Slope geological disaster multi-mode early warning method and system

Abstract

The invention discloses a slope geological disaster multi-mode early warning method and system, and relates to the technical field of slope monitoring, and the method comprises the steps: laying a distributed sensor network, and collecting slope multi-source monitoring data; the multi-source monitoring data comprises displacement data, stress data and vibration frequency data; feature extraction is performed on the multi-source monitoring data by using a graph neural network, and the feature extraction comprises capturing spatial correlation among sensor nodes and identifying an abnormal mode, identifying a potential instability area of the slope according to the abnormal mode, and outputting features of the potential instability area; and generating a slope stability risk grade assessment report based on the characteristics of the potential instability region and a disaster evolution graph constructed by combining historical disaster data. According to the invention, the comprehensive monitoring of the slope from the outside to the inside and from the static state to the dynamic state can be realized, the abnormal mode and the potential instability area can be accurately identified, and the accurate assessment and timely early warning of the slope risk can be realized based on the historical data.

Description

Technical Field

[0001] The present invention relates to the technical field of slope monitoring, and specifically to a multi-modal early warning method and system for slope geological disasters. Background Art

[0002] Slope stability monitoring and risk assessment are important research topics in the field of geotechnical engineering, and are of crucial significance for preventing and mitigating geological disasters such as landslides. Traditional slope monitoring methods mainly rely on manual inspections and fixed-point monitoring based on single sensors, lacking intelligent analysis capabilities. The manual inspection method is limited by human, material resources and observation conditions, and it is difficult to achieve real-time and comprehensive understanding of the slope state; although the fixed-point monitoring based on single sensors can obtain high-precision data, the monitoring range is limited, it is difficult to reflect the overall deformation and failure characteristics of the slope, and there is no effective fusion mechanism between different types of sensor data, making it difficult to form a unified understanding of the slope state.

[0003] In recent years, with the rapid development of artificial intelligence, especially machine learning technology, slope monitoring technology has gradually developed towards automation, networking and intelligence. The application of distributed sensor networks, combined with advanced machine learning algorithms, makes it possible to conduct large-scale, multi-parameter real-time monitoring and intelligent analysis of slopes. However, how to effectively fuse multi-source heterogeneous monitoring data through advanced machine learning methods such as graph neural networks, extract key feature information from it, and accurately identify abnormal patterns and potential instability regions of slopes remains a huge challenge in the field of slope monitoring.

[0004] Existing slope monitoring data analysis methods mainly include traditional mathematical statistics methods, numerical simulation methods and machine learning methods. Mathematical statistics methods establish empirical models or semi-empirical models through statistical analysis of monitoring data to predict the deformation trend of slopes, but such methods usually assume that the data follows a certain specific distribution and it is difficult to consider the complex non-linear process of slope instability. Numerical simulation methods establish a mechanical model of the slope and simulate the deformation and failure process of the slope under different working conditions, and can quantitatively evaluate the slope stability, but such methods have high requirements for the accuracy of model parameters and large computational workload, making it difficult to achieve real-time early warning.

[0005] Traditional machine learning methods, such as support vector machine SVM, traditional neural networks, etc., have been applied to slope stability prediction, but these methods usually regard monitoring data as independent samples, ignoring the spatial correlation between sensors, and it is difficult to effectively process multi-source heterogeneous data. In recent years, the graph neural network GNN based on deep learning, as an advanced machine learning model capable of processing graph-structured data, has shown great potential in dealing with spatial relationship modeling and multi-source data fusion. This technology can construct decision trees and support ensemble learning methods, significantly improving the prediction performance.

[0006] However, the application of advanced machine learning technologies such as graph neural networks in the field of slope monitoring is still in its infancy. Most of the existing research only focuses on a single type of monitoring data, or simply concatenates different types of data, failing to fully utilize the complementary information and spatial correlation between multi-source data. In addition, most of the existing methods only focus on the classification or prediction of slope states, lacking in-depth analysis and modeling of the slope instability evolution process, making it difficult to accurately assess slope risks and issue timely warnings through intelligent algorithms. Summary of the Invention

[0007] In view of the above problems, the present invention is proposed.

[0008] Therefore, the present invention provides a multi-modal early warning method for slope geological disasters, which can solve the problems mentioned in the background art.

[0009] To solve the above technical problems, the present invention provides the following technical solution: A multi-modal early warning method for slope geological disasters, comprising: deploying a distributed sensor network to collect multi-source monitoring data of the slope; the multi-source monitoring data includes displacement data, stress data, and vibration frequency data; using a graph neural network to perform feature extraction on the multi-source monitoring data, the feature extraction includes capturing the spatial correlation between sensor nodes and identifying abnormal patterns, identifying potential unstable areas of the slope according to the abnormal patterns, and outputting the features of the potential unstable areas; generating a slope stability risk level assessment report based on the features of the potential unstable areas and the disaster evolution map constructed by combining historical disaster data.

[0010] As a preferred embodiment of the multi-modal early warning method for slope geological disasters according to the present invention, the method includes: extracting features from the multi-source monitoring data using a graph neural network, where the feature extraction includes capturing the spatial correlation between sensor nodes and identifying abnormal patterns, identifying potential unstable areas of the slope based on the abnormal patterns, and outputting the features of the potential unstable areas, which includes the following steps: constructing a slope monitoring graph model, taking each node where a sensor in the distributed sensor network is located as a graph node, constructing edges based on the spatial position relationship of the graph nodes to form a graph structure, and inputting the graph structure into the graph neural network; through the node relationship modeling module of the graph neural network, capturing the spatial correlation between the graph nodes based on the graph structure; through the abnormal pattern classifier of the graph neural network, identifying the abnormal patterns of the graph nodes based on the spatial correlation and the multi-source monitoring data; identifying the potential unstable areas of the slope according to the spatial positions of the abnormal nodes output by the abnormal pattern classifier and the categories of the abnormal patterns, and outputting the features of the potential unstable areas, specifically identifying the area covered by abnormal nodes with adjacent spatial positions and the same category of the abnormal pattern as the potential unstable area; where the features of the potential unstable area include the spatial positions of the abnormal nodes, the categories of the abnormal patterns, and the multi-source monitoring data corresponding to the abnormal nodes.

[0011] As a preferred embodiment of the multi-modal early warning method for slope geological disasters according to the present invention, the abnormal pattern classifier uses a multi-layer perceptron to extract the correlation features between the spatial correlation features and each dimension data in the multi-source monitoring data, uses a one-dimensional convolutional neural network to extract the time series features of the multi-source monitoring data, and fuses the correlation features and the time series features for abnormal pattern classification to output the category of the abnormal pattern of the graph node; the categories of the abnormal patterns include abrupt abnormal, gradual abnormal, and normal.

