A landslide crack prediction method fusing LSTM-VAE and gradient boosting decision tree

Abstract

This invention discloses a landslide crack prediction method integrating LSTM-VAE and gradient boosting decision tree, comprising the following steps: collecting landslide-related monitoring data; preprocessing the monitoring data; constructing a landslide time-series data sequence; constructing a joint time-series data augmentation model based on LSTM and VAE, the LSTM-VAE model; training the LSTM-VAE model using the VAE loss function; generating synthetic data using the LSTM-VAE model and constructing an augmented dataset; dividing the augmented dataset into training, validation, and test sets according to a set ratio; establishing a landslide crack prediction GBDT model and training the GBDT model using the training set; evaluating the predictions of the GBDT model to obtain the optimal landslide crack prediction model. This invention, while ensuring the accuracy of the crack prediction model, solves the problems of traditional machine learning models' difficulty in effectively capturing the temporal characteristics of data and their high requirements for data quality, providing a scientific basis and technical support for landslide risk management at engineering sites.

Description

Technical Field

[0001] This invention relates to the field of landslide disaster monitoring and prediction, specifically to a landslide crack prediction method that integrates the LSTM-VAE data augmentation method and the gradient boosting decision tree (GBDT) algorithm. Background Technology

[0002] Currently, research in the field of landslide crack prediction mainly focuses on the application of traditional machine learning methods, such as Support Vector Machines (SVM), Random Forests (RF), and Decision Trees (DT). These methods typically rely on historical monitoring data and establish predictive models through statistical analysis. However, the prediction accuracy of these models is often limited by the data scale and the effectiveness of feature extraction, making it difficult to fully capture the complex nonlinear characteristics of landslide crack evolution.

[0003] In recent years, deep learning models, such as Long Short-Term Memory (LSTM) networks and Variational Autoencoders (VAEs), have been introduced into the field of landslide monitoring and early warning, demonstrating their potential in data processing and feature extraction. However, these methods typically require large amounts of high-quality monitoring data, while in practical applications, landslide crack monitoring data often suffers from problems such as small scale, insufficient annotation, and outlier interference. Furthermore, traditional machine learning and deep learning models each have their limitations; the former struggles to effectively capture temporal features and nonlinear relationships, while the latter may suffer from insufficient generalization ability, high data requirements, and poor interpretability. Therefore, it is necessary to develop a landslide crack prediction method that integrates the advantages of deep learning and traditional machine learning to improve the accuracy and reliability of predictions. Summary of the Invention

[0004] The purpose of this invention is to provide a landslide crack prediction method that integrates LSTM-VAE time-series data augmentation and gradient boosting decision tree (GBDT), aiming to solve the problems of insufficient data, difficulty in effectively extracting time-series features, and low prediction accuracy in the processing of landslide crack monitoring data in existing technologies.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A landslide crack prediction method integrating LSTM-VAE and gradient boosting decision tree includes the following steps:

[0007] S1: Collect data on cracks and related environmental conditions at the landslide site to form initial monitoring data;

[0008] S2: Perform data preprocessing on the initial monitoring data;

[0009] S3: Construct a continuous landslide time series data sequence based on the preprocessed data;

[0010] S4: Construct a joint time-series data augmentation model based on LSTM and VAE, namely the LSTM-VAE model;

[0011] S5: Train the LSTM-VAE model using the standard VAE loss function;

[0012] S6: Using the trained LSTM-VAE model, randomly sample and generate synthetic data containing the temporal characteristics of landslide crack evolution in the latent space, and construct an augmented dataset containing real data and synthetic data.

[0013] S7: Divide the augmented dataset into training, validation and test sets according to the proportions for training and evaluation of the GBDT model;

[0014] S8: Establish a GBDT model for landslide crack prediction based on gradient boosting decision tree, and train the GBDT model for landslide crack prediction using the training set;

[0015] S9: The prediction accuracy of the GBDT model for landslide crack prediction is comprehensively evaluated through the validation set and test set to obtain the optimal model parameters and determine the final landslide crack width prediction model.

