Landslide displacement prediction method and system based on Graph-Multi-TimesBlock model
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-23
- Publication Date
- 2026-04-03
Smart Images
Figure CN121787209A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of time series analysis technology in machine learning, and particularly relates to a landslide displacement prediction method and system based on the Graph-Multi-TimesBlock model. Background Technology
[0002] With the development and transformation of artificial intelligence technology, research on landslide displacement prediction has sprung up like mushrooms after rain, resulting in a large number of excellent technologies and methods. In particular, data-driven models have become a research hotspot in this field. Data-driven models include mathematical statistics, machine learning, and deep learning. Prediction methods based on mathematical statistics, such as grey system models, Verhulst, and ARIMA models, are favored because of their simple modeling process, intuitive expressions, and easily solvable structures. However, these models often ignore the nonlinear relationships in landslide displacement data, leading to poor prediction performance. Machine learning methods have attracted much attention for their ability to capture the nonlinear dynamic behavior of landslides. For example, random forests support large-scale data and handling missing values; support vector regression (SVR) has good data fitting capabilities; XGBoost supports parallel computing and regularization to prevent overfitting; and ELM has faster learning speed and generalization ability. Although the advantages of the above machine learning methods can enable landslide displacement prediction, landslide displacement data has complex correlations in the time dimension, and these methods cannot obtain more implicit dependencies in the time domain, thus limiting prediction performance. Long Short-Term Memory (LSTM) neural networks and Gated Recurrent Units (GRUs), representative of deep learning, possess a "memory block" structure, enabling them to pass on historical information and capture the temporal dependencies in displacement sequences, thus being widely used for landslide displacement prediction. However, while these models capture dependencies through temporal recursion, they are limited by the vanishing gradient problem, resulting in weak modeling capabilities for long-distance temporal features. More importantly, the sequential structure of these recurrent network models and their inability to represent feature weights prevent them from characterizing dynamic effects across different time stamps and from fully learning and interpreting the varying impacts of historical information at different time stamps on the degree of displacement.
[0003] Recently, the self-attention mechanism and the ability to support parallel computation of feature interactions across all time steps of the Transformer have revolutionized the sequential structure of RNNs, sparking a research boom among scholars. The Transformer can capture nonlinear correlations across multiple time steps, focusing on long-distance temporal dependencies by calculating the interaction weights of any pair of positions within a channel through self-attention. Based on this, researchers have studied and improved the Transformer from different perspectives. LiteTransNet is an optimized version of the Transformer in a specific domain (landslide prediction). Through local attention, a lightweight architecture, and an end-to-end design, it solves the efficiency and interpretability problems of traditional Transformers in long-sequence prediction, while retaining the core advantages of the attention mechanism. However, the displacement changes of loess landslides are complex and nonlinear, influenced by various environmental factors, resulting in multiple trends such as daily cycles due to warm days and cold nights, weekly cycles of farmland irrigation, monthly cycles of periodic reservoir storage and release, and annual cycles driven by freeze-thaw cycles, as well as various temporal patterns including trend and seasonality. These mixed temporal patterns deeply obscure the temporal dependencies in landslide sequences. LiteTransNet primarily focuses on point-to-point dependencies between different time points in a sequence, failing to characterize the hidden temporal dependencies within complex sequence structures and struggling to distinguish which are truly critical temporal dependencies. More importantly, loess landslide displacement is influenced by various environmental triggering factors, exhibiting a highly complex dynamic evolution process. Loess particles are predominantly silt-sized, with large pores exhibiting strong water sensitivity and collapsibility, making them susceptible to influences from environmental temperature, groundwater level, and soil moisture content. During the winter-spring transition, temperature changes lead to loess frost heave and thaw settlement; this freeze-thaw cycle damages soil structure, reduces shear strength, and increases landslide risk. Furthermore, during the freeze-thaw period, groundwater seepage zones freeze and accumulate, and as temperature rises, soil moisture content increases, causing a sharp decrease in soil strength and generating a freeze-water retention effect that induces slope instability. During the dry or rainy season, the groundwater level rises rapidly due to farmland irrigation or heavy rainfall. The collapsibility of loess causes its structure to be rapidly destroyed after water immersion, increasing porosity, saturating soil moisture content, and drastically reducing soil strength, making it prone to instability and failure. Moreover, the failure mode gradually develops from rear edge tensile cracking to "tensile cracking-slip" composite failure, and the slip surface morphology evolves from linear to arc-shaped.
[0004] While deep learning can capture the dependencies between different points in time through self-attention mechanisms, it cannot capture the varying degrees of influence of different factors on displacement within a given time point, nor the interactions between these factors.
[0005] Based on the above analysis, the problems and shortcomings of the existing technology are as follows:
[0006] (1) Public reports on landslide prediction are common and have achieved remarkable results and excellent performance, but most of the research focuses on reservoir-type landslides. For loess landslide prediction, most of the methods used are mathematical statistics and physical models, or the study of slope instability; there are few reports on the research results of using deep learning for loess landslide prediction.
[0007] (2) Loess landslides exhibit highly complex nonlinear evolution characteristics due to the influence of external inducing factors and their own properties, presenting different temporal variation patterns. The mixture of various temporal patterns makes the temporal dependencies deeply blurred. Existing deep learning methods are all based on time series and learn the dependencies between various time points in the sequence through self-attention mechanisms. However, they are difficult to distinguish the fuzzy temporal dependencies in this mixed temporal pattern, resulting in a lack of predictive ability.
[0008] (3) Various external environmental factors such as temperature and groundwater, as well as soil properties such as soil moisture content, will have different degrees of influence on loess landslide displacement, and these influencing factors are coupled with each other. Although deep learning can learn the different degrees of influence of the external environment on loess landslide displacement, it cannot clearly characterize the degree of influence and the coupling effect between various influencing factors. Summary of the Invention
[0009] To address the problems existing in the prior art, this invention provides a landslide displacement prediction method based on the Graph-Multi-TimesBlock model.
[0010] This invention is implemented as follows: a landslide displacement prediction method based on a Graph-Multi-TimesBlock model includes:
[0011] Step 1: Collect data on landslide displacement, temperature, soil moisture content, and groundwater level in the Heifangtai area to construct a landslide characteristic database;
[0012] Step 2: Construct the Graph-Multi-TimesBlock model;
[0013] Step 3: Input the landslide characteristic data into the time decomposition module, and obtain the different period transformations of the landslide displacement sequence through fast Fourier transform;
[0014] Step 4: The landslide displacement sequence is transformed by different periods through the time aggregation module to aggregate the intra-period and inter-period changes of different periods to obtain a one-dimensional time series containing time dependencies.
