Slope safety hidden danger identification and early warning method and device and electronic equipment

Abstract

The application provides a kind of slope safety hidden danger identification and early warning method, device and electronic equipment, related to engineering safety monitoring and geological disaster prevention technical field, the method is by different edge computing nodes to the multi-source monitoring data of slope area is preprocessed, by the local model of each edge computing node to the corresponding preprocessed data is dynamically spatio-temporal correlation feature extraction, wherein local model is obtained by center server using federated learning mode to train the multi-scale spatio-temporal graph model of corresponding edge computing node, by each edge computing node based on preset physical constraint information and corresponding preprocessed data, corresponding dynamic spatio-temporal correlation feature is detected, and corresponding risk classification early warning is carried out. Using the present application can improve the accuracy of slope hazard identification, effectively reduce data transmission delay, realize model coordination optimization, reduce data transmission cost, improve the pertinence and effectiveness of early warning response.

Description

Technical Field

[0001] This invention relates to the field of engineering safety monitoring and geological disaster prevention technology, and in particular to a method, device and electronic equipment for identifying and warning of slope safety hazards. Background Technology

[0002] Timely identification and accurate early warning of slope safety hazards are the core links of slope engineering safety control. At present, the traditional slope safety monitoring methods mainly have the following problems: (1) Data is collected by a single type of sensor, which has the problems of single monitoring dimension and insufficient data correlation analysis, making it difficult to capture the complex spatiotemporal evolution characteristics of slope instability; (2) Data processing is mostly concentrated in the cloud, which has the defects of high transmission delay and weak real-time response capability at the edge, and cannot meet the needs of rapid early warning in the event of sudden slope disasters; (3) Early warning mechanisms mostly adopt fixed early warning thresholds and strategies, lack quantitative assessment of slope risk level, and the early warning response is not targeted and dynamic enough, which easily leads to over-warning or delayed early warning. Summary of the Invention

[0003] In view of this, the purpose of the present invention is to provide a method, device and electronic device for identifying and warning of slope safety hazards, so as to alleviate the above-mentioned problems existing in related technologies.

[0004] In a first aspect, embodiments of the present invention provide a method for identifying and warning of slope safety hazards, comprising: preprocessing multi-source monitoring data of a slope area through different edge computing nodes; extracting dynamic spatiotemporal correlation features of the corresponding preprocessed data through the local model of each edge computing node; wherein the local model is obtained by training the multi-scale spatiotemporal graph model of the corresponding edge computing node through a federated learning method by a central server; and detecting safety hazards based on the corresponding dynamic spatiotemporal correlation features by each edge computing node based on preset physical constraint information and the corresponding preprocessed data, and performing corresponding risk classification and warning.

[0005] Secondly, embodiments of the present invention also provide a slope safety hazard identification and early warning device, comprising: a preprocessing module for preprocessing multi-source monitoring data of a slope area through different edge computing nodes; an extraction module for extracting dynamic spatiotemporal correlation features of the corresponding preprocessed data through the local model of each edge computing node; wherein the local model is obtained by training the multi-scale spatiotemporal graph model of the corresponding edge computing node through a federated learning method by a central server; and a detection and early warning module for detecting safety hazards of the corresponding dynamic spatiotemporal correlation features based on preset physical constraint information and the corresponding preprocessed data by each edge computing node, and performing corresponding risk classification and early warning.

[0006] Thirdly, embodiments of the present invention also provide an electronic device, including a processor and a memory, wherein the memory stores computer-executable instructions that can be executed by the processor, and the processor executes the computer-executable instructions to implement the slope safety hazard identification and early warning method described in the first aspect above.