[0012] As a preferred solution of the multi-modal early warning method for slope geological disasters described in the present invention, wherein: the determination rules for the abnormal patterns include: performing short-time Fourier transform on the displacement data, stress data, and vibration frequency data of each of the graph nodes to obtain the time-frequency domain characteristics of each of the graph nodes; if there is at least one frequency band in the time-frequency domain characteristics of the graph node whose energy exceeds K times the standard deviation of the historical energy of this frequency band, and the central frequency of this frequency band matches any one of the frequencies in the set of slope instability characteristic frequencies, then it is determined that the graph node is a candidate node for abrupt abnormal type; if within the most recent M sampling periods, the absolute value of the first-order difference sequence of at least one of the displacement data, stress data, or vibration frequency data of the graph node is greater than L times the standard deviation of its historical first-order difference sequence, and the sign of the first-order difference sequence is the same as the sign of the theoretical change trend of this item of data calculated based on the slope instability model at the corresponding position and corresponding time, then it is determined that the graph node is a candidate node for gradual abnormal type; if the graph node simultaneously meets the determination conditions for both the candidate node for abrupt abnormal type and the candidate node for gradual abnormal type, then it is determined that the graph node is a pending abnormal node.

[0013] As a preferred solution of the multi-modal early warning method for slope geological disasters described in the present invention, wherein: if the graph node is a pending abnormal node, then perform a secondary analysis and judgment: calculate the similarity matrix of the change trends of the displacement data, stress data, and vibration frequency data of the pending abnormal node and its adjacent nodes in the past N sampling periods; where N sampling periods should be greater than M sampling periods; if there is at least one element in the similarity matrix that is greater than the preset similarity threshold, and at least one of the two nodes corresponding to this element shows an accelerating change trend in the displacement data, stress data, or vibration frequency data within the N sampling periods, then it is determined that the pending abnormal node is of the gradual abnormal type; otherwise, it is determined that the pending abnormal node is of the abrupt abnormal type; if the graph node is not any one of the candidate node for abrupt abnormal type, the candidate node for gradual abnormal type, and the pending abnormal node, then it is determined that the graph node is normal.

[0014] As a preferred solution of the multi-modal early warning method for slope geological disasters described in the present invention, wherein: the risk levels of the slope include high risk level, medium risk level, and low risk level; based on the characteristics of the potential instability area, combined with the historical disaster data to construct a disaster evolution map, and generate a slope stability risk level assessment report, including the following steps: construct a disaster evolution map; based on the disaster evolution map and the characteristics of the potential instability area, construct a risk level assessment model, and output the risk level of the potential instability area through the risk level assessment model; based on the risk level of the slope output by the risk level assessment model, generate the slope stability risk level assessment report.

[0015] As a preferred solution of the multi-modal early warning method for slope geological disasters described in the present invention, wherein: the construction of the disaster evolution map includes the following steps: extracting the characteristics of the potential instability areas at different times before each disaster event in the historical disaster data and the disaster level of each disaster event, and constructing a historical disaster feature database; taking the characteristics of the potential instability areas at different times before each historical disaster event as the disaster map nodes of the disaster evolution map, connecting the disaster map nodes in chronological order to form multiple historical disaster evolution sub-graphs, and taking the disaster level of the historical disaster event as the label of the historical disaster evolution sub-graph; fusing the multiple historical disaster evolution sub-graphs into a disaster evolution map.

[0016] To further solve the above technical problems, the present invention provides the following technical solution: a multi-modal early warning system for slope geological disasters, including: a data acquisition module for deploying a distributed sensor network to collect multi-source monitoring data of the slope; a feature extraction module for using a graph neural network to extract features from the multi-source monitoring data, the feature extraction including capturing the spatial correlation between sensor nodes and identifying abnormal patterns, identifying potential instability areas of the slope according to the abnormal patterns, and outputting the characteristics of the potential instability areas; a risk assessment module for analyzing the characteristics of the potential instability areas in combination with historical disaster data to generate a slope stability risk level assessment report.

[0017] A computer device includes a memory and a processor, the memory stores a computer program, and is characterized in that when the processor executes the computer program, the steps of the multi-modal early warning method for slope geological disasters described above are implemented.

[0018] A computer-readable storage medium stores a computer program, and is characterized in that when the computer program is executed by a processor, the steps of the multi-modal early warning method for slope geological disasters described above are implemented.

[0019] Advantages of the present invention: The present invention realizes the synchronous acquisition of multi-source monitoring data by constructing a distributed sensor network, uses graph neural networks to extract the spatial correlation between sensor nodes and identify abnormal patterns, and constructs a disaster evolution map based on historical disaster data for risk assessment, solving the problems in the prior art such as the difficulty in effectively fusing multi-source heterogeneous data, ignoring spatial correlation, low accuracy in abnormal pattern recognition, and lack of spatio-temporal evolution information in risk assessment. Compared with the prior art, the focus of the present invention lies in constructing a multi-dimensional data acquisition system that can comprehensively reflect the slope state; effectively capturing and quantifying the spatial correlation between sensor nodes through a graph attention network; proposing an abnormal pattern recognition method that combines time-frequency domain analysis and instability pattern matching, and designing a secondary analysis and judgment mechanism to improve the recognition accuracy; constructing a disaster evolution map, converting historical disaster data into graph-structured data, and using a deep learning model to realize the risk level assessment based on historical evolution patterns. The present invention realizes the comprehensive perception, accurate diagnosis, and risk early warning of the slope state, providing more reliable technical support for slope safety management. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following described drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0021] Figure 1 It is a schematic diagram of the overall process of a multi-modal early warning method for slope geological disasters proposed by the present invention;

[0022] Figure 2 It is a schematic diagram of the overall structure of a multi-modal early warning system for slope geological disasters proposed by the present invention;

[0023] Figure 3 It is a diagram of a computer device in a multi-modal early warning method for slope geological disasters proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0024] In order to make the above objects, features, and advantages of the present invention more obvious and understandable, the following will describe the specific embodiments of the present invention in detail with reference to the accompanying drawings of the specification. Obviously, the described embodiments are some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0025] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Persons skilled in the art may make similar generalizations without departing from the spirit of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0026] Example 1, referring to Figure 1 and Figure 2 , which is an embodiment of the present invention, provides a multi-modal early warning method for slope geological disasters.

[0027] Figure 1 Fig. shows a schematic diagram of the overall process of a multi-modal early warning method for slope geological disasters, including the following steps:

[0028] S100: Deploy a distributed sensor network to collect multi-source monitoring data of the slope;

[0029] S200: Use a graph neural network to extract features from the multi-source monitoring data. Feature extraction includes capturing the spatial correlation between sensor nodes and identifying abnormal patterns, identifying potential unstable areas of the slope according to the abnormal patterns, and outputting the features of the potential unstable areas;

[0030] S300: Based on the features of the potential unstable areas, combine with the disaster evolution map constructed from historical disaster data to generate a slope stability risk level assessment report.

