Landslide displacement prediction method, electronic equipment and storage medium

Abstract

The invention discloses a landslide displacement prediction method, electronic equipment and a storage medium, and the method comprises the steps: carrying out the abnormal value detection and elimination of original landslide displacement monitoring time series data, and obtaining clean displacement data; decomposing the clean displacement data into a plurality of modal components by adopting a self-adaptive modal decomposition algorithm; screening out a target modal component containing a random item, a fluctuation item and a trend item through a clustering algorithm; constructing a dual-drive prediction architecture, extracting local spatial features, and modeling long-term time dependence to obtain a predicted value of each modal component; and carrying out superposition and recombination on the predicted values to finally obtain a landslide total displacement prediction sequence. The method can accurately predict the landslide displacement.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of highway landslide, and particularly relates to a landslide displacement prediction method, an electronic device and a storage medium. BACKGROUND

[0002] Highway slope stability monitoring is the core link to ensure road operation safety. With the development of monitoring technology, landslide displacement prediction has become a key technical support for the slope safety warning system. In recent years, machine learning models such as extreme learning machine (ELM), support vector machine (SVM), light gradient boosting machine (LightGBM) and extreme gradient boosting (XGBoost) have been widely used in displacement prediction research. However, these traditional algorithms have obvious limitations: first, they are weak in handling high-dimensional and large-scale data, and the simple model structure makes it difficult to capture complex data patterns; second, they have low computational efficiency, high parameter sensitivity and poor prediction robustness. For example, ELM has the defects of slow learning rate and easy to fall into local minimum, SVM is highly sensitive to parameter settings and has poor robustness to noise data, and LightGBM and XGBoost have the disadvantages of high memory occupation and weak dynamic modeling ability.

[0003] The recurrent neural network variants represented by long short-term memory network (LSTM) and gated recurrent unit (GRU) can capture long-term dependencies of sequences, but in practical applications, they are found to have problems such as long training time, gradient vanishing or explosion. Especially in the environment of highway slope which is strongly disturbed by human engineering activities (such as blasting excavation and mechanical vibration), displacement monitoring data often have the characteristics of frequent occurrence of abnormal values and significant sequence mutation, and directly using such models will lead to a decrease in prediction stability. In addition, the slope displacement data has nonlinear and non-stationary characteristics, and if the original displacement sequence is directly predicted, the model will have difficulty in effectively extracting key features, resulting in limited prediction accuracy.

[0004] In order to improve the prediction accuracy, researchers often use signal decomposition techniques to preprocess the displacement data. Wavelet transform (WT), empirical mode decomposition (EMD), ensemble empirical mode decomposition (EEMD) and variational mode decomposition (VMD) are widely used, but each method has its own inherent defects: WT needs to pre-determine the basis function and has poor adaptability; EMD has strong adaptability but has mode mixing and end effect; EEMD alleviates mode mixing but is time-consuming and has residual white noise; VMD can avoid mode mixing and has strong noise resistance, but the number of modal components obtained after decomposition may exceed the actual needs, and some components lack physical meaning, increasing the analysis complexity.

[0005] The evolution of highway slope displacement is a complex process of coupling of multiple factors such as geology, hydrology and human activities, but the existing prediction models mostly use single factor input (only displacement data) or simple multi-factor superposition, and cannot effectively fuse the space-time characteristics. Especially in the displacement mutation stage, the traditional model is difficult to synchronously capture the local spatial characteristics and long-term time dependence, resulting in insufficient prediction reliability of the mutation point. In addition, the existing method does not sufficiently excavate the physical meaning of the displacement component, lacks a targeted prediction strategy for the trend item, fluctuation item and random item, and is difficult to meet the dual requirements of prediction accuracy and real-time performance of long-distance monitoring of highways.

[0006] In summary, the existing highway landslide displacement prediction method mainly has three core problems:

[0007] First, the data quality is seriously interfered by noise and outliers, affecting the reliability of model input;

[0008] Second, the physical meaning of the modal component after signal decomposition is not clear, which is difficult to effectively guide engineering interpretation;

[0009] Third, the spatio-temporal feature fusion is insufficient, which is difficult to balance the relationship between local feature extraction and long-term dependence modeling.

[0010] Therefore, it is urgent to develop a landslide displacement prediction method that can comprehensively solve the problems of data preprocessing, modal decomposition optimization and feature fusion to meet the actual needs of highway slope safety monitoring. SUMMARY

[0011] The purpose of the present application is to provide a landslide displacement prediction method, electronic equipment and storage medium.

