Landslide displacement prediction method using Markov attention mechanism

Abstract

The invention discloses a landslide displacement prediction method using a Markov attention mechanism, and relates to the field of landslide displacement feature extraction and high-precision prediction. The method comprises the following steps: acquiring a landslide displacement time sequence of a research area, and inputting the landslide displacement time sequence into a landslide displacement prediction model; in the landslide displacement prediction model, constructing a landslide displacement Markov matrix based on the landslide displacement time sequence; the more later the elements in the landslide displacement time sequence appear in the landslide displacement Markov matrix; performing feature extraction on the landslide displacement Markov matrix through a Markov matrix information extraction model to obtain landslide displacement change features; adding the landslide displacement time sequence and the landslide displacement change characteristics to obtain an output sequence after the landslide displacement time sequence is processed by the Markov attention mechanism; and inputting the output sequence into the nonlinear model to obtain a landslide displacement predicted value of the research area. The method can improve the prediction precision of the landslide displacement.

Description

Technical Field

[0001] The present invention relates to the field of landslide displacement feature extraction and high-precision prediction, and in particular to landslide displacement prediction using a Markov attention mechanism. Background Art

[0002] Landslide displacement prediction is a crucial part of landslide disaster analysis. It is mainly based on the regression idea in time series analysis. A regression model is established between influencing factor data or historical landslide displacement data and future data. The model parameters are determined using optimization methods such as least squares and gradient descent to obtain a landslide displacement prediction model. The landslide process is relatively complex. Taking other influencing factors such as rainfall and temperature into account and combining external influencing factors to establish a landslide displacement prediction model is called multivariate landslide displacement prediction. However, different landslide scenes, different internal structures, and different environments will make the generalization performance of the multivariate model in different landslide displacement prediction scenarios poor.

[0003] At present, landslide displacement prediction methods can be divided into two categories: traditional prediction methods and data-driven methods. Traditional prediction methods are mainly based on gray prediction models, which have low data requirements and high ultra-short-term prediction accuracy. Therefore, data-driven models are the most commonly used models in landslide displacement prediction. Commonly used data-driven models are fitting models, traditional machine learning models, and deep learning models. The most widely used method in landslide displacement prediction is the deep learning prediction method, which converts the high-demand feature engineering into automated neural learning, improving the generalization performance and accuracy of model prediction.

[0004] However, the existing landslide displacement prediction methods mainly optimize the model itself without analyzing and modeling the characteristics of landslide displacement, resulting in low prediction accuracy of landslide displacement. Summary of the invention

[0005] Based on this, it is necessary to provide a landslide displacement prediction method using Markov attention mechanism to address the above technical issues. This method can improve the prediction accuracy of landslide displacement.

[0006] The present invention adopts the following technical solutions:

[0007] The present invention provides a landslide displacement prediction method using a Markov attention mechanism, comprising:

[0008] Obtain the landslide displacement time series in the study area and input the landslide displacement time series into the landslide displacement prediction model; the landslide displacement prediction model includes a Markov matrix information extraction model and a nonlinear model;

[0009] In the landslide displacement prediction model, the landslide displacement Markov matrix is ​​constructed based on the landslide displacement time series; the later the time element in the landslide displacement time series, the more times it appears in the landslide displacement Markov matrix;

[0010] The Markov matrix information extraction model is used to extract the features of the landslide displacement Markov matrix and obtain the landslide displacement change features; the Markov matrix information extraction model is constructed by multiple serially connected Markov attention mechanisms;

[0011] The landslide displacement time series and the landslide displacement change characteristics are added together to obtain the output sequence after the landslide displacement time series is processed by the Markov attention mechanism;

[0012] The output sequence is input into the nonlinear model to obtain the predicted values ​​of landslide displacement in the study area.

[0013] Preferably, the training process of the landslide displacement prediction model includes:

[0014] Obtaining the historical landslide displacement time series, and inputting the historical landslide displacement time series into the initial landslide displacement prediction model;

[0015] In the initial landslide displacement prediction model, the historical landslide displacement Markov matrix is ​​constructed based on the historical landslide displacement time series;

[0016] The historical landslide displacement Markov matrix is ​​extracted and processed by connecting multiple Markov attention mechanisms in series to obtain the historical extraction results.

[0017] The historical landslide displacement time series and the historical extraction results are added together to obtain the historical output sequence after the Markov attention mechanism processes the historical landslide displacement time series;

[0018] Input the historical output sequence into the nonlinear model to obtain the prediction value of the initial landslide displacement prediction model for the historical landslide displacement time series;

[0019] Get the observations of the historical output sequence;

[0020] The mean square error between the predicted value and the observed value is calculated, and the parameters of the initial landslide displacement prediction model are adjusted until the error is minimized to obtain the landslide displacement prediction model.

