Landslide monitoring data prediction method fusing time-delay reconstruction and attention mechanism

Abstract

The invention relates to the technical field of landslide monitoring, in particular to a landslide monitoring data prediction method fusing time-lag reconstruction and an attention mechanism, which takes landslide displacement historical monitoring data and external induction factors as modeling objects, and aims at nonlinear and time-lag characteristics of a landslide deformation process. Constructing a time-delay driving feature set through time-delay cross correlation analysis, and identifying a guiding-following causal relationship between landslide displacement and an induction factor; and a multi-source time-lag feature expression fused with an attention mechanism is constructed, and dynamic prediction modeling is performed on the landslide displacement through a gating recurrent neural network, so that accurate prediction of the landslide displacement trend and periodic change is realized. Experimental data prove that the method is superior to an existing traditional method in the aspects of landslide deformation prediction precision and robustness, the prediction precision, the feature utilization rate and the model interpretability are effectively improved, and the method has good engineering practicability and popularization prospects.

Description

Technical Field

[0001] The present invention relates to the technical field of landslide monitoring, and in particular to a landslide monitoring data prediction method integrating time-delay reconstruction and attention mechanism. Background Art

[0002] Landslide displacement prediction is a critical step in landslide disaster early warning. Most existing methods construct time series models based on cumulative landslide displacement sequences. While this improves prediction accuracy to a certain extent, it generally ignores the dynamic influence of inducing factors. In reality, landslide deformation is controlled by internal factors such as geological structure and soil properties, as well as external factors such as rainfall, vibration, and temperature, exhibiting significant nonlinearity and time lag characteristics.

[0003] Some current studies have used empirical mode decomposition (EMD) to analyze landslide displacement and precipitating factor sequences, using these components as covariates to input deep models for prediction. However, these methods often rely on fixed decomposition parameters, lack adaptive mechanisms, and fail to clearly quantify the impact of precipitating factors on landslide deformation. They rely solely on model learning features, resulting in insufficient interpretability and limited generalization capabilities.

[0004] In addition, the sole use of displacement sequences is prone to overfitting, while the simple introduction of multiple variables may reduce the prediction accuracy. Summary of the Invention

[0005] The purpose of the present invention is to provide a landslide monitoring data prediction method that integrates time-delay reconstruction and attention mechanism, aiming to solve the problems of existing landslide prediction methods that fail to fully quantify the time-delay impact of inducing factors, have weak model interpretability, and have limited prediction accuracy, thereby improving the reliability and intelligence level of landslide displacement prediction.

[0006] To achieve the above objectives, the present invention provides a landslide monitoring data prediction method integrating time-delay reconstruction and attention mechanism, comprising the following steps:

[0007] Step 1: The Grey Wolf Optimization Algorithm is introduced to optimize the multi-parameter improved ensemble empirical mode decomposition, and then the landslide displacement series and inducing factors are decomposed to extract the trend term and periodic disturbance term, thus completing the feature enhancement of multi-source data.

[0008] Step 2: Analyze the landslide displacement history series using the partial autocorrelation function and Spearman correlation coefficient method to identify the significant correlation of the target variable itself at different time lags, gradually construct the lag feature set of the target variable, and select the key lag step length;

[0009] Step 3: For the decomposed inducing factor components, determine the lag window range through time-lagged cross-correlation analysis, use the information guidance index to evaluate the guidance ability of each lag step on the target variable, and construct the covariate lag feature set;

[0010] Step 4: Calculate the target variable lag characteristic Z obtained from steps 2 and 3 Y and the covariate lagged characteristic Z X , and combine the two to construct a complete feature set Z i (t), as the set of input variables for landslide displacement prediction;

[0011] Step 5: By designing an attention mechanism module, the lagged feature set Z i Each lag feature in (t) is dynamically weighted to generate a weighted feature representation To strengthen the focus on key lag factors;

[0012] Step 6: The feature vector after attention weighting As input, combined with the historical hidden state, the gated recurrent unit network is used to dynamically predict the current landslide displacement and obtain the prediction result y i (t), thus realizing the prediction of landslide monitoring data.