[0015] Step 5: The landslide feature data is processed through the influencing factor dependency graph learning module to obtain the influencing factor dependency graph structure, which reflects the different degrees of influence of each influencing factor on the landslide displacement and the coupling relationship between each influencing factor.
[0016] Step 6: Aggregate the one-dimensional time series obtained in Step 4 and the influencing factor dependency graph structure obtained in Step 5 through the dual dependency aggregation module to obtain a feature vector containing time dependency and inter-channel dependency.
[0017] Step 7: Calculate the final displacement prediction value by using linear regression on the feature vector obtained in Step 6.
[0018] Furthermore, the Graph-Multi-TimesBlock model consists of four modules: a time decomposition module, a time aggregation module, an influencing factor dependency graph learning module, and a dual dependency aggregation module.
[0019] The complete architecture of the Graph-Multi-TimesBlock model is as follows: First, landslide displacement, soil moisture content, groundwater level, and ambient temperature are collected to obtain landslide displacement time series data. These data are then processed through a time decomposition and aggregation module, as well as an influencing factor dependency graph learning module, to extract time-dependent feature vectors and capture the influence of triggering factors on the degree of landslide displacement and the interactions between various influencing factors. Second, a dual-dependency aggregation module is used to fuse the two dependencies to obtain feature vectors containing both time dependence and the interaction between influencing factors and landslide displacement. Finally, the dual-dependency feature vectors are used to calculate the final predicted value through linear regression.
[0020] Furthermore, in step three, the time decomposition module is composed of a TimseBlock module and the landslide displacement period is segmented using Fast Fourier Transform (FFT).
[0021] For landslide sequences that include soil moisture content, temperature, groundwater level, and landslide displacement (N represents the length of time, Performing a Fast Fourier Transform (FFT) on the channel number (representing the number of channels) yields different frequencies and their corresponding periods; this can be expressed as:
[0022]
[0023] Where Amp refers to the amplitude value after FFT calculation. This indicates that the average value is calculated, where A represents the amplitude of each frequency, which is obtained by averaging the values of each dimension.
[0024] Based on the sparsity of the frequency domain, and to prevent meaningless high-frequency noise, the frequencies corresponding to the K largest amplitudes (argTopK) are selected. ; can be represented as:
[0025]
[0026] Based on the selected frequency Obtain the corresponding amplitude and the length of the corresponding K periods ,then:
[0027]
[0028] The time series is zero-extended along the time dimension and based on the results... , Segmentation yields a two-dimensional tensor ;then
[0029]
[0030] Padding(·) performs zero-expansion of the time series along the time dimension, and Reshape(·) represents the reconstruction operation of transforming a one-dimensional tensor into a two-dimensional tensor.
[0031] Furthermore, in step four, the time aggregation module aggregates the intra-period and inter-period variations of the landslide displacement sequence to obtain a one-dimensional time series containing time dependencies; this module uses a convolutional neural network (CNN) to aggregate the time information within and during the landslide displacement period; and the two-dimensional tensor of the intra-period variation reconstructed in step three is used to... We use a CNN to aggregate the intra-period variations (columns) and inter-period variations (rows); thus:
[0032]
[0033] Since amplitude A can represent the relative importance of each frequency and period, and thus the importance of each transformed two-dimensional tensor, the amplitude A is first subjected to a Softmax transform to obtain... .based on Each one-dimensional tensor Aggregate into Specifically, it can be expressed as:
[0034]
[0035] .
[0036] Furthermore, in step five, the influencing factor dependency graph learning module clearly characterizes and represents the interactions of various inducing factors and their impact on landslide displacement from a new perspective—graph structure learning. In this module, a gravity graph learning model is used to capture global multi-dimensional dependency information, construct inter-channel dependencies, and automatically assign different weights to each channel, effectively reflecting the interactions between various influencing factors and their degree of influence on landslide displacement, thus realizing the graph structure association and representation of the disaster-causing mechanism. Inducing factors (temperature, soil moisture content, groundwater level) and landslide displacement are constructed as nodes in the graph, and their complex relationships are constructed into the graph structure as edges. By learning the characteristics of the nodes and edges in the graph... The system uses features and weights to capture the complex interactions between these multidimensional factors. For example, temperature responds to landslide displacement, establishing a boundary relationship (freeze-thaw cycles lead to slope instability). Simultaneously, temperature and soil moisture content have a boundary relationship (warm days and cold nights cause changes in soil moisture content), and soil moisture saturation or a surge triggers landslides, indicating that soil moisture content and landslide displacement nodes also have a boundary relationship. The weights of these boundaries are automatically updated based on changes in node characteristics (the weights of environmental factors affecting landslide displacement change with time and seasons), thus realistically depicting this implicit chain of "inducing factors—multidimensional interaction—disaster-pregnancy association—displacement deformation," avoiding the "black box" defects of traditional data-driven models.
[0037] First, a multilayer perceptron (MLP) is used to transform the landslide sequence to obtain feature sequences with multiple dimensions. In order to capture multidimensional information of landslide sequences; therefore:
[0038]
[0039] in, , , , For trainable weights, Representing nonlinear functions
[0040] In each dimension, each channel after feature transformation Treating it as a physical node with mass, mass for:
[0041]
[0042] in Represents a node The quality, N is the node The number of features, Represents a node The value of feature k; the numerator is the absolute value of the node feature to ensure that the quality is positive; the denominator is the sum of the absolute values of all node features to take into account the absolute size of the feature.
[0043] Gravity in two physical nodes Dot product similarity is used to measure the interaction between each influencing factor and the landslide displacement after feature changes. It can be represented as:
[0044]
[0045] in This represents the distance between two points. Represents two nodes in the i-th dimension , The similarity between them;
[0046] To focus attention on the most relevant influencing relationships and partially suppress those that might introduce redundant interference, the TopK(k) function is used to retain the top k most relevant relationships for each node, while setting the rest to 0. Then, the gravitational similarity is transformed into multiple dimensions. The final learned gravitational similarity B is obtained by merging the results, which is used to represent the influence relationship between the influencing factors and the landslide displacement.
[0047]
[0048] in This represents the attention weight of the i-th level gravitational similarity matrix computed within the self-attention mechanism.
[0049] Furthermore, in step six, the dual-dependency aggregation module aggregates the obtained one-dimensional time series and the influencing factor dependency graph structure obtained in step five to obtain a feature vector containing time dependency and inter-channel dependency; this module uses a graph convolutional neural network (GCN) to aggregate the time dependency vector Z obtained in step four and the inter-channel dependency B obtained in step five to obtain a dual-dependency vector F containing time dependency and channel dependency.
[0050] The channel features of the one-dimensional time vector Z are used as the feature vector of each node in the graph, and the channel dependency relationship B is used as the edge and weight relationship in the graph. Two layers of GCN are used to aggregate the surrounding node information and edge weight information to obtain the dual dependency vector F. Finally, the final displacement prediction value is obtained by linear regression calculation.
