Strip mine slope displacement prediction model and method based on InSAR and ensemble learning

Abstract

The invention provides an InSAR and integrated learning-based strip mine slope displacement prediction model and method. The method comprises the following steps of data acquisition and preprocessing, multi-model integrated prediction framework, time series data processing, and model tuning and verification. According to the method, the InSAR is adopted to carry out mine slope displacement time sequence data extraction, the moving average method is utilized to decompose displacement data into trend term data and periodic term data, and model input factor screening is carried out. And then, based on a sliding window method and a Stacking frame, predicting the displacement of a periodic term, and predicting the displacement of a trend term by using an SVR algorithm to finally obtain a total displacement prediction result. According to the model, the displacement change of the mine slope can be accurately predicted, effective technical support is provided for mine slope disaster early warning and safety monitoring, and the model has high practical application value.

Description

Technical Field

[0001] The present invention belongs to the technical field of slope displacement, and particularly relates to an open-pit mine slope displacement prediction model and method based on InSAR and ensemble learning. Background Art

[0002] With the continuous increase in the demand for mineral resources, the mining scale and depth of open-pit mines are gradually expanding, and the slope stability problem has gradually become the focus of mine safety management. Due to its unique geological conditions and construction environment, open-pit mining is prone to slope instability. Especially during the mining process, due to factors such as construction interference and rainfall, the occurrence frequency and intensity of slope displacement are also increasing day by day. As important parameters for the development of slope deformation, the monitoring data of mine slope displacement is of great significance for accurately evaluating slope stability and timely predicting the occurrence of disasters such as landslides.

[0003] Traditional slope displacement prediction methods mainly rely on physical models and engineering experience. These methods usually require a large amount of on-site monitoring data, and when facing complex geological conditions, deep soil layers, and high-steep slopes, the prediction effects are often not ideal. Although some physical models and numerical simulation methods have certain applications in specific environments, they cannot completely solve the prediction accuracy problems caused by factors such as complex actual geological conditions and insufficient data. With the continuous development of machine learning technology, more and more scholars have begun to apply these technologies to slope displacement prediction and achieved remarkable progress.

[0004] In recent years' research, machine learning, especially deep learning, has been widely applied to the prediction of slope displacement. For example, deep learning models based on recurrent neural network (RNN), long short-term memory network (LSTM), convolutional neural network (CNN), etc. have become one of the main technical means for slope displacement prediction. These methods can discover the potential laws and complex non-linear relationships in the data by learning a large amount of historical monitoring data, thus providing new solutions for the dynamic monitoring and prediction of slope deformation. However, a single model often encounters problems of underfitting or overfitting when dealing with complex non-linear relationships and large-scale data, resulting in a decrease in prediction accuracy. Therefore, the ensemble learning method is proposed as a solution to improve the stability and accuracy of prediction by fusing the prediction results of multiple deep learning models. Summary of the Invention

[0005] To address the issues such as insufficient capture of complex spatio-temporal variations and accuracy problems in single-model prediction in traditional slope displacement prediction methods in the prior art, the present invention provides an open-pit mine slope displacement prediction model and method based on InSAR and ensemble learning, which can accurately predict the slope displacement of open-pit mines. Especially under the influence of factors such as mining activities and geological changes, the prediction results have higher reliability and accuracy, providing effective support for mine safety production and landslide warning. The present invention combines the high spatio-temporal resolution of InSAR data and the powerful expression ability of deep learning models, providing more scientific and accurate technical support for the prediction and warning of slope deformation.

[0006] The present invention realizes the solution of its technical problems by adopting the following technical solutions:

[0007] The first object of the present invention is to provide an open-pit mine slope displacement prediction method based on InSAR and ensemble learning, including:

[0008] S1. Data collection and preprocessing: Use InSAR technology to extract the displacement time series data of the open-pit mine slope from remote sensing images, and use the moving average method to decompose the original displacement time series data into a trend term and a periodic term to enhance the interpretability of the data;

[0009] S2. Multi-model ensemble prediction framework: Combine four deep learning models, namely RNN, LSTM, CNN, and DNN, to construct a base learner layer; use the Stacking ensemble learning framework to separately predict the trend term and the periodic term through the base learners, and then fuse the outputs of the base learners through the SVR model to optimize the prediction results; and screen the model input factors;

[0010] S3. Time series data processing: Use the sliding window method to divide the long sequence into subsequences of a fixed size, and each subsequence is used as an independent data window for training. The training set and the test set are divided in a ratio of 4:1 to predict the displacement of the periodic term and ensure the verification of the model generalization ability;

[0011] S4. Model tuning and verification: Use the grid search method to tune the parameters of the deep learning model and the SVR model to obtain the total displacement prediction results; use the double indicators of RMSE and MAPE to evaluate the prediction effect of the model, and combine the measured data to verify the model accuracy.