[0013] Optionally, in step 1, for the original monitoring signal of landslide displacement and the signal of inducing factors, the wolf pack position is first initialized, the standard deviation of random noise and the number of trials are used as a parameter combination, fitness evaluation is performed by minimizing the signal reconstruction error, and the wolf pack position is dynamically updated. Finally, the improved ensemble empirical mode decomposition result under the optimal parameters is obtained, the displacement signal is decomposed into trend terms and periodic terms, and the triggering factor signal is decomposed into low-frequency terms and high-frequency terms, providing input signals for noise suppression and feature enhancement for subsequent modeling;

[0014] The execution process of step 1 includes the following steps:

[0015] Step 1.1: Initialize the wolf pack positions of the gray wolf optimization algorithm, use the noise standard deviation and the number of trials to form a parameter combination, and define the maximum number of iterations and population size;

[0016] Step 1.2: For each parameter combination, perform the improved ensemble empirical mode decomposition algorithm, calculate the reconstruction error and evaluate it as the fitness function;

[0017] Step 1.3: Classify the gray wolf individuals into α, β, or δ according to their fitness values, update their positions, and proceed to the next iteration;

[0018] Step 1.4: When the convergence condition is met or the maximum number of iterations is reached, the optimal decomposition parameters are output and the decomposition processing of the trend term and periodic term, or the high-frequency term and low-frequency term of the landslide displacement and inducing factor signal is completed.

[0019] Optionally, the execution process of step 2 includes the following steps:

[0020] Step 2.1: Calculate the target variable y using the partial autocorrelation function i The optimal lag order p of (t);

[0021] Step 2.2: In τ y ≤p range, calculate the current value y i (t) and hysteresis value y i (t-τ y ) between the Spearman correlation coefficient ρ;

[0022] Step 2.3: If the current correlation coefficient ρ>ρ max , then update the optimal correlation coefficient ρ=ρ max , and the corresponding hysteresis characteristic y i (t-τ y ) Add the target variable lag feature set;

[0023] Step 2.4: Iteratively increase τ y , until all lag orders are traversed, and finally the target variable lag feature set Z is output Y .

[0024] Optionally, the process of constructing the covariate lagged feature set in step 3 includes the following steps:

[0025] Step 3.1: For the covariate x j (t) and the target variable y i (t) Perform time-lagged cross-correlation analysis to determine the optimal lag order upper limit q ij ;

[0026] Step 3.2: In τ x ≤q ij Calculate the lagged covariate x ij (t-τ x ) and the target variable y i The information between (t) guides the metric γ max ;

[0027] Step 3.4: If γ≤γ max , then update γ=γ max , and the lag term y i (t-τ x ) joins set Z' X ;

[0028] Step 3.5: Iteratively increase τ x , until all lag orders of the covariate are traversed;

[0029] Step 3.6: Map the current covariate to the lag set Merge into the total feature set Z X , that is, ZX =Z X ∪Z' X , and proceed to the next covariate;

[0030] Step 3.7: After all covariates are traversed, the covariate lag feature set Z is output X .

[0031] Optional, Step 5.1: Construct query vector Q i , the historical value of the target variable is passed through the learnable parameter matrix Mapping generation;

[0032] Step 5.2: Construct a key vector for each lagged feature in the feature set With value vector

[0033] Step 5.3: Calculate the attention score using the dot product attention mechanism Then normalize the attention weight by the Softmax function

[0034] Step 5.4: Weighted sum of the value vector according to the attention weight and output weighted feature representation

[0035] Optionally, the gated recurrent unit network prediction process in step 6 includes the following steps:

[0036] Step 6.1: Weighted feature vector As input at time step t, the joint historical hidden state Input to the gated recurrent unit network;

[0037] Step 6.2: Complete the state transfer through the reset gate and update gate in the gated recurrent unit network structure and calculate the current hidden state

[0038] Step 6.3: Use the output layer to linearly map Mapped to the predicted landslide displacement at the current moment