[0031] Next, this embodiment elaborates on steps S100 to S300 one by one in detail:

[0032] S100: Deploy a distributed sensor network to collect multi-source monitoring data of the slope.

[0033] Specifically, the multi-source monitoring data includes displacement data, stress data, and vibration frequency data.

[0034] S110: Select a first type of sensor for collecting displacement data, a second type of sensor for collecting stress data, and a third type of sensor for collecting vibration frequency data. Deploy the first type of sensor, the second type of sensor, and the third type of sensor on the slope to be monitored to construct a distributed sensor network.

[0035] Specifically, the first type of sensor is at least one of a GNSS receiver, a total station, or an InSAR radar; the second type of sensor is at least one of a fiber Bragg grating strain gauge, an earth pressure cell, a resistance strain gauge, or a differential pressure strain gauge; the third type of sensor is at least one of a MEMS vibration sensor, an accelerometer, a velocity sensor, or a piezoelectric ceramic sensor. The first type of sensor, the second type of sensor, and the third type of sensor are connected by wired communication or wireless communication to form a distributed sensor network.

[0036] Among them, the GNSS receiver uses the satellite navigation system to obtain high-precision three-dimensional displacement information; the total station determines the displacement of the monitoring point by measuring angles and distances; the InSAR radar uses synthetic aperture radar interferometry to obtain large-scale surface deformation information; the fiber Bragg grating strain gauge measures strain through the wavelength change of the grating in the optical fiber; the earth pressure cell reflects the stress change by measuring the pressure of the soil mass on the cell body; the resistance strain gauge measures the change in resistance using a Wheatstone bridge and thus measures strain; the differential pressure strain gauge measures strain by measuring the pressure difference of liquid or gas; the MEMS vibration sensor senses vibration through microelectromechanical system technology; the accelerometer obtains vibration frequency and displacement information by measuring acceleration and integrating; the velocity sensor reflects vibration frequency data by measuring velocity; the piezoelectric ceramic sensor measures strain using the piezoelectric effect.

[0037] In addition to the above-listed types of sensors, the first type of sensor can also be a laser rangefinder or an inclinometer. The laser rangefinder determines the distance by measuring the round-trip time of the laser, and the inclinometer reflects the displacement by measuring the tilt angle relative to the direction of gravity. The second type of sensor can also be a multi-point displacement meter or a piezometer. The multi-point displacement meter reflects the internal deformation of the slope by measuring the relative displacement between multiple anchor points, and the piezometer indirectly reflects the stress state of the slope by measuring the pore water pressure. The third type of sensor can also be a seismograph, which reflects the vibration of the slope by picking up the vibration signal on the ground.

[0038] In a possible embodiment, for the communication method of the sensors, the wired communication method can be optical fiber communication, RS485 bus or Ethernet; the wireless communication method can be ZigBee, LoRa, NB-IoT or 5G. These communication methods have their own advantages and disadvantages, and the specific choice of which method needs to be comprehensively considered according to factors such as the actual conditions of the slope site, data transmission rate requirements, and power consumption limitations.

[0039] S120: Based on the geometric shape, geological conditions, and historical instability data of the slope to be monitored, determine the spatial layout scheme of the first type of sensor, the second type of sensor, and the third type of sensor in the distributed sensor network.

[0040] Specifically, the spatial layout scheme includes the layout positions and layout densities of the first type of sensor, the second type of sensor, and the third type of sensor at the toe, crest, slope surface, and potential slip surface of the slope to be monitored; in the areas prone to instability indicated by the historical instability data, increase the layout density of the first type of sensor, the second type of sensor, and the third type of sensor. Among them, the potential slip surface refers to the area within the slope where sliding may occur as determined based on geological exploration data and numerical simulation analysis; the historical instability data refers to the data records of slopes that have experienced instability in history or the data records of slopes with precursors of instability.

[0041] In a possible embodiment, the installation locations may also include boreholes inside the slope, the surface of the support structure, or both sides of the cracks. Installing the sensors in the boreholes can obtain the deformation and stress information deep inside the slope; installing them on the surface of the support structure can monitor the stress state of the support structure; and installing them on both sides of the cracks can monitor the expansion of the cracks. The installation density can be determined according to the results of slope stability analysis, the monitoring range of the sensors, and the accuracy requirements. For example, for areas with poor stability, the installation density can be appropriately increased; for sensors with a small monitoring range, the installation density can also be appropriately increased; and for monitoring items with high accuracy requirements, the installation density also needs to be increased. In addition, the installation density of the sensors can be optimized and determined through numerical simulation methods. By establishing a three-dimensional numerical model of the slope, the monitoring effects under different installation densities can be simulated, and the optimal installation plan can be selected. The historical instability data may include historical landslide data, historical deformation data, historical stress data, or historical vibration data, and all these data can provide references for sensor installation.

[0042] S130: Using a clock synchronization protocol, control the first type of sensors, the second type of sensors, and the third type of sensors in the distributed sensor network to synchronously collect displacement data, stress data, and vibration frequency data at a preset sampling frequency, and add timestamps to the displacement data, stress data, and vibration frequency data and then transmit them to the data processing center;

[0043] Specifically, the clock synchronization protocol is one of the Network Time Protocol (NTP), the IEEE 1588 Precision Time Protocol, GPS time synchronization, or Beidou time synchronization. The timestamp is used to identify the precise moment of data collection and is used for subsequent data fusion and time series analysis.

[0044] In addition to the above-mentioned clock synchronization protocols, a custom synchronization protocol can also be adopted. The custom synchronization protocol can be designed according to the specific sensor type and communication method to meet specific synchronization accuracy requirements. The preset sampling frequency can be determined based on the slope deformation rate, sensor response time, and data processing capacity. For slopes with a faster deformation rate, a higher sampling frequency is required; for sensors with a longer response time, the sampling frequency needs to be appropriately reduced; for systems with limited data processing capacity, the sampling frequency also needs to be appropriately reduced. In addition, the preset sampling frequency can be variable and adjusted according to the changes in the monitoring data. For example, when significant changes occur in the displacement or stress data during monitoring, the sampling frequency can be automatically increased to more precisely capture the dynamic changes of the slope. The data processing center can be a local server or a cloud server. The local server is deployed near the slope site and can achieve real-time data processing and analysis; the cloud server has more powerful computing and storage capabilities and can achieve the storage and analysis of massive data. In a possible implementation manner, the displacement data, stress data, and vibration frequency data can be first stored in the edge computing node, and after data preprocessing, the preprocessed data is then transmitted to the data processing center. Among them, the edge computing node refers to a device deployed at the slope site with certain data storage and computing capabilities, such as an embedded system or an industrial control computer.