[0012] The technical scheme adopted by the present application is as follows:

[0013] A landslide displacement prediction method, comprising the following steps:

[0014] Step 1: outlier detection and elimination, performing outlier detection and elimination on the original landslide displacement monitoring time series data to obtain clean displacement data;

[0015] Step 2: adaptive modal decomposition, using an adaptive modal decomposition algorithm to decompose the clean displacement data into a plurality of modal components;

[0016] Step 3: modal component clustering and screening, screening the plurality of modal components by a clustering algorithm, and retaining target modal components having physical meaning, the target modal components including a random item, a fluctuation item and a trend item;

[0017] Step 4: Dual-driving architecture prediction, a dual-driving prediction architecture containing a spatial feature extraction network and a time-dependent modeling network is constructed, the local spatial features of the target modal components are extracted by using the spatial feature extraction network, and the local spatial features are input into the time-dependent modeling network to model long-term time dependence, to obtain the prediction values of each target modal component respectively;

[0018] Step 5: Prediction value superposition and reorganization, the prediction values of each target modal component are superimposed and reorganized to obtain a landslide total displacement prediction sequence.

[0019] Further, it further includes step 6: prediction performance evaluation, using evaluation indexes to evaluate the landslide total displacement prediction sequence. Further, in step 1, the outlier detection and elimination uses an isolation forest algorithm, and the specific process includes:

[0020] Step 1.1: Construct an isolation forest, set the number of trees and the maximum depth of each tree;

[0021] Step 1.2: Calculate the path length, calculate the path length of each original landslide displacement monitoring time series data point in all trees in the isolation forest;

[0022] Step 1.3: Calculate the anomaly score, calculate the anomaly score of each original landslide displacement monitoring time series data point according to the path length;

[0023] Step 1.4: Outlier elimination, set an anomaly score threshold, and determine the original landslide displacement monitoring time series data points with an anomaly score higher than the anomaly score threshold as outliers and eliminate them, and the remaining data is the clean displacement data.

[0024] Further, in step 2, the adaptive modal decomposition algorithm is a variational modal decomposition algorithm, and the specific process includes:

[0025] Step 2.1: Parameter pre-setting, pre-setting the number of modes and the initial center frequency of each modal component;

[0026] Step 2.2: Analytic signal generation, Hilbert transform is performed on the clean displacement data signal to generate an analytic signal;

[0027] Step 2.3: Frequency spectrum translation, translating the frequency spectrum of the analytic signal to the base frequency band;

[0028] Step 2.4: Variational model construction, based on the signal after frequency spectrum translation, a constrained variational model is constructed to minimize the sum of the bandwidths of each modal component;

[0029] Step 2.5: Model transformation and solution, introducing a quadratic penalty factor and a Lagrange multiplier, transforming the constrained variational model into an unconstrained variational model, and iteratively updating the modal components, central frequency and Lagrange multiplier using the multiplier alternating direction method;

[0030] Step 2.6: Convergence criterion, when the iteration meets the preset convergence condition, stop the iteration, and output the final multiple modal components.

[0031] Further, in step 3, the clustering algorithm is the fuzzy C-means algorithm, and the specific process includes:

[0032] Step 3.1: Parameter setting, set the number of clusters, fuzzy factor, maximum iteration number and convergence threshold;

[0033] Step 3.2: Initialization, initialize the membership matrix and cluster center;

[0034] Step 3.3: Distance calculation, calculate the Euclidean distance of each modal component to each cluster center;

[0035] Step 3.4: Objective function construction, construct the objective function based on membership and Euclidean distance;

[0036] Step 3.5: Iterative update, iteratively update the membership matrix and cluster center according to the objective function;

[0037] Step 3.6: Cluster termination, when the number of iterations reaches the maximum iteration number or the change of the objective function is less than the convergence threshold, stop the iteration, output the clustering result, and obtain three types of target modal components: random term, fluctuation term and trend term.

[0038] Further, in step 4, in the double-drive prediction architecture, the spatial feature extraction network is a convolutional neural network, and the time series dependence modeling network is a gated recurrent unit, and the specific process includes:

[0039] Step 4.1: Data preprocessing, normalizing the target modal components to obtain standardized data;

[0040] Step 4.2: Spatial feature extraction, inputting the standardized data into the convolutional neural network, automatically extracting the local spatial features of the target modal components through the convolutional layer, pooling layer and fully connected layer of the convolutional neural network;

[0041] Step 4.3: Time series dependence modeling, inputting the local spatial features into the gated recurrent unit, modeling the long-term time dependence of the target modal components through the reset gate and update gate of the gated recurrent unit;

[0042] Step 4.4: Component prediction output, the output of the gated recurrent unit is the predicted value of the corresponding target modal component.

[0043] Further, in step 5, the specific process of superimposing and reorganizing the predicted values is: linearly superimposing the random item predicted value, the fluctuation item predicted value and the trend item predicted value according to the corresponding time nodes to obtain the total displacement predicted value of each time node, and the total displacement predicted values of all time nodes constitute the total displacement prediction sequence of the landslide.