[0039] This invention provides a landslide monitoring data prediction method that integrates time-lag reconstruction and an attention mechanism. Using historical landslide displacement monitoring data and external inducing factors as modeling objects, and addressing the nonlinear and time-lag characteristics of landslide deformation processes, this method constructs a time-lag driven feature set through time-lag cross-correlation analysis and identifies the leading-following causal relationship between landslide displacement and inducing factors. Furthermore, a multi-source time-lag feature representation is constructed that incorporates the attention mechanism. Dynamic predictive modeling of landslide displacement is performed using a gated recurrent neural network, enabling accurate prediction of landslide displacement trends and cyclical variations. Experimental data demonstrates that this method outperforms existing traditional methods in both accuracy and robustness in landslide deformation prediction, effectively improving prediction accuracy, feature utilization, and model interpretability. The method has excellent engineering practicality and potential for widespread application. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0041] Figure 1 This is a schematic flow chart of the steps of a landslide monitoring data prediction method that integrates time-delay reconstruction and attention mechanism of the present invention.

[0042] Figure 2 1 is a schematic flow chart of the GWO-ICCEMDAN steps of the present invention.

[0043] Figure 3 It is a schematic diagram of the landslide displacement decomposition according to a specific embodiment of the present invention.

[0044] Figure 4 It is a schematic diagram of the decomposition of the inducing factor (rainfall) of a specific embodiment of the present invention.

[0045] Figure 5 It is a schematic diagram of the PACF analysis diagram of the trend term and the period term of the landslide monitoring points ZG01 to ZG03 according to a specific embodiment of the present invention.

[0046] Figure 6 1 is a schematic diagram of a TLCC correlation analysis diagram of the ZG01 trend item and the period item with different high-frequency / low-frequency rainfall factors according to a specific embodiment of the present invention.

[0047] Figure 7 It is a comparison chart of the prediction effects of ZG01_Trend under various prediction models according to a specific embodiment of the present invention.

[0048] Figure 8It is a comparison chart of the prediction effects of ZG01_Periodic under various prediction models according to a specific embodiment of the present invention.

[0049] Figure 9 It is a comparison diagram of the prediction effects of the cumulative displacement of ZG01 under various prediction models according to a specific embodiment of the present invention.

[0050] Figure 10 It is a comparison chart of evaluation indicators of ZG01_Trend in trend item prediction according to a specific embodiment of the present invention.

[0051] Figure 11 It is a comparison chart of evaluation indicators in the ZG01_Periodic periodic item prediction of a specific embodiment of the present invention.

[0052] Figure 12 It is a comparison chart of evaluation indicators in the cumulative displacement prediction of ZG01 in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0053] The following describes embodiments of the present invention in detail, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and are not to be construed as limiting the present invention.

[0054] The following is an explanation of the Chinese and English abbreviations of the relevant terms in the present invention. Some of the text below uses English abbreviations to refer to the relevant terms:

[0055] Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN);

[0056] Grey Wolf Optimizer (GWO);

[0057] Gated Recurrent Unit (GRU);

[0058] Attention mechanism;

[0059] Time-Lagged Cross-Correlation (TLCC);

[0060] Mutual Information (MI);

[0061] Partial Autocorrelation Function (PACF);

[0062] Autocorrelation Function (ACF);

[0063] Information-guided Metric (IGM);

[0064] Dynamic Lag Feature Selection (DLFS);

[0065] Exponential Smoothing State Space Model (ETS);

[0066] Vector AutoRegression (VAR) model;

[0067] Seasonal AutoRegressiveIntegrated MovingAverage with exogenous regressors (SARIMAX);

[0068] Autoregressive Integrated Moving Average (ARIMA) model;

[0069] Long Short-Term Memory (LSTM)

[0070] Bidirectional Long Short-Term Memory (BILSTM)

[0071] A hybrid model of convolutional neural network and long short-term memory network (CNN-LSTM);

[0072] Temporal Convolutional Network (TCN);

[0073] Multi-Layer Perceptron (MLP).

[0074] The present invention provides a landslide monitoring data prediction method integrating time-delay reconstruction and attention mechanism, comprising the following steps:

[0075] Step 1: The Grey Wolf Optimization Algorithm is introduced to optimize the multi-parameter improved ensemble empirical mode decomposition, and then the landslide displacement series and inducing factors are decomposed to extract the trend term and periodic disturbance term, thus completing the feature enhancement of multi-source data.

