A landslide displacement space-time prediction method, device, system and storage medium

Abstract

The application provides a landslide displacement space-time prediction method, device and system and a storage medium, and relates to the technical field of engineering geology. The landslide displacement space-time prediction method comprises the following steps: acquiring historical landslide multi-field time series monitoring data of each monitoring point in a to-be-measured area, constructing a monitoring point position map based on historical spatial correlation data, and constructing a time series feature matrix according to a historical displacement time sequence, a historical landslide external disturbance factor time sequence and the monitoring point position map; inputting the time series feature matrix and the monitoring point position map into a trained landslide displacement space-time prediction model to obtain a landslide displacement space-time prediction result; and the landslide displacement space-time prediction model is constructed based on an improved GCN network and an improved BiLSTM network. Through the combination of spatial correlation data and external disturbance factors and the use of the deep learning model, the space-time prediction accuracy of landslide displacement can be significantly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of engineering geology, in particular to a landslide displacement space-time prediction method, device, system and storage medium. BACKGROUND

[0002] As a common geological disaster, the internal geological environment and external environmental factors of landslide are extremely complex. Under the action of various external disturbance factors, landslide is prone to instability and damage, which poses a major threat to the life safety and property of surrounding residents. With the advancement of technology, many large landslide areas have built a three-dimensional monitoring and early warning system. The massive landslide displacement time series monitoring data obtained through these systems have become an important basis for studying landslide behavior. Landslide deformation is a key indicator of landslide safety monitoring. Real-time monitoring of the deformation process of landslide can record the whole process information before instability and damage, and extract the precursor characteristics before landslide, thereby providing a reliable basis for landslide disaster warning. Therefore, establishing an efficient landslide prediction model and using massive data sets to accurately predict the deformation trend of landslide not only has important theoretical significance, but also has significant practical application value.

[0003] In recent years, with the rapid development of machine learning theory, nonlinear science and system science, the nonlinear characteristics of landslide displacement time series have been widely recognized. Machine learning models based on nonlinear theory have gradually become a hot spot in landslide displacement prediction research due to their good applicability and generalization ability. In particular, artificial intelligence models such as artificial neural network (ANN), convolutional neural network (CNN), recurrent neural network (RNN), and long short-term memory neural network (LSTM) are widely used. However, these traditional prediction methods mainly focus on the temporal relationship of monitoring data and analyze data from a single monitoring point, which makes it difficult to fully consider the complex spatio-temporal relationship between multiple landslide monitoring points, thereby leading to poor accuracy of the final prediction results. That is, these methods have limitations in judging the overall deformation trend of landslide, and also face challenges in terms of prediction efficiency and time cost of landslide multi-point monitoring time series data. SUMMARY

[0004] The problem solved by the present application is how to improve the accuracy of landslide displacement prediction based on the spatio-temporal relationship of multiple landslide monitoring points, and thereby improve the monitoring and early warning and risk management capabilities of landslide overall deformation.

[0005] To solve the above problems, the present application provides a landslide displacement space-time prediction method and device.

[0006] In a first aspect, the present application provides a landslide displacement space-time prediction method, comprising:

[0007] Obtain historical landslide multi-field time series monitoring data of each monitoring point of the to-be-tested region, the historical landslide multi-field time series monitoring data including historical spatial correlation data, a historical displacement time sequence, and a historical landslide external disturbance factor time sequence;

[0008] Construct a monitoring point bitmap based on the historical spatial correlation data, and construct a time sequence feature matrix according to the historical displacement time sequence, the historical landslide external disturbance factor time sequence, and the monitoring point bitmap;

[0009] Input the time sequence feature matrix and the monitoring point bitmap into a trained landslide displacement space-time prediction model to obtain a landslide displacement space-time prediction result, the landslide displacement space-time prediction model being constructed based on an improved GCN network and an improved BiLSTM network.

[0010] Optionally, the inputting the time sequence feature matrix and the monitoring point bitmap into the trained landslide displacement space-time prediction model to obtain the landslide displacement space-time prediction result includes:

[0011] The time sequence feature matrix and the monitoring point bitmap are processed through the improved GCN network to obtain a spatial time sequence feature matrix;

[0012] The spatial time sequence feature matrix is predicted through the improved BiLSTM network to obtain the landslide displacement space-time prediction result.

[0013] Optionally, the improved GCN network includes an original GCN network and a multi-head attention mechanism, and the processing the time sequence feature matrix and the monitoring point bitmap through the improved GCN network to obtain the spatial time sequence feature matrix includes:

[0014] The time sequence feature matrix is fused and processed through the multi-head attention mechanism to obtain multi-head attention fusion output data;

[0015] The multi-head attention fusion output data and the monitoring point bitmap are reconstructed through the original GCN network to obtain the spatial time sequence feature matrix.

[0016] Optionally, the reconstructing the multi-head attention fusion output data and the monitoring point bitmap through the original GCN network to obtain the spatial time sequence feature matrix includes:

[0017] A weighted adjacency matrix of a spatial monitoring point is calculated based on the monitoring point bitmap;

[0018] The multi-head attention fusion output data is reconstructed based on the weighted adjacency matrix to obtain the spatial time sequence feature matrix.

[0019] Optionally, the improved BiLSTM network comprises a time attention mechanism; and the construction process of the improved BiLSTM network comprises:

[0020] an original BiLSTM network is acquired, the original BiLSTM network comprising a forward LSTM unit and a corresponding backward LSTM unit;

[0021] the time attention mechanism is added to both the forward LSTM unit and the backward LSTM unit, to obtain the improved BiLSTM network, the time attention mechanism being constructed based on a full connection layer and a softmax function.

[0022] Optionally, the prediction of the spatial and temporal feature matrix by the improved BiLSTM network to obtain the landslide displacement spatiotemporal prediction result comprises:

[0023] the spatial and temporal feature matrix is extracted by the original BiLSTM network to obtain a hidden state matrix;

[0024] the hidden state matrix is processed by the time attention mechanism to obtain a temporary hidden state matrix;

[0025] the landslide displacement spatiotemporal prediction result is obtained according to the temporary hidden state matrix.