[0012] Further, the InSAR technology obtains data through Sentinel-1A satellite images, and after interference processing, extracts the displacement time series data of the open-pit mine slope.

[0013] Furthermore, import the data obtained from Sentinel-1A satellite images in ENVI-SARscape, crop the study area, and generate a connection map; then select a super master image, establish the master-slave relationship for all images according to the set baseline threshold, and then perform interference processing on each pair of image pairs according to the connection relationship of the image pairs.

[0014] Furthermore, the interference processing refers to registering each pair of image pairs with a master-slave relationship, generating an interferogram, flattening the interferogram, and calculating the amplitude deviation index, then performing two PS inversions on the data, and finally geocoding the PS results and projecting them into the map system.

[0015] Furthermore, the preprocessing refers to using the monitoring data after interference processing as the research object, expanding it into a dataset using the interpolation method, and then decomposing the dataset using the moving average method to form a trend term and a periodic term.

[0016] Furthermore, the trend term represents the long-term change trend of slope displacement, and the periodic term reflects seasonal fluctuations and the influence of external factors.

[0017] Furthermore, the prediction model framework of the trend term is constructed using the SVR algorithm.

[0018] Furthermore, the prediction model framework of the periodic term adopts the Stacking ensemble algorithm. In the first-layer algorithm, the base learner dataset is divided into four folds to obtain four mutually exclusive datasets, and the sliding window method is used to process the datasets; in the second-layer algorithm, the SVR algorithm is used as the meta-learner to divide the training set and the validation set.

[0019] Furthermore, when the sliding window method processes the datasets, use the first-fold dataset as the training set and the second-fold dataset as the validation set to obtain the first base model; use the first two-fold datasets as the training set and the third-fold dataset as the validation set to obtain the second base model; use the first three-fold datasets as the training set and the fourth-fold dataset as the validation set to obtain the third base model, and process each base learner dataset separately.

[0020] Furthermore, the selection of slope displacement time series data factors is screened using the grey relational analysis method. When the grey relational coefficient is higher, it indicates a stronger correlation between the two variables, and the factors with a Pearson correlation coefficient higher than this are excluded.

[0021] Furthermore, the method for model tuning is: establish a deep learning model using the Keras framework, search for the best parameters of the batch size, number of neurons, and number of model layers of the SVR model using the network search method, and then use the network search method again to determine the best parameters of the number of iterations.

[0022] The second object of the invention is to provide an open-pit mine slope displacement prediction model based on InSAR and integrated learning, including:

[0023] A data extraction module for extracting historical InSAR monitoring data;

[0024] A preprocessing module for preprocessing historical InSAR monitoring data;

[0025] A factor screening module for screening candidate factors for slope displacement;

[0026] A training module for establishing a deep learning model based on the preprocessed historical InSAR monitoring data and training according to the optimal parameters of the model;

[0027] A prediction module for inputting the to-be-processed InSAR monitoring data into the trained deep learning model for slope displacement prediction;

[0028] An evaluation and verification module for evaluating the prediction results and comparing and verifying them with the measured data.

[0029] Compared with the prior art, the beneficial technical effects of the present invention are as follows:

[0030] The slope displacement prediction system of the present invention combining InSAR and the Stacking integrated deep learning algorithm can assign the same or different weights to different factors, thus providing a new method for slope deformation prediction in complex geological environments. This method is not only applicable to the prediction of slope displacement by a single factor, but also can perform dynamic synchronous prediction for a combination of multiple factors, and can monitor the deviation between the prediction result and the actual displacement data in real time, more truly restoring the deformation law of the slope under different external influences. Especially in the prediction of long-term series data of mine slopes, the present invention can optimize the existing model, provide higher-precision prediction results, and provide an important theoretical basis for the construction of the safety monitoring and disaster warning system of mine slopes. The present invention optimizes various influencing factors, adopts a sliding window method to dynamically process slope displacement data, performs synchronous prediction of a single factor and a combination of multiple factors, and dynamically monitors the displacement change in real time during the prediction process. Through this innovative method, the present invention can more accurately reflect the deformation law of mine slopes in complex environments, provide a scientific basis for the early warning and prevention of slope disasters, and further optimize the slope stability evaluation and monitoring plan.