[0076] Step 2: Analyze the landslide displacement history series using the partial autocorrelation function and Spearman correlation coefficient method to identify the significant correlation of the target variable itself at different time lags, gradually construct the lag feature set of the target variable, and select the key lag step length;

[0077] Step 3: For the decomposed inducing factor components, determine the lag window range through time-lagged cross-correlation analysis, use the information guidance index to evaluate the guidance ability of each lag step on the target variable, and construct the covariate lag feature set;

[0078] Step 4: Calculate the target variable lag characteristic Z obtained from steps 2 and 3 Y and the covariate lagged characteristic Z X , and combine the two to construct a complete feature set Z i (t), as the set of input variables for landslide displacement prediction;

[0079] Step 5: By designing an attention mechanism module, the lagged feature set Z i Each lag feature in (t) is dynamically weighted to generate a weighted feature representation To strengthen the focus on key lag factors;

[0080] Step 6: The feature vector after attention weighting As input, combined with the historical hidden state, the gated recurrent unit network is used to dynamically predict the current landslide displacement and obtain the prediction result y i (t), thus realizing the prediction of landslide monitoring data.

[0081] Specific implementation steps and processes are as follows: Figure 1 As shown, the following is further explained in conjunction with specific embodiments and execution steps:

[0082] The execution process of step 1 can be divided into the following steps:

[0083] (1.1) Signal Initialization: GPS landslide monitoring data was selected as the target signal, and data on its main inducing factors were introduced to construct the original time series. Key parameters in the ICEEMDAN algorithm, such as the noise standard deviation range (noise_std), the number of noise trials (trials), the maximum number of iterations (T_max), and the number of gray wolves (wolf pack size), were set to lay the foundation for subsequent optimization.

[0084] Specifically, the inducing factors include daily rainfall, cumulative rainfall in the previous 15 days, cumulative rainfall in the previous 30 days, rainfall in the previous 45 days, rainfall in the previous 60 days, reservoir water level, reservoir water level change, etc. There is a certain lag in the response of inducing factors to landslide displacement.

[0085] (1.2) Within the set range, the position of each "grey wolf" is initialized by a random strategy. Each position corresponds to a parameter combination, namely (noise_std, trials), which constitutes the initial parameter candidate set.

[0086] (1.3) Reconstruction error calculation: Perform the ICEEMDAN decomposition operation on each parameter combination to decompose the original signal into a set of intrinsic mode functions (IMFs) and residual terms, and reconstruct the signal through each component And compared with the original signal x(t), the reconstruction error is calculated as the fitness evaluation index, which is in the form of:

[0087]

[0088] (1.4) Gray wolf position evaluation and update: The fitness of the current parameter combinations is sorted according to the reconstruction error, and the three optimal sets of parameters are selected as α, β, and δ. The remaining individuals adjust their positions according to the update strategy of the gray wolf optimization algorithm to converge to the optimal solution.

[0089] (1.5) Convergence judgment: Determine whether the preset maximum number of iterations is reached or the error change meets the convergence condition. If not, continue iterative optimization; otherwise, terminate the search.

[0090] (1.6) Optimal parameter selection and final decomposition: Extract the global optimal parameters during the iterative process, perform the final ICEEMDAN decomposition, extract the trend term and periodic disturbance term of the original landslide signal and the inducing factor signal, and complete the feature enhancement processing of the multi-source signal, providing basic input for the subsequent time-lag feature construction and predictive modeling.

[0091] Furthermore, during the execution of step 1, Figure 2 Schematic diagram of the steps of the GWO-ICCEMDAN algorithm, which introduces the multi-parameter improved ensemble empirical mode decomposition algorithm of the Grey Wolf optimization algorithm. Figure 3 and Figure 4They are the decomposition result diagrams, among which ZG01, ZG02 and ZG03 are three GNSS monitoring points, RF_S15, RF_S30, RF_S45 and RF_S60 are the cumulative rainfall in the previous 15 days, the cumulative rainfall in the previous 30 days, the cumulative rainfall in the previous 45 days and the cumulative rainfall in the previous 60 days respectively.