[0026] Optionally, the processing of the hidden state matrix by the time attention mechanism to obtain a temporary hidden state matrix comprises:

[0027] the hidden state matrix is processed by the full connection layer to obtain time attention weights, and the time attention weights are normalized according to the softmax function to obtain an attention weight vector;

[0028] the temporary hidden state matrix is calculated according to the attention weight vector and the hidden state matrix.

[0029] In a second aspect, the present application provides a landslide displacement spatiotemporal prediction device, comprising:

[0030] an acquisition unit configured to acquire historical landslide multi-field time series monitoring data of each monitoring point in a to-be-measured region, the historical landslide multi-field time series monitoring data comprising historical spatial correlation data, a historical displacement time series and a historical landslide external disturbance factor time series;

[0031] a processing unit configured to construct a monitoring point bitmap based on the historical spatial correlation data, and construct a time series feature matrix according to the historical displacement time series, the historical landslide external disturbance factor time series and the monitoring point bitmap;

[0032] The acquisition unit is further configured to input the time sequence feature matrix and the monitoring point graph into a trained landslide displacement space-time prediction model to obtain a landslide displacement space-time prediction result.

[0033] In a third aspect, the present application provides a landslide displacement space-time prediction system, comprising a memory and a processor.

[0034] The memory is configured to store a computer program.

[0035] The processor is configured to implement the landslide displacement space-time prediction method according to the first aspect when executing the computer program.

[0036] In a fourth aspect, the present application provides a computer readable storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the landslide displacement space-time prediction method according to the first aspect is implemented.

[0037] The landslide displacement space-time prediction method and device have the following advantages:

[0038] First, the historical data of all monitoring points of the to-be-tested landslide are acquired, including historical spatial correlation data, i.e., historical spatial coordinates, historical displacement time series, i.e., for each monitoring point, the displacement change at different time points is recorded, including horizontal and vertical displacement. Historical landslide external disturbance factor time series, such as monthly displacement, monthly average rainfall, monthly average reservoir water level change value, etc. time series data; then, according to the collected historical spatial correlation data, a landslide monitoring point graph is generated. Generally, the monitoring point graph is represented as a graph structure, wherein the edges represent the spatial relationship between the monitoring points, and the nodes represent the monitoring points.

[0039] The historical displacement time series and the historical landslide external disturbance factor time series are combined to extract features that affect the landslide displacement. The dimension of the feature matrix is usually the product of the number of monitoring points and the number of time points, including the relevant change value and the timestamp.

[0040] Then, the time sequence feature matrix and the monitoring point graph are input into a trained landslide displacement space-time prediction model to obtain a landslide displacement space-time prediction result. The improved graph convolution network (GCN) is used to process the spatial relationship and correlation between the monitoring points, which can effectively capture the complex relationship between the spatial features and the local features. In the time sequence processing part of the model, the improved bidirectional long short-term memory network (BiLSTM) is used to learn the dynamic changes of the time sequence from the historical displacement and the external disturbance factor, which can better capture the time sequence features.

[0041] In summary, the application combines spatial correlation data and external disturbance factors, uses the capabilities of deep learning models, and can significantly improve the spatio-temporal prediction accuracy of landslide displacement. This method can also provide real-time prediction of new monitoring data and timely feedback on landslide risks, providing an important basis for emergency management. The improved GCN and BiLSTM networks can adapt to different types of landslide events, have strong adaptability, can handle complex time series data features, and are beneficial for landslide prediction and monitoring in different regions. The combination of various types of data makes the prediction model not only focus on a single factor, but also fully consider the spatial features and correlations between multiple landslide monitoring nodes, effectively improving the prediction efficiency and accuracy of landslide multi-point monitoring data. It can also consider the influence of multiple external disturbances to provide more scientific analysis and decision-making basis. This method can be widely applied to landslide, debris flow, collapse and other geological disaster monitoring and risk management, improving the scientificity and effectiveness of natural disaster prevention and evaluation and prevention. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 A flowchart of a landslide displacement spatio-temporal prediction method according to an embodiment of the application;

[0043] Figure 2 An improved GCN graph convolution network structure diagram according to an embodiment of the application;

[0044] Figure 3 A multi-head attention mechanism structure diagram according to an embodiment of the application;

[0045] Figure 4 An original BiLSTM model structure diagram according to an embodiment of the application;

[0046] Figure 5 A time attention mechanism structure diagram according to an embodiment of the application;

[0047] Figure 6 A structure diagram of a landslide displacement spatio-temporal prediction device according to an embodiment of the application;

[0048] Figure 7 A structure diagram of a landslide displacement spatio-temporal prediction system according to an embodiment of the application. DETAILED DESCRIPTION

[0049] In order to make the above objectives, characteristics and advantages of the present application more obvious and comprehensible, specific embodiments of the present application are described in detail below with reference to the drawings. Although some embodiments of the present application are shown in the drawings, it should be understood that the present application can be implemented in various forms, and should not be interpreted as being limited to the embodiments described herein, but rather, these embodiments are provided to make the present application more thorough and complete. It should be understood that the drawings and embodiments of the present application are only for exemplary purposes, and are not intended to limit the scope of protection of the present application.

[0050] It should be understood that each step described in the method embodiments of the present application can be performed in different orders and / or in parallel. In addition, the method embodiments can include additional steps and / or omit the steps shown. The scope of the present application is not limited in this respect.

[0051] The term "comprising" and variations thereof as used herein are open-ended, that is "including, but not limited to"; the term "based on" is "based, at least in part, on"; the term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments"; the term "optional" means "optional in at least some embodiments". Related definitions are given throughout the description. It should be noted that the concepts "first", "second", etc. mentioned in the present application are only used to distinguish different devices, modules or units, and are not intended to limit the order or interdependence of the functions performed by these devices, modules or units.