[0044] The application further discloses an electronic device, including a memory, a processor and a computer program stored on the memory and executable on the processor, and the processor implements the steps of the landslide displacement prediction method when executing the computer program.

[0045] The application further discloses a computer readable storage medium, and the computer readable storage medium stores a computer program, and the computer program implements the steps of the landslide displacement prediction method when executed by a processor.

[0046] In summary, due to the adoption of the above technical solutions, the application has the following beneficial effects:

[0047] The application realizes progressive steps of "abnormal value detection and elimination, adaptive modal decomposition, modal component clustering and screening, double-driven architecture prediction, and prediction value superimposition and reorganization", purifies the original landslide displacement monitoring time series data first, then decomposes the clean displacement data into multiple modal components and screens out random items, fluctuation items and trend items with physical significance, then predicts each component through the double-driven architecture of the fusion spatial feature extraction network and the time series dependent modeling network, and finally superimposes and reorganizes to obtain the total displacement prediction sequence.

[0048] The reliability of the model input data is ensured by eliminating the abnormal values in the original data first; the physical meaning of the modal components is clarified through clustering and screening, avoiding meaningless components from interfering with the prediction; the double-driven architecture realizes the synchronous capture of local spatial features and long-term time dependence, solving the problem of insufficient spatio-temporal feature fusion of traditional models; finally, the prediction contributions of each physical component are integrated through component superimposition and reorganization, significantly improving the prediction accuracy and stability of the total displacement prediction sequence of the highway landslide, effectively dealing with the nonlinear and non-stationary characteristics of the slope displacement data. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 The flowchart of the application. DETAILED DESCRIPTION

[0050] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations.

[0051] Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0052] Example 1

[0053] As shown in the drawings, Figure 1 The present application discloses a landslide displacement prediction method, which can be applied to mountain landslide, highway landslide and other fields. The present application takes highway landslide as an example to illustrate the landslide displacement prediction method.

[0054] The landslide displacement prediction method comprises the following steps:

[0055] Step 1: Abnormal value detection and elimination, abnormal value detection and elimination are performed on the original landslide displacement monitoring time series data to obtain clean displacement data;

[0056] Step 2: Adaptive modal decomposition, the clean displacement data is decomposed into a plurality of modal components by using an adaptive modal decomposition algorithm;

[0057] Step 3: Modal component clustering and screening, the plurality of modal components are screened by a clustering algorithm, and target modal components with physical significance are retained, the target modal components including random terms, fluctuation terms and trend terms;

[0058] Step 4: Dual-drive architecture prediction, a dual-drive prediction architecture including a spatial feature extraction network and a time series dependence modeling network is constructed, local spatial features of the target modal components are extracted by using the spatial feature extraction network, and the local spatial features are input into the time series dependence modeling network to model long-term time dependence, and prediction values of each target modal component are obtained respectively;

[0059] Step 5: Prediction value superposition and reorganization, the prediction values of each target modal component are superimposed and reorganized to obtain a highway landslide total displacement prediction sequence.

[0060] The present application carries out purification treatment on the original landslide displacement monitoring time series data through the progressive steps of "outlier detection and elimination→adaptive modal decomposition→modal component clustering screening→double-drive architecture prediction→prediction value superposition and reorganization", and then decomposes the clean displacement data into multiple modal components and screens out random items, fluctuation items and trend items with physical significance, then predicts each component through the double-drive architecture of the fusion of spatial feature extraction network and time series dependent modeling network, and finally superimposes and reorganizes to obtain the total displacement prediction sequence.

[0061] The reliability of the model input data is ensured by eliminating the outliers in the original data in advance, the physical meaning of the modal components is clarified through clustering screening, the meaningless component interference is avoided, the local spatial features and long-term time dependence are simultaneously captured through the double-drive architecture, the problem of insufficient time-space feature fusion of the traditional model is solved, the prediction accuracy and stability of the total displacement prediction sequence of the highway landslide are significantly improved through the prediction contribution of the integration of each physical component, and the nonlinear and non-stationary characteristics of the slope displacement data are effectively dealt with.

[0062] Further, the landslide displacement prediction method of the present application further comprises step 6: prediction performance evaluation, which uses evaluation indexes to evaluate the total displacement prediction sequence.

[0063] The reliability and accuracy of the total displacement prediction sequence can be quantitatively verified through the prediction performance evaluation, and it is clear whether the prediction result meets the actual engineering demand; at the same time, it provides a targeted basis for subsequent method optimization, avoids the engineering application risk caused by the unquantitative verification of the prediction result, and guarantees the practicality and rigor of the prediction method.