[0092] The execution process of step 2 includes the following steps:

[0093] (2.1) Target variable selection: The landslide displacement trend item extracted in step S1 is used as the target variable for prediction, denoted as y(t). It represents the evolution characteristics of the landslide under the medium- and long-term trend and has strong physical significance and stability. The upper bound of the correlation is set as ρ max ∈[0.8,1.0].

[0094] (2.2) Constructing the initial lag feature set: Perform a lag operation on the target variable y(t) to construct a set of lag feature sequences {y(t-1), y(t-2),…, y(tk)}, where k is the set maximum lag order, aiming to explore the historical dependence of the target variable itself.

[0095] (2.3) Partial Autocorrelation Analysis (PACF): Perform a partial autocorrelation function analysis on the above set of lagged features to assess the direct correlation between each lagged variable and the current variable. By observing the lag orders in the PACF plot that fall outside the confidence interval, we can screen out significant lagged features as candidate input variables.

[0096]

[0097] (2.4) Spearman rank correlation calculation: To further measure the nonlinear correlation between the lagged features and the target variable, the Spearman rank correlation coefficient method is introduced to evaluate the rank correlation between the lagged features and y(t) and obtain the correlation coefficient value for each lag order.

[0098] ρ=ρ s (y i (t)y i (t-τ y ))

[0099] (2.5) Lagged feature screening: Combined with the results of PACF and Spearman analysis, double-constrained screening of lagged features is performed. Only lagged variables that have both significant partial autocorrelation and a high-rank correlation with the target variable are retained to form the final target variable lagged feature set for subsequent joint modeling with induced factors.

[0100] At this point, the construction of the lag structure based on its own information has been completed in step 2, providing a comparison benchmark and structural reference for the subsequent construction of the lag characteristics of exogenous variables.

[0101] Figure 5 The figure is a schematic diagram of the lag order of the target variable based on PACF analysis. It can be seen that ZG01_Trend has a stronger impact after lag of 1 to 3 days, and ZG01_Periodic has obvious cyclical autocorrelation, which is significant at multiple lag orders (such as 1, 7, 13, etc.).

[0102] Step 3 can be broken down into the following steps:

[0103] (3.1) Definition and preparation of inducing factor variables

[0104] Assume that the set of inducing factor variables is:

[0105] X={x1(t),x2(t),…,x j (t)}|

[0106] Among them, x j (t) represents the observed value of the j-th inducing factor at time t (such as daily rainfall in the landslide area, reservoir water level changes, etc.), the target variable is the trend term y(t) obtained in the previous step, and the maximum lag window is set to L.

[0107] (3.2) Calculate the time delay cross correlation (TLCC)

[0108] The time-lag cross-correlation (TLCC) method is used to traverse all lag steps of each inducing factor in the lag interval τ∈[0,L] and calculate the correlation coefficient between it and the target variable y(t):

[0109]

[0110] Among them, corr(·) can use Pearson or Spearman correlation method.

[0111] (3.3) Determine the optimal time delay step size

[0112] For each predisposing factor x j (t), find the lag point with the largest absolute value of correlation with the target variable in the corresponding lag step:

[0113]

[0114] The optimal delay step size will serve as the critical time response to this eliciting factor.

[0115] (3.4) Calculation of Information Guided Metric (IGM)

[0116] In order to measure the guiding strength of the inducing factor at this lag point, the information guidance metric (IGM) is introduced:

[0117]

[0118] This indicator combines the strength of correlation and the strength of information fluctuation, and is used to characterize the degree to which the covariate guides the target variable.

[0119] (3.5) Screening and constructing the hysteresis feature set of inducing factors

[0120] Set the IGM threshold (screen the top N factors based on the cumulative contribution rate), retain the lag features of the inducing factors whose IGM values ​​are greater than the threshold, and construct the covariate lag feature set:

[0121]

[0122] The final output covariate set Z X It will serve as a key feature input with a time-lag driving effect and participate in the subsequent joint modeling with the lagged characteristics of the target variable.