[0045] Preferably, the key point of step S100 of the present invention is to construct a distributed sensor network and collect multi-source monitoring data, providing a comprehensive and reliable data basis for subsequent slope stability analysis. By reasonably selecting GNSS receivers, total stations or InSAR radars (the first type of sensors), fiber Bragg grating strain gauges, earth pressure cells, resistance strain gauges or differential pressure strain gauges (the second type of sensors), and MEMS vibration sensors, accelerometers, velocity sensors or piezoelectric ceramic sensors (the third type of sensors) for combination, and providing the possibility of selecting various sensor types, it realizes the all-round monitoring of the slope from the surface to the interior and from static to dynamic. Compared with the traditional single-type sensor scheme, the obtained information is more comprehensive and accurate, and this information cannot be achieved by a simple combination of existing technologies. Determine the sensor spatial layout scheme (including layout positions and layout densities) based on the geometric shape, geological conditions and historical instability data of the slope, and increase the layout density in the areas prone to instability, and provide various methods for determining layout positions and layout densities, avoiding false alarms or missed alarms caused by unreasonable arrangement of monitoring points. Use a unified clock synchronization protocol (NTP, IEEE 1588 Precision Time Protocol, GPS time synchronization or Beidou time synchronization) to realize the synchronous acquisition of multi-source data, and provide the possibility of customizing the synchronization protocol, and ensure the time comparability of data by adding time stamps, providing a key basis for subsequent multi-dimensional data fusion and time series analysis, enabling the tracking of the change trend of the slope state over time, and further improving the utilization value of the monitoring data. In addition, various implementation methods such as variable sampling frequency and edge computing nodes for data preprocessing are also provided, enhancing the flexibility and adaptability of the system. Although some of the above technical means belong to conventional technologies, their ingenious application and combination in the slope monitoring scenario of the present invention have produced unexpected effects.

[0046] S200: Use a graph neural network to extract features from the multi-source monitoring data. The feature extraction includes capturing the spatial correlation between sensor nodes and identifying abnormal patterns, identifying potential unstable areas of the slope according to the abnormal patterns, and outputting the features of the potential unstable areas.

[0047] S210: Construct a slope monitoring graph model, take each sensor (including the first type of sensors, the second type of sensors and the third type of sensors) in the distributed sensor network as a graph node, construct edges based on the spatial position relationship of the graph nodes to form a graph structure, and input the graph structure into the graph neural network.

[0048] Specifically, the graph structure includes a node feature matrix and an adjacency matrix. The node feature matrix contains displacement data, stress data, and vibration frequency data collected for each graph node; the adjacency matrix represents the spatial proximity relationship between graph nodes and is constructed using the K-nearest neighbor algorithm or a method weighted by the reciprocal of the distance. Among them, the adjacency matrix is a matrix used to represent the connection relationship between nodes in a graph; the K-nearest neighbor algorithm means that for each graph node, the K graph nodes closest to it are selected as its neighbors; the method weighted by the reciprocal of the distance means that the connection weight between graph nodes is inversely proportional to the distance between graph nodes.

[0049] It should be noted that step S210 of the present invention provides a basic data structure for subsequent graph neural network processing. By converting the sensor network into a graph structure, the slope monitoring problem can be transformed into a graph data processing problem, thereby utilizing the powerful relationship learning ability of the graph neural network to capture the spatial correlation between sensor nodes, which is difficult to achieve by traditional data processing methods. In addition, this step completely characterizes the state of the slope monitoring system through the node feature matrix and the adjacency matrix, providing comprehensive data support for subsequent feature extraction and abnormal pattern recognition.

[0050] S220: Through the inter-node relationship modeling module of the graph neural network, based on the graph structure, capture the spatial correlation between graph nodes.

[0051] Specifically, the inter-node relationship modeling module uses a graph attention network (GAT). The input of the graph attention network (GAT) is the graph structure. The attention coefficient between each graph node and its adjacent graph nodes is calculated through an attention mechanism. The attention coefficient represents the influence degree of adjacent graph nodes on the graph node; based on the attention coefficient and the initial feature vectors of adjacent graph nodes, the initial feature vectors of adjacent graph nodes are weighted and averaged to obtain the spatial correlation features of the graph node.

[0052] Specifically, the process of calculating the attention coefficient is as follows:

[0053] 1. Concatenate the initial feature vectors of each graph node and the initial feature vectors of adjacent graph nodes, and perform feature transformation through a shared fully connected layer;

[0054] Among them, the initial feature vector is the vector corresponding to the graph node in the node feature matrix.

[0055] 2. Input the result of the feature transformation into a single-layer perceptron to calculate the scalar attention score;

[0056] 3. Perform Softmax normalization on the attention score to obtain the attention coefficient between the graph node and each of its adjacent graph nodes.

[0057] Among them, the fully connected layer is used to extract the correlation information between node features; the single-layer perceptron is used to map the feature vector to a scalar attention score; and the Softmax normalization is used to convert the attention score into a probability distribution, representing the importance weight of adjacent nodes to the current node.

[0058] It should be noted that in this step, GAT is applied to the slope monitoring scenario. The key lies in using the attention mechanism to automatically learn and quantify the spatial correlation between sensor nodes. This correlation is an important basis for slope stability analysis. Compared with traditional methods based on fixed weights or distances, GAT can capture the mutual influence between graph nodes more flexibly and accurately, providing more reliable features for subsequent abnormal pattern recognition. By weighted-averaging the initial feature vectors of adjacent graph nodes, the spatial correlation features of graph nodes are obtained. This feature integrates the information of surrounding nodes, providing a more comprehensive perspective for subsequent analysis.

[0059] S230: Based on the spatial correlation and multi-source monitoring data, identify the abnormal patterns of graph nodes through the abnormal pattern classifier of the graph neural network.

[0060] It should be noted that in this implementation, spatial correlation is a concept or property. It refers to the mutual influence and correlation existing in the monitoring data (displacement data, stress data, and vibration frequency data) between sensor nodes (i.e., graph nodes) at different positions on the slope due to geographical proximity and geological condition similarity. This correlation may be positively correlated (when the data of one node increases, the other also tends to increase), negatively correlated (when the data of one node increases, the other tends to decrease), or there may be more complex relationships. And the spatial correlation feature is the result of quantifying and feature extracting the spatial correlation through a graph neural network (specifically GAT). It is a numerical representation of the spatial relationship around each graph node.

[0061] Furthermore, the abnormal pattern classifier adopts a model combining a multi-layer perceptron (MLP) and a one-dimensional convolutional neural network (1D-CNN). First, use the MLP to extract the correlation features between the spatial correlation features and each dimension data in the multi-source monitoring data, then use the 1D-CNN to extract the time series features of the multi-source monitoring data, and fuse the correlation features and time series features to perform abnormal pattern classification, outputting the abnormal pattern category of the graph node. Among them, the multi-layer perceptron (Multilayer Perceptron, MLP) is a classic feedforward neural network model; the one-dimensional convolutional neural network (1D-CNN) is good at extracting local features of time series data.