[0031] The above description is only an overview of the technical solution of the present invention. In order to be able to understand the technical means of the present invention more clearly, it can be implemented according to the content of the specification. And in order to make the above and other objects, features and advantages of the present invention more obvious and understandable, the specific embodiments of the present invention are given below. Description of the Drawings

[0032] Figure 1 This is the prediction flowchart of a slope displacement prediction model and method for open-pit mines based on InSAR and ensemble learning according to the present invention.

[0033] Figure 2 This is the flowchart of the traditional Stacking model.

[0034] Figure 3 This is the flowchart of the Stacking model optimized by the sliding window method for a slope displacement prediction model and method for open-pit mines based on InSAR and ensemble learning according to the present invention.

[0035] Figure 4 This is the operation flowchart of ENVI-SARscape in a slope displacement prediction model and method for open-pit mines based on InSAR and ensemble learning according to the present invention.

[0036] Figure 5 This is the historical image of the mine slope in 2018 and 2019 and the cumulative displacement trend chart of the slope in 2019 for a slope displacement prediction model and method for open-pit mines based on InSAR and ensemble learning according to the present invention, (a) Historical image of the mine slope in 2018; (b) Historical image of the mine slope in 2019; (c) Cumulative displacement trend chart of the slope in 2019.

[0037] Figure 6 This is the displacement decomposition result of the slope monitoring points in a slope displacement prediction model and method for open-pit mines based on InSAR and ensemble learning according to the present invention.

[0038] Figure 7 This is the curve graph of the displacement of slope monitoring points and rainfall monitoring data in a slope displacement prediction model and method for open-pit mines based on InSAR and ensemble learning according to the present invention.

[0039] Figure 8 This is the Pearson correlation coefficient in a slope displacement prediction model and method for open-pit mines based on InSAR and ensemble learning according to the present invention. Detailed implementation manners

[0040] The technical solution of the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following embodiments are only for illustrative and explanatory purposes of the present invention, and should not be construed as limiting the protection scope of the present invention. All technologies implemented based on the above content of the present invention are covered within the scope of protection intended by the present invention.

[0041] In addition, unless otherwise specifically stated, various raw materials, reagents, instruments, and equipment used in the present invention can be obtained through market purchases or prepared by existing methods.

[0042] A method for predicting the displacement of a mine slope based on the integration of InSAR and Stacking deep learning, specifically including the following steps:

[0043] S1. Data collection and preprocessing

[0044] 1. Extraction of mine slope displacement data based on InSAR

[0045] First, Sentinel-1A data and precise orbit files were downloaded, with the date ranging from January 10, 2019, to September 19, 2019, and one scene every 12 days. Then, the Sentinel-1A data was imported into ENVI-SARscape and the study area was cropped; next, a connection map was generated: a super master image was selected, and the master-slave relationship was established for all images according to the set baseline threshold. Then, according to the connection relationship of the image pairs, the interferometric workflow was processed for each pair of image pairs. Four processes were carried out in this step, namely registration, interferogram generation, interferogram flattening, and calculation of the amplitude deviation index. Subsequently, the first PS inversion was performed: the rate and elevation correction values were obtained, and the second PS inversion was performed: the atmospheric phase was estimated and removed. Finally, geocoding was carried out: all PS-related results (such as deformation rate, height residual, deformation sequence, KML, vector file, etc.) were projected into the map system. The schematic diagram of the ENVI-SARscape operation process is as Figure 4 shown. Finally, in combination with the collected meteorological data, we carried out a correlation analysis on the input rainfall factor and the output displacement, and selected a characteristic point for prediction research. Figure 5 It is the historical image of the mine slope in this mining area from 2018 to 2019 and the cumulative displacement trend map of the slope in 2019.

[0046] 2. Establishment and preprocessing of the dataset

[0047] 1) Dataset establishment.

[0048] The monitoring data of the slope monitoring points from January 10, 2019, to September 19, 2019, were selected as the research object. Among them, the data monitored in real time by InSAR were 10 time-step data, and the interpolation method was used to expand the dataset to form a dataset of 60 time steps.

[0049] 2) Decomposition of displacement data.

[0050] First, the original data of the slope displacement was decomposed using the moving average method. After decomposition, the first 48 time-step data were used as the fitting set data, and the last 12 time-step data were used as the prediction set to evaluate the ability of the model. The specific situation of the decomposition of the slope cumulative displacement is as Figure 6 shown.

[0051] The original displacement time series data is decomposed into a trend term and a periodic term using the moving average method to enhance data interpretability.

[0052] S2. Multi-model integrated prediction framework

[0053] 1. Trend term prediction framework.

[0054] The SVR algorithm is used for trend term prediction. The first 36 time series are the training set, the middle 12 time series are the validation set, and the last 12 time series are the test set.