[0052] It should be noted that the modification of "one" or "multiple" mentioned in the present application is illustrative rather than limiting, and those skilled in the art should understand that, unless otherwise explicitly indicated in the context, it should be understood as "one or more".

[0053] The names of the messages or information exchanged between the devices in the embodiments of the present application are only for illustrative purposes, and are not intended to limit the scope of the messages or information.

[0054] The present embodiment provides a landslide displacement space-time prediction method, device and system.

[0055] As shown in Figure 1 The landslide displacement space-time prediction method provided by the present embodiment comprises:

[0056] S100, historical landslide multi-field time series monitoring data of each monitoring point in a to-be-measured area is acquired, the historical landslide multi-field time series monitoring data comprising historical space correlation data, a historical displacement time sequence and a historical landslide external disturbance factor time sequence.

[0057] Specifically, historical spatial data mainly includes the spatial coordinates of multiple monitoring points in the area to be measured; and historical displacement time series, which records the displacement changes of each monitoring point within a certain time range, usually plotted with time on the x-axis and displacement on the y-axis. These time series data can reflect the dynamic changes of landslides and help researchers identify the evolution patterns and potential instability trends of landslides.

[0058] This dataset of historical landslide external disturbance factors mainly includes external factors affecting landslides, such as rainfall, temperature changes, seismic activity, and construction disturbances. In some embodiments, the time series of historical landslide external disturbance factors includes monitoring data such as monthly displacement at landslide monitoring points, monthly average rainfall, and monthly average reservoir water level changes. These disturbance factors have a significant impact on the occurrence and development of landslides. By analyzing the historical data of these factors, we can better understand their role in landslide stability.

[0059] By acquiring multiple time-series monitoring data, landslides can be comprehensively analyzed from multiple dimensions, taking into account spatial characteristics, historical displacement, and external disturbance factors, thus providing a deeper understanding of the landslide occurrence mechanism.

[0060] S200, construct a monitoring point map based on the historical spatial correlation data, and construct a time series feature matrix based on the historical displacement time series, the historical landslide external disturbance factor time series, and the monitoring point map.

[0061] Specifically, a monitoring point map is constructed based on historical spatial correlation data, that is, a monitoring point map G representing spatial relationships is constructed based on the spatial coordinates of multiple monitoring points. Then, based on time-related data, namely historical displacement time series and historical landslide external disturbance factor time series, a time-series feature matrix X is established in conjunction with the monitoring point map G. The displacement data and external disturbance factor data of each monitoring point are organized according to the time series.

[0062] By constructing a monitoring point map and a time-series feature matrix, scattered monitoring data can be systematized, facilitating comprehensive spatial and temporal analysis. Furthermore, the time-series feature matrix provides rich input features for subsequent machine learning models, enhancing the model's understanding of landslide occurrence mechanisms and improving prediction accuracy. Simultaneously, the construction of the monitoring point map makes the spatial distribution of landslide monitoring points readily apparent, enabling decision-makers and researchers to intuitively understand potential landslide risk areas. Moreover, by integrating displacement time series and time series of external disturbance factors, it is possible to analyze the relationships between multiple variables, revealing the complex relationship between landslide behavior and external factors.

[0063] Based on the constructed time-series feature matrix and monitoring point map, a scientific basis can be provided for landslide monitoring, early warning and disaster management, helping relevant departments to formulate more reasonable disaster prevention and mitigation measures.

[0064] S300, the time-series feature matrix and the monitoring point location map are input into the trained landslide displacement spatiotemporal prediction model to obtain the landslide displacement spatiotemporal prediction result; the landslide displacement spatiotemporal prediction model is constructed based on the improved GCN network and the improved BiLSTM network.

[0065] Specifically, the constructed time-series feature matrix and monitoring point location map are appropriately formatted to ensure they can be correctly input into the landslide displacement spatiotemporal prediction model. The time-series feature matrix contains historical displacement and external disturbance factor information for each monitoring point, while the monitoring point location map provides the spatial coordinate information of each monitoring point. The time-series feature matrix and monitoring point location map are input into the trained landslide displacement spatiotemporal prediction model to obtain the landslide displacement spatiotemporal prediction results. An improved GCN (Graph Convolutional Network) is used to capture the spatial relationships between monitoring points. GCN can effectively process graph-structured data, extracting the spatial features and interrelationships of monitoring points. By modeling the connectivity of monitoring points, GCN can understand the spatial distribution pattern of the landslide-affected area. An improved BiLSTM (Bidirectional Long Short-Term Memory) network is designed to capture the dynamic characteristics of time-series data. By processing time-series information bidirectionally, BiLSTM can learn the temporal dependencies of the past and future, improving the prediction ability of landslide displacement.

[0066] Furthermore, during the training phase, the model was trained using historical landslide displacement data, temporal feature matrices, and monitoring point location maps to learn the spatiotemporal variation patterns of landslides. In the prediction phase, the temporal feature matrix and monitoring point location maps were input into the trained landslide displacement spatiotemporal prediction model. The model then uses a combination of GCN and BiLSTM to predict future landslide displacements, generating spatiotemporal prediction results for landslide displacement.

[0067] Improved GCN and BiLSTM models that combine spatial and temporal information can provide a more comprehensive understanding of landslide behavior and improve prediction accuracy. This is of great significance for landslide monitoring and early warning. Furthermore, by integrating geospatial information and time-series data, the models can effectively capture the spatiotemporal characteristics of landslides and reveal the real-time dynamic changes in landslide deformation.

[0068] Furthermore, the improved GCN and BiLSTM networks can better handle the effects of noise and missing data, making the model more robust and adaptable to various complex situations in real-world environments. This allows the generated spatiotemporal predictions of landslide displacement to provide crucial decision support for the monitoring, assessment, and management of landslide disasters, helping to formulate more effective prevention and emergency response measures.