[0021] Preferably, the landslide displacement time series is:

[0022] X={x 1 ,x 2 ,x 3 ,…,x n};

[0023] Where X is the landslide displacement time series, x irepresents the cumulative landslide displacement at the i-th moment, where i ranges from 1 to n;

[0024] The landslide displacement Markov matrix is:

[0025]

[0026] Among them, MarkovMatrix X is the landslide displacement Markov matrix, x i It represents the cumulative landslide displacement at the i-th moment, and the value of i ranges from 1 to n.

[0027] Preferably, the Markov attention mechanism includes one-dimensional temporal convolution, one-dimensional pooling, and an activation function. The implementation process of the Markov attention mechanism specifically includes:

[0028] Input the input information of the Markov attention mechanism into the one-dimensional temporal convolution module to obtain the output of the one-dimensional temporal convolution module;

[0029] Input the output of the one-dimensional temporal convolution module into the one-dimensional pooling module to obtain the pooling result;

[0030] The pooling result is input into the activation function to obtain the output result of the Markov attention mechanism.

[0031] Preferably, the nonlinear model is:

[0032] NLinear:R M →R N ;

[0033] Among them, M represents the length of the nonlinear model input sequence, and N represents the length of the nonlinear model prediction sequence.

[0034] The present invention provides a landslide displacement prediction device using a Markov attention mechanism, comprising:

[0035] An acquisition module is used to acquire the landslide displacement time series in the study area and input the landslide displacement time series into a landslide displacement prediction model; the landslide displacement prediction model includes a Markov matrix information extraction model and a nonlinear model;

[0036] A construction module is used to construct a landslide displacement Markov matrix based on the landslide displacement time series in the landslide displacement prediction model; the later the time element in the landslide displacement time series, the more times it appears in the landslide displacement Markov matrix;

[0037] An extraction module is used to extract features from the landslide displacement Markov matrix through a Markov matrix information extraction model to obtain the landslide displacement change features; the Markov matrix information extraction model is constructed by multiple serially connected Markov attention mechanisms;

[0038] The first determination module is used to add the landslide displacement time series and the landslide displacement change characteristics to obtain an output sequence after the landslide displacement time series is processed by the Markov attention mechanism;

[0039] The second determination module is used to input the output sequence into the nonlinear model to obtain the predicted value of the landslide displacement in the study area.

[0040] The present invention provides a computer-readable storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the landslide displacement prediction method using the Markov attention mechanism is implemented.

[0041] The present invention provides a computer device, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the above-mentioned landslide displacement prediction method using the Markov attention mechanism is implemented.

[0042] At least one of the above technical solutions adopted by the present invention can achieve the following beneficial effects:

[0043] Firstly, by constructing a Markov matrix based on the landslide displacement time series, the information contained is gradually reduced from the back to the front according to the time, ensuring that the data closer to the prediction point contributes more information to the landslide displacement prediction model, which greatly improves the ability of the landslide displacement prediction model to extract the information near the prediction point. Secondly, multiple Markov attention mechanisms are used to extract the Markov matrix based on the landslide displacement time series, and the extraction results are added to the landslide displacement time series to obtain the output of the Markov attention mechanism. The single representation learning and reasoning prediction are combined to establish a landslide displacement prediction model that combines the Markov attention mechanism and the nonlinear model. This method can improve the prediction accuracy of landslide displacement prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0045] Figure 1 A flow chart of a landslide displacement prediction method using a Markov attention mechanism provided by the present invention;

[0046] Figure 2 The Markov attention mechanism workflow diagram provided by the present invention;

[0047] Figure 3 A flowchart of a specific embodiment provided by the present invention;

[0048] Figure 4 This is a schematic diagram of data after preprocessing of the landslide data at point HF05 in a specific embodiment of the present invention;

[0049] Figure 5 This is a schematic diagram of data after preprocessing of the landslide data at point HF08 in a specific embodiment of the present invention;

[0050] Figure 6 This is a data schematic diagram of the landslide data of point HF09 after preprocessing in a specific embodiment of the present invention;

[0051] Figure 7 A comparison chart of the prediction performance of the MANLinear model provided by the present invention in the displacement mutation stage;

[0052] Figure 8 A prediction residual graph of different landslide displacement prediction models provided by the present invention in the case of sudden landslide displacement prediction;