[0064] Further, in step 1, the outlier detection and elimination uses the isolated forest algorithm, and the specific process includes:

[0065] Step 1.1: Construct an isolated forest, set the number of trees and the maximum depth of each tree;

[0066] Step 1.2: Calculate the path length, for each original landslide displacement monitoring time series data point, calculate its path length in all trees in the isolated forest;

[0067] Step 1.3: Calculate the anomaly score, calculate the anomaly score of each original landslide displacement monitoring time series data point according to the path length;

[0068] Step 1.4: Outlier elimination, set an anomaly score threshold, determine the original landslide displacement monitoring time series data points with anomaly scores higher than the anomaly score threshold as outliers and eliminate them, and the remaining data is the clean displacement data.

[0069] Step 1 uses the Isolation Forest algorithm (IF) to realize anomaly detection and elimination. Through the specific process of "building an isolated forest → calculating path length → calculating anomaly score → setting threshold to eliminate abnormal values", the high sensitivity of isolated forest to abnormal data is utilized, the abnormality degree of data points is quantified by path length, and abnormal values are accurately identified and eliminated.

[0070] Compared with traditional anomaly processing methods, the Isolation Forest algorithm does not need to preset data distribution assumptions and can efficiently process abnormal points in landslide displacement monitoring time series data (such as sudden data caused by human engineering activities disturbance); by accurately eliminating abnormal values to obtain clean displacement data, the interference of abnormal values on subsequent modal decomposition and prediction models is avoided, and the reliability of input data is guaranteed from the source, laying a data foundation for subsequent prediction accuracy improvement.

[0071] The calculation of the anomaly score in step 1.3 is as follows:

[0072]

[0073] Among them:

[0074] is the anomaly score of a single original landslide displacement monitoring time series data point;

[0075] is a single original landslide displacement monitoring time series data point;

[0076] is the total number of samples in the original landslide displacement monitoring time series data set;

[0077] is the average value of the path length of the original landslide displacement monitoring time series data point in all trees of the isolated forest;

[0078] is the number of splits of the original landslide displacement monitoring time series data point from the root node to the leaf node, i.e. the path length described in step 1.2;

[0079] is the expected value of the average path length.

[0080] The calculation of the expected value of the average path length is as follows:

[0081] where H is the harmonic series.

[0082] Further, in step 2, the adaptive modal decomposition algorithm is a variational modal decomposition algorithm, and the specific process includes:

[0083] ​Step 2.1: parameter presetting, presetting the number of modes and the initial center frequency of each mode component;

[0084] Step 2.2: analytic signal generation, Hilbert transforming the clean displacement data signal to generate an analytic signal;

[0085] Step 2.3: spectrum translation, translating the spectrum of the analytic signal to the base frequency band;

[0086] Step 2.4: variational model construction, based on the signal after spectrum translation, constructing a constrained variational model to minimize the sum of the bandwidths of the mode components;

[0087] Step 2.5: model conversion and solving, introducing a quadratic penalty factor and a Lagrange multiplier to convert the constrained variational model into an unconstrained variational model, and using the multiplier alternating direction method to iteratively update the mode components, the center frequency and the Lagrange multiplier;

[0088] Step 2.6: convergence determination, when the iteration meets the preset convergence condition, stopping the iteration and outputting the final multiple mode components.

[0089] Step 2 uses the Variational Mode Decomposition (VMD) algorithm as an adaptive mode decomposition method, through the process of "parameter presetting → analytic signal generation → spectrum translation → constrained variational model construction → unconstrained model conversion and solving → convergence determination", with the goal of minimizing the sum of the bandwidths of the mode components, and through iterative updating to realize the adaptive and non-recursive decomposition of the signal.

[0090] The VMD algorithm overcomes the defects of traditional wavelet transform (which requires presetting of the base function) and empirical mode decomposition (EMD, which has mode aliasing), can adaptively decompose the clean displacement data into multiple mode components with clear frequency characteristics, and effectively avoids mode aliasing and end effect; the mode components obtained by decomposition are concentrated in frequency and have strong regularity, providing a high-quality decomposition basis for subsequent clustering to select target components with physical significance, and reducing the difficulty of subsequent feature extraction.

[0091] The preset number of modes K and the quadratic penalty factor a in step 2.1 are determined by particle swarm optimization.

[0092] The calculation of the analytic signal in step 2.2 is as follows:

[0093]

[0094] Wherein:

[0095] : impulse function;

[0096] : time variable;

[0097] : imaginary unit (satisfying

[0098] : pi (value is about 3.1416);

[0099] : time domain signal of the kth modal component (k is the serial number of the modal component, k = 1, 2, …, K, K is the preset modal number described in step 2.1).