[0016] Furthermore, step S2, which preprocesses the initial monitoring data, includes: removing obvious outliers using the interquartile range method, filling in continuous and random missing values ​​using the random forest algorithm, analyzing the correlation between features using the Pearson correlation coefficient, removing completely correlated redundant features, and normalizing the data.

[0017] Furthermore, step S4, constructing the LSTM-VAE model, a joint time-series data augmentation model based on LSTM and VAE, includes:

[0018] S30: The model encoder uses an LSTM network to encode the input landslide time series data, extract sequence information, and retrieve the hidden state at the last time step. As a sequence feature;

[0019] S31: Regarding h T Normalize:

[0020]

[0021] Where, μ batch and These are the mean and variance of the current batch, respectively; γ and β are learnable parameters; and ∈ is a small constant to prevent division by zero.

[0022] S32: For the normalized vector Perform regularization:

[0023]

[0024] Where, r i via Bernoulli(p) keep Independent sampling, p keep = 1 - dropout rate;

[0025] S33: The regularized vector h norm Mapped to the latent space parameter vector via a fully connected layer:

[0026] y = FC(h) norm )=W·h norm +b, where,

[0027] S34: Divide y into two parts to obtain the mean and log-variance of the latent variables:

[0028] μ = y 1:L ,logσ 2 =y L+1:2L ;

[0029] S35: Sample the latent vector z from the latent distribution using a reparameterization method:

[0030]

[0031] S36: The sampled latent vector z∈R^L is expanded into a sequence of length T using the model decoder, and then the original input is reconstructed using LSTM and fully connected layers to finally obtain the reconstructed sequence.

[0032] Further, step S36 includes the following steps:

[0033] S361: Copy the latent vector z T times to form an extended sequence:

[0034]

[0035] S362: Expand the sequence Input the data into an LSTM network to obtain the hidden state at each time step.

[0036] {h1,h2,…,h T} = LSTM(z expanded );

[0037] S363: For each hidden state h t Apply a fully connected layer:

[0038]

[0039] in, and These are the mapping matrix and the bias, d, respectively. input The original input feature dimension;

[0040] S64: Obtain the reconstructed sequence:

[0041] Furthermore, the loss function in step S5 includes two parts: reconstruction loss and KL divergence. The reconstruction loss is used to measure the reconstructed data x. recon The KL divergence, used to measure the difference between the encoder output distribution q(z|x) and the prior distribution p(z), is defined as the loss function:

[0042]

[0043] in, To reconstruct the loss, Let KL divergence be the KL divergence. This represents an overall loss.

[0044] Furthermore, in step S7, the augmented dataset is divided into a training set, a validation set, and a test set according to a ratio of 70%, 15%, and 15%.

[0045] Furthermore, the iterative update formula for the landslide crack prediction GBDT model in step S8 is as follows:

[0046] F m (x)=F m-1 (x)+γ m h m (x)

[0047] Among them, F m (x) represents the prediction result after the m-th iteration, h m (x) is the decision tree obtained from the fitting residuals of this round, γ m The step size coefficient is determined by line search in each iteration. The landslide crack prediction GBDT model minimizes the error between the predicted crack width and the actual monitored value through iterative learning.

[0048] Furthermore, in step S9, the prediction accuracy of the GBDT model for landslide crack prediction is comprehensively evaluated, including by measuring the root mean square error (RMSE), mean absolute error (MAE), and goodness of fit (R²). 2 Three indicators were used to comprehensively evaluate the prediction accuracy of the GBDT model for landslide crack prediction.