[0053] Fig. 9 A comparison chart of the improvement rate of neural network accuracy of the MANLinear model provided by the present invention under prediction tasks with different data volumes;

[0054] Fig.10 A comparison chart of weight results at different positions of the neural network after training provided by the present invention is completed;

[0055] Fig.11 This is a data sensitivity analysis result diagram of the MANLinear model provided by the present invention;

[0056] Fig.12 A schematic diagram of a landslide displacement prediction device using a Markov attention mechanism provided by the present invention;

[0057] Fig.13 A schematic diagram of a computer device for implementing a landslide displacement prediction method using a Markov attention mechanism provided by the present invention. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0059] Devices such as desktop computers, servers, and notebook computers that can execute the solution of the present invention. For the sake of convenience, the following description will only take the server as the execution subject.

[0060] The technical solutions provided by various embodiments of the present invention are described in detail below in conjunction with the accompanying drawings.

[0061] Figure 1 The present invention is a flow chart of a landslide displacement prediction method using a Markov attention mechanism, which specifically includes the following steps:

[0062] S101: Obtain the landslide displacement time series in the study area, and input the landslide displacement time series into a landslide displacement prediction model; the landslide displacement prediction model includes a Markov matrix information extraction model and a nonlinear model.

[0063] In an exemplary embodiment, the landslide displacement time series in the study area is shown in formula (1):

[0064] X={x 1 ,x 2 ,x 3 ,…,x n} (1);

[0065] Where X is the landslide displacement time series, x i It represents the cumulative landslide displacement at the i-th moment, where i ranges from 1 to n.

[0066] Specifically, landslide displacement has a very obvious autocorrelation characteristic. There will be different landslide-inducing factors under different geological and environmental conditions. Therefore, the study of the autocorrelation of landslide displacement is more fundamental and has better universality.

[0067] In an exemplary embodiment, the training process of the landslide displacement prediction model includes: obtaining a historical landslide displacement time series, and inputting the historical landslide displacement time series into an initial landslide displacement prediction model; in the initial landslide displacement prediction model, constructing a historical landslide displacement Markov matrix based on the historical landslide displacement time series; extracting and processing the historical landslide displacement Markov matrix through a plurality of Markov attention mechanisms connected in series to obtain a historical extraction result; adding the historical landslide displacement time series and the historical extraction result to obtain a historical output sequence after the historical landslide displacement time series is processed by the Markov attention mechanism; inputting the historical output sequence into a nonlinear model to obtain a predicted value of the initial landslide displacement prediction model for the historical landslide displacement time series; obtaining an observed value of the historical output sequence; calculating the mean square error between the predicted value and the observed value, and adjusting the parameters of the initial landslide displacement prediction model until the error is minimized to obtain a landslide displacement prediction model.

[0068] The mean square error between the predicted value and the observed value is calculated as shown in formula (2):

[0069]

[0070] Among them, Y is the observed value of landslide displacement, is the predicted value of the landslide displacement model, is the mean square error between the predicted and observed landslide displacements, N is the number of observed landslide displacements, With Y i They represent the model predicted value of the i-th landslide displacement and the observed value of the i-th landslide displacement respectively.

[0071] Specifically, the historical landslide displacement time series is used as a training set; the initial landslide displacement prediction model is trained using the training set, and the mean square error between the predicted values ​​of the initial landslide displacement prediction model and the observed values ​​of the historical output sequence is calculated; the landslide displacement prediction model is obtained by adjusting the parameters of the initial landslide displacement prediction model until the mean square error is minimized.

[0072] Specifically, in an embodiment of the present invention, the error calculation formula between the predicted value and the true value of the landslide displacement prediction model for the historical landslide displacement sequence is calculated as formula (2). The prediction accuracy of the landslide displacement prediction model is evaluated using the mean square error, and then the model parameters are adjusted using the Adam optimization algorithm including learning rate adaptation and momentum gradient descent, the convergence of the model error is completed on the training set, and the model generalization error is tested on the test set. After the landslide displacement prediction model is trained, the model parameters can be fixed and deployed in the actual landslide displacement prediction.

[0073] S102: In the landslide displacement prediction model, a landslide displacement Markov matrix is ​​constructed based on the landslide displacement time series; the later the time element in the landslide displacement time series, the more times it appears in the landslide displacement Markov matrix.

[0074] In an exemplary embodiment, the landslide displacement Markov matrix is ​​shown in formula (3):

[0075]

[0076] Among them, MarkovMatrix X is the landslide displacement Markov matrix, x i It represents the cumulative landslide displacement at the i-th moment, and the value of i ranges from 1 to n.