[0062] Specifically, the categories of abnormal patterns include mutation-type abnormalities, gradual-change type abnormalities, and normal.

[0063] The determination rules for the abnormal mode include:

[0064] Perform short-time Fourier transform (STFT) on the displacement data, stress data, and vibration frequency data of each graph node to obtain the time-frequency domain characteristics of each graph node. Among them, the short-time Fourier transform (STFT) is an analysis method that decomposes a signal into the variation of different frequency components over time.

[0065] If there is at least one frequency band in the time-frequency domain characteristics of the graph node whose energy exceeds K times the standard deviation of the historical energy of this frequency band, and the center frequency of this frequency band matches any frequency in the set of slope instability characteristic frequencies, then the graph node is determined as a candidate node for mutant abnormal;

[0066] Among them, K in K times the standard deviation is obtained by training with a large amount of historical measured data and numerical simulation data. The historical measured data and numerical simulation data include the monitoring data and simulation results of multiple slopes under different working conditions; the set of slope instability characteristic frequencies is a set of frequency ranges closely related to slope instability determined according to the natural frequency of the slope, geological conditions, historical instability data, and finite element analysis. Among them, finite element analysis refers to the method of performing stability analysis on the slope through numerical calculation to obtain the instability characteristic frequencies of the slope under different working conditions.

[0067] If within the most recent M sampling periods, the absolute value of the first-order difference sequence of at least one of the displacement data, stress data, or vibration frequency data of the graph node is greater than L times the standard deviation of its historical first-order difference sequence, and the sign of the first-order difference sequence is the same as the sign of the theoretical change trend of this item of data calculated based on the slope instability model at the corresponding position and corresponding time, then the graph node is determined as a candidate node for gradual abnormal;

[0068] Among them, the M sampling periods are determined according to the characteristic time of the progressive instability process of the slope, and can be several hours, several days, or several weeks. The first-order difference sequence refers to the sequence formed by the difference between the data value at the current moment and the data value at the previous moment. L in L times the standard deviation is obtained by training with a large amount of historical measured data and numerical simulation data. The slope instability model refers to a mechanical model established based on the geometric shape, geological conditions, material parameters, etc. of the slope, and is used to simulate the deformation and failure process of the slope under different working conditions.

[0069] If the graph node simultaneously meets the determination conditions of the candidate node for mutant abnormal and the candidate node for gradual abnormal, then the graph node is determined as a pending abnormal node.

[0070] If the graph node is a pending abnormal node, then perform a secondary analysis and judgment:

[0071] Calculate the similarity matrix of the change trends of the displacement data, stress data, and vibration frequency data of the pending abnormal node and its adjacent nodes in the past N sampling periods;

[0072] Among them, the N sampling periods should be greater than the M sampling periods to include information over a longer time; the elements of the similarity matrix represent the similarity of the specified data change trends between any two nodes, and the similarity calculation can use the DTW distance, cross-correlation coefficient, etc.

[0073] If there is at least one element in the similarity matrix greater than the preset similarity threshold, and at least one of the two nodes corresponding to this element shows an accelerating change trend in the displacement data, stress data, or vibration frequency data within the N sampling periods, then the to-be-determined abnormal node is determined to be a gradually changing type of abnormality; otherwise, the to-be-determined abnormal node is determined to be a suddenly changing type of abnormality.

[0074] Among them, the accelerating change trend refers to the trend that the change rate increases with time. The preset similarity threshold is set based on a large amount of historical data and expert experience. Specifically, by calculating the similarity of the change trends between adjacent sensor nodes in the historical data and the simulated data (using the DTW distance or cross-correlation coefficient), analyzing its distribution characteristics, and selecting the quantile that can effectively distinguish the gradually changing type of abnormality and the suddenly changing type of abnormality as the preferred value of the preset similarity threshold. The specific value of the preset similarity threshold is related to the slope type, sensor layout, and data quality, and needs to be adjusted and optimized in actual applications.

[0075] If the graph node does not conform to any of the above three situations (suddenly changing type of abnormal candidate node, gradually changing type of abnormal candidate node, to-be-determined abnormal node), then the graph node is determined to be a normal node.

[0076] It should be noted that in the actual application scenario of slope instability warning, compared with the prior art, step S230 has the following advantages: First, step S230 focuses on classifying the abnormal patterns before slope instability into suddenly changing type, gradually changing type, and to-be-determined type (requiring secondary analysis), and designing determination rules respectively. This classification method fully considers the diversity of slope instability patterns and avoids the limitations of traditional methods that only focus on a single abnormal pattern (such as exceeding the threshold). For the suddenly changing type of abnormality, the short-time Fourier transform (STFT) is used to extract time-frequency domain features and match them with the set of slope instability characteristic frequencies. This can effectively identify the precursors of slope instability caused by sudden events such as earthquakes and blasting, and solve the problem that the prior art fails to respond to sudden events in a timely manner. For the gradually changing type of abnormality, by comparing the first-order difference sequence of the monitoring data with its historical data and combining the theoretical change trend calculated based on the slope instability model, the precursors of progressive instability caused by creep, cumulative deformation, etc. can be effectively identified, avoiding misjudgment caused by the fluctuations of the monitoring data itself.

[0077] Secondly, in step S230, a secondary analysis and judgment mechanism is introduced for graph nodes that simultaneously meet the mutant and gradual change abnormal candidate conditions. This secondary analysis and judgment mechanism fully considers the possible situations during the slope instability process, that is, some instability processes may have both mutant and gradual change characteristics, or show different characteristics at different stages. By calculating the similarity of the change trends of the to-be-determined abnormal node and its adjacent nodes within a relatively long time range, and combining the judgment of the accelerated change trend, the final abnormal type can be judged more accurately, avoiding misjudging the short-term fluctuations during the gradual change process as mutations, or misjudging the continuous deformation after mutations as gradual changes, thus improving the accuracy and refinement degree of abnormal pattern recognition.

[0078] Finally, the judgment rules in step S230 make full use of the prior knowledge of the slope and multi-source monitoring data. Whether it is the set of slope instability characteristic frequencies used in the mutant abnormal judgment or the slope instability model used in the gradual change abnormal judgment, they are all established based on the inherent characteristics of the slope, geological conditions, historical instability data, etc., and have clear physical meanings, rather than simply empirical thresholds, improving the reliability of the present invention. In addition, by fusing spatial correlation features, the monitoring information of adjacent nodes can be effectively utilized, reducing noise interference and improving the robustness of abnormal judgment, which is also difficult to achieve by traditional methods.