[0069] In this embodiment, historical data of all monitoring points of the landslide to be monitored is first acquired, including historical spatial correlation data, i.e., historical spatial coordinates, historical displacement time series (i.e., for each monitoring point, the displacement changes at different time points are recorded, including horizontal and vertical displacements), and historical landslide external disturbance factor time series data, such as monthly displacement, monthly average rainfall, and monthly average reservoir water level changes. Then, based on the collected historical spatial correlation data, a landslide monitoring point location map is generated. Generally, the monitoring point location map is represented as a graph structure, where edges represent the spatial relationships between monitoring points, and nodes represent individual monitoring points.

[0070] By combining historical displacement time series with historical landslide external disturbance factor time series, features influencing landslide displacement are extracted. The dimension of the feature matrix is ​​typically the product of the number of monitoring points and the number of time points, including relevant change values ​​and timestamps.

[0071] The temporal feature matrix and monitoring point location map are then input into the trained landslide displacement spatiotemporal prediction model to obtain the landslide displacement spatiotemporal prediction results. An improved graph convolutional network (GCN) is used to handle the spatial relationships and correlations between monitoring points, effectively capturing the complex relationships between spatial and local features. In the model's time series processing part, an improved bidirectional long short-term memory network (BiLSTM) is used to learn the dynamic changes of the time series from historical displacements and external disturbance factors, enabling better capture of temporal features.

[0072] In summary, this embodiment, by combining spatially relevant data with external disturbance factors and leveraging the capabilities of deep learning models, significantly improves the spatiotemporal prediction accuracy of landslide displacement. Furthermore, this method can perform real-time predictions of new monitoring data, providing timely feedback on landslide risks and offering crucial information for emergency management. The improved GCN and BiLSTM networks are adaptable to different types of landslide events, demonstrating strong adaptability and the ability to handle complex time-series data characteristics, which is beneficial for landslide prediction and monitoring in different regions. Moreover, by combining multiple data types, the prediction model not only focuses on single factors but also fully considers the spatial characteristics and correlations between multiple landslide monitoring nodes, effectively improving the prediction efficiency and accuracy of multi-point landslide monitoring data. It can also comprehensively consider the impact of multiple external disturbances, providing a more scientific basis for analysis and decision-making. This method can be widely applied to the monitoring and risk management of geological disasters such as landslides, debris flows, and collapses, improving the scientific rigor and effectiveness of natural disaster prevention assessment and control.

[0073] Optionally, the step of inputting the time-series feature matrix and the monitoring point location map into the trained landslide displacement spatiotemporal prediction model to obtain the landslide displacement spatiotemporal prediction result includes:

[0074] The spatial temporal feature matrix is ​​obtained by processing the temporal feature matrix and the monitoring point map using the improved GCN network.

[0075] The improved BiLSTM network is used to predict the spatial temporal feature matrix to obtain the spatiotemporal prediction result of the landslide displacement.

[0076] Specifically, the integrated temporal feature matrix and monitoring point location map are used as inputs to the model. These two datasets combine the spatial location and temporal evolution characteristics of landslide monitoring points, providing comprehensive information for the model. The input temporal feature matrix (containing time series of historical displacement and external disturbance factors) is combined with the monitoring point location map (showing the spatial relationships between monitoring points).

[0077] In the improved GCN network, spatial correlations between monitoring points are extracted by performing graph convolution operations on the connections between them. The improved GCN network effectively captures spatial dependencies between landslide monitoring points by aggregating neighborhood information of nodes. After processing, a new spatial-temporal feature matrix is ​​output, which contains the fusion results of spatial relationships and time-series information.

[0078] The improved spatial-temporal feature matrix, processed by the enhanced GCN, is used as input to the improved BiLSTM network. The improved BiLSTM model extracts appropriate temporal features from this matrix, capturing the sequential relationships within the time series through bidirectional learning, thus better understanding the impact of temporal changes on landslide displacement. Finally, the BiLSTM network generates spatiotemporal predictions of landslide displacement, identifying potential landslide risks in advance by predicting displacement changes at future times.

[0079] The improved GCN network fuses spatial information with time-series data, providing a multi-dimensional information analysis method that enhances the model's understanding and capture of landslide behavior. Furthermore, the improved spatial-temporal feature matrix extracted by GCN accurately reflects the relationships between monitoring points and their interaction with landslide displacement, laying the foundation for BiLSTM spatiotemporal prediction. Finally, the temporal dependencies captured by bidirectional LSTM significantly improve the model's prediction accuracy for landslide displacement, especially for dynamic changes with temporal continuity.

[0080] The improved GCN and BiLSTM network design enhances the model's robustness in handling complex and irregular data, adapting it to landslide prediction under diverse geographical and environmental conditions. This integrated prediction method enables real-time monitoring and early warning of landslide displacement, helping relevant departments respond quickly and implement disaster prevention and mitigation measures, thereby improving public safety.

[0081] Optionally, the improved GCN network includes the original GCN network and a multi-head attention mechanism; the step of processing the temporal feature matrix and the monitoring point map through the improved GCN network to obtain the spatial temporal feature matrix includes:

[0082] The temporal feature matrix is ​​fused using the multi-head attention mechanism to obtain multi-head attention fusion output data.

[0083] The spatial temporal feature matrix is ​​obtained by reconstructing the multi-head attention fusion output data and the monitoring point map using the original GCN network.

[0084] Optionally, the step of reconstructing the multi-head attention fusion output data and the monitoring point map using the original GCN network to obtain the spatial-temporal feature matrix includes:

[0085] Calculate the weighted adjacency matrix of the spatial monitoring points based on the monitoring point location map;

[0086] Based on the weighted adjacency matrix, the multi-head attention fusion output data is reconstructed to obtain the spatial temporal feature matrix.