[0051] Another objective of this invention is to provide a landslide displacement prediction system based on a Graph-Multi-TimesBlock model, comprising:
[0052] The time decomposition module is used to decompose the fuzzy time changes in the landslide sequence that are mixed with multiple time patterns and trends, providing key guidance for the model to learn time dependencies.
[0053] The time aggregation module is used to aggregate multiple time patterns and trends of landslides to obtain a vector representation containing the time dependencies of landslides;
[0054] The factor dependency graph learning module is used to learn the different degrees of influence of each factor on landslide displacement and the coupling relationships between the factors; the dual dependency aggregation module...
[0055] The dual dependency aggregation module is used to aggregate time dependencies and channel dependencies;
[0056] Finally, the Graph-Multi-TimesBlock model was used to learn the Heifangtai landslide displacement dataset, enabling accurate prediction of landslide displacement.
[0057] Another object of the present invention is to provide a computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the landslide displacement prediction method based on the Graph-Multi-TimesBlock model.
[0058] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the landslide displacement prediction method based on the Graph-Multi-TimesBlock model.
[0059] Another objective of this invention is to provide an information data processing terminal for implementing the landslide displacement prediction system based on the Graph-Multi-TimesBlock model.
[0060] Based on the above technical solutions and the technical problems solved, please analyze the advantages and positive effects of the technical solution to be protected by this invention from the following aspects:
[0061] First, addressing the technical problems existing in the prior art and the difficulty in solving them, this paper closely analyzes, in conjunction with the technical solution to be protected by this invention and the results and data obtained during the research and development process, how the technical solution of this invention solves the technical problems, and the inventive technical effects brought about by solving these problems. The specific description is as follows:
[0062] This invention proposes a Graph-Multi-TimesBlock model, which for the first time addresses the problem of loess landslide displacement prediction and the significant shortcomings of current neural network-based reservoir landslide displacement prediction models in terms of predictive capability. It constructs a graph structure representing the coupling relationships of influencing factors and different influences on landslides from a new perspective—graph learning. Furthermore, it employs TimesBlock to deconstruct the complex time dependencies of landslides from different periodic and inter-period perspectives. Compared to traditional neural networks such as LSTM and GRU, this network can improve accuracy by 30%-45%, meeting the requirements for accurate and timely landslide displacement prediction in practical applications.
[0063] To address the challenge of distinguishing ambiguous time dependencies in complex time patterns, a time dependency learning module was designed. This module extracts multiple periodic variations from the landslide sequence through a time decomposition module, providing a clear visual representation of the different displacement changes across multiple periods. For each period, a time aggregation module combines the intra-period and inter-period representations of the landslide sequence, enabling a deeper understanding of the complex time patterns and uncovering hidden time dependencies, thus achieving effective differentiation of key time dependencies.
[0064] To clearly represent the coupling effects among influencing factors and their key impact on landslide displacement, an influencing factor dependency graph learning module was designed. Graph structure learning is used to visualize various influencing relationships. Nodes in the graph represent landslide displacement monitoring points and various environmental factors, while edges represent their interrelationships, thus constructing a complex graph structure. By learning the features and weights of the nodes and edges in the graph, the complex interactions between these multidimensional factors can be captured more accurately.
[0065] The landslide displacement prediction based on the Graph-Multi-TimesBlock model of this invention can fully learn the complex time dynamic pattern of landslide displacement, which is multi-scale, multi-period, nonlinear, event-driven (freeze-thaw cycle drives slope instability), and strongly coupled with external factors. It can realistically depict the implicit chain of "inducing factors - multi-dimensional interaction - disaster-pregnancy correlation - displacement deformation", avoiding the "black box" defects of traditional data-driven models, and providing support and decision-making for landslide prediction, early warning, disaster prevention and mitigation.
[0066] To address the aforementioned problems in existing technologies, this invention proposes a Graph-Multi-TimesBlock intelligent prediction method for loess landslide displacement that integrates multi-view temporal dependencies and graph structure environment modeling, and realizes loess landslide displacement prediction, thus overcoming the shortcomings of existing reservoir landslide prediction models.
[0067] This invention employs a novel perspective—graph structure learning—to co-model various inducing factors and landslide displacement into computable and correlated graph structure features. While preserving physical priors, it captures the interactions between various inducing factors and their nonlinear interaction characteristics on landslide displacement. It addresses the problem that traditional methods cannot capture the varying degrees of influence of different factors on displacement within a given timeframe, as well as the interactions between these factors. This enhances the comprehensive understanding of landslide causes by the Graph-Multi-TimesBlock model, providing a new paradigm for landslide displacement prediction that combines physical mechanisms with data-driven approaches. Furthermore, it utilizes multiple stacked TimesBlocks and a multi-period identification mechanism to deeply capture multi-scale features in landslide displacement. The lower-level TimesBlocks capture more subtle, higher-frequency (shorter-period) change patterns (such as diurnal variations, responses to irrigation and soil moisture content), while higher-level TimesBlocks integrate lower-level information to capture coarser-grained, lower-frequency (longer-period) patterns (such as seasonal trends and long-term creep). This multi-scale representation capability can uncover hidden temporal features in complex sequence structures, effectively distinguish key time dependencies, and provide a supporting environment for defuzzifying time dependencies in mixed temporal patterns. Ultimately, the dual-dependency aggregation module not only captures the spatial dependencies between nodes (the interactions of various triggering factors and their impact on landslide displacement) but also models the temporal dependencies of their own states (such as the historical behavior of landslide displacement), achieving a deep fusion of spatial neighborhood information and time series information. The synergistic effect of multiple modules enables the model to achieve better results in the field of landslide displacement prediction.
[0068] Secondly, as supplementary evidence of the inventive step of the claims of this invention, it is also reflected in the following important aspects:
[0069] (1) The expected benefits and commercial value of the technical solution of this invention after transformation are as follows:
[0070] After the technology transfer of this invention, it possesses extremely high commercial and practical value for predicting loess landslide displacement. Firstly, experiments have demonstrated that this invention achieves high-precision prediction results on the Heifangtai landslide dataset. Secondly, it plays a positive role in time series prediction projects, exhibiting good portability in various time series prediction tasks such as weather, traffic flow, and transformer temperature prediction, with a wide range of applications. Researchers in various fields can use this invention to achieve their different purposes. Furthermore, after the technology transfer, this invention also has the following advantages: it is simple and easy to operate, allowing different users to quickly acquire the necessary skills. Additionally, given the current context where landslide hazard early warning has become an important issue, the emergence of this invention aligns with the trend of the times. Based on the above points, it can be said that this invention has extremely high commercial value after technology transfer.