[0079] S240: According to the spatial positions of the abnormal nodes output by the abnormal pattern classifier and the categories of the abnormal patterns, identify the potential slope instability regions and output the characteristics of the potential instability regions.

[0080] Specifically, the region covered by abnormal nodes with adjacent spatial positions and the same category of abnormal patterns is identified as the potential instability region; the characteristics of the potential instability region include the spatial positions of the abnormal nodes, the categories of the abnormal patterns, and the multi-source monitoring data corresponding to the abnormal nodes.

[0081] It should be noted that in step S240, by defining the region covered by abnormal nodes with adjacent spatial positions and the same category of abnormal patterns as the potential instability region, the subjectivity and inefficiency of manually demarcating the region are avoided. At the same time, the characteristics of the output potential instability region (including the spatial positions of the abnormal nodes, the categories of the abnormal patterns, and the multi-source monitoring data corresponding to the abnormal nodes) provide comprehensive information support for subsequent risk assessment and decision-making. Although identifying potential instability regions is one of the goals of slope monitoring, this step realizes automatic region identification based on the abnormal detection results of the graph neural network, improving the efficiency and accuracy.

[0082] S300: Based on the characteristics of the potential instability region and combined with the disaster evolution map constructed from historical disaster data, generate a slope stability risk level assessment report.

[0083] Specifically, the risk levels of the slope include high risk level, medium risk level, and low risk level.

[0084] S310: Construct a disaster evolution map.

[0085] Specifically, first, extract the characteristics of the potentially unstable regions at different times before each disaster event in the historical disaster data and the disaster levels of each disaster event, and construct a historical disaster feature database.

[0086] Among them, the historical disaster feature database includes the spatial positions of abnormal nodes, the categories of abnormal patterns, the multi-source monitoring data corresponding to the abnormal nodes, and the disaster levels of each disaster event at different times before each disaster event.

[0087] Then, use the characteristics of the potentially unstable regions at different times before each historical disaster event as the disaster map nodes of the disaster evolution map, connect the disaster map nodes in chronological order to form multiple historical disaster evolution sub-graphs, and use the disaster levels of the historical disaster events as the labels of the historical disaster evolution sub-graphs. The disaster levels include high disaster level, medium disaster level, and low disaster level.

[0088] Finally, fuse multiple historical disaster evolution sub-graphs into a disaster evolution map.

[0089] Among them, the node characteristics of the disaster evolution map are the characteristics of the potentially unstable regions at different times before the historical disaster events, and the edges of the disaster evolution map represent the evolution relationships between the states at different times.

[0090] Preferably, in step S310 of the present invention, the originally independent historical disaster events are associated in terms of time and space by constructing a disaster evolution map. Traditional slope risk assessment methods often only focus on current monitoring data or conduct simple statistical analysis on historical data, making it difficult to effectively utilize the rich information contained in historical disaster data. However, in this step, by constructing a disaster evolution map, the slope states (represented by the characteristics of potential instability regions) at different times before each disaster event are abstracted into graph nodes, and these nodes are connected in chronological order to form a historical disaster evolution sub-graph, thereby transforming historical disaster data into a graph-structured data with spatio-temporal correlation. This data organization form can clearly show the evolution process of slope instability, laying a foundation for subsequent in-depth analysis using graph neural networks and solving the problem that it is difficult for the prior art to effectively utilize historical disaster data. In addition, traditional risk assessment methods are often based on empirical formulas or numerical simulations, making it difficult to comprehensively consider the complexity and diversity of slope instability. However, the disaster evolution map constructed in this step can incorporate slope instability events of different types, scales, and inducing factors into the same framework for analysis. Through the learning of the graph neural network in the subsequent steps, it is expected to discover the commonalities and laws between different instability events, thereby achieving more accurate risk prediction.

[0091] S320: Based on the characteristics of the disaster evolution map and the potential instability region, construct a risk level assessment model.

[0092] Specifically, use the disaster evolution map as training data to train a deep learning model for risk level assessment. This deep learning model includes an encoder and a decoder. Among them, the encoder adopts a graph convolutional network (GCN) or a graph attention network (GAT) to map the disaster graph nodes of the disaster evolution map to a low-dimensional vector space to obtain the embedding vectors of the disaster graph nodes. The decoder infers the risk level of the slope corresponding to the disaster graph node according to the embedding vector of the disaster graph node and the label of the historical disaster evolution sub-graph.

[0093] The risk level assessment model is trained by minimizing the difference between the inferred risk level of the slope and the label of the historical disaster evolution sub-graph; input the characteristics of the potential instability region into the trained deep learning model to obtain the risk level of the slope.

[0094] It should be noted that in step S310 of the present invention, the construction of the disaster evolution map is based on historical disaster data. The node characteristics of the disaster map nodes in the disaster evolution map are derived from the characteristics of potential instability regions at different times before the occurrence of historical disaster events. From the perspective of data sources, the disaster map nodes only contain information about the potential instability regions of the slopes. However, in step S320, the "historical risk level of the disaster map nodes" inferred by the decoder is an estimation of the risk level of the slope represented by the disaster map nodes at the corresponding time. Since the historical disaster data is a complete record of the slopes, each historical disaster evolution sub-graph uniquely corresponds to the evolution process of a slope over a period of time until the occurrence of the disaster. Each historical disaster evolution sub-graph is composed of a series of disaster map nodes connected in chronological order. Each disaster map node corresponds to a specific time before the occurrence of the disaster event on the slope, and the node characteristics of the disaster map nodes at each time can reflect the overall state of the specific slope at that time.

[0095] Further analysis shows that as a representation of the slope state at different times before the occurrence of historical disaster events, the time attribute of the disaster map nodes is crucial. Each disaster map node represents the state of a specific slope at a specific time, and the disaster map nodes connected in chronological order constitute the historical evolution trajectory of the slope, that is, the historical disaster evolution sub-graph. The label "disaster level" of the historical disaster evolution sub-graph is an objective description of the overall risk level of the corresponding slope in this disaster event. Therefore, in step S320, the "historical risk level of the disaster map nodes" inferred by the decoder is essentially to infer the risk level of the slope represented by the disaster map nodes at the corresponding time. Although the disaster map nodes only contain the characteristics of the potential instability regions, the selection and extraction of the characteristics of the potential instability regions (as described in step S200) are sufficient to represent the state of a specific slope at a specific time and can be used for the inference of the risk level. In other words, each disaster event in the historical disaster data corresponds to a specific slope, and the disaster map nodes are snapshots of the states of the specific slope at different times before the occurrence of the disaster event. Therefore, through the disaster map nodes, the corresponding specific slope can be traced back.