[0087] Specifically, a time-series feature matrix containing historical landslide data and external influencing factors is input into a multi-head attention mechanism. The time-series feature matrix is ​​divided into multiple heads, each capturing a specific information segment. Attention weights for each head are calculated, and the importance of features is obtained through weighted averaging. The results from all heads are concatenated to generate multi-head attention fusion output data, carrying multi-dimensional feature information. This multi-head attention fusion output data, along with a monitoring point location map, is input into the original GCN network. Using the monitoring point location map, a weighted adjacency matrix is ​​calculated. The weighted adjacency matrix represents the spatial relationships between monitoring points, and the weights can be set based on distance, historical relevance, or other factors. The weighted relationships statistically analyze the mutual influence of monitoring points, thus enabling more accurate network construction.

[0088] The calculated weighted adjacency matrix is ​​used to perform graph convolution on the multi-head attention fusion output data. The original GCN combines the features of each monitoring point with the features of its neighboring monitoring points, strengthening the node features through aggregation operations. After graph convolution processing, the generated spatial-temporal feature matrix not only contains temporal information but also comprehensively considers the spatial relationships between monitoring points.

[0089] The multi-head attention mechanism effectively integrates temporal features, making the model more accurate in information extraction by focusing on the importance of different features. The weighted adjacency matrix reasonably reflects the mutual influence and spatial dependence between monitoring points, enhancing the expressive power of the GCN. Compared to a simple adjacency matrix, weighted connections better capture the complexity of the local network. The original GCN network processes the multi-head attention output and adaptively aggregates neighborhood information using the weighted adjacency matrix, making the resulting spatial temporal feature matrix more representative and effective, thus promoting accurate landslide displacement prediction.

[0090] By focusing on and integrating multiple features, the improved GCN network exhibits stronger robustness to uncertainties and noise in the input data, consistently providing accurate predictions in complex environments. Furthermore, the generation of the spatial-temporal feature matrix provides dynamic decision support for landslide monitoring and early warning, improving the timeliness and accuracy of landslide risk assessment. Through this mechanism, the improved GCN network significantly enhances its performance in landslide displacement prediction, better handling complex spatiotemporal data and more effectively supporting disaster early warning and management.

[0091] In some embodiments, such as Figure 2 and Figure 3 As shown, where Figure 2 To improve the schematic diagram of the GCN graph convolutional network structure, Figure 3 This is a schematic diagram of a multi-head attention mechanism; based on the spatial coordinates of multiple monitoring points, a monitoring point map G representing the spatial relationships is constructed, expressed as follows:

[0092] G = (V, E, W);

[0093] Where V∈{v1,v2…v N Let W be a set of N monitoring points, each connected by an adjacent edge E. The adjacent edge E can be obtained from the spatial coordinates of the monitoring point, where W∈[0,1]. N×N Let W represent a weighted adjacency matrix, where the weights w are... ij This indicates the spatial correlation between monitoring points.

[0094] The weight w is calculated using the Gaussian similarity function. ij The expression is as follows:

[0095]

[0096] Among them, dist ij For monitoring point v i With v j The Euclidean distance, where σ is the standard deviation of the distance;

[0097] The expression for the weighted adjacency matrix W is as follows:

[0098]

[0099] Based on time-related data, namely historical displacement time series and historical landslide external disturbance factor time series, and combined with the monitoring point location map G, a time-series feature matrix X is established. The expression for the time-series feature matrix X is as follows:

[0100] X∈R N×S×M ;

[0101] Where N is the number of monitoring points, S is the number of monitoring features, M is the length of the historical time series data for each monitoring point, and X is... t ∈R N×S×t Let be the historical time series feature matrix of each monitoring point at time t.

[0102] Then, a graph convolutional neural network (GCN) (original GCN network) is used, and a multi-head attention mechanism (GAT) is introduced to construct an improved GCN network (spatial module) to capture the spatial correlation of landslide monitoring points and reconstruct the temporal feature matrix of landslide spatial correlation.

[0103] The implementation process of the multi-head attention mechanism GAT is as follows:

[0104] X i Let X be the time series feature matrix, i.e., the historical time series feature matrix X = {X1, X2, ..., X}. N The characteristic matrix X of a certain central node in} i ∈R S×M X j To find the characteristic matrix of the nodes adjacent to the central node, we can define the neighboring node matrix X. j The query matrix Q is obtained after linear transformation. ij , central node matrix X i After linear deformation bond matrix K i Value matrix V i The calculation process is as follows:

[0105]

[0106] in, They represent Q respectively ij K i V i The weight matrix, N i This represents the set of nodes adjacent to node i, and S represents the dimension of the weight matrix, i.e. the number of monitored features.

[0107] Using the query matrix Q ij Key matrix K iPerform dot product scaling, and then calculate the weight coefficients w using softmax normalization. ij Based on the weight coefficients, the value matrix V i After performing a weighted summation, the self-attention weight output of node j with respect to node i is finally obtained, and the calculation formula is as follows;

[0108]

[0109] in, Represents the bond matrix K i In terms of dimensions, the softmax function can yield self-attention weights that sum to 1;

[0110] The multi-head attention mechanism GAT includes multiple self-attention heads. These heads learn the correlations between different representation subspaces, and the final self-attention value is obtained by concatenating and linearly transforming the multiple self-attention values. Taking an n-head self-attention mechanism as an example, the feature vector X of the input center node i is... i Divide into X i / / There are n sub-feature sequences. Each sub-feature sequence calculates its own self-attention. Finally, they are concatenated to obtain the new feature matrix X after the center node i has been learned by the multi-head attention mechanism. i The splicing process is as follows:

[0111] Multi-Head(X i = concat(z1, z2, ... z) n );

[0112]

[0113] Where n represents the total number of attention points in the multi-head attention mechanism. This indicates that the k-th attention head corresponds to Q. ij K i V i The weight matrix, N i This indicates traversing the set of nodes adjacent to node i. The `concat` function is used to concatenate the output values ​​of each attention point. k This represents the output value of the k-th attention head, i.e., the sub-feature sequence learned by the attention mechanism;

[0114] Repeat the above steps N times to output the historical time-series feature matrix of all nodes on the monitoring point map G after learning by the multi-head attention mechanism. The multi-head attention fusion output data is X′={X1′,X2′,…,X N ′},X′∈R N×S×M .