[0077] Specifically, the data with earlier time nodes in the landslide displacement time series appear less frequently in the landslide displacement Markov matrix, while the data with later time nodes appear more frequently in the landslide displacement Markov matrix and provide more information.

[0078] S103: extracting features from the landslide displacement Markov matrix through a Markov matrix information extraction model to obtain landslide displacement change features; the Markov matrix information extraction model is constructed by multiple serially connected Markov attention mechanisms.

[0079] In an exemplary embodiment, the Markov attention mechanism includes one-dimensional temporal convolution, one-dimensional pooling, and an activation function. The implementation process of the Markov attention mechanism specifically includes: inputting the input information of the Markov attention mechanism into the one-dimensional temporal convolution module to obtain the output of the one-dimensional temporal convolution module; inputting the output of the one-dimensional temporal convolution module into the one-dimensional pooling module to obtain the pooling result; inputting the pooling result into the activation function to obtain the output result of the Markov attention mechanism.

[0080] Specifically, the input of the first Markov attention mechanism in the Markov matrix information extraction model is the landslide displacement Markov matrix. In turn, the Markov matrix is ​​processed by multiple serially connected Markov attention mechanisms, and the output of the previous Markov attention mechanism is the input of the next Markov attention mechanism.

[0081] One-dimensional temporal convolution is shown in formula (4):

[0082]

[0083] Among them, x is the given input sequence, w represents the convolution kernel, K is the convolution kernel length, y is the output of the one-dimensional temporal convolution, y is recorded as Conv1d(), i is the convolution kernel index, and t is the time point of the output sequence.

[0084] Pooling is an important means to reduce data dimension and prevent overfitting, which mainly includes mean pooling and maximum pooling.

[0085] The maximum pooling is shown in formula (5):

[0086] y j =max(x j ,x j+1 ,…,x j+K-1 ) (5);

[0087] Among them, y j is the output value after the pooling operation, that is, the maximum value of the pooling window, x j is the starting value of the pooling window, j is the starting position of the pooling window, and K is the length of the pooling window.

[0088] Mean pooling is shown in formula (6):

[0089]

[0090] Where x is the given input time series, K is the pooling window length, and y is the pooling result. For the convenience of representation, the pooling result is recorded as Pooling1d(x).

[0091] The activation function is the key to the nonlinear mapping of the neural network. The commonly used ReLU function is shown in formula (7):

[0092] ReLU(x)=max(0,x) (7);

[0093] Among them, ReLU is the activation function and x is the given input time series.

[0094] The Markov attention mechanism is shown in formula (8) and formula (9):

[0095] G=ReLU(CnnBlock(CnnBlock(MarkovMatrix X )) (8);

[0096] CnnBlock(x)=Pooling1d(ReLU(Conv1d(x))) (9);

[0097] Among them, G is the extraction result, ReLU is the activation function, and MarkovMatrix X is the landslide displacement Markov matrix, Pooling1d is the pooling result, and Conv1d is the one-dimensional temporal convolution operation.

[0098] S104: Add the landslide displacement time series and the landslide displacement change feature to obtain an output sequence after the landslide displacement time series is processed by the Markov attention mechanism.

[0099] Specifically, G in formula (8) is the extraction result, which is added to the landslide displacement time series to obtain the output sequence after the Markov attention mechanism processes the landslide displacement time series, as shown in formula (10):

[0100]

[0101] in, is the output sequence, G is the extraction result, and X is the landslide displacement time series.

[0102] Specifically, the overall calculation process of the Markov attention mechanism is as follows Figure 2 As shown, the landslide displacement input sequence is the landslide displacement time series mentioned in the present invention. A Markov matrix is ​​constructed based on the landslide displacement input sequence. The constructed Markov matrix is ​​subjected to multi-layer 1-D convolution, activation function ReLU, and 1-D pooling operations to obtain an output sequence after the landslide displacement time series is processed by the Markov attention mechanism.

[0103] S105: Input the output sequence into the nonlinear model to obtain the predicted value of the landslide displacement in the study area.

[0104] In an exemplary embodiment, the nonlinear model is shown in formula (11):

[0105] NLinear:R M →R N (11);

[0106] Among them, M is the length of the nonlinear model input sequence, and N is the length of the nonlinear model prediction sequence.

[0107] Specifically, the nonlinear model is a mapping from input data dimension to output data dimension. In the present invention, the input sequence of the nonlinear model is the output sequence of the Markov attention mechanism, and the output sequence of the nonlinear model is the predicted value of the landslide displacement time series.