[0055] 2. Periodic term prediction framework.

[0056] The Stacking integration algorithm is used for periodic term prediction. The influencing factors selected are the rainfall on the current day, the rainfall on the day before yesterday, and the rainfall two days ago. In the first-layer algorithm, the fitting set is first divided into 4 folds to obtain 4 mutually exclusive data sets. The base learners use four deep learning algorithms, namely LSTM (long-dependency capture), RNN (time series modeling), CNN (local feature extraction), and DNN (nonlinear fitting). The sliding window method is used to process the fitting set data. Taking LSTM as an example, the first fold data set is used as the training set and the second fold data set is used as the validation set to obtain the first base model; the first two fold data sets are used as the training set and the third fold data set is used as the validation set to obtain the second base model; the first three fold data sets are used as the training set and the fourth fold data set is used as the validation set to obtain the third base model. The same applies to the other base learners. The test set uses the weighted average based on the MAE index. In the second-layer algorithm, the SVR algorithm is used as the meta-learner, and the first 24 time series are selected as the training set, the middle 12 time series are the validation set, and the last 12 time series are the test set.

[0057] 3. Factor screening

[0058] 1) Candidate factors for slope displacement.

[0059] Rainfall-related factors are one of the important triggering factors for open-pit mine slopes. In the embodiments of the present invention, several factors related to rainfall changes are selected as candidate factors for slope displacement, namely the rainfall on the current day, the rainfall on the previous day, and the rainfall two days ago, and are marked as f1 - f3 in sequence. In the factor screening stage, correlation is usually used for factor screening. Figure 7 It is the graph of cumulative displacement of slope points versus rainfall.

[0060] 2) The grey relational analysis method is used to screen the candidate factors.

[0061] When the grey correlation coefficient is higher, it indicates a stronger correlation between two variables. The grey correlation coefficients between the candidate factors and the periodic displacement of the monitoring points are shown in Table 1. The results in the table show that there is a strong correlation between the selected candidate factors and the periodic displacement.

[0062] Table 1 Grey Correlation Coefficient Table

[0063]

[0064] 3) Perform Pearson correlation coefficient analysis on the candidate factors.

[0065] By screening the input factors that exceed the standard, the prediction accuracy of the model can be improved. Generally speaking, a correlation coefficient higher than 0.6 indicates high collinearity and should be excluded. As can be seen from the figure, the candidate factors selected in this section have low collinearity with other factors and do not need to be excluded. The Pearson correlation diagram between the candidate factors and the periodic displacement is as Figure 8 shown.

[0066] S3. Time Series Data Processing

[0067] 1. Normalization and inverse normalization.

[0068] To avoid information leakage, the boundary values of normalization and inverse normalization must be obtained from the fitting dataset instead of the entire dataset, because the fitting dataset is regarded as known information while the prediction dataset is unknown.

[0069] 2) Hyperparameters of the deep learning model and parameters of the SVR model. Use the Keras framework to build the deep learning model, use TensorFlow as the backend, and write the code in Python language.

[0070] The SVR model runs in the Python language environment, and the grid search method (GS) is used to search for the best parameters of the SVR model. During the grid search process, the search range of Gamma is [0.05, 17.05] with a step size of 0.05; the search range of C is [1, 5] with a step size of 1. The parameters of the SVR model for predicting the trend term are shown in Table 2.

[0071] Table 2 Parameters of the Trend Term Prediction Model

[0072]

[0073] S4. Model Tuning and Validation

[0074] The deep learning model runs in the Python language environment. To build the model, the batch size, the number of neurons, and the number of model layers are determined in sequence by the grid search (GS) method. During the GS process, the range of the batch size of the grid is [1, 20] with a step size of 1; the range of the number of neurons is [1, 20] with a step size of 1. During the algorithm process, the best number of training epochs is obtained by setting the early stopping method, and the patience parameter of the early stopping method is set to 50, which means that when the result does not improve within 50 steps, the algorithm process will stop. After determining the three parameters of the batch size, the number of neurons, and the number of neural network layers, the grid search method is used again to determine the number of iterations, with a range of [50, 120] and a step size of 1. The best hyperparameters of each group of base models are shown in Table 3, where "TR n" represents the base model corresponding to each base learner, and the best parameters of the second-layer algorithm are shown in Table 4.

[0075] Table 3 Hyperparameters of the periodic term prediction model

[0076]

[0077]

[0078] Table 4 Parameters of the second-layer model

[0079]

[0080] The prediction effect of the model is evaluated using two indicators, RMSE (quantifying the absolute error) and MAPE (evaluating the percentage of relative error), and the model accuracy is verified by combining the measured data.