[0096] It should be noted that the deep learning model for risk level assessment is different from the graph neural network model for feature extraction in step S200. The graph neural network model in step S200 focuses on extracting spatial correlation features and identifying abnormal patterns from multi-source monitoring data, while the deep learning model in this step focuses on risk level assessment based on the disaster evolution map and the characteristics of potential instability regions. Its input and output are different from those of the graph neural network in step S200. The deep learning model in this step adopts an encoder-decoder structure, which can learn the long-term dependencies and evolution patterns in the disaster evolution map, so as to achieve more accurate risk prediction.

[0097] Furthermore, calculate the evolution path of the embedding vector obtained by the feature input encoder in the potentially unstable area, i.e., the first evolution path, and the evolution rate of the first evolution path. Calculate the matching degree between the first evolution path and the evolution paths of the historical disaster evolution subgraphs respectively.

[0098] If the matching degree level between the first evolution path and the evolution path of the historical disaster evolution subgraph with a high disaster level label is high, and the evolution rate of the first evolution path is greater than the third quartile of the evolution rate distribution of all historical disaster evolution subgraphs in the disaster evolution map, then the risk level of the slope output by the risk level assessment model is a high risk level;

[0099] If the matching degree level between the first evolution path and the evolution path of the historical disaster evolution subgraph with a medium disaster level label is high, and the evolution rate of the first evolution path is greater than the first quartile of the evolution rate distribution of all historical disaster evolution subgraphs in the disaster evolution map, then the risk level of the slope output by the risk level assessment model is a medium risk level;

[0100] If the matching degree level between the first evolution path and the evolution path of the historical disaster evolution subgraph with a low disaster level label is high, or the matching degree levels with the evolution paths of all historical disaster evolution subgraphs are all low, then the risk level of the slope output by the risk level assessment model is a low risk level.

[0101] If the matching degree level between the first evolution path and the evolution path of the historical disaster evolution subgraph with any label is medium, then the risk level of the slope output by the risk level assessment model is a medium risk level.

[0102] Among them, the evolution rate refers to the rate at which the features of the potentially unstable area transfer between the disaster map nodes in the disaster evolution map. The evolution rate is determined by calculating the average time required for the embedding vector obtained by the feature input encoder of the potentially unstable area to transfer between adjacent disaster map nodes in the disaster evolution map. The dynamic time warping (DTW) algorithm is used to calculate the matching degree of the evolution path. The matching degree levels include three levels: high, medium, and low. Specifically, the DTW distances are sorted from small to large, and the matching degree levels corresponding to the first 33% of the DTW distances are high, the matching degree levels corresponding to the middle 33% of the DTW distances are medium, and the matching degree levels corresponding to the last 33% of the DTW distances are low. The first quartile refers to the value at the 25% position after sorting a set of data from small to large; the third quartile refers to the value at the 75% position after sorting a set of data from small to large.

[0103] Among them, the evolution path refers to the sequence formed by arranging the embedding vectors corresponding to the disaster map nodes of the disaster evolution map in time series.

[0104] Preferably, step S320 of the present invention solves the problems in the prior art that it is difficult to capture the evolution process of slope instability and the risk assessment results are not accurate and objective enough. Compared with the prior art, it has the following advantages: First, step S320 adopts an encoder-decoder model based on a graph convolutional network or a graph attention network, which can effectively learn the complex spatio-temporal relationships and evolution patterns contained in the disaster evolution map. Traditional risk assessment methods are difficult to capture the dynamic evolution process of slope instability, while the model constructed in this step can extract the embedding vectors of the nodes in the disaster map through the encoder and learn the mapping relationship between the embedding vectors and the risk levels through the decoder, so as to realize the modeling of the evolution process of slope instability, which is difficult for traditional methods to achieve. Second, this step compares the evolution path, evolution rate of the current potentially unstable area with the matching degree and evolution rate of the historical disaster evolution sub-map, and dynamically adjusts the risk level of the slope. This dynamic adjustment mechanism is not based on artificially set thresholds, but is achieved by comparing with historical disaster data, making the risk assessment results more objective and reliable. In particular, by comparing the matching degree of the evolution paths, the historical disaster events most similar to the current situation can be identified, so as to draw on their evolution trends for risk assessment; by comparing the evolution rates, it can be judged whether the deterioration speed of the current slope state exceeds the high-risk events in history, so as to issue an alarm more timely.

[0105] S330: Generate a slope stability risk level assessment report based on the risk level of the slope output by the risk level assessment model.

[0106] In summary, the present invention realizes the synchronous acquisition of multi-source monitoring data by constructing a distributed sensor network, uses a graph neural network to extract the spatial correlation between sensor nodes and identify abnormal patterns, and constructs a disaster evolution map based on historical disaster data for risk assessment, solving the problems in the prior art that it is difficult to effectively fuse multi-source heterogeneous data, ignore spatial correlation, have low accuracy in abnormal pattern recognition, and lack spatio-temporal evolution information in risk assessment. Compared with the prior art, the focus of the present invention is to construct a multi-dimensional data acquisition system that can comprehensively reflect the slope state; effectively capture and quantify the spatial correlation between sensor nodes through a graph attention network; propose an abnormal pattern recognition method that combines time-frequency domain analysis and instability pattern matching, and design a secondary analysis and judgment mechanism to improve the recognition accuracy; construct a disaster evolution map, convert historical disaster data into graph-structured data, and use a deep learning model to realize the risk level assessment based on historical evolution patterns. The present invention realizes the comprehensive perception, accurate diagnosis and risk warning of the slope state, providing a more reliable technical support for slope safety management.

[0107] Embodiment 2, referring to Figure 2 , is an embodiment of the present invention, which provides a multi-modal early warning system for slope geological disasters.

[0108] As shown Figure 2 in the figure, it is a schematic diagram of the overall structure of the system, including:

[0109] A data acquisition module for deploying a distributed sensor network to collect multi-source monitoring data of the slope;

[0110] A feature extraction module for using a graph neural network to extract features from the multi-source monitoring data. The feature extraction includes capturing the spatial correlation between sensor nodes and identifying abnormal patterns, identifying potential unstable areas of the slope according to the abnormal patterns, and outputting the features of the potential unstable areas;

[0111] A risk assessment module for generating an assessment report on the slope stability risk level based on the features of the potential unstable areas and combining with a disaster evolution map constructed from historical disaster data.

[0112] Example 3, referring to Figure 3 , is an embodiment of the present invention. The difference from the previous embodiment is that if the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or a part of this technical solution can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. And the aforementioned storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes.

[0113] The logic and / or steps represented in the flowchart or described in other ways herein, for example, can be considered as a definite sequence list of executable instructions for implementing a logical function, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch instructions from the instruction execution system, apparatus, or device and execute the instructions), or used in combination with these instruction execution systems, apparatus, or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device.