[0115] In the original Graph Convolutional Network (GCN) model, the composition of the input layer is crucial for landslide displacement prediction. The input layer mainly consists of a weighted adjacency matrix W and multi-head attention fused output data. Together, these provide the model with the necessary spatial relationships and temporal information.

[0116] First, the weighted adjacency matrix W is used to describe the spatial relationships between monitoring points, reflecting their mutual influence in the landslide prediction process. The corresponding weights represent the degree of correlation between different monitoring points, and are set based on factors such as geographical location and historical data.

[0117] Secondly, the multi-head attention fusion output data includes feature data collected from each monitoring point over a historical time period. This data (such as displacement, rainfall, etc.) is organized into a matrix, with each row representing the temporal characteristics of a monitoring point and each column corresponding to the feature value at a specific time point. This matrix provides the model with important information in the time dimension, enabling it to analyze the changing trends of the monitoring points over time.

[0118] In the structural design, the input layer is connected to the first hidden convolutional layer by fusing the output data through a weighted adjacency matrix W and multi-head attention. Each hidden convolutional layer is combined with a corresponding ReLU (Rectified Linear Unit) activation unit to introduce nonlinearity, thereby enhancing the model's expressiveness and adaptability.

[0119] Finally, the last ReLU activation unit is connected to the output layer of the GCN. The output layer is responsible for generating a temporal feature matrix containing spatially relevant information about landslides. This output matrix not only comprehensively reflects the temporal characteristics of the monitoring points but also highlights their spatial correlations, thus providing a solid data foundation for landslide event prediction.

[0120] In summary, the GCN model, through its carefully designed input layer and hierarchical structure, effectively captures the spatiotemporal characteristics of landslide displacement, providing important support for disaster prediction and management.

[0121] The propagation process between adjacent convolutional layers in the original GCN network is performed according to the following formula:

[0122]

[0123] D ii =∑ j W ij ;

[0124] O (0) =X′;

[0125] Among them, O (l+1) G represents the output of the l-th convolutional layer, σ is the activation function, and G... (l)For the characteristic matrix O (l) The weight matrix undergoes linear transformation, where l represents the number of convolutional layers. This represents the adjacency matrix after symmetric normalization of the weighted adjacency matrix W. The degree matrix D is calculated from the weighted adjacency matrix W. (0) This represents the input to the first hidden convolutional layer, i.e., the multi-head attention fusion output data.

[0126] Optionally, the improved BiLSTM network includes a temporal attention mechanism; the improved BiLSTM network construction process includes:

[0127] Obtain the original BiLSTM network, which includes forward LSTM units and their corresponding backward LSTM units;

[0128] The improved BiLSTM network is obtained by adding the temporal attention mechanism to both the forward LSTM unit and the backward LSTM unit. The temporal attention mechanism is constructed based on a fully connected layer and a softmax function.

[0129] Optionally, the step of predicting the spatial-temporal feature matrix using the improved BiLSTM network to obtain the landslide displacement spatiotemporal prediction result includes:

[0130] The hidden state matrix is ​​obtained by extracting features from the spatial temporal feature matrix using the original BiLSTM network.

[0131] The hidden state matrix is ​​processed by the aforementioned temporal attention mechanism to obtain a temporary hidden state matrix;

[0132] The spatiotemporal prediction result of the landslide displacement is obtained based on the temporary hidden state matrix.

[0133] Optionally, the step of processing the hidden state matrix through the temporal attention mechanism to obtain a temporary hidden state matrix includes:

[0134] The fully connected layer processes the hidden state matrix to obtain temporal attention weights, and then normalizes the temporal attention weights according to the softmax function to obtain an attention weight vector.

[0135] The temporary hidden state matrix is ​​calculated based on the attention weight vector and the hidden state matrix.

[0136] Specifically, such as Figure 4As shown, the original Bidirectional Long Short-Term Memory (BiLSTM) network model consists of two independent sets of forward and backward LSTM units, designed to extract forward and reverse sequence features from temporal information. Through these units, the model can fully capture the temporal dynamics of landslide displacement.

[0137] Specifically, the forward LSTM unit processes the forward-order input data, while the backward LSTM unit processes the reverse-order input data. These two sets of LSTM units operate independently within the model, each generating its own hidden state. Subsequently, the model uses an output layer to weightedly sum the hidden states calculated by the forward and backward LSTM units, forming a fused and concatenated hidden state at each time step. These concatenated hidden states are then input into a fully connected layer to generate the final landslide displacement prediction. This process effectively improves data utilization and prediction accuracy.

[0138] In some embodiments, the original Bidirectional Long Short-Term Memory (BiLSTM) network model comprises a forward LSTM unit and a corresponding backward LSTM unit. Each LSTM unit internally includes an input gate, a forget gate, an output gate, and a cell state. These components work together to determine the flow of information within the network and the retention of its importance. Specifically, the landslide spatially correlated temporal feature matrix is ​​passed as input to each LSTM unit, ultimately outputting the hidden state matrix for each time step.

[0139] The main calculation formulas are as follows:

[0140]

[0141] Where, x t The input at time step t represents the value vector of the spatially correlated temporal feature matrix of the landslide at time step t, h. t-1 It is the hidden state matrix at time step t-1, c t-1 It is the cell state at time step t-1, i t f t , c t o t h t W represents the input gate, forget gate, cell state update, cell state, output gate, and hidden state at time step t, respectively. xi W xf W xo W xc With W hi W hf W ho Whc These are the three gates mentioned above, as well as the input and hidden matrices of the cell state, W. ci W cf W co For the weight matrix involving long-term memory in the aforementioned three gates, b i b f b o b c These are the corresponding bias vectors, where σ is the sigmoid function and tanh is the hyperbolic tangent function.