[0081] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0082] The embodiments of the present invention have been described above in conjunction with the accompanying drawings. However, the present invention is not limited to the above specific embodiments. The above specific embodiments are merely illustrative and not restrictive. Under the inspiration of the present invention, those of ordinary skill in the art can also make many forms without departing from the spirit and scope protected by the present invention and the claims. These all fall within the protection scope of the present invention.

Claims

1. A method for predicting slope displacement in open-pit mines based on InSAR and ensemble learning, characterized in that: include: S1. Data acquisition and preprocessing: InSAR technology is used to extract the displacement time series data of the open-pit mine slope from remote sensing images, and the moving average method is used to decompose the original displacement time series data into trend terms and periodic terms to enhance data interpretability; S2, multi-model integrated prediction framework: Combine the four deep learning models of RNN, LSTM, CNN, and DNN to build a base learner layer; use the Stacking integrated learning framework to predict trend items and period items respectively through the base learners, and then fuse the base learner outputs through the SVR model to optimize the prediction results; And the model input factors are screened; S3, Time series data processing: Use the sliding window method to divide the long sequence into subsequences of fixed size. Each subsequence is trained as an independent data window. The training set and the test set are divided in a 4:1 ratio to predict the displacement of periodic items and ensure the generalization ability of the model. S4. Model tuning and verification: The grid search method is used to tune the parameters of the deep learning model and the SVR model to obtain the prediction results of the total displacement; the RMSE and MAPE dual indicators are used to evaluate the model prediction effect, and the model accuracy is verified by combining the measured data.

2. The method for predicting open-pit mine slope displacement based on InSAR and ensemble learning as claimed in claim 1, characterized in that: The interference processing refers to the registration of each pair of images that establish a master-slave relationship, the generation of interference patterns, the curvature of interference patterns, and the calculation of amplitude deviation index, followed by two PS inversions of the data, and finally geocoding the PS results and projecting them into the map system.

3. The method for predicting open-pit mine slope displacement based on InSAR and ensemble learning as claimed in claim 1, characterized in that: The preprocessing refers to taking the monitoring data after interference processing as the research object, expanding it by interpolation method to form a data set, and then decomposing the data set by moving average method to form trend items and period items.

4. The method for predicting open-pit mine slope displacement based on InSAR and ensemble learning as claimed in claim 3, characterized in that: The prediction model framework of the trend item is constructed using the SVR algorithm.

5. The method for predicting open-pit mine slope displacement based on InSAR and ensemble learning as claimed in claim 3, characterized in that: The prediction model framework of the periodic item adopts the Stacking integration algorithm. In the first layer of the algorithm, the base learner data set is divided into four folds to obtain four mutually exclusive data sets, and the data set is processed using the sliding window method; in the second algorithm, the SVR algorithm is used as the meta-learner to divide the training set and the validation set.

6. The method for predicting open-pit mine slope displacement based on InSAR and ensemble learning as claimed in claim 1, characterized in that: When the sliding window method is used to process the data set, the first fold data set is used as the training set and the second fold data set is used as the validation set to obtain the first base model; the first two fold data sets are used as the training set and the third fold data set is used as the validation set to obtain the second base model; the first three fold data sets are used as the training set and the fourth fold data set is used as the validation set to obtain the third base model, and each base learner data set is processed separately.

7. The method for predicting open-pit mine slope displacement based on InSAR and ensemble learning as claimed in claim 1, characterized in that: The factors of slope displacement time series data are selected by grey correlation analysis. The higher the grey correlation coefficient is, the stronger the correlation is between the two variables. The factors with a higher Pearson correlation coefficient are eliminated.

8. The method for predicting open-pit mine slope displacement based on InSAR and ensemble learning as claimed in claim 1, characterized in that: The method of model tuning is as follows: a deep learning model is established using the Keras framework, a network search method is used to search for the optimal parameters of the SVR model batch size, number of neurons, and number of model layers, and then the network search method is used again to determine the optimal parameter for the number of iterations.

9. An open-pit mine slope displacement prediction model based on InSAR and ensemble learning, characterized in that: include: Data extraction module, used to extract historical InSAR monitoring data; Preprocessing module, used to preprocess historical InSAR monitoring data; Factor screening module, used to screen candidate factors for slope displacement; The training module is used to establish a deep learning model based on the preprocessed historical InSAR monitoring data and train it according to the optimal parameters of the model; The prediction module is used to input the InSAR monitoring data to be processed into the trained deep learning model for slope displacement prediction; The evaluation and verification module is used to evaluate the prediction results and compare and verify them with the measured data.

Images