[0007] This invention provides a method, device, and electronic device for identifying and warning of slope safety hazards. First, multi-source monitoring data of the slope area is preprocessed through different edge computing nodes. Then, dynamic spatiotemporal correlation features are extracted from the corresponding preprocessed data through the local model of each edge computing node. The local model is obtained by training the multi-scale spatiotemporal graph model of the corresponding edge computing node through federated learning by the central server. Finally, each edge computing node detects safety hazards based on preset physical constraint information and corresponding preprocessed data, and performs corresponding risk classification and warning. By employing the aforementioned technologies, dynamic spatiotemporal correlation features of multi-source slope monitoring data can be extracted using multi-scale spatiotemporal mapping models. This, combined with physical constraints, enables real-time and accurate identification of slope safety hazards, making hazard identification more consistent with the actual failure mechanisms of slope soil and rock. Compared to existing technologies, this significantly improves the accuracy of hazard identification. By deploying data preprocessing, feature extraction, hazard detection, and risk classification early warning on edge computing nodes, real-time assessment of slope safety status is achieved, effectively reducing data transmission latency and solving the problem of data lag in cloud processing in traditional slope safety monitoring methods. This facilitates rapid response in the event of sudden slope disasters. Distributed model training through federated learning reduces data transmission costs while achieving coordinated model optimization. Risk classification early warning based on slope safety hazard identification results effectively avoids over-warning and delayed warnings, improving the targeting and effectiveness of early warning responses compared to existing technologies.

[0008] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention.

[0009] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0010] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0011] Figure 1 This is a flowchart illustrating a method for identifying and warning of slope safety hazards in an embodiment of the present invention; Figure 2 This is an overall flowchart of the slope safety hazard identification and early warning method in this embodiment of the invention; Figure 3 This is a schematic diagram of the structure of a slope safety hazard identification and early warning device according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] At present, the traditional slope safety monitoring methods mainly have the following problems: (1) Data is collected by a single type of sensor, which has the problems of single monitoring dimension and insufficient data correlation analysis, making it difficult to capture the complex spatiotemporal evolution characteristics of slope instability; (2) Data processing is mostly concentrated in the cloud, which has the defects of high transmission delay and weak real-time response capability at the edge, and cannot meet the needs of rapid early warning in the event of sudden slope disasters; (3) Early warning mechanisms mostly adopt fixed early warning thresholds and strategies, lack quantitative assessment of slope risk level, and the early warning response is not targeted and dynamic enough, which easily leads to over-warning or delayed early warning.

[0014] Based on this, the present invention provides a method, device and electronic equipment for identifying and warning of slope safety hazards, which can alleviate the above-mentioned problems existing in related technologies.

[0015] To facilitate understanding of this embodiment, a detailed description of the slope safety hazard identification and early warning method disclosed in this embodiment of the invention will be provided first. (See [link to relevant documentation]). Figure 1 As shown, the method may include the following steps: Step S102: Preprocess the multi-source monitoring data of the slope area using different edge computing nodes.

[0016] Step S104: Dynamic spatiotemporal correlation features are extracted from the corresponding preprocessed data using the local model of each edge computing node.

[0017] The local model can be obtained by training the multi-scale spatiotemporal graph model of the corresponding edge computing nodes through federated learning on the central server.

[0018] Step S106: Based on preset physical constraint information and corresponding preprocessed data, each edge computing node performs security hazard detection on the corresponding dynamic spatiotemporal correlation features and performs corresponding risk classification and early warning.

[0019] This invention provides a method for identifying and warning of slope safety hazards. It utilizes a multi-scale spatiotemporal model to extract dynamic spatiotemporal correlation features from multi-source monitoring data of slopes, and then combines this with physical constraints to achieve real-time and accurate identification of slope safety hazards. This makes the identification of slope safety hazards more consistent with the actual failure mechanism of slope soil and rock, significantly improving the accuracy of slope hazard identification compared to existing technologies. By deploying data preprocessing, feature extraction, safety hazard detection, and risk classification and warning operations on edge computing nodes, it achieves near-real-time judgment of slope safety status, effectively reducing data transmission latency and solving the problem of data lag in cloud processing in traditional slope safety monitoring methods. This facilitates rapid response to sudden slope disasters. Distributed model training is achieved through federated learning, reducing data transmission costs while achieving coordinated model optimization. Risk classification and warning based on slope safety hazard identification results effectively avoids over-warning and delayed warnings, improving the targeting and effectiveness of warning responses compared to existing technologies.