[0049] Compared with the prior art, the present invention has the following advantages:

[0050] This invention uses a fusion of LSTM-VAE time-series data augmentation and Gradient Boosting Decision Tree (GBDT) algorithm for landslide crack prediction. While ensuring the accuracy of the crack prediction model, it addresses the limitations of traditional machine learning models in effectively capturing the temporal characteristics of data and their high data quality requirements. By introducing LSTM-VAE to generate synthetic data containing real landslide crack evolution characteristics, the dataset is effectively expanded, significantly improving the model's generalization performance and prediction accuracy. Furthermore, in the data preprocessing process, this invention employs random forest filling, interquartile range outlier handling, and feature correlation analysis methods to effectively improve data quality and reduce the negative impact of outliers and missing values ​​on model accuracy. Simultaneously, in the landslide crack prediction model construction process, the advantages of the Gradient Boosting Decision Tree algorithm are leveraged to achieve efficient modeling and high-precision prediction, effectively improving the timeliness, reliability, and stability of landslide crack prediction and early warning.

[0051] This invention represents a significant breakthrough in model structure design, data augmentation methods, and prediction accuracy, outperforming existing technologies and achieving more accurate, stable, and real-time landslide crack prediction and early warning effects. It provides a scientific basis and technical support for landslide risk management at engineering sites. Attached Figure Description

[0052] Figure 1 This is a flowchart of the landslide crack prediction method that integrates LSTM-VAE and gradient boosting decision tree according to the present invention.

[0053] Figure 2 This is a heatmap of the characteristic Pearson correlation coefficient of an embodiment of the present invention;

[0054] Figure 3 This is a comparison chart of the prediction results of the original data model and the actual data in an embodiment of the present invention;

[0055] Figure 4 This is a comparison chart of the model prediction results after data augmentation and the actual data in an embodiment of the present invention. Detailed Implementation

[0056] The following is in conjunction with the appendix Figure 1-4 The specific embodiments of the present invention will be described in detail below. These embodiments are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. Obviously, the embodiments described in this invention are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0057] In this specification, the terms “comprising,” “including,” “having,” and variations thereof mean “including, but not limited to,” unless otherwise specifically emphasized.

[0058] A landslide crack prediction method integrating LSTM-VAE and gradient boosting decision tree includes the following steps:

[0059] Step 1: Use automated monitoring equipment to collect real-time data on cracks and related environmental conditions at the landslide site, including crack width, surface displacement, groundwater level, and ambient temperature and humidity, to form initial monitoring data.

[0060] Step 2: Data preprocessing of the initial monitoring data, including outlier identification and removal, missing value imputation, feature correlation analysis, and redundant feature removal. Specifically, this includes removing obvious outliers using the interquartile range method, imputing continuous and random missing values ​​using the random forest algorithm, analyzing the correlation between features using the Pearson correlation coefficient, removing fully correlated redundant features, and normalizing the data. Data preprocessing significantly improves data quality and the reliability of model input.

[0061] Step 3: Based on the preprocessed monitoring data, construct time-series data samples and form a continuous time-series data sequence using the sliding window method:

[0062] Z t ={z t-T+1 ,z t-T+2 ,…,z t}

[0063] Where T is the window length, Z t Let z be the time series data sequence corresponding to time t. t Let be the eigenvector at time t.

[0064] Step 4: Construct a joint temporal data augmentation model based on Long Short-Term Memory (LSTM) networks and Variational Autoencoders (VAEs), namely the LSTM-VAE model. The LSTM-VAE model consists of three parts: encoder, reparameterization, and decoder. Then, dropout regularization and batch normalization are applied to enhance the stability of the LSTM-VAE model and improve the quality of generated data, fundamentally overcoming the technical bottleneck of traditional data augmentation methods' insufficient enhancement of temporal features. The specific operations are as follows:

[0065] The model encoder uses an LSTM network to encode the input temporal data, extract sequence information, and retrieve the hidden state at the last time step. As a sequence feature;

[0066] For hT Normalize:

[0067]

[0068] Where, μ batch and These are the mean and variance of the current batch, respectively; γ and β are learnable parameters; and ∈ is a small constant to prevent division by zero.