[0108] In an exemplary embodiment, the predicted value of the landslide displacement time series by the trained landslide displacement prediction model is shown in formula (12):

[0109]

[0110] in, is the predicted value of the landslide displacement time series by the trained landslide displacement model, is the output sequence of the Markov attention mechanism of the landslide displacement time series, and NLinear is a nonlinear model.

[0111] Specifically, the output sequence of the Markov attention mechanism is predicted by a nonlinear model to obtain the predicted value of the landslide displacement time series by the landslide displacement model.

[0112] The present invention innovatively models the Markov characteristics of the landslide displacement change process, so that the landslide displacement prediction model can more directly focus on the information of the points near the prediction point in the input information. Combined with the subsequent prediction model, a higher-precision landslide displacement prediction is achieved. The present invention realizes the modeling of the Markov process by designing a Markov attention mechanism, and combines it with the existing neural network prediction model, which greatly improves the prediction accuracy of the landslide displacement model, and significantly improves the accuracy and stability of the prediction in the stage of rapid change of landslide displacement.

[0113] The present invention greatly improves the prediction accuracy of landslide displacement by considering the Markov characteristics of the landslide displacement change process. According to the high-precision landslide displacement prediction value, it can be further combined with landslide threshold judgment and other work, and can be actually deployed in dangerous slopes or slopes with hidden dangers, providing a scientific basis for economic development, urban and rural construction, land use planning, geological disaster prevention and control, etc.

[0114] In an exemplary embodiment, Figure 3 The specific workflow diagram of the scheme of the present invention is provided as shown. First, the accumulated time-series monitoring displacement data of the landslide is obtained, and the displacement data is preprocessed to obtain the preprocessed data. The preprocessing specifically includes gross error elimination, trend extraction, and data resampling. Because there are still a large number of significant gross errors in the original data due to the acquisition method, gross error elimination is required; at the same time, due to the limitations of the positioning method itself, the result has obvious fluctuations, so it is necessary to extract the trend; the sampling interval of the original data is 5 minutes, and the dense sampling interval leads to significant data redundancy. At the same time, the time interval cannot meet the needs of landslide displacement warning and forecasting, so it is necessary to resample the data; the preprocessed data is segmented to obtain a test set, a validation set, and a training set; then, a landslide displacement prediction model (Markov AttentionNLinear, MANLinear) using a Markov attention mechanism proposed in the present invention is trained according to the training set, and the MANLinear model is optimized using the validation set to obtain a trained MANLinear model; finally, the test set is input into the trained MANLinear model to obtain the prediction result of the MANLinear model for the test set.

[0115] In another exemplary embodiment, the experimental area selected by the present invention is a typical loess landslide area. In order to better grasp the movement and deformation conditions of the typical landslide body, a high-precision Beidou and Global Navigation Satellite System (GNSS) monitoring network was deployed in the area to obtain the long-term deformation monitoring information of the typical landslide body. The landslide data of three collapse events were selected: the HF08 point landslide on March 26, 2019, the HF09 point landslide on January 29, 2021, and the HF05 point landslide data on July 8, 2021. In view of the complexity of the study area environment and the existence of gross errors in the monitoring time series displacement, the acquired landslide displacement monitoring data contains significant gross errors. At the same time, in order to verify the test results of the model at different time scales, the original data is subjected to data preprocessing such as gross error elimination, trend extraction and resampling. Figure 4 is the preprocessed data of the HF05 point landslide, such as Figure 5 is the preprocessed data of the HF08 point landslide, such as Figure 6This is the preprocessed data of the landslide at point HF09.

[0116] In this embodiment, the split ratio of the training set, validation set and test set is 8:1:1. In order to ensure sufficient information input and consider the complexity of the model, 7 historical landslide displacement data are selected to predict the future displacement data of 1 landslide. Table 1 shows the number of training sets of different landslide monitoring points at different time intervals obtained after organizing the data sets of HF05, HF08 and HF09. The prediction performance of different models is evaluated by calculating the root mean square error (RMSE) between the prediction results and the true values ​​in the test set.

[0117] Table 1

[0118]

[0119] In order to verify the superiority of MANLinear proposed in the present invention, the prediction accuracy of different models is compared and analyzed. In addition to MANLinear and the basic model (NLinear), time series prediction models in the field of deep learning are selected, including non-stationary Transformer, frequency domain enhanced decomposition model (FEDformer), autocorrelation decomposition model (Autoformer) and efficient long sequence prediction model (Informer), as well as traditional machine learning algorithms including support vector regression (SVR), extreme learning machines (ELM), long short-term memory network (LSTM) and gated recurrent unit (GRU). The prediction accuracy of different models is evaluated by calculating the RMSE of the above models in the test set.

