results of each IMF component learner as the input of the XGBoost regression model, divide the training set, validation set and test set according to the ratio of 8:1:1 to train the model and obtain the final slope displacement prediction result. Step S7 includes the following specific steps: S701: The XGBoost regression model is an ensemble learning algorithm based on gradient boosting. It treats each IMF component learner as a regression tree, and progressively stacks multiple regression trees through an additive model. The output of each tree is used as a residual correction term, and the final prediction is the weighted sum of the outputs of all trees. ； In the formula: Represents the final predicted value; K m It represents the number of regression trees; η k Indicates the first k The learning rate of each regression tree; f k ( x i ) is the first k The output of the regression tree.

2. The landslide displacement dual-layer fusion prediction method according to claim 1, characterized in that: Step S1 includes the following specific steps: S101: Let the original displacement time series be... x ( t The goal of decomposition is to obtain k IMF components and final residuals r ( t Define the modality extraction operator. This indicates the extraction of the first signal from the signal. k The decomposition process of the first-order IMF components is described as follows: When calculating each IMF component, it is necessary to perform... N Round iteration, generating N Grouped noisy residual sequences: ； In the formula: Indicates that during the first k The first IMF component i The noisy residual sequence generated during round iteration, where ; r k-1 ( t ) indicates the first k The residuals of each IMF component, and r 0 ( t )= x ( t ); β k-1 It is the first k Noise amplitude coefficient of each IMF component; It is the first i Gaussian white noise is added during round iteration; This indicates the extraction of the first noise from the added noise. k- First-order IMF components; S102: Process the generated N sets of noisy residual sequences to obtain the first... k One IMF component: ； In the formula: IMF k ( t ) indicates the first k One IMF component; N This represents the number of iterations. This indicates the extraction of the first [item] from the noisy residual sequence. k IMF components of order; S103: Update residuals: ； In the formula: r k ( t ) indicates the first k+ The residual of one IMF component; S104: Repeat S101-S103, when the residual signal... r k ( t If the number of extreme points is less than 3 or the number of IMF components decomposed reaches the preset upper limit, the decomposition process will be terminated immediately.

3. The landslide displacement dual-layer fusion prediction method according to claim 1, characterized in that: Step S2 includes the following specific steps: S201: Using a dynamic window mechanism, the window size is dynamically configured according to the frequency of IMF components, and features are extracted from the data in the sliding window to construct nine features: window mean, standard deviation, skewness, kurtosis, dominant frequency, spectral entropy, approximate entropy, trend slope, and zero-crossing rate. S202: Use a random forest regressor to perform feature importance analysis and remove features whose contribution is close to zero in all components.

4. The landslide displacement dual-layer fusion prediction method according to claim 1, characterized in that: Step S3 includes the following specific steps: S301: Divide all the feature-engineered data into training, validation, and test sets in a 7:1:2 ratio; first, calculate the mean of each feature. μ and standard deviation σ : ， ； Then, for each feature x i Standardize using the following formula: ； In the formula: μ The characteristic average; n The number of samples; x i For training set x The Middle i The original sequence of each feature, ; σ Standard deviation; z i The new sequence after transformation, ; S302: Use the mean and standard deviation calculated from the training set to perform the same standardization process on the validation and test sets.

5. The landslide displacement dual-layer fusion prediction method according to claim 1, characterized in that: Step S5 includes the following specific steps: S501: The core parameters of the CPO algorithm are set as follows: initial population size 100, early stopping patience value 30, maximum number of iterations 35, defense factor 0.7, and attack factor 0.5; S502: The search space of CPO-CNN-BiLSTM consists of four dimensions: the number of convolutional kernels (16, 32, 64), the number of LSTM units (32, 64, 128), the learning rate (0.0001-0.01), and the dropout rate (0.1-0.5). First, the population is initialized based on the search space parameter combinations, using the following formula: ； In the formula: It is the first i The initial position of each individual; It is the lower bound vector of the search space; It is the upper bound vector of the search space; It is a random vector in the range [0,1], which controls the random distribution of the initial solution; It is the Hadamard product symbol, representing a matrix operation based on element-wise multiplication; S503: Parameter updates are performed through visual intimidation, auditory intimidation, olfactory attacks, and physical attacks; Visual scare tactics represent the process by which an individual moves closer to the current optimal solution, thus expanding the search range. ， ； In the formula: It is the first t +1 iterations i The location of each individual; It is the first t In the first iteration i The location of each individual; Indicates the position of the current best individual; This indicates the predator's location; , and These are three randomly selected individual locations; τ 1, τ 2, τ 3 is a random number within [0,1], controlling the exploration direction and step size; Voice intimidation represents an information exchange between individuals, enhancing group diversity: ； In the formula: It is a binary vector, where each element takes the value 0 or 1, which determines whether to avoid the predator; , and These are three randomly selected individual locations; τ 4 Use a random number within [0,1] to adjust the movement range; Odor attack represents the process by which an individual converges to a local optimum: ； In the formula: and These are two randomly selected individual locations; τ 5 Use random numbers within [0,1] to adjust the local search range; Physical attacks represent the process by which an individual rapidly approaches the global optimum: ； In the formula: β It is the attenuation coefficient, which controls the speed of convergence towards the optimal solution; T max Represents the maximum number of iterations; S504: By updating the solution set through group collaboration, the solution is gradually approached to the global optimum until the predetermined number of iterations is reached or the stopping condition is met, thus obtaining the optimal combination of hyperparameters.

6. The landslide displacement dual-layer fusion prediction method according to claim 1, characterized in that: Step S7 further includes S702: evaluating the prediction model using MAE (mean absolute error), RMSE (root mean square error), and R² (coefficient of determination); MAE represents the mean absolute deviation between the predicted value and the true value, reflecting the accuracy of the prediction. The closer its value is to 0, the higher the degree of agreement between the predicted value and the true value. RMSE reflects the degree of dispersion of predicted values ​​relative to true values. The closer its value is to 0, the smaller the average deviation of predicted values ​​and the better the model performance. R² describes the model's ability to explain data fluctuations. The closer its value is to 1, the stronger the model's fitting ability. ， ， ； In the formula: n The number of samples; y i For the first i The true value of each sample; It is the first i Predicted values ​​for each sample; This is the mean of the true values.

7. A two-layer fusion prediction model for landslide displacement, characterized in that: The landslide displacement dual-layer fusion prediction model is constructed using a landslide displacement dual-layer fusion prediction method as described in any one of claims 1 to 6.

Images