[0069] For the normalized vector Perform Dropout processing:

[0070]

[0071] Where, r i ~Bernoulli(p keep Independent sampling, p keep = 1 - dropout rate, in this embodiment of the invention the dropout rate is 0.2, then p keep =0.8;

[0072] The regularized vector h norm Mapped to the latent space parameter vector via a fully connected layer:

[0073] y = FC(h) norm )=W·h norm +b, where,

[0074] Then, y is split into two parts to obtain the mean and log-variance of the latent variables:

[0075] μ = y 1:L ,logσ 2 =y L+1:2L ;

[0076] The latent vector z is sampled from the latent distribution using the reparameterization technique:

[0077]

[0078] The model decoder will sample the latent vectors. The input is expanded into a sequence of length T, and then reconstructed using an LSTM network and a fully connected layer. The specific steps are as follows:

[0079] Copy z T times to form an extended sequence:

[0080]

[0081] The expanded sequence Inputting the LSTM yields the hidden state at each time step.

[0082] {h1,h2,…,h T} = LSTM(z expanded );

[0083] Then for each hidden state h t Apply a fully connected layer:

[0084]

[0085] in, and These are the mapping matrix and the bias, d, respectively. input The original input feature dimension;

[0086] The final reconstructed sequence is:

[0087]

[0088] Step 5: Train the LSTM-VAE model using the standard VAE loss function. The loss function consists of two parts: reconstruction loss and KL divergence. The reconstruction loss is represented by mean squared error (MSE) and is used to measure the reconstructed data x. recon The difference between the encoder output distribution q(z|x) and the prior distribution p(z) is measured by the KL divergence, which measures the difference between the encoder output distribution q(z|x) and the prior distribution p(z). The loss function is defined as:

[0089]

[0090] in, To reconstruct the loss, Let KL divergence be the KL divergence. This represents an overall loss.

[0091] Step 6: Using the trained LSTM-VAE model, random sampling is performed in the latent space to generate synthetic data containing the temporal characteristics of landslide crack evolution. Then, an augmented dataset containing real and synthetic data is constructed, which significantly increases the scale and representativeness of the training data, effectively solves the problem of existing models being sensitive to data scale, and improves the generalization ability of the prediction model.

[0092] Step 7: Divide the augmented dataset into training, validation, and test sets according to a ratio of 70%, 15%, and 15%, respectively, for the training and evaluation of the Gradient Boosting Landslide Crack Prediction GBDT model.

[0093] Step 8: Predict the crack development trend by fusing the Gradient Boosting Decision Tree (GBDT) model. Leveraging GBDT's efficient nonlinear fitting and generalization capabilities, the prediction accuracy is significantly improved. An iterative update formula for a landslide crack prediction GBDT model based on Gradient Boosting Decision Tree (GBDT) is established:

[0094] F m (x)=F m-1 (x)+γ m h m (x),

[0095] Among them, F m (x) represents the prediction result after the m-th iteration, h m (x) is the decision tree obtained from the fitting residuals of this round, γ m The step size coefficient is determined by line search in each iteration. The GBDT model for landslide crack prediction minimizes the error between the predicted crack width and the actual monitored value through iterative learning.

[0096] Step 9: Analyze the root mean square error (RMSE), mean absolute error (MAE), and goodness of fit (R²). 2 The prediction accuracy of the GBDT model for landslide crack prediction is comprehensively evaluated using three indicators to obtain the optimal model parameters and determine the final landslide crack prediction model.

[0097] Depend on Figures 2 to 3 As shown, compared with the original data model, the augmented training set obtained by the present invention through joint data construction, joint LSTM-VAE model training and data augmentation achieves better generalization effect on the prediction model. Its RMSE, MAE and R^2 indicators are significantly improved. This shows that the synthetic data not only effectively expands the number of training samples, but also provides more potential feature information, thereby improving the model's generalization ability to new data, and making the final prediction model have higher robustness and prediction accuracy under new data.