[0120] In this embodiment, experiments were conducted on different time resolutions of landslide displacement monitoring points HF05, HF08 and HF09, and the RMSE errors of different models on the test set were compared as shown in Table 2.

[0121] Table 2

[0122]

[0123] For the HF05 landslide monitoring point with sufficient data, the RMSE of the MANLinear model is better than all other compared models. Compared with the basic model NLinear before improvement, the accuracy is improved by 91.96%, 74.59% and 63.68% at 6-hour, 12-hour and 24-hour time intervals, respectively. For the HF08 data set with insufficient data, the RMSE error of the MANLinear model is also lower than that of all other compared models. Compared with the basic model NLinear, the accuracy is improved by 91.73%, 82.08% and 78.75% at 6-hour, 12-hour and 24-hour time intervals, respectively. For the prediction of the HF09 landslide monitoring point, in the case of insufficient training data, MANLinear still has the smallest RMSE compared with all models. Compared with the basic model NLinear, the accuracy is improved by 64.24%, 53.82% and 42.92% at 6-hour, 12-hour and 24-hour time intervals, respectively.

[0124] Figure 7 The prediction performance of the MANLinear model in the displacement mutation stage is intuitively displayed. The test results of HF05, HF08 and HF09 at a 6-hour time interval are visualized. In order to show the comparison results more clearly, in addition to the model proposed in the present invention, only the test results of the two models with the second-best prediction accuracy in Table 2 are selected for comparison. For HF05, HF08 and HF09 points, the other two models with the second-best prediction accuracy are LSTM and GRU, LSTM and ELM, and LSTM and ELM, respectively.

[0125] Figure 8 The prediction residuals of different landslide displacement prediction models under the condition of landslide displacement mutation are given. The different prediction performances of different models under the condition of landslide displacement mutation in places with large data volume are further analyzed. Figure 8 Figure (a) shows the central mutation of HF05. Figure 8 Figure (b) shows the end mutation of HF05. Figure 8 Figure (c) shows the mutation at the end of HF08, and the absolute value of the prediction residual of each model is calculated. Figure 8 The results show that the MANLinear model is significantly better than all the comparison models in the prediction of landslide displacement mutation. Its residual mean and residual range are significantly lower than those of the selected comparison models, which are about 1 / 2 to 1 / 4 of the comparison models. This shows that the model proposed in the present invention has more superior prediction accuracy and stability in the case of landslide displacement mutation. At the same time, this result also verifies the correctness of the design of the Markov attention mechanism.

[0126] Fig. 9 A comparison chart of the improvement rate of neural network accuracy of the MANLinear model proposed in the present invention under prediction tasks with different data amounts is given. Fig. 9 It shows that for the HF08 landslide monitoring point, the prediction accuracy improvement rate of the model at different time intervals is significantly higher than that of the HF05 landslide monitoring point, especially at the 24-hour time interval, the difference in accuracy improvement rate between HF08 and HF05 before and after model improvement is about 20%. The above characteristics show that the Markov attention mechanism can still greatly improve the accuracy of the NLinear model when data is relatively scarce; the accuracy improvement rate of the HF09 landslide monitoring point at different time intervals is significantly lower than that of HF08 and HF05, indicating that when the amount of data is too small, it will limit the prediction effect of the improved algorithm, but the MANLinear model can still achieve a considerable accuracy improvement.

[0127] Fig.10 A comparison chart of the weights of different positions of the neural network is given. In order to verify the assumption that the Markov attention mechanism pays more attention to the nearby information of the prediction point, a two-layer linear neural network without hidden layers is designed. The HF05 dataset with the largest amount of data is used for the 6-hour time interval. Fig.10 Figure (a) and the 3-hour time interval are shown in Fig.10 The training is performed using the graph (b) in Figure 1. Fig.10 As shown in the figure, there are obvious differences in the weights of different positions. The positions closer to the predicted point (the further back) have larger weights, while the positions farther from the predicted point (the closer to the front) have smaller weights, and the weights are close to 0. This result is consistent with the existing related research results, verifying the starting point of the design of the Markov attention mechanism.