[0100] The calculation of the spectrum shift described in step 2.3 is as follows:

[0101] Wherein:

[0102] : impulse function;

[0103] : time variable;

[0104] : imaginary unit (satisfying

[0105] : pi (value is about 3.1416);

[0106] : time domain signal of the kth modal component (k is the serial number of the modal component, k = 1, 2, …, K, K is the preset modal number described in step 2.1).

[0107] : natural constant;

[0108] : center frequency of the kth modal component (is the value of the initial center frequency described in step 2.1 of claim 4 after iterative update); The construction of the constraint variation model described in step 2.4 is as follows:

[0109]

[0110] The constraint condition is:

[0111] Wherein:

[0112]

[0113] ​​: The kth modal component (k is the modal component number, k=1,2,...,K, K is the preset number of modes mentioned in step 2.1, and this modal component is a component that has not been clustered and filtered);

[0114] : The center frequency of the k-th modal component;

[0115] Sum all modal components of k from 1 to K;

[0116] : The partial derivative with respect to the time variable 𝑡;

[0117] Impulse function;

[0118] Time variable;

[0119] Imaginary unit (satisfying 𝑗²=−1);

[0120] Pi (π), with a value of approximately 3.1416;

[0121] : Natural constant, with a value of approximately 2.7183;

[0122] L2 norm (Euclidean norm);

[0123] Time-domain signal representation of cleanroom displacement data.

[0124] The transformation of the unconstrained variational model described in step 2.5 is shown in the following equation:

[0125] in,

[0126] The objective function of an unconstrained variational model;

[0127] The k-th modal component (k is the modal component index, k=1,2,...,K, K is the preset number of modes mentioned in step 2.1, and this modal component is a component that has not undergone clustering screening);

[0128] : The center frequency of the k-th modal component;

[0129] Lagrange multipliers;

[0130] Secondary penalty factor;

[0131] Sum all modal components of k from 1 to K;

[0132] : The partial derivative with respect to the time variable 𝑡;

[0133] : Impact function;

[0134] t: time variable;

[0135] j: Imaginary unit (satisfying j) 2 =-1);

[0136] Pi (π), with a value of approximately 3.1416;

[0137] e: natural constant, with a value of approximately 2.7183;

[0138] L2 norm (Euclidean norm);

[0139] Time-domain signal representation of cleanroom displacement data;

[0140] The Lagrange multiplier at time t;

[0141] Inner product operation.

[0142] The iterative update formulas for the modal components and center frequencies mentioned in step 2.5 are as follows:

[0143]

[0144]

[0145] in:

[0146] : No. The frequency domain signal estimate of the k-th modal component at the next iteration;

[0147] n: Number of iterations;

[0148] Frequency variable;

[0149] The frequency domain signal estimate of the clean displacement data as described in claim 1 (obtained by Fourier transform of its time domain signal);

[0150] Summing over all modal components except the k-th one;

[0151] The frequency domain signal estimate of the k-th modal component during the nth iteration;

[0152] The frequency domain signal estimate of the Lagrange multipliers at the nth iteration;

[0153] Secondary penalty factor;

[0154] The center frequency of the k-th modal component during the nth iteration;

[0155] : No. The center frequency of the k-th modal component during the next iteration;

[0156] Absolute value operation;

[0157] For frequency variables Integral operations from 0 to infinity;

[0158] : Integral variable differentiation.

[0159] The preset convergence condition mentioned in step 2.6 is shown in the following formula:

[0160]

[0161] in:

[0162] Sum all modal components of k from 1 to K; The k-th modal component during the (n+1)th iteration;

[0163] The k-th modal component during the nth iteration;

[0164] L2 norm (Euclidean norm);

[0165] : Preset convergence threshold;

[0166] n: Number of iterations;

[0167] The original data source for the modal components is the clean displacement data described in step 1.

[0168] Furthermore, in step 3, the clustering algorithm is a fuzzy C-means algorithm, and the specific process includes:

[0169] Step 3.1: Parameter setting, setting the number of clusters, fuzzy factor, maximum number of iterations, and convergence threshold;

[0170] Step 3.2: Initialization, initialize the membership matrix and cluster centers;

[0171] Step 3.3: Distance calculation, calculate the Euclidean distance from each modal component to each cluster center;

[0172] Step 3.4: Construct the objective function, and construct an objective function based on membership degree and Euclidean distance;

[0173] Step 3.5: Iterative update, update the membership matrix and cluster centers iteratively according to the objective function;

[0174] Step 3.6: Clustering Termination. When the number of iterations reaches the maximum number of iterations or the change in the objective function is less than the convergence threshold, the iteration stops, the clustering results are output, and three types of objective modal components are obtained: random terms, fluctuation terms, and trend terms.