[0142] In some preferred embodiments, such as Figure 5 As shown, the implementation process of the time attention mechanism is as follows:

[0143] By adding a fully connected layer and a softmax function as an attention layer to each original LSTM unit to implement a temporal attention mechanism, the weights of hidden states can be selected throughout the time series memory cycle. This allows for dynamic adjustment of the importance of the input sequence at different time steps, thereby improving the model's accuracy.

[0144] The input data for the time attention mechanism is the hidden state matrix H = [h1, h2, ..., h] corresponding to all T time steps output by the LSTM unit. T ] T×S h T Let T be the hidden state corresponding to time step T, and S be the size of each hidden state, i.e., the number of monitored features. The hidden state matrix is ​​processed through a fully connected neural network to obtain the temporal attention weights. These weights are then normalized using a softmax function to obtain the attention weight vector α. Multiplying the attention weight vector α by the hidden state matrix yields the hidden state matrix H after the temporal attention mechanism has assigned its weights. at The weighting formula and final output are as follows:

[0145] α = softmax(A) t (H))=[α1,α2,…,α T ] 1×T ;

[0146]

[0147] Among them, A t Indicates a fully connected neural network (fully connected layer), α T Here are the attention weights corresponding to time step T. This represents the hidden state (temporary hidden state matrix) of the LSTM unit at time T.

[0148] It should be noted that there is no limit to the number of forward LSTM units; the specific number can be determined according to the actual situation. The number of backward LSTM units corresponds to the number of forward LSTM units.

[0149] In some embodiments, during the training process of the landslide displacement spatiotemporal prediction model, the landslide displacement time series prediction value is often evaluated based on three error evaluation indicators: mean absolute error (MAE), mean absolute proportional error (MASE), and root mean square error (RMSE) to obtain the prediction evaluation result.

[0150] Specifically, compare the actual landslide displacement monitoring sequence (spatial-temporal monitoring results of landslide displacement) Y=[y1,y2,…,y T ] 1×T The displacement time-series prediction results of the model (spatial-temporal prediction results of landslide displacement) are evaluated using three indicators: root mean square error (MAE), mean absolute error (MASE), and mean relative absolute error (RMSE). The formulas for calculating the root mean square error (MAE), mean absolute error (MASE), and mean relative absolute error (RMSE) are as follows:

[0151]

[0152] Where T is the length of the time prediction sequence, y i and The monitored and predicted values ​​of the landslide displacement at time i are respectively.

[0153] It should be noted that all data used in model training and practical application are historical data. For example, the model training process can use historical data from 5 years ago, while in practical application, historical data from 1 year ago can be used to predict the spatiotemporal displacement of future landslides.

[0154] like Figure 6 As shown in the figure, an embodiment of the present invention provides a landslide displacement spatiotemporal prediction device, comprising:

[0155] The acquisition unit is used to acquire historical landslide multi-time series monitoring data of each monitoring point in the area to be measured. The historical landslide multi-time series monitoring data includes historical spatial correlation data, historical displacement time series and historical landslide external disturbance factor time series.

[0156] The processing unit is used to construct a monitoring point map based on the historical spatial correlation data, and to construct a time series feature matrix based on the historical displacement time series, the historical landslide external disturbance factor time series, and the monitoring point map;

[0157] The acquisition unit is also used to input the time-series feature matrix and the monitoring point map into the trained landslide displacement spatiotemporal prediction model to obtain the landslide displacement spatiotemporal prediction result; the landslide displacement spatiotemporal prediction model is constructed based on the improved GCN network and the improved BiLSTM network.

[0158] The landslide displacement spatiotemporal prediction device of this embodiment is used to implement the landslide displacement spatiotemporal prediction method as described above. Its advantages over the prior art are the same as the advantages of the landslide displacement spatiotemporal prediction method over the prior art, and will not be repeated here.

[0159] like Figure 7 As shown, an embodiment of the present invention provides a landslide displacement spatiotemporal prediction system 700, including a memory 710 and a processor 720; the memory 710 is used to store a computer program; the processor 720 is used to implement the landslide displacement spatiotemporal prediction method as described above when the computer program is executed.

[0160] Alternatively, a landslide displacement spatiotemporal prediction system 700 includes a memory 710 and a processor 720 coupled to the memory 710; the memory 710 is configured to store a computer program; the processor 720 is configured to perform the following operations when the computer program is executed:

[0161] Acquire historical landslide multi-field time-series monitoring data of each monitoring point in the area to be tested. The historical landslide multi-field time-series monitoring data includes historical spatial correlation data, historical displacement time series and historical landslide external disturbance factor time series.

[0162] A monitoring point map is constructed based on the historical spatial correlation data, and a time series feature matrix is ​​constructed based on the historical displacement time series, the historical landslide external disturbance factor time series, and the monitoring point map;

[0163] The time-series feature matrix and the monitoring point location map are input into the trained landslide displacement spatiotemporal prediction model to obtain the landslide displacement spatiotemporal prediction result; the landslide displacement spatiotemporal prediction model is constructed based on the improved GCN network and the improved BiLSTM network.

[0164] This invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the landslide displacement spatiotemporal prediction method as described above.

[0165] Alternatively, a non-volatile computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the following operations:

[0166] Acquire historical landslide multi-field time-series monitoring data of each monitoring point in the area to be tested. The historical landslide multi-field time-series monitoring data includes historical spatial correlation data, historical displacement time series and historical landslide external disturbance factor time series.

[0167] A monitoring point map is constructed based on the historical spatial correlation data, and a time series feature matrix is ​​constructed based on the historical displacement time series, the historical landslide external disturbance factor time series, and the monitoring point map;

[0168] The time-series feature matrix and the monitoring point location map are input into the trained landslide displacement spatiotemporal prediction model to obtain the landslide displacement spatiotemporal prediction result; the landslide displacement spatiotemporal prediction model is constructed based on the improved GCN network and the improved BiLSTM network.