[0123] See also Figure 6 , Figure 6 This is a schematic diagram of the lag period of the covariates based on the TLCC analysis. RF_S15_L, RF_S30_L, RF_S45_L, and RF_S60_L represent the low-frequency component of the cumulative rainfall over the preceding 15 days, 30 days, 45 days, and 60 days, respectively. RF_S15_H, RF_S30_H, RF_S45_H, and RF_S60_H represent the high-frequency component of the cumulative rainfall over the preceding 15 days, 30 days, 45 days, and 60 days, respectively. Step 3 completes the construction of the lag structure based on the covariates.

[0124] The execution process of step 4 is divided into the following steps:

[0125] (4.1) Target variable lagged feature set input

[0126] The target variable lag feature set output from step S2 is received and recorded as:

[0127] Z Y ={y(t-τ1),y(t-τ2),…,y(t-τ p )}

[0128] Among them, y(t) is the landslide trend term or period term, τi It represents the significant lag order after double screening by PACF and Spearman.

[0129] (4.2) Covariate lagged feature set input

[0130] The lag feature set of the inducing factor covariate outputted in step S3 is received and recorded as:

[0131]

[0132] where x j (t) is the jth inducing factor (such as rainfall, reservoir water level), is its optimal lag step length, which satisfies the significant IGM value.

[0133] (4.3) Feature set integration and alignment

[0134] Z Y With Z X Perform horizontal splicing to form a complete multi-source joint feature set:

[0135] Z=Z Y ∪Z X

[0136] To ensure the alignment of feature time series, all time series need to be clipped according to the minimum common valid time window, that is:

[0137]

[0138] (4.4) Standardization

[0139] Standardization (z-score) is performed on all variables in the joint feature set Z to improve the stability of model training.

[0140] (4.5) Construct the final feature matrix

[0141] Get the final multi-source joint feature input matrix Z∈R T×d , where T represents the effective time step, d represents the total dimension of the joint feature (target lag + covariate lag), and each row Z t As the input feature vector of the subsequent model at time t.

[0142] Step 4 completes the construction of the multi-source joint lagged feature matrix.

[0143] The execution process of step 5 includes the following steps:

[0144] (5.1) Feature matrix input

[0145] The input is the joint feature matrix Z∈R T×d, where T is the time step, and d is the multi-source lag feature dimension, including the lag term of the target variable and the lag term of the inducing factor.

[0146] (5.2) Attention weight calculation

[0147] Adopting a learnable attention mechanism, the attention weight α of each feature is automatically calculated according to the influence of the feature on the predicted target. i , and normalized by the Softmax function to satisfy:

[0148]

[0149] (5.3) Weighted feature representation generation

[0150] Apply the attention weights to the original features to obtain the final weighted feature vector:

[0151]

[0152] in is the weighted feature representation at time t, which serves as the input of the subsequent prediction model.

[0153] This process realizes the automatic identification and dynamic weighting of key features, providing a more expressive input basis for modeling and prediction of landslide displacement.

[0154] At this point, step 5 completes the weighted features after integrating the attention mechanism.

[0155] The execution process of step 6 includes the following steps:

[0156] This step uses the Gated Recurrent Unit (GRU) to build a landslide displacement prediction model, and performs time series modeling and prediction on the weighted features obtained after integrating the attention mechanism. Specifically:

[0157] (6.1) Model input

[0158] Receive the weighted feature sequence output from step 5 As input, it is used to capture the evolution trend of time series features in the time dimension.

[0159] (6.2) GRU network structure design

[0160] A recurrent neural network consisting of three layers of GRU units is constructed to mine long-term dependencies in time series. Compared to traditional RNN structures, GRU has stronger memory capacity and lower computational complexity, making it suitable for modeling nonlinear and non-stationary time series such as landslides.

[0161] (6.3) Prediction output

[0162] After calculation through the GRU network, the prediction results are output It represents the predicted value of the landslide displacement trend term at the next time step.

[0163] (6.4) Training and Optimization

[0164] Historical data is used for model training, and mean square error (MSE) is used as the loss function. The model performance can be optimized by combining early stopping strategies and anti-overfitting regularization terms.

[0165] At this point, step 6 completes the construction of a landslide displacement prediction model using a gated recurrent unit (GRU), and performs time series modeling and prediction on the weighted features obtained after integrating the attention mechanism.