[0175] Step 3 employs the Fuzzy C-means Algorithm (FCM) clustering algorithm to screen modal components. The process involves “parameter setting → initializing the membership matrix and cluster centers → calculating the Euclidean distance → constructing and iteratively updating the objective function → clustering termination”. Based on the feature similarity of the modal components, clustering is performed to accurately screen out three types of target modal components with clear physical meaning: random terms, fluctuation terms, and trend terms.

[0176] This study addresses the issues of redundant modal components and lack of physical meaning in some components after VMD decomposition. By using FCM clustering, redundant components without engineering significance can be eliminated, allowing subsequent predictions to focus on the core components that reflect the essential laws of landslide displacement. At the same time, the physical attributes of each target component are clarified (random terms reflect accidental disturbances, fluctuation terms reflect periodic effects, and trend terms reflect long-term evolution trends), providing a basis for constructing targeted prediction models, reducing the complexity of prediction models, and improving the logicality of predictions.

[0177] The Euclidean distance mentioned in step 3.3 is calculated as follows:

[0178]

[0179] in:

[0180] : The Euclidean distance from the i-th modal component to the j-th cluster center;

[0181] : The eigenvector of the i-th modal component (i is the modal component index, (where n is the total number of modal components).

[0182] : The feature vector of the j-th cluster center (j is the cluster center index, (corresponding to the random term, fluctuation term, and trend term, respectively).

[0183] Euclidean norm.

[0184] The expression for the objective function mentioned in step 3.4 is:

[0185]

[0186] Where: F: the objective function of the fuzzy C-means algorithm;

[0187] Sum all modal components of i from 1 to n;

[0188] : Sum all cluster centers from j to c (c is the number of clusters described in step 3.1, with a value of 3, corresponding to the random term, fluctuation term, and trend term);

[0189] : The membership degree of the i-th modal component to the j-th cluster center;

[0190] m: fuzziness factor, with a value of 2;

[0191] : The Euclidean distance from the i-th modal component to the j-th cluster center.

[0192] The iterative formula for the membership matrix mentioned in step 3.5 is:

[0193]

[0194] The iterative formula for cluster centers is:

[0195]

[0196] in:

[0197] : The membership degree of the i-th modal component to the j-th cluster center;

[0198] : Sum the cluster centers of all clusters from k to c;

[0199] : The i-th modal component To the j-th cluster center Euclidean distance;

[0200] : The i-th modal component To the kth cluster center Euclidean distance;

[0201] m: fuzzy factor;

[0202] Exponentiation; : The feature vector of the j-th cluster center;

[0203] Sum all modal components of i from 1 to n;

[0204] : The eigenvector of the i-th modal component;

[0205] : The m-th power.

[0206] Furthermore, in step 4, in the dual-drive prediction architecture, the spatial feature extraction network is a convolutional neural network, and the temporal dependency modeling network is a gated recurrent unit. The specific process includes:

[0207] Step 4.1: Data preprocessing, normalizing the target modal components to obtain standardized data;

[0208] Step 4.2: Spatial feature extraction. Standardized data is input into a convolutional neural network. Through the convolutional layers, pooling layers, and fully connected layers of the convolutional neural network, local spatial features of the target modal components are automatically extracted.

[0209] Step 4.3: Temporal dependency modeling. The local spatial features are input into the gated loop unit. The long-term temporal dependency of the target modal components is modeled through the reset gate and update gate of the gated loop unit.

[0210] Step 4.4: Component prediction output. The output of the gated loop unit is the predicted value of the corresponding target modal component.

[0211] Step 4 constructs a dual-drive prediction architecture of "Convolutional Neural Networks (CNN) + Gated Recurrent Unit (GRU)". Through the process of "data normalization preprocessing → CNN extracting local spatial features → GRU modeling long-term temporal dependencies → outputting component prediction values", the convolution and pooling layers of CNN are used to efficiently extract the local spatial features of the target modality components, and the reset gate and update gate of GRU are used to accurately capture the long-term temporal dependencies of the components.

[0212] Overcoming the limitations of traditional single models (such as LSTM and ELM) in simultaneously taking into account spatial features and temporal dependencies, the synergistic effect of CNN and GRU enables the model to capture both the local detailed features of the target components (such as the local fluctuations of displacement abrupt changes) and accurately model the long-term temporal evolution patterns. At the same time, GRU has fewer parameters and higher training efficiency than LSTM. Combined with data normalization preprocessing, it further improves the accuracy of the predicted values ​​of each target modal component and the model training efficiency, providing core support for subsequent superposition and recombination to obtain a high-quality total displacement prediction sequence.