[0128] Fig.11 The data sensitivity analysis results of the MANLinear model proposed in this paper are given. Models based on data-driven tend to improve the prediction accuracy of the model as the amount of data increases, that is, as the amount of training data increases, the model can more accurately learn the inherent changes in the data. Therefore, for the same prediction task, the change in error under different data amounts reflects the data-driven performance of the model. Fig.11 Figure (a) in Fig.11 Figure (b) and Fig.11 Figure (c) shows the accuracy change of different data volumes of HF05, HF08 and HF09 landslide monitoring points before and after model improvement. Fig.11It can be seen that as the amount of data increases, its accuracy does not steadily decrease, and even increases instead of decreasing. Therefore, the increase in the amount of data cannot directly improve the accuracy of the basic model prediction, and may even make the model worse, indicating that the basic model is not an efficient data-driven model. Due to its single structure, the basic model cannot fully extract the information contained in the data after increasing the amount of data, so it will reduce the prediction accuracy of the model. The improved model proposed in the present invention not only significantly improves the prediction accuracy of the basic model, but also shows good data-driven characteristics, that is, as the amount of data increases, its prediction accuracy also steadily increases.

[0129] The MANLinear model proposed in the present invention first constructs a time-series Markov matrix so that the information it contains is gradually reduced from the back to the front, ensuring that the data closer to the prediction point contributes more information to the model, and then uses a stacked one-dimensional convolutional network to perform sliding convolution on the Markov matrix, and finally adds a residual connection to the convolution result to obtain the output of the Markov attention mechanism. At the same time, it combines single representation learning with reasoning prediction to establish a landslide displacement prediction model that combines the Markov attention mechanism and the NLinear model. The new model greatly improves the model's ability to extract information near the prediction point, greatly improves the accuracy of landslide displacement prediction in obvious Markov processes such as landslide displacement, and significantly improves the accuracy of displacement prediction in the rapid change stage of landslides.

[0130] Traditional landslide displacement prediction models usually use "historical landslide displacement time series" as model learning data. However, the landslide displacement historical time series does not include landslide displacement mutations, that is, landslide events. Therefore, the landslide displacement prediction model trained in historical data has low prediction accuracy when encountering landslide displacement mutations, because the deep learning model can only fit the features it has seen in the training stage. The Markov attention mechanism provided in the present invention can extract the information of the input sequence with emphasis, because the landslide displacement process is a Markov process, and there is a strong correlation between adjacent data. The features extracted by the Markov attention mechanism are used for landslide displacement prediction, which can greatly improve the prediction accuracy in the landslide displacement mutation stage.

[0131] When applying the landslide displacement prediction method using the Markov attention mechanism provided by the present invention, it is not necessary to Figure 1 The steps are executed in the order shown. The specific execution order of the steps can be determined according to needs, and the present invention does not limit this.

[0132] The above is a landslide displacement prediction method using the Markov attention mechanism provided by one or more embodiments of the present invention. Based on the same idea, the present invention also provides a corresponding landslide displacement prediction device using the Markov attention mechanism, such as Fig.12 shown.

[0133] Fig.12 A schematic diagram of a landslide displacement prediction device using a Markov attention mechanism provided by the present invention includes:

[0134] The acquisition module 1201 is used to acquire the landslide displacement time series in the study area and input the landslide displacement time series into the landslide displacement prediction model; the landslide displacement prediction model includes a Markov matrix information extraction model and a nonlinear model.

[0135] The construction module 1202 is used to construct a landslide displacement Markov matrix based on the landslide displacement time series in the landslide displacement prediction model; the later the time element in the landslide displacement time series, the more times it appears in the landslide displacement Markov matrix.

[0136] The extraction module 1203 is used to extract features from the landslide displacement Markov matrix through a Markov matrix information extraction model to obtain landslide displacement change features; the Markov matrix information extraction model is constructed by multiple serially connected Markov attention mechanisms.

[0137] The first determination module 1204 is used to add the landslide displacement time series and the landslide displacement change feature to obtain an output sequence after the landslide displacement time series is processed by the Markov attention mechanism.

[0138] The second determination module 1205 is used to input the output sequence into the nonlinear model to obtain the predicted value of the landslide displacement in the study area.

[0139] The specific definition of a landslide displacement prediction device using a Markov attention mechanism can be found in the above definition of a landslide displacement prediction method using a Markov attention mechanism, which will not be repeated here. Each module in the above-mentioned landslide displacement prediction device using a Markov attention mechanism can be implemented in whole or in part by software, hardware, and a combination thereof. The above-mentioned modules can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0140] The present invention also provides a computer-readable storage medium, which stores a computer program, which can be used to execute the above Figure 1 A landslide displacement prediction method using Markov attention mechanism is provided.