[0071] (2) The technical solution of this invention fills a technical gap in the industry both domestically and internationally:
[0072] Currently, most landslide prediction methods for loess landslides rely on mathematical statistics and physical models, or focus on slope instability. Research using deep learning for loess landslide prediction is lacking. While deep learning-based reservoir landslide prediction models can predict loess landslide displacement, they have several limitations: ① Loess landslide displacement is complex and nonlinear, influenced by various environmental factors, resulting in multiple periodic variations at different time scales. These include daily cycles due to warm days and cold nights in the Loess Plateau region, weekly cycles of farmland irrigation, monthly cycles of reservoir water storage and release, and annual cycles driven by freeze-thaw cycles, exhibiting various temporal patterns (such as trend-based and seasonal patterns). This mixture of trends and temporal patterns deeply obscures the temporal dependencies within landslide sequences. Existing deep learning-based landslide displacement prediction models primarily focus on point-to-point dependencies between different time points in the sequence, failing to characterize the hidden temporal dependencies within complex sequence structures and making it difficult to distinguish which are truly critical temporal dependencies. ② It is impossible to clearly characterize the different impacts of various environmental factors on the degree of displacement triggered by loess landslides, nor can it characterize the coupling effect between various influencing factors. Therefore, a landslide displacement prediction based on the Graph-Multi-TimesBlock model is proposed to fully characterize the mixed time patterns and the different degrees of influence of triggering factors, as well as the interaction between various influencing factors, so as to achieve timely and accurate landslide displacement prediction.
[0073] (3) Whether the technical solution of the present invention solves the technical problem that people have long wanted to solve but have never been able to solve successfully:
[0074] Current research on loess landslide displacement prediction is limited, with most studies focusing on landslide characteristics and slope instability mechanisms. Traditional machine learning models cannot capture more implicit dependencies in the time domain, limiting their predictive performance. Improved Transformer-based models or ensemble prediction deep learning models fail to fully deconstruct the mixed temporal patterns in landslide sequences, do not effectively link the model with the nonlinear evolution mechanism of landslides, and do not elucidate the intrinsic connection between the dynamic evolution mechanism of landslide displacement under the coupling effect of environmental factors and the algorithm mechanism.
[0075] The landslide displacement prediction method based on the Graph-Multi-TimesBlock model proposed in this invention can fully characterize the different degrees of influence of mixed time patterns and inducing factors, as well as the interactions between these factors, achieving timely and accurate landslide displacement prediction. It can play a role in landslide early warning and disaster prevention and mitigation, helping to identify and warn of potential landslide hazards, and prompting disaster management to shift from "passive response" to "proactive prevention," achieving maximum economic recovery with minimal population loss. Attached Figure Description
[0076] Figure 1 This is a flowchart of the landslide displacement prediction method based on the Graph-Multi-TimesBlock model provided in this embodiment of the invention.
[0077] Figure 2 This is a block diagram of the landslide displacement prediction system based on the Graph-Multi-TimesBlock model provided in an embodiment of the present invention.
[0078] Figure 3 This is a sample image of the Heifangtai landslide dataset provided in an embodiment of the present invention.
[0079] Figure 4 This is an overall architecture diagram of the Graph-Multi-TimesBlock model provided in this embodiment of the invention.
[0080] Figure 5 This is a structural diagram of the time decomposition module provided in an embodiment of the present invention.
[0081] Figure 6 This is a structural diagram of the time aggregation module provided in an embodiment of the present invention.
[0082] Figure 7 This is a structural diagram of the influencing factor dependency graph learning module provided in an embodiment of the present invention.
[0083] Figure 8 This is a structural diagram of the dual-dependency aggregation module provided in an embodiment of the present invention.
[0084] Figure 9 These are the prediction curves of each landslide displacement prediction model provided in this embodiment of the invention on the Heifangtai landslide dataset, as well as the error graphs at each time point.
[0085] Figure 10 This is a comparison chart of ablation experimental results of each module of the Graph-Multi-TimesBlock model provided in the embodiments of the present invention.
[0086] Figure 11 This is a heat map showing the correlation between various influencing factors and landslide displacement provided in the embodiments of the present invention. Detailed Implementation
[0087] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0088] The implementation process of this invention is mainly divided into three aspects: the time decomposition module is used to decompose landslide displacement and extract intra-period and inter-period changes in different periods of landslide displacement; the time aggregation module is used to integrate intra-period and inter-period changes and learn the time dependence relationship of deep fuzziness in landslide displacement; the influencing factor graph learning module is used to learn the different degrees of influence of each influencing factor on landslide displacement and the coupling relationship between each influencing factor.
[0089] refer to Figure 1 The landslide displacement prediction based on the Graph-Multi-TimesBlock model provided in this embodiment of the invention includes the following steps:
[0090] S101, collect data on landslide displacement, temperature, soil moisture content, groundwater level, etc. in the Heifangtai area to construct a landslide characteristic database;
[0091] S102, Construct the Graph-Multi-TimesBlock model;
[0092] S103, input the landslide characteristic data into the time decomposition module, and obtain different periodic transformations of the landslide displacement sequence through fast Fourier transform;
[0093] S104, the landslide displacement sequence is transformed by different periods through the time aggregation module, and the intra-period and inter-period changes of different periods are aggregated to obtain a one-dimensional time series containing time dependence;
[0094] S105, the landslide feature data is processed through the influencing factor dependency graph learning module to obtain the influencing factor dependency graph structure, which reflects the different degrees of influence of each influencing factor on the landslide displacement and the coupling relationship between each influencing factor.
[0095] S106, The one-dimensional time series obtained in S104 and the influence factor dependency graph structure obtained in step 5 are aggregated through the dual dependency aggregation module to obtain a feature vector containing time dependency and inter-channel dependency.
[0096] S107. The feature vector obtained in S106 is used to calculate the final displacement prediction value through linear regression.
[0097] As a preferred embodiment, such as Figure 2As shown, an embodiment of the present invention provides a landslide displacement prediction system based on a Graph-Multi-TimesBlock model, comprising:
[0098] The time decomposition module is used to decompose the fuzzy time changes in the landslide sequence that are mixed with multiple time patterns and trends, providing key guidance for the model to learn time dependencies.
[0099] The time aggregation module is used to aggregate multiple time patterns and trends of landslides to obtain a vector representation containing the time dependencies of landslides;
[0100] The factor dependency graph learning module is used to learn the different degrees of influence of each factor on landslide displacement and the coupling relationships between the factors; the dual dependency aggregation module...
[0101] The dual dependency aggregation module is used to aggregate time dependencies and channel dependencies;
[0102] Finally, the Graph-Multi-TimesBlock model was used to learn the Heifangtai landslide displacement dataset, enabling accurate prediction of landslide displacement.