[0213] The convolution operation of the convolutional layer in the convolutional neural network described in step 4.2 is shown in the following formula:

[0214]

[0215] in:

[0216] : Convolutional layer outputs feature maps;

[0217] f: Activation function, selected from ReLU or Sigmoid;

[0218] : Convolution kernel weights;

[0219] : Convolution operator;

[0220] : Input data for convolutional layers;

[0221] Kernel bias.

[0222] The calculations for the reset gate, update gate, hidden layer candidate state, and final output of the gated loop unit in step 4.3 are as follows:

[0223] Remastered Gate:

[0224] Update Gate:

[0225] Hidden layer candidate states:

[0226] Final output:

[0227] in:

[0228] The gate output is reset at time t.

[0229] Update the gate output at time t;

[0230] : Sigmoid activation function;

[0231] , Reset the weight matrix of the gate;

[0232] , Update the gate weight matrix;

[0233] Local spatial features input at time t;

[0234] The hidden state of the gated loop unit at time t-1;

[0235] , : Reset the door, update the door offset;

[0236] Candidate states of the hidden layer at time t

[0237] tanh: Hyperbolic tangent activation function;

[0238] W, U: Weight matrices from the input layer to the hidden layer;

[0239] Hadamard product;

[0240] The final output of the gated loop unit at time t;

[0241] t: Time node number.

[0242] Furthermore, in step 5, the specific process of superimposing and recombining the predicted values ​​is as follows: the predicted values ​​of the random term, the predicted values ​​of the fluctuation term, and the predicted values ​​of the trend term are linearly superimposed according to the corresponding time nodes to obtain the total displacement predicted value of each time node. The total displacement predicted values ​​of all time nodes constitute the landslide total displacement prediction sequence.

[0243] Step 5 adopts the predicted value superposition and recombination process of "linear superposition according to corresponding time nodes → integration into a total sequence". Based on the physical synergy of random, fluctuation and trend components, the predicted values ​​of the three components are matched and superimposed one by one according to time nodes. Then, the superposition results of all time nodes are integrated in chronological order to form the total displacement prediction sequence.

[0244] It avoids the prediction bias caused by traditional methods such as "direct prediction of the original sequence" or "simple splicing of components". Through linear superposition with precise time node matching, it can fully integrate the prediction contributions of the three types of target components and completely restore the comprehensive evolution law of highway landslide displacement (including both long-term trends and periodic fluctuations and accidental disturbances). The final total displacement prediction sequence is logically coherent and data is complete, with a significantly higher accuracy than single component prediction or non-targeted superposition methods.

[0245] Furthermore, in step 6, the evaluation index includes one or more of the following: correlation coefficient, mean absolute percentage error, and root mean square error.

[0246] Step 6 uses one or more of the following as evaluation indicators: correlation coefficient (Pearson Correlation Coefficient, R), mean absolute percentage error (MAPE), and root mean square error (RMSE). It quantifies the deviation between the total displacement prediction sequence and the actual monitoring data from different dimensions such as "degree of linear correlation", "relative error" and "absolute error".

[0247] Compared with a single evaluation index, multi-dimensional indicators can more comprehensively and objectively reflect the accuracy level of the total displacement prediction sequence, avoiding the one-sidedness of single index evaluation. By quantifying the evaluation results, the applicable scenarios and performance boundaries of the prediction method can be clarified, providing targeted basis for the engineering application verification and parameter optimization of the method, and ensuring the reliability of the prediction method in actual highway slope monitoring.

[0248] The correlation coefficient, mean absolute percentage error, and root mean square error are calculated as follows:

[0249] Correlation coefficient:

[0250]

[0251] Mean absolute percentage error:

[0252]

[0253] Root mean square error:

[0254]

[0255] Where: R: correlation coefficient;

[0256] MAPE: Mean Absolute Percentage Error;

[0257] RMSE: Root Mean Square Error;

[0258] N: The total number of time points included in the landslide total displacement prediction sequence;

[0259] : The measured displacement value at the i-th time node;

[0260] : The average value of the measured displacement at all time points;

[0261] : The predicted displacement value at the i-th time node;

[0262] : The average value of the predicted displacement at all time points.

[0263] Example 2

[0264] The present invention also discloses an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the landslide displacement prediction method described in Embodiment 1.

[0265] Example 3

[0266] The present invention also discloses a computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the landslide displacement prediction method described in Embodiment 1.