[0169] The landslide displacement spatiotemporal prediction system 700, which can serve as a server or client of the present invention, is described below as an example of a hardware device applicable to various aspects of the present invention. The landslide displacement spatiotemporal prediction system 700 is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The landslide displacement spatiotemporal prediction system 700 can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0170] The landslide displacement spatiotemporal prediction system 700 includes a computing unit that can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) or a computer program loaded from a storage unit into random access memory (RAM). The RAM can also store various programs and data required for device operation. The computing unit, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0171] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc. In this application, the units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of the present invention according to actual needs. Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units can be implemented in hardware or as software functional units.

[0172] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A method for spatiotemporal prediction of landslide displacement, characterized in that, include: Acquire historical landslide multi-field time-series monitoring data of each monitoring point in the area to be tested. The historical landslide multi-field time-series monitoring data includes historical spatial correlation data, historical displacement time series and historical landslide external disturbance factor time series. A monitoring point map is constructed based on the historical spatial correlation data, and a time series feature matrix is ​​constructed based on the historical displacement time series, the historical landslide external disturbance factor time series, and the monitoring point map; The time-series feature matrix and the monitoring point location map are input into the trained landslide displacement spatiotemporal prediction model to obtain the landslide displacement spatiotemporal prediction result; The landslide displacement spatiotemporal prediction model is constructed based on an improved GCN network and an improved BiLSTM network. The improved GCN network includes the original GCN network and a multi-head attention mechanism. The step of inputting the temporal feature matrix and the monitoring point location map into the trained landslide displacement spatiotemporal prediction model to obtain the landslide displacement spatiotemporal prediction result includes: The spatial temporal feature matrix is ​​obtained by processing the temporal feature matrix and the monitoring point map using the improved GCN network. The improved BiLSTM network is used to predict the spatial temporal feature matrix to obtain the spatiotemporal prediction result of the landslide displacement. The improved BiLSTM network includes a temporal attention mechanism to obtain the original BiLSTM network, which includes a forward LSTM unit and a corresponding backward LSTM unit. The improved BiLSTM network is obtained by adding the temporal attention mechanism to both the forward LSTM unit and the backward LSTM unit. The temporal attention mechanism is constructed based on a fully connected layer and a softmax function.

2. The landslide displacement spatiotemporal prediction method according to claim 1, characterized in that, The improved GCN network includes the original GCN network and a multi-head attention mechanism; The process of processing the temporal feature matrix and the monitoring point map through the improved GCN network to obtain the spatial temporal feature matrix includes: The temporal feature matrix is ​​fused using the multi-head attention mechanism to obtain multi-head attention fusion output data. The spatial temporal feature matrix is ​​obtained by reconstructing the multi-head attention fusion output data and the monitoring point map using the original GCN network.

3. The landslide displacement spatiotemporal prediction method according to claim 2, characterized in that, The process of reconstructing the multi-head attention fusion output data and the monitoring point map using the original GCN network to obtain the spatial-temporal feature matrix includes: Calculate the weighted adjacency matrix of the spatial monitoring points based on the monitoring point location map; Based on the weighted adjacency matrix, the multi-head attention fusion output data is reconstructed to obtain the spatial temporal feature matrix.

4. The landslide displacement spatiotemporal prediction method according to claim 1, characterized in that, The step of predicting the spatial-temporal feature matrix using the improved BiLSTM network to obtain the spatiotemporal prediction result of the landslide displacement includes: The hidden state matrix is ​​obtained by extracting features from the spatial temporal feature matrix using the original BiLSTM network. The hidden state matrix is ​​processed by the aforementioned temporal attention mechanism to obtain a temporary hidden state matrix; The spatiotemporal prediction result of the landslide displacement is obtained based on the temporary hidden state matrix.

5. The landslide displacement spatiotemporal prediction method according to claim 4, characterized in that, The step of processing the hidden state matrix through the time attention mechanism to obtain a temporary hidden state matrix includes: The fully connected layer processes the hidden state matrix to obtain temporal attention weights, and then normalizes the temporal attention weights according to the softmax function to obtain an attention weight vector. The temporary hidden state matrix is ​​calculated based on the attention weight vector and the hidden state matrix.

6. A landslide displacement spatiotemporal prediction device, characterized in that, include: The acquisition unit is used to acquire historical landslide multi-time series monitoring data of each monitoring point in the area to be measured. The historical landslide multi-time series monitoring data includes historical spatial correlation data, historical displacement time series and historical landslide external disturbance factor time series. The processing unit is used to construct a monitoring point map based on the historical spatial correlation data, and to construct a time series feature matrix based on the historical displacement time series, the historical landslide external disturbance factor time series, and the monitoring point map; The acquisition unit is also used to input the time series feature matrix and the monitoring point map into the trained landslide displacement spatiotemporal prediction model to obtain the landslide displacement spatiotemporal prediction result; The landslide displacement spatiotemporal prediction model is constructed based on an improved GCN network and an improved BiLSTM network. The improved GCN network includes the original GCN network and a multi-head attention mechanism. The step of inputting the temporal feature matrix and the monitoring point map into the trained landslide displacement spatiotemporal prediction model to obtain the landslide displacement spatiotemporal prediction result includes: processing the temporal feature matrix and the monitoring point map through the improved GCN network to obtain a spatial temporal feature matrix; and predicting the spatial temporal feature matrix through the improved BiLSTM network to obtain the landslide displacement spatiotemporal prediction result. The improved BiLSTM network includes a temporal attention mechanism. The original BiLSTM network is obtained, comprising forward LSTM units and corresponding backward LSTM units. The temporal attention mechanism is added to both the forward and backward LSTM units to obtain the improved BiLSTM network, which is constructed based on fully connected layers and a softmax function.

7. A landslide displacement spatiotemporal prediction system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the landslide displacement spatiotemporal prediction method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the landslide displacement spatiotemporal prediction method as described in any one of claims 1 to 5.

Images