[0103] The landslide displacement prediction based on the Graph-Multi-TimesBlock model provided in this embodiment of the invention specifically includes the following steps:
[0104] Step 1: Use temperature sensors, landslide displacement monitoring integrated machines, pore water pressure gauges, and soil moisture sensors to collect data such as temperature, landslide displacement, groundwater level, and soil moisture content to construct a landslide characteristic database.
[0105] Figure 3 This dataset contains soil moisture content, temperature, landslide displacement, and groundwater level changes from November 2016 to November 2017 in the Heifangtai landslide dataset. Monitoring was performed every 1 second, followed by data resampling every 1 hour, resulting in a total of 8577 data points. The dataset was divided into training, validation, and test sets in a 7:2:1 ratio.
[0106] Step 2: Construct the Graph-Multi-TimesBlock model.
[0107] Figure 4The complete architecture of the Graph-Multi-TimesBlock model is presented. First, landslide displacement, soil moisture content, groundwater level, and ambient temperature are collected to obtain time-series data of landslide displacement. These data are then processed through a time decomposition and aggregation module, and an influencing factor dependency graph learning module, respectively, to extract time-dependent feature vectors and capture the influence of triggering factors on the degree of landslide displacement and the interactions between these factors. Next, a dual-dependency aggregation module is used to fuse the two dependencies, resulting in feature vectors that include both time dependence and the interaction between influencing factors and landslide displacement. Finally, the dual-dependency feature vectors are used to calculate the final predicted value through linear regression.
[0108] Step 3: Input the landslide feature data into the time decomposition module, and obtain different periodic transformations of the landslide displacement sequence through fast Fourier transform.
[0109] Figure 5 This is a structural diagram of the time decomposition module. This module consists of TimseBlock modules and uses Fast Fourier Transform (FFT) to segment the landslide displacement period.
[0110] For landslide sequences that include soil moisture content, temperature, groundwater level, and landslide displacement (N represents the length of time, Performing a Fast Fourier Transform (FFT) on the channel number (representing the number of channels) yields different frequencies and their corresponding periods. This can be expressed as:
[0111]
[0112] Here, Amp refers to the amplitude value calculated by FFT, and A represents the amplitude of each frequency, which is obtained by averaging each dimension.
[0113] Based on the sparsity of the frequency domain, and to prevent meaningless high-frequency noise, the frequencies corresponding to the K largest amplitudes (argTopK) are selected. ; can be represented as:
[0114]
[0115] Based on the selected frequency Obtain the corresponding amplitude and the length of the corresponding K periods ,then:
[0116]
[0117] The time series is zero-extended along the time dimension and based on the results... , Segmentation yields a two-dimensional tensor .then
[0118]
[0119] Padding(·) performs zero-expansion of the time series along the time dimension, and Reshape(·) represents the reconstruction operation of transforming a one-dimensional tensor into a two-dimensional tensor.
[0120] Step 4: The landslide displacement sequence is transformed by different periods through the time aggregation module to aggregate the intra-period and inter-period changes of different periods to obtain a one-dimensional time series containing time dependencies.
[0121] Figure 6 A time aggregation module is described, which aggregates temporal information within and during landslide displacement cycles using a convolutional neural network (CNN). This module converts the reconstructed two-dimensional tensor of changes within each cycle from step three. We use a CNN to aggregate the intra-period variations (columns) and inter-period variations (rows). Thus:
[0122]
[0123] Since amplitude A can represent the relative importance of each frequency and period, and thus the importance of each transformed two-dimensional tensor, we first perform a Softmax transform on amplitude A to obtain... .based on Each one-dimensional tensor Aggregate into Specifically, it can be expressed as:
[0124]
[0125]
[0126] Step 5: The landslide feature data is processed through the influencing factor dependency graph learning module to obtain the influencing factor dependency graph structure, which reflects the different degrees of influence of each influencing factor on the landslide displacement and the coupling relationship between the influencing factors.
[0127] like Figure 7As shown, the influencing factor dependency graph learning module clearly characterizes and represents the interactions of various inducing factors and their impact on landslide displacement from a new perspective—graph structure learning. This module uses a gravity graph learning model to capture global multi-dimensional dependency information, construct inter-channel dependencies, and automatically assign different weights to each channel, effectively reflecting the interactions between various influencing factors and their degree of influence on landslide displacement, thus achieving graph structure association and representation of the disaster-causing mechanism. Inducing factors (temperature, soil moisture content, groundwater level) and landslide displacement are constructed as nodes in the graph, and their complex relationships are constructed into the graph structure as edges. By learning the features and weights of the nodes and edges in the graph, the complex interactions between these multi-dimensional factors are captured. For example, temperature responds to landslide displacement, establishing an edge relationship (freeze-thaw cycles lead to slope instability); simultaneously, temperature and soil moisture content also have an edge relationship (warm days and cold nights lead to changes in soil moisture content), and soil moisture saturation or a surge triggers landslides, indicating that soil moisture content and landslide displacement nodes also have an edge relationship; and the weight of the edge is automatically updated according to the changes in node characteristics (as time and season change, the weight of environmental factors affecting landslide displacement changes accordingly), thus truly depicting this implicit chain of "inducing factors - multidimensional interaction - disaster-causing association - displacement deformation", avoiding the "black box" defects of traditional data-driven models.
[0128] First, a multilayer perceptron (MLP) is used to transform the landslide sequence to obtain feature sequences with multiple dimensions. This is to capture multidimensional information about the landslide sequence. Therefore:
[0129]
[0130] in, , , , For trainable weights, This represents a nonlinear function.
[0131] In each dimension, each channel after feature transformation Treating it as a physical node with mass, mass for:
[0132]
[0133] in Represents a node The quality, N is the node The number of features, Represents a node The value at feature k. The numerator is the absolute value of the node feature to ensure the quality is positive. The denominator is the sum of the absolute values of all node features to account for the absolute size of the feature.
[0134] Gravity in two physical nodes Dot product similarity is used to measure the interaction between each influencing factor and the landslide displacement after feature changes. It can be represented as:
[0135]
[0136] in This represents the distance between two points. Represents two nodes in the i-th dimension , The similarity between them.
[0137] To focus attention on the most relevant influencing relationships and partially suppress potentially redundant interference, the TopK(k) function is used to retain the top k most relevant relationships for each node, while setting the rest to 0. Subsequently, the gravitational similarity is obtained through transformation across multiple dimensions. The final learned gravitational similarity B is obtained by merging the results, which is used to represent the influence relationship between the influencing factors and the landslide displacement.
[0138]
[0139] in This represents the attention weight of the i-th level gravitational similarity matrix computed within the self-attention mechanism.
[0140] Step 6: The obtained one-dimensional time series and the influencing factor dependency graph structure obtained in Step 5 are aggregated through the dual dependency aggregation module to obtain a feature vector containing time dependency and inter-channel dependency.