[0020] As one possible implementation method, the above-mentioned slope safety hazard identification and early warning method may further include: collecting multi-source monitoring data by deploying different multi-source sensor networks in the slope area, and sending the collected multi-source monitoring data to different edge computing nodes; wherein each edge computing node corresponds to at least one multi-source sensor network deployed in the corresponding monitoring area of ​​the slope area.

[0021] In practical applications, multi-source sensor networks include, but are not limited to: Displacement-type sensor networks are used to monitor displacement changes in the surface and deep layers of slopes in order to assess the surface and deep deformation of slopes. Stress-based sensor networks are used to monitor soil pressure and stress state of support structures within slopes, in order to capture signals of mechanical imbalance within the slope and predict the risk of instability caused by abnormal stress. Environmental sensor networks are used to monitor environmental parameters around slopes (such as rainfall, temperature, humidity, wind speed, etc.). Since environmental parameters around slopes are important external conditions that can induce slope instability, they can provide environmental correlation data for anomaly detection. Vibration sensor networks are used to monitor vibration signals generated by earthquakes, construction disturbances, rock weathering, etc., on slopes in order to identify the impact of abnormal vibrations on slope stability and prevent slope collapses caused by sudden disturbances.

[0022] Continuing from the previous example, an edge computing node can manage multiple sensor networks (specifically, at least one multi-source sensor network) within a certain geographical area (as the corresponding monitoring area); a complete multi-scale spatiotemporal map model of the slope can be deployed on the central server (as the global model), and a corresponding version obtained by the multi-scale spatiotemporal map model of the slope after appropriate lightweight processing can be deployed on each edge computing node (as the local model for local inference).

[0023] Within the slope monitoring area, the deployment of multiple edge computing nodes and multi-source sensor networks exhibits a "zonal jurisdiction" topology. For example, the slope is divided into three monitoring zones: upper, middle, and lower. Displacement, stress, and environmental sensors (forming the multi-source sensor network of that zone) within each zone are connected to the edge computing nodes in that zone via wired / wireless connections. At the model level, a complete multi-scale spatiotemporal map model covering the entire slope is deployed on the central server for global analysis and cross-regional correlation mining. To achieve local real-time inference, each edge computing node loads and runs a local model. This local model only contains the spatiotemporal correlations between different sensors within the monitoring zone under its jurisdiction, enabling independent slope safety hazard identification and risk classification early warning even when the network connection with the central server is interrupted.

[0024] As one possible implementation, step S102 (i.e., preprocessing the multi-source monitoring data of the slope area through different edge computing nodes) may include: filtering, smoothing and normalizing the multi-source monitoring data corresponding to the monitoring area through each edge computing node.

[0025] Following the previous example, the data preprocessing process performed by edge computing nodes on the corresponding multi-source monitoring data can mainly include: Edge computing nodes employ a weighted moving average method to filter multi-source monitoring data, removing noise and outliers. The formula for the weighted moving average method is as follows:

[0026] in, It is a weighted moving average; The number of periods for the moving average; For the first The weights corresponding to each period of data are usually ; For the first The actual data values ​​for each period.

[0027] Edge computing nodes employ exponential smoothing to smooth the filtered data, further reducing random fluctuations in the signal and ensuring the reliability of subsequent analysis. The exponential smoothing formula is as follows:

[0028] The iterative calculation process is as follows:

[0029]

[0030]

[0031] in, For the first Data after smoothing for one period; For the first The baseline level for each period (the baseline level after removing seasonality and noise can be used as the long-term trend baseline). For the first The trend increment value of each period (the trend increment within a unit period). For the first The seasonal component of each cycle (the seasonal deviation of the current cycle). For the first The original, unsmoothed data for each period (i.e., filtered data, which typically contains noise and seasonal fluctuations). These are the smoothing factors for horizontal, trend, and seasonality, respectively, and are typically between 0 and 1. To smooth out the number of future seasonal cycles; This represents the length of the seasonal cycle. Indicates the first The seasonal component of the cycle; Indicates the first The seasonal component of the cycle; Indicates the first The baseline level value for the cycle; Indicates the first The trend increment value of the cycle.