[0166] In step 6, Figure 7-12 This is a schematic diagram of landslide displacement prediction and performance evaluation, where Measured represents the actual monitoring value of the monitoring point, ETS, SARIMAX, BILSTM, GRU, CNN-LSTM, TCN, VAR, LSTM, MLP, and ARIMA are comparison models, and MyModel represents the method proposed in the present invention.

[0167] Compared with the existing technical solutions, the present invention has the following beneficial effects:

[0168] The present invention provides a landslide monitoring data prediction method that integrates time-lag reconstruction and attention mechanism, which has obvious innovation and practical value. In the feature extraction stage, in order to solve the problem that traditional signal decomposition methods rely on manually set parameters and the decomposition results are unstable, the present invention introduces the gray wolf optimization algorithm to perform global optimization search on key parameters such as noise standard deviation and number of trials in the ICEEMDAN decomposition process, realizes parameter adaptive selection and optimal trend-disturbance separation, and improves the decomposition quality and stability of landslide displacement and inducing factors (such as rainfall, reservoir water level); in terms of feature construction, the partial autocorrelation function (PACF) and the time-lag correlation analysis method based on TLCC-IGM are used comprehensively, starting from the target variable itself and external inducing factors, respectively, to accurately By extracting significant lag features and constructing a more physically interpretable multi-source time-lag feature system, the model's ability to identify key driving factors is significantly enhanced. Furthermore, by introducing an attention mechanism to weighted model the multi-source lag features, the contribution of different features to the evolution of landslide displacement at different time steps can be dynamically evaluated, solving the problem of fixed lag feature processing and difficulty in quantifying importance in traditional methods. In the predictive modeling link, a GRU neural network is used to perform time series modeling on weighted features, fully considering the nonlinearity, non-stationarity, and long-term dependencies in the landslide evolution process, thereby improving the expressive power and generalization performance of the predictive model. Overall, the method of the present invention not only improves the accuracy and stability of landslide displacement prediction, but also has good adaptability and scalability. It is suitable for a variety of complex monitoring scenarios such as different landslide types, different combinations of inducing factors, and multi-site collaborative modeling, and has broad engineering application prospects and promotion value.

[0169] The above disclosure is merely one or more preferred embodiments of the present invention, and certainly cannot be used to limit the scope of the present invention. A person skilled in the art can understand that all or part of the processes of the above embodiments and equivalent changes made in accordance with the claims of the present invention still fall within the scope of the invention.

Claims

1. A landslide monitoring data prediction method integrating time-delay reconstruction and attention mechanism, characterized in that: The following steps are involved: Step 1: The Grey Wolf Optimization Algorithm is introduced to optimize the multi-parameter improved ensemble empirical mode decomposition, and then the landslide displacement series and inducing factors are decomposed to extract the trend term and periodic disturbance term, thus completing the feature enhancement of multi-source data. Step 2: Analyze the landslide displacement history series using the partial autocorrelation function and Spearman correlation coefficient method to identify the significant correlation of the target variable itself at different time lags, gradually construct the lag feature set of the target variable, and select the key lag step length; Step 3: For the decomposed inducing factor components, determine the lag window range through time-lagged cross-correlation analysis, use the information guidance index to evaluate the guidance ability of each lag step on the target variable, and construct the covariate lag feature set; Step 4: Calculate the target variable lag characteristic Z obtained from steps 2 and 3 Y and the covariate lagged characteristic Z X , and combine the two to construct a complete feature set Z i (t), as the set of input variables for landslide displacement prediction; Step 5: By designing an attention mechanism module, the lagged feature set Z i Each lag feature in (t) is dynamically weighted to generate a weighted feature representation To strengthen the focus on key lag factors; Step 6: The feature vector after attention weighting As input, combined with the historical hidden state, the gated recurrent unit network is used to dynamically predict the current landslide displacement and obtain the prediction result y i (t), thus realizing the prediction of landslide monitoring data.