[0141] like Figure 8 As shown, the dual dependency aggregation module uses a graph convolutional neural network (GCN) to aggregate the time dependency vector Z obtained in step four and the inter-channel dependency B obtained in step five, resulting in a dual dependency vector F containing both time dependency and inter-channel dependency.
[0142] The channel features of the one-dimensional time vector Z are used as the feature vector of each node in the graph, and the channel dependency relationship B is used as the edge and weight relationship in the graph. Two layers of GCN are used to aggregate the surrounding node information and edge weight information to obtain the dual dependency vector F.
[0143] Step 7: The time dependency vector F obtained in Step 6 is used to calculate the final displacement prediction value through linear regression.
[0144] To verify the inventiveness and technicality of the present invention, the technical solution of the claims is described herein in conjunction with specific application embodiments.
[0145] The landslide displacement prediction method based on the Graph-Multi-TimesBlock model designed in this invention can predict future changes in landslide displacement in a timely and accurate manner. It has already demonstrated excellent performance on the Heifangtai landslide dataset, and this landslide displacement prediction method based on the Graph-Multi-TimesBlock model can also achieve advanced prediction performance in other prediction tasks. For example, its model can achieve good prediction results in weather and traffic flow prediction.
[0146] Figure 4 The Graph-Multi-TimesBlock loess landslide displacement prediction model established by the method described in this section is presented. Figure 9 This table shows the prediction results of the model on the Heifangtai landslide dataset with a 96-hour input and a 192-hour prediction setting, along with the absolute error value at each time point. Table 1 also presents the prediction performance of each model on the Heifangtai landslide dataset; the bolded portion indicates that Graph-Multi-TimesBlock achieved the best result in this metric. It can be seen that the five comparative models—SAU-DTAE, VMD-ARIMA-IM, LiteTransNet, VMD-CNN-Attention-BiGRU, and GWO-VMD-DES-infomer—exhibit varying degrees of random fluctuation in their prediction results during the gradual creep period (0-85 hours, 110-140 hours, 160-192 hours) and inaccuracies in their prediction results during the step change period (85-110 hours, 140-160 hours). The Graph-Multi-TimesBlock prediction curves are closer to the actual displacement curves in both the slow creep period and the step deformation period, which verifies that the Graph-Multi-TimesBlock model can effectively solve the random fluctuations in the prediction results of landslide displacement during the slow creep period and the inaccuracy of the prediction results when step deformation occurs.
[0147] Table 1
[0148] Model MSE MAE <![CDATA[R 2 ]]> SAU-DTAE 0.216 0.223 0.909 VMD-ARIMA-IM 0.235 0.233 0.889 LiteTransNet 0.221 0.213 0.892 VMD-CNN-Attention-BiGRU 0.214 0.212 0.934 GWO-VMD-DES-infomer 0.221 0.230 0.897 Graph-Multi-TimesBlock 0.177 0.194 0.956
[0149] Specifically, compared with existing landslide identification technologies, the present invention also has the following technical advantages:
[0150] 1. Graph-Multi-TimesBlock addresses the significant limitations of current neural network-based reservoir landslide displacement prediction models in predicting landslide displacement. It employs a novel perspective—graph learning—to construct graph structures representing the coupling relationships between influencing factors and their different impacts on landslides. Furthermore, it utilizes TimesBlock to deconstruct the complex temporal dependencies of landslides from different periodic and inter-period perspectives. Compared to traditional neural networks, this network achieves higher accuracy, meeting the requirements for precise and timely landslide displacement prediction in practical applications.
[0151] 2. To address the challenge of distinguishing ambiguous time dependencies under mixed time patterns, a time dependency learning module was designed. This module extracts multiple periodic variations from the landslide sequence through a time decomposition module, providing a clear visual representation of the different displacement changes across multiple periods. For each period, a time aggregation module combines the intra-period and inter-period representations of the landslide sequence, enabling a deeper understanding of the complex time patterns and uncovering hidden time dependencies, thus achieving effective differentiation of key time dependencies.
[0152] 3. To clearly represent the coupling effects among influencing factors and their key impact on landslide displacement, an influencing factor dependency graph learning module was designed. Graph structure learning is used to visualize various influencing relationships. Nodes in the graph represent landslide displacement monitoring points and various environmental factors, and edges represent their interrelationships, thus constructing a complex graph structure. By learning the features and weights of the nodes and edges in the graph, the complex interactions between these multidimensional factors can be captured more accurately.
[0153] The landslide displacement prediction based on the Graph-Multi-TimesBlock model of this invention can fully learn the complex time dynamic pattern of landslide displacement, which is multi-scale, multi-period, nonlinear, event-driven (freeze-thaw cycle drives slope instability), and strongly coupled with external factors. It can realistically depict the implicit chain of "inducing factors - multi-dimensional interaction - disaster-pregnancy association - displacement deformation", avoiding the "black box" defects of traditional data-driven models. It can play a certain role in landslide early warning, forecasting and disaster prevention and mitigation, and can effectively identify landslides, thus promoting disaster management from "passive response" to "proactive prevention and control".
[0154] Example 1
[0155] In this embodiment, the Heifangtai landslide monitoring area is taken as the research object. Multi-source monitoring data from 2015 to 2024 were collected, including multi-dimensional time series data such as displacement of landslide displacement monitoring points, surface temperature, shallow soil moisture content, and groundwater level. The data sampling interval was 1 hour, and the monitoring period was 365 days. A landslide feature database was established for model input. By normalizing and denoising the data, equipment errors and outliers were eliminated to ensure the consistency and comparability of the input features.
[0156] A Graph-Multi-TimesBlock model was used for modeling and analysis. Multi-source time series data were input into the time decomposition module and the influencing factor dependency graph learning module, respectively, to obtain time-dependent features and coupling features of inducing factors. Then, the two features were input into a dual-dependency aggregation module. Finally, a linear regression layer was used to obtain the landslide displacement prediction results. The mean squared error (MSE) of the prediction accuracy was 0.177 mm, indicating that the model can accurately capture the multi-scale time characteristics and multi-factor coupling relationships of landslide displacement.
[0157] Example 2
[0158] The Graph-Multi-TimesBlock model constructed in this embodiment adopts a four-layer modular structure: a time decomposition module for periodic feature extraction, a time aggregation module for temporal dependency modeling, an influencing factor dependency graph learning module for multi-dimensional factor coupling modeling, and a dual dependency aggregation module for final feature fusion. The modules are connected in the form of tensor transfer to ensure stable gradient backpropagation. The input dimension of the model is (N×d), where N is the time step (720) and d is the number of feature channels (4).