[0141] The present invention also provides Fig.13 The structural diagram of the computer device shown in FIG. Fig.13As shown in the figure, at the hardware level, the computer device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory, and may also include other hardware required for the business. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to achieve the above Figure 1 A landslide displacement prediction method using Markov attention mechanism is provided.

[0142] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0143] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of the present invention.

Claims

1. A landslide displacement prediction method using Markov attention mechanism, characterized in that: include: Obtaining a landslide displacement time series in a study area, and inputting the landslide displacement time series into a landslide displacement prediction model; the landslide displacement prediction model includes a Markov matrix information extraction model and a nonlinear model; In the landslide displacement prediction model, a landslide displacement Markov matrix is ​​constructed based on the landslide displacement time series; the later the time element in the landslide displacement time series, the more times it appears in the landslide displacement Markov matrix; Extracting features from the landslide displacement Markov matrix using the Markov matrix information extraction model to obtain landslide displacement change features; The Markov matrix information extraction model is constructed by multiple serially connected Markov attention mechanisms; Adding the landslide displacement time series and the landslide displacement change feature to obtain an output sequence after the landslide displacement time series is processed by the Markov attention mechanism; The output sequence is input into the nonlinear model to obtain the predicted value of landslide displacement in the study area.

2. The method according to claim 1, characterized in that The training process of the landslide displacement prediction model includes: Acquire a historical landslide displacement time series, and input the historical landslide displacement time series into an initial landslide displacement prediction model; In the initial landslide displacement prediction model, a historical landslide displacement Markov matrix is ​​constructed based on the historical landslide displacement time series; The historical landslide displacement Markov matrix is ​​extracted and processed by connecting a plurality of Markov attention mechanisms in series to obtain a historical extraction result; Adding the historical landslide displacement time series and the historical extraction result to obtain a historical output sequence after the historical landslide displacement time series is processed by the Markov attention mechanism; Inputting the historical output sequence into a nonlinear model to obtain a predicted value of the initial landslide displacement prediction model for the historical landslide displacement time series; Obtaining observation values ​​of the historical output sequence; The mean square error between the predicted value and the observed value is calculated, and the parameters of the initial landslide displacement prediction model are adjusted until the error is minimized to obtain the landslide displacement prediction model.

3. The method according to claim 1, characterized in that The landslide displacement time series is: X={x1,x2,x3,…,x n }; Where X is the landslide displacement time series, x i It represents the cumulative landslide displacement at the i-th moment, where i ranges from 1 to n; The landslide displacement Markov matrix is: Among them, MarkovMatrix X is the landslide displacement Markov matrix, x i It represents the cumulative landslide displacement at the i-th moment, and the value of i ranges from 1 to n.

4. The method according to claim 1, characterized in that The Markov attention mechanism includes one-dimensional temporal convolution, one-dimensional pooling, and an activation function. The implementation process of the Markov attention mechanism specifically includes: Inputting the input information of the Markov attention mechanism into the one-dimensional temporal convolution module to obtain the output of the one-dimensional temporal convolution module; Input the output of the one-dimensional temporal convolution module into the one-dimensional pooling module to obtain the pooling result; The pooling result is input into the activation function to obtain the output result of the Markov attention mechanism.

5. The method according to claim 1, characterized in that The nonlinear model is: NLinear:R M →R N ; Among them, M represents the length of the nonlinear model input sequence, and N represents the length of the nonlinear model prediction sequence.

6. A landslide displacement prediction device using Markov attention mechanism, characterized in that: include: An acquisition module is used to acquire a landslide displacement time series in a study area and input the landslide displacement time series into a landslide displacement prediction model; the landslide displacement prediction model includes a Markov matrix information extraction model and a nonlinear model; A construction module is used to construct a landslide displacement Markov matrix based on the landslide displacement time series in the landslide displacement prediction model; the later the time of the elements in the landslide displacement time series, the more times they appear in the landslide displacement Markov matrix; An extraction module is used to extract features from the landslide displacement Markov matrix through a Markov matrix information extraction model to obtain landslide displacement change features; The Markov matrix information extraction model is constructed by multiple serially connected Markov attention mechanisms; A first determination module is used to add the landslide displacement time series and the landslide displacement change feature to obtain an output sequence after the landslide displacement time series is processed by a Markov attention mechanism; The second determination module is used to input the output sequence into a nonlinear model to obtain a predicted value of landslide displacement in a study area.

7. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

8. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method according to any one of claims 1 to 5 is implemented.

Images