[0267] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A landslide displacement prediction method, characterized in that, Includes the following steps: Step 1: Outlier detection and removal. Outlier detection and removal are performed on the original landslide displacement monitoring time series data to obtain clean displacement data. Step 2: Adaptive Mode Decomposition. The cleanroom displacement data is decomposed into multiple modal components using an adaptive mode decomposition algorithm. Step 3: Modal component clustering and screening. The multiple modal components are screened using a clustering algorithm to retain target modal components with physical meaning. The target modal components include random terms, fluctuation terms, and trend terms. Step 4: Dual-drive architecture prediction. Construct a dual-drive prediction architecture that includes a spatial feature extraction network and a temporal dependency modeling network. Use the spatial feature extraction network to extract the local spatial features of the target modal components, and input the local spatial features into the temporal dependency modeling network to model long-term temporal dependencies, thereby obtaining the predicted values ​​of each target modal component. Step 5: Superimpose and recombine the predicted values. Superimpose and recombine the predicted values ​​of each target modal component to obtain the total landslide displacement prediction sequence.

2. The method according to claim 1, characterized in that: It also includes step 6: prediction performance evaluation, which uses evaluation indicators to evaluate the total landslide displacement prediction sequence.

3. The method according to claim 1, characterized in that: In step 1, the outlier detection and removal employs the Isolation Forest algorithm, and the specific process includes: Step 1.1: Construct an isolated forest, setting the number of trees and the maximum depth of each tree; Step 1.2: Calculate the path length. For each original landslide displacement monitoring time series data point, calculate its path length among all trees in the isolated forest. Step 1.3: Calculate the anomaly score. Calculate the anomaly score for each original landslide displacement monitoring time series data point based on the path length. Step 1.4: Outlier removal. Set an outlier score threshold. Identify and remove original landslide displacement monitoring time series data points with outlier scores higher than the outlier score threshold. The remaining data is the clean displacement data.

4. The method according to claim 1, characterized in that: In step 2, the adaptive mode decomposition algorithm is a variational mode decomposition algorithm, and the specific process includes: Step 2.1: Parameter preset, preset the number of modes and the initial center frequency of each modal component; Step 2.2: Analytical signal generation. Perform Hilbert transform on the clean displacement data signal to generate an analytical signal; Step 2.3: Spectrum shift, shifting the spectrum of the analyzed signal to the fundamental frequency band; Step 2.4: Variational model construction. Based on the spectrum-shifted signal, a constrained variational model is constructed to minimize the sum of the bandwidths of each modal component. Step 2.5: Model transformation and solution. Introduce the quadratic penalty factor and Lagrange multipliers to transform the constrained variational model into an unconstrained variational model. Use the alternating direction method of multipliers to iteratively update the modal components, center frequencies and Lagrange multipliers. Step 2.6: Convergence determination. When the iteration meets the preset convergence condition, stop the iteration and output the final multiple modal components.

5. The method according to claim 1, characterized in that: In step 3, the clustering algorithm is the fuzzy C-means algorithm, and the specific process includes: Step 3.1: Parameter setting, setting the number of clusters, fuzzy factor, maximum number of iterations, and convergence threshold; Step 3.2: Initialization, initialize the membership matrix and cluster centers; Step 3.3: Distance calculation, calculate the Euclidean distance from each modal component to each cluster center; Step 3.4: Construct the objective function, and construct an objective function based on membership degree and Euclidean distance; Step 3.5: Iterative update, update the membership matrix and cluster centers iteratively according to the objective function; Step 3.6: Clustering Termination. When the number of iterations reaches the maximum number of iterations or the change in the objective function is less than the convergence threshold, the iteration stops, the clustering results are output, and three types of objective modal components are obtained: random terms, fluctuation terms, and trend terms.

6. The method according to claim 1, characterized in that: In step 4, in the dual-drive prediction architecture, the spatial feature extraction network is a convolutional neural network, and the temporal dependency modeling network is a gated recurrent unit. The specific process includes: Step 4.1: Data preprocessing, normalizing the target modal components to obtain standardized data; Step 4.2: Spatial feature extraction. Standardized data is input into a convolutional neural network. Through the convolutional layers, pooling layers, and fully connected layers of the convolutional neural network, local spatial features of the target modal components are automatically extracted. Step 4.3: Temporal dependency modeling. The local spatial features are input into the gated loop unit. The long-term temporal dependency of the target modal components is modeled through the reset gate and update gate of the gated loop unit. Step 4.4: Component prediction output. The output of the gated loop unit is the predicted value of the corresponding target modal component.

7. The method according to claim 1, characterized in that: In step 5, the specific process of superimposing and recombining the predicted values ​​is as follows: the predicted values ​​of the random term, the predicted values ​​of the fluctuation term, and the predicted values ​​of the trend term are linearly superimposed according to the corresponding time nodes to obtain the total displacement predicted value of each time node. The total displacement predicted values ​​of all time nodes constitute the landslide total displacement prediction sequence.

8. The method according to claim 2, characterized in that: In step 6, the evaluation index includes one or more of the following: correlation coefficient, mean absolute percentage error, and root mean square error.

9. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the landslide displacement prediction method as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the landslide displacement prediction method as described in any one of claims 1 to 8.

Images