[0159] The model was trained using the Adam optimizer with an initial learning rate of 0.001 and 200 training epochs. The loss function was mean squared error (MSE), and an EarlyStopping strategy was employed to prevent overfitting. Results show that compared to advanced landslide displacement prediction models such as Lightweight Transformer (LiteTransNet), GWO-VMD-DES-infomer, VMD-ARIMA-LSTM, and VMD-CNN-Attention-BiGRU, this model achieved a maximum improvement of 24.68% in MSE, 16.73% in MAE, and 7.5% in R² on the test set, validating the effectiveness of the temporal and graph dependency fusion design in the model architecture.
[0160] Example 3
[0161] In this embodiment, the time decomposition module uses Fast Fourier Transform (FFT) to perform frequency domain decomposition on the landslide displacement time series. Taking a displacement sequence length of N=720 as an example, the amplitude spectrum is obtained after performing FFT. By selecting the frequency components corresponding to the first K=5 largest amplitude values, the main periods are extracted. Based on the frequency and sampling period, the period length set {p1,p2,p3,p4,p5} is calculated, and then a two-dimensional time tensor is reconstructed based on each period length.
[0162] During the reconstruction process, the time series is padded to an integer multiple of the period length using a zero-padding method to prevent information loss caused by truncation. The reconstructed two-dimensional tensor contains intra-period variations and inter-period trend features, providing input for subsequent convolutional aggregation. Experiments show that the time decomposition module can significantly improve the model's ability to distinguish seasonal fluctuations and sudden responses to landslides, especially reducing the prediction error by approximately 18% during the spring thaw and rainy season displacement surge phases.
[0163] Example 4
[0164] In this embodiment, the time aggregation module is based on a one-dimensional convolutional neural network structure, using a 3×1 convolutional kernel to aggregate the changing features within a period, and using multi-layer convolution to achieve cross-period feature fusion. The two-dimensional tensor of each period is convolved and max-pooled to obtain a one-dimensional time feature Y, and then weighted and summed according to the amplitude Aᶦ obtained by FFT to obtain the time dependency vector Z.
[0165] This aggregation strategy adaptively adjusts the weights based on the importance of different frequency components, allowing high-amplitude frequencies to dominate overall trend modeling while low-amplitude frequencies are used for detail compensation. Experiments demonstrate that, compared to unweighted convolution aggregation, the time-dependent vectors in this module can reduce prediction bias by more than 25% in samples with significant periodic changes (such as during periods of heavy rainfall), thus improving the physical interpretability of time modeling.
[0166] Example 5
[0167] In this embodiment, the influencing factor dependency graph learning module maps four channel features—temperature, soil moisture content, groundwater level, and landslide displacement—using a multilayer perceptron (MLP) to generate multidimensional node features. The edge weights between nodes are calculated using the gravitational similarity formula, where the node mass is determined by the normalization of the channel feature magnitude, and the distance between nodes is obtained by inverse calculation using cosine similarity.
[0168] Subsequently, the TopK function was used to select the three most relevant neighboring nodes for each node, constructing a sparse dependency graph to reduce noisy edges. This graph structure reflects the physical connections between inducing factors, such as the strong coupling between groundwater level and soil moisture content. Weighting coefficients αᶦ calculated using a self-attention mechanism were used to adjust the contributions of dependencies in different dimensions. Experiments show that this module can automatically identify key catastrophic pathways such as freeze-thaw cycles and water saturation effects, demonstrating good interpretability.
[0169] Example 6
[0170] In this embodiment, the dual-dependency aggregation module employs a two-layer graph convolutional network structure, using the time dependency vector Z as the node feature input and the influencing factor dependency matrix B as the edge weight input. The first layer of graph convolution achieves local dependency aggregation, and the second layer of graph convolution achieves global dependency fusion. Each layer is followed by ReLU activation and normalization to prevent gradient vanishing.
[0171] The fused dual-dependency feature vector F achieves feature interaction in both time and channel dimensions, enabling it to simultaneously capture the complex relationship between the dynamic temporal characteristics of landslides and environmental factors. The mean absolute error (MAE) of the predicted results output by the linear regression layer compared with measured data is 0.194 mm, which is better than the 0.233 mm of the traditional temporal regression model, verifying the significant improvement effect of the dual-dependency aggregation design.
[0172] Example 7
[0173] The landslide displacement prediction system provided in this embodiment includes four core functional modules: a time decomposition module, a time aggregation module, an influencing factor dependency graph learning module, and a dual dependency aggregation module. The system is implemented using Python and the PyTorch framework and runs on a computing platform equipped with an NVIDIA RTX 3090 graphics card and 32 GB of memory.
[0174] The system input interface can receive monitoring data in CSV or database format, while the output interface provides predicted displacement, trend graphs, and early warning level assessment results. The system can automatically perform model training and inference processes and supports visualization and interpretability of model parameters, enabling real-time prediction and visual analysis of landslide displacement.
[0175] Example 8
[0176] The computer device provided in this embodiment includes a central processing unit, a memory, and a communication interface module. The memory contains a landslide prediction program. When the central processing unit executes the program, it calls the four functional modules of the Graph-Multi-TimesBlock model to perform periodic decomposition, feature extraction, and prediction output on the input landslide monitoring data.
[0177] This device can be installed in the data processing terminal of a landslide monitoring station, combining with IoT sensor nodes to receive monitoring data in real time and periodically output displacement prediction curves. Field deployment tests showed that the system latency is less than 2 seconds and the prediction update cycle is 10 minutes, meeting the real-time requirements for early warning and response in landslide monitoring.
[0178] Example 9
[0179] In this embodiment, the computer-readable storage medium is a solid-state drive or a cloud server storage unit, containing a computer program file. This program includes a data acquisition module, a model building module, and a prediction output module. When this program is executed, landslide displacement prediction can be achieved on any device with a Python runtime environment.
[0180] By deploying this program in the cloud, parallel analysis and centralized management of data from multiple landslide areas are possible. The program file structure follows a modular architecture, supporting parameter fine-tuning and retraining. Experiments show that this media-stored program maintains over 90% prediction accuracy after transfer learning on multi-regional datasets, demonstrating high reusability.
[0181] Example 10
[0182] The information data processing terminal in this embodiment includes a display screen, a data processing module, and a communication module. The terminal can connect to the landslide monitoring data center via a wireless network, retrieve monitoring data from the database in real time, and execute a Graph-Multi-TimesBlock model for prediction.
[0183] The prediction results are displayed in chart form on the terminal interface, including displacement-time curves, trend prediction lines, and risk level indicators. Users can view the prediction results and influencing factor weights for different monitoring points via touch operation. The terminal supports customizable warning thresholds; when the predicted displacement exceeds the set threshold, it automatically issues an audible and visual alarm, realizing the visualization and intelligentization of landslide monitoring.
[0184] This invention can analyze various time patterns in landslide displacement, the coupling relationship between various influencing factors, and the different degrees of influence on landslide displacement.