[0098] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A landslide crack prediction method integrating LSTM-VAE and gradient boosting decision tree, characterized in that, Includes the following steps: S1: Collect data on cracks and related environmental conditions at the landslide site to form initial monitoring data; S2: Perform data preprocessing on the initial monitoring data; S3: Construct a continuous landslide time series data sequence based on the preprocessed data; S4: Construct a joint time-series data augmentation model based on LSTM and VAE, namely the LSTM-VAE model; S5: Train the LSTM-VAE model using the standard VAE loss function; S6: Using the trained LSTM-VAE model, randomly sample and generate synthetic data containing the temporal characteristics of landslide crack evolution in the latent space, and construct an augmented dataset containing real data and synthetic data. S7: Divide the augmented dataset into training, validation and test sets according to the proportions; S8: Establish a GBDT model for landslide crack prediction based on gradient boosting decision tree, and train the GBDT model for landslide crack prediction using the training set; S9: The prediction accuracy of the GBDT model for landslide crack prediction is comprehensively evaluated through the validation set and test set to obtain the optimal model parameters and determine the final landslide crack width prediction model. Step S4 specifically includes the following steps: S40: The model encoder encodes the input landslide time series data using an LSTM network, extracts sequence information, and selects the hidden state at the last time as the sequence feature as a sequence feature; S41: performing normalization on the first feature vector and the second feature vector to obtain a first normalized feature vector and a second normalized feature vector. performing ， wherein, with are the current batch mean and variance, respectively, with are learnable parameters, is a small constant to prevent division by zero; S42: normalizing the vector performing a regularization process: ， wherein independent sampling, ; S43: Regularize the vector Map to the latent space parameter vector through a fully connected layer: wherein , ; S44: obtaining the mean and log variance of the latent variable by splitting into two parts S44: obtaining the mean and log variance of the latent variable by splitting into two parts ； S45: Sample a latent vector from the latent distribution using a reparameterization method : ； S46: The sampled latent vector z is expanded into a sequence of length T using the model decoder, z∈R^L, and then the original input is reconstructed using an LSTM network and a fully connected layer to finally obtain the reconstructed sequence; specifically including the following steps: S461: the latent vector copying at each time, the extended sequence is constituted: ； S462: extend the sequence input the LSTM network to get the hidden state at each time step : ； S463: for each hidden state Apply the fully connected layer: ， wherein, with are a mapping matrix and a bias, respectively, is the original input feature dimension; S464: Obtain the reconstituted sequence: .

2. The landslide crack prediction method of claim 1, wherein, Step S2, which involves data preprocessing of the initial monitoring data, includes: removing obvious outliers using the interquartile range method, filling in continuous and random missing values ​​using the random forest algorithm, analyzing the correlation between features using the Pearson correlation coefficient, removing completely correlated redundant features, and normalizing the data. 3.The landslide crack prediction method of fusing LSTM-VAE and gradient boosting decision tree according to claim 1, characterized in that, In step S5, the loss function includes two parts: reconstruction loss and KL divergence. The reconstruction loss is used to measure the reconstructed data. Compared to real input The difference between them, the KL divergence is used to measure the encoder output distribution. With prior distribution The difference between them, the loss function is defined as: ； ； ； wherein, is the reconstruction loss, is the KL divergence, is the overall loss.

4. The landslide crack prediction method of claim 3, wherein, In step S7, the augmented dataset is divided into a training set, a validation set, and a test set according to a ratio of 70%, 15%, and 15%.

5. The landslide crack prediction method of claim 4, wherein, The iterative update formula for the landslide crack prediction GBDT model in step S8 is as follows: ， in, For the first Prediction results after rounds of iteration This is the decision tree obtained from the fitting residuals in this round. The step size coefficient is determined by line search in each iteration.

6. The landslide crack prediction method of claim 5, wherein, The method for comprehensively evaluating the prediction accuracy of the landslide crack prediction GBDT model in the step S9 includes evaluating the prediction accuracy of the landslide crack prediction GBDT model by using the root mean square error RMSE, the mean absolute error MAE, and the goodness of fit The three indexes are used for comprehensively evaluating the prediction accuracy of the landslide crack prediction GBDT model.

Images