[0114] More specific examples (a non-exhaustive list) of computer-readable media include the following: electrical connections (electronic devices) having one or more wirings, portable computer disk cartridges (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber devices, and portable compact disc read-only memory (CDROM). Additionally, the computer-readable media can even be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other media, then editing, interpreting, or otherwise processing it as appropriate, and then storing it in a computer memory.

[0115] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), and the like.

[0116] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A multi-modal early warning method for slope geological disasters, characterized in that, Including: Deploy a distributed sensor network to collect multi-source monitoring data of the slope; The multi-source monitoring data includes displacement data, stress data, and vibration frequency data; Use a graph neural network to extract features from the multi-source monitoring data. The feature extraction includes capturing the spatial correlation between sensor nodes and identifying abnormal patterns, identifying potential unstable areas of the slope according to the abnormal patterns, and outputting the features of the potential unstable areas; Based on the features of the potential unstable areas and combined with the disaster evolution map constructed from historical disaster data, generate a slope stability risk level assessment report.

2. The multi-modal early warning method for slope geological disasters according to claim 1, wherein: Using a graph neural network to extract features from the multi-source monitoring data. The feature extraction includes capturing the spatial correlation between sensor nodes and identifying abnormal patterns, identifying potential unstable areas of the slope according to the abnormal patterns, and outputting the features of the potential unstable areas, including the following steps: Construct a slope monitoring graph model, take each node where a sensor in the distributed sensor network is located as a graph node, construct edges based on the spatial position relationship of the graph nodes to form a graph structure, and input the graph structure into the graph neural network; Through the node relationship modeling module of the graph neural network, capture the spatial correlation between the graph nodes based on the graph structure; Through the abnormal pattern classifier of the graph neural network, identify the abnormal patterns of the graph nodes based on the spatial correlation and the multi-source monitoring data; According to the spatial positions of the abnormal nodes output by the abnormal pattern classifier and the categories of the abnormal patterns, identify the potential unstable areas of the slope and output the features of the potential unstable areas. Specifically, the area covered by abnormal nodes with adjacent spatial positions and the same category of abnormal patterns is identified as the potential unstable area; Among them, the features of the potential unstable areas include the spatial positions of the abnormal nodes, the categories of the abnormal patterns, and the multi-source monitoring data corresponding to the abnormal nodes.

3. The multi-modal early warning method for slope geological disasters according to claim 2, wherein: The abnormal pattern classifier uses a multi-layer perceptron to extract the correlation features between the spatial correlation features and each dimension data in the multi-source monitoring data, uses a one-dimensional convolutional neural network to extract the time series features of the multi-source monitoring data, and fuses the correlation features and the time series features to perform abnormal pattern classification and output the categories of the abnormal patterns of the graph nodes; The categories of the abnormal patterns include mutant anomalies, gradual anomalies, and normal.

4. The multimodal early warning method for slope geological disasters according to claim 3, wherein: The determination rules of the abnormal patterns include: Perform short-time Fourier transform on the displacement data, stress data, and vibration frequency data of each graph node to obtain the time-frequency domain features of each graph node; If the energy of at least one frequency band in the time-frequency domain features of the graph node exceeds K times the standard deviation of the historical energy of this frequency band, and the center frequency of this frequency band matches any frequency in the slope instability characteristic frequency set, then determine that the graph node is a mutant anomaly candidate node; If, within the most recent M sampling periods, the absolute value of the first-order difference sequence of at least one of the displacement data, stress data, or vibration frequency data of the graph node is greater than L times the standard deviation of the absolute value of its historical first-order difference sequence, and the sign of the first-order difference sequence is the same as the sign of the theoretical change trend of this item of data calculated based on the slope instability model at the corresponding position and corresponding time, then the graph node is determined to be a candidate node for gradual change type anomaly; If the graph node simultaneously meets the determination conditions for both candidate nodes for sudden change type anomaly and candidate nodes for gradual change type anomaly, then the graph node is determined to be a pending anomaly node.

5. The multi-modal early warning method for slope geological disasters according to claim 4, characterized in that: If the graph node is a pending anomaly node, then perform a secondary analysis and judgment: Calculate the similarity matrix of the change trends of the displacement data, stress data, and vibration frequency data of the pending anomaly node and its adjacent nodes within the past N sampling periods; where, the N sampling periods should be greater than the M sampling periods; If there is at least one element in the similarity matrix greater than the preset similarity threshold, and at least one of the two nodes corresponding to this element shows an accelerating change trend in the displacement data, stress data, or vibration frequency data within the N sampling periods, then the pending anomaly node is determined to be a gradual change type anomaly; otherwise, the pending anomaly node is determined to be a sudden change type anomaly; If the graph node is not any one of the candidate nodes for sudden change type anomaly, candidate nodes for gradual change type anomaly, and pending anomaly nodes, then the graph node is determined to be normal.

6. The multi-modal early warning method for slope geological disasters according to claim 5, wherein: The risk levels of the slope include high risk level, medium risk level, and low risk level; Based on the characteristics of the potential instability area and combined with the historical disaster data to construct a disaster evolution map, generate a slope stability risk level assessment report, including the following steps: Construct a disaster evolution map; Based on the disaster evolution map and the characteristics of the potential instability area, construct a risk level assessment model, and output the risk level of the potential instability area through the risk level assessment model; Generate the slope stability risk level assessment report based on the risk level of the slope output by the risk level assessment model.

7. The multimodal early warning method for slope geological disasters according to claim 6, characterized in that: The construction of the disaster evolution map includes the following steps: Extract the characteristics of the potential instability area at different times before each disaster event in the historical disaster data and the disaster level of each disaster event, and construct a historical disaster feature database; Take the characteristics of the potential instability area at different times before each historical disaster event as the disaster graph nodes of the disaster evolution map, connect the disaster graph nodes in chronological order to form multiple historical disaster evolution sub-graphs, and use the disaster level of the historical disaster event as the label of the historical disaster evolution sub-graph; Fuse the multiple historical disaster evolution sub-graphs into a disaster evolution map.

8. A multi-modal early warning system for slope geological disasters, based on the multi-modal early warning method for slope geological disasters according to any one of claims 1 to 7, characterized in that: Including, A data acquisition module for deploying a distributed sensor network to collect multi-source monitoring data of the slope; A feature extraction module for using a graph neural network to extract features from the multi-source monitoring data. The feature extraction includes capturing the spatial correlation between sensor nodes and identifying abnormal patterns, identifying the potential instability area of the slope according to the abnormal patterns, and outputting the characteristics of the potential instability area; A risk assessment module, which is used to generate a slope stability risk level assessment report based on the characteristics of the potential instability area and combined with the disaster evolution map constructed from historical disaster data.

9. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the slope geological disaster multi-modal early warning method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the slope geological disaster multi-modal early warning method according to any one of claims 1 to 7.

Images