[0185] 1. The time decomposition module defuzzifies the mixed time patterns in landslide displacement.
[0186] The time decomposition module transforms a one-dimensional time series into multiple two-dimensional sequences using Fourier transform, intuitively reflecting the different changes in landslide displacement over multiple periods. Therefore, we visualized the multi-period landslide changes identified by the time decomposition module and analyzed them in the form of a heatmap, as shown in the figure. Specifically, the landslide displacement sequence ( Figure 10 a) Projected onto the frequency domain via Fourier transform ( Figure 10(b) Taking the two points with the largest amplitudes (red and green dots) as examples, we can characterize the heat map performance of the two-dimensional changes in the landslide sequence on the corresponding period, such as... Figure 10 As shown in the figure, each column in the heatmap represents the trend of landslide displacement within a period, and each row represents the trend of landslide displacement during the period. This allows for better identification and differentiation of the trend of landslide displacement under deeply blurred time dependencies.
[0187] exist Figure 10 In the data, the values in each column first rise and then fall, indicating that the landslide displacement has a peak within each cycle. The different fluctuations in each row of data indicate that the peak displacement varies across different cycles (15 peak values). Figure 10 Overall, each column of data first decreases and then increases, exhibiting troughs, reflecting the trend of landslide displacement changing from low to high within each cycle. The fluctuations in each row of data also reflect the variation of the troughs in landslide displacement (14 troughs). These two data variations are consistent with the changing trends of the landslide sequence in the time domain, indicating that the time dependency learning module can effectively distinguish between different time variation trends within and between cycles in the mixed time pattern of the landslide sequence. It can deeply mine the time dependencies within the landslide sequence and achieve defuzzification of time dependencies under mixed time patterns.
[0188] 2. The influencing factor dependency graph learning module visualizes the influence relationships between channels.
[0189] The Graph-Multi-TimesBlock model, through a novel perspective—graph structure learning—clearly characterizes and represents the coupling effects of various environmental factors and their influence on loess landslide displacement. This allows for a better interpretation of the intrinsic mechanisms of loess landslide dynamic evolution, achieving graph-structured association and representation of the disaster-causing mechanism. Here, we input the landslide sequence data from monitoring point ZK4 in May 2017 into the Graph-Multi-TimesBlock model to visualize the interactions of various triggering factors and their different impacts on the degree of landslide displacement, such as thermal factors. Figure 11 It can be seen that, compared with temperature, the increase in soil moisture content and the rise in groundwater level due to irrigation have a significant impact on the landslide displacement, with values of 0.82 and 0.86 respectively. Furthermore, there is a mutual influence between changes in groundwater level and soil moisture content, with a value of 0.81.
[0190] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.
[0191] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A landslide displacement prediction method based on a Graph-Multi-TimesBlock model, characterized in that, Includes the following steps: (1) Collect multi-source data such as displacement, temperature, soil moisture content, and groundwater level in the landslide monitoring area to construct a landslide characteristic database; (2) Construct the Graph-Multi-TimesBlock model; (3) Input the landslide characteristic data into the time decomposition module and separate the periodic components of the landslide displacement sequence through fast Fourier transform; (4) Aggregate intra-period and inter-period changes of different periods through the time aggregation module to generate a one-dimensional time series containing time dependencies; (5) Construct the influence factor dependency graph structure through the influence factor dependency graph learning module to characterize the influence of each factor on landslide displacement and their interaction relationship; (6) Input the time-dependent features and the influencing factor dependency graph structure into the dual-dependency aggregation module, fuse the time dependency and channel dependency relationships, and generate feature vectors; (7) Input the fused feature vector into the linear regression layer and output the predicted landslide displacement value.
2. The landslide displacement prediction method as described in claim 1, characterized in that, The Graph-Multi-TimesBlock model comprises four modules: a time decomposition module, a time aggregation module, an influencing factor dependency graph learning module, and a dual dependency aggregation module. These modules are connected sequentially to form a feature extraction and fusion structure, which is used to jointly learn the temporal variation pattern of landslides and the coupling effect of environmental factors.
3. The landslide displacement prediction method as described in claim 1, characterized in that, The time decomposition module performs frequency analysis on the landslide sequence based on the Fast Fourier Transform, extracting the periods corresponding to the first few frequencies with the largest amplitudes. The period length is calculated based on the frequency, and the landslide sequence is divided into two-dimensional tensors according to the period. Time series are expanded and transformed through filling and reconstruction operations to preserve the variation characteristics within and between periods.
4. The landslide displacement prediction method as described in claim 1, characterized in that, The time aggregation module extracts time features within a period and during a week using a convolutional neural network. The features are then weighted and aggregated based on the amplitude weights corresponding to each frequency. Generate time feature vectors that contain time dependencies.
5. The landslide displacement prediction method as described in claim 1, characterized in that, The influencing factor-dependent graph learning module establishes the correlation between inducing factors and landslide displacement based on graph structure learning; Multidimensional features are extracted using a multilayer perceptron, each influencing factor is represented as a node, and the similarity between nodes is represented as edge weights. Calculate the gravitational similarity matrix between nodes to characterize the coupling strength between factors; By weighting the similarity matrices of different dimensions using a self-attention mechanism, a comprehensive dependency matrix is obtained.
6. The landslide displacement prediction method as described in claim 1, characterized in that, The dual-dependency aggregation module fuses temporal dependency features and channel dependency features based on a graph convolutional network. Using time feature vectors as node features and the dependency matrix of influencing factors as edge weight matrix, By performing multi-layer graph convolution operations, the temporal and channel dependency information is jointly aggregated to obtain a dual-dependency fused feature vector.
7. A landslide displacement prediction system for implementing the method according to any one of claims 1 to 6, characterized in that, include: The time decomposition module is used to decompose the periodic characteristics of landslide displacement sequences; The time aggregation module is used to aggregate changes within and outside a period and extract time-dependent features; The influencing factor dependency graph learning module is used to construct graph-structured dependencies between inducing factors; The dual-dependency aggregation module is used to fuse time dependency and channel dependency to generate predicted feature vectors; The system output module is used to perform linear regression on the fused features and output the landslide displacement prediction results.
8. A computer device, characterized in that, Including processor and memory, The memory contains computer programs. When the processor executes the program, it performs the landslide displacement prediction method according to any one of claims 1 to 6.
9. A computer-readable storage medium having a computer program stored thereon, When the program is executed by the processor, it performs all the steps of the landslide displacement prediction method according to any one of claims 1 to 6.
10. An information data processing terminal, characterized in that, The terminal includes a display unit and a computing module. Used to invoke the landslide displacement prediction system as described in claim 7 It enables the input of landslide monitoring data, the output of prediction results, and their visualization.