2. The landslide monitoring data prediction method integrating time-delay reconstruction and attention mechanism according to claim 1 is characterized in that: In step 1, the wolf pack positions are initialized for the original monitoring signals of landslide displacement and the triggering factor signals. The standard deviation of random noise and the number of trials are used as parameter combinations. Fitness is evaluated by minimizing the signal reconstruction error, and the wolf pack positions are dynamically updated. Finally, the improved ensemble empirical mode decomposition results under the optimal parameters are obtained. The displacement signal is decomposed into trend terms and periodic terms, and the triggering factor signal is decomposed into low-frequency terms and high-frequency terms, providing input signals for noise suppression and feature enhancement for subsequent modeling. The execution process of step 1 includes the following steps: Step 1.1: Initialize the wolf pack positions of the gray wolf optimization algorithm, use the noise standard deviation and the number of trials to form a parameter combination, and define the maximum number of iterations and population size; Step 1.2: For each parameter combination, perform the improved ensemble empirical mode decomposition algorithm, calculate the reconstruction error and evaluate it as the fitness function; Step 1.3: Classify the gray wolf individuals into α, β, or δ according to their fitness values, update their positions, and proceed to the next iteration; Step 1.4: When the convergence condition is met or the maximum number of iterations is reached, the optimal decomposition parameters are output and the decomposition processing of the trend term and periodic term, or the high-frequency term and low-frequency term of the landslide displacement and inducing factor signal is completed.

3. The landslide monitoring data prediction method integrating time-delay reconstruction and attention mechanism according to claim 2 is characterized in that: The execution process of step 2 includes the following steps: Step 2.1: Initialize the lag step index τ y =1, defines the target variable lag feature set Set the upper limit of the maximum lag order p and the upper limit of the optimal correlation coefficient ρ max ; Step 2.2: Calculate the target variable y using the partial autocorrelation function i The optimal lag order p of (t); Step 2.3: In τ y ≤p range, calculate the current value y i (t) and hysteresis value y i (t-τ y ) between the Spearman correlation coefficient ρ; Step 2.4: If the current correlation coefficient ρ>ρ max , then update the optimal correlation coefficient ρ=ρ max , and the corresponding hysteresis characteristic y i (t-τ y ) Add the target variable lag feature set; Step 2.5: Iteratively increase τ y , until all lag orders are traversed, and finally the target variable lag feature set Z is output Y .

4. The landslide monitoring data prediction method integrating time-delay reconstruction and attention mechanism according to claim 3 is characterized in that: The process of constructing the covariate lagged feature set in step 3 includes the following steps: Step 3.1: For the covariate x j (t) and the target variable y i (t) Perform time-lagged cross-correlation analysis to determine the optimal lag order upper limit q ij ; Step 3.2: In τ x ≤q ij Calculate the lagged covariate x ij (t-τ x ) and the target variable y i The information between (t) guides the metric γ max ; Step 3.4: If γ≤γ max , then update γ=γ max , and the lag term y i (t-τ x ) joins set Z' X ; Step 3.5: Iteratively increase τ x , until all lag orders of the covariate are traversed; Step 3.6: Map the current covariate to the lag set Merge into the total feature set Z X , that is, Z X =Z X ∪Z' X , and proceed to the next covariate; Step 3.7: After all covariates are traversed, the covariate lag feature set Z is output X .

5. The landslide monitoring data prediction method integrating time-delay reconstruction and attention mechanism according to claim 4 is characterized in that: The feature weighting process of the attention mechanism module in step 5 includes the following steps: Step 5.1: Construct query vector Q i , the historical value of the target variable is passed through the learnable parameter matrix Mapping generation; Step 5.2: Construct a key vector for each lagged feature in the feature set With value vector Step 5.3: Calculate the attention score using the dot product attention mechanism Then normalize the attention weight by the Softmax function Step 5.4: Weighted sum of the value vector according to the attention weight and output weighted feature representation 6. The landslide monitoring data prediction method integrating time-delay reconstruction and attention mechanism according to claim 5 is characterized in that: The gated recurrent unit network prediction process in step 6 includes the following steps: Step 6.1: Weighted feature vector As input at time step t, the joint historical hidden state Input to the gated recurrent unit network; Step 6.2: Complete the state transfer through the reset gate and update gate in the gated recurrent unit network structure and calculate the current hidden state Step 6.3: Use the output layer to linearly map Mapped to the predicted landslide displacement at the current moment

Images