[0032] Edge computing nodes perform normalization on the smoothed data to unify data from different sources to the same standard range. The normalization formula is as follows:

[0033] in, This refers to the normalized data from the corresponding source; The original, unnormalized data (i.e., the smoothed data) comes from the corresponding source. , These are the maximum and minimum values ​​in the smoothed dataset from the corresponding sources.

[0034] As one possible implementation, the above-mentioned multi-scale spatiotemporal graph model can adopt a preset graph structure corresponding to the monitoring area. In the preset graph structure, nodes can represent sensors and edges between nodes can represent the spatiotemporal correlation between data collected by different sensors. Based on this, step S104 (i.e., dynamic spatiotemporal correlation feature extraction of the corresponding preprocessed data by calculating the local model of each edge node) may include: Step A1: For each local model, a corresponding node feature matrix is ​​generated for the preprocessed data based on the corresponding preset graph structure.

[0035] For example, the preset graph structure may include an adjacency matrix representing the corresponding spatiotemporal relationships, and the local model may include a corresponding graph convolutional network model. For each local model, the corresponding preprocessed data can be iteratively calculated using the corresponding graph convolutional network model based on the corresponding node feature matrix and adjacency matrix to generate the corresponding node feature matrix.

[0036] Step A2 involves estimating the temporal state of the corresponding node feature matrix for each local model based on a preset time step.

[0037] For example, a local model may include a corresponding long short-term memory network model; for each local model, corresponding time-series input data can be generated for the corresponding node feature matrix according to a preset time step, and the corresponding time-series state can be estimated by the corresponding long short-term memory network model based on the corresponding time-series input data.

[0038] Step A3: By fusing the corresponding node feature matrix with its estimated temporal state through each local model, the corresponding comprehensive features are obtained.

[0039] For example, for each local model, the corresponding node feature matrix and temporal state can be fused into the corresponding comprehensive feature through the local model.

[0040] Step A4: Classify the corresponding comprehensive features through each local model to obtain the corresponding dynamic spatiotemporal correlation features.

[0041] For example, a local model may include a corresponding multi-level risk perception model; for each local model, the corresponding hidden danger category can be predicted as the corresponding dynamic spatiotemporal correlation feature based on the corresponding comprehensive features through the corresponding multi-level risk perception model.

[0042] Continuing from the previous example, the above multi-scale spatiotemporal graph model is a model constructed based on preprocessed multi-source monitoring data. The core of this multi-scale spatiotemporal graph model lies in forming a spatiotemporal graph through the representation of nodes and edges. Specifically, the nodes of the spatiotemporal graph represent the monitoring data collected by each sensor network (composed of sensors), while the edges of the spatiotemporal graph are used to connect different nodes to represent the spatiotemporal correlation between the monitoring data collected by different sensor networks.

[0043] The construction and computation of the above multi-scale spatiotemporal graph model mainly includes the following steps: First, the graph structure is constructed. In a multi-scale spatiotemporal graph model, each node represents an independent sensor data stream, while edges represent the spatiotemporal correlations between different sensor data streams. The adjacency matrix of the spatiotemporal graph is used to quantify these connections between nodes. The elements of the adjacency matrix are expressed as follows:

[0044] in, The elements of the adjacency matrix represent the connection relationships between nodes. Through these connections, the spatiotemporal correlation of sensor data can be effectively captured, and important safety hazards and stress distribution characteristics of slopes can be identified.

[0045] Secondly, Graph Convolutional Networks (GCNs) are used to aggregate and update node features on the spatiotemporal graph structure. The input to the GCN is the node feature matrix of the current layer. and the adjacency matrix predefined based on the connection relationships between sensors (This matrix remains fixed during model training, representing the spatial topological relationships between sensors.) GCN uses a message-passing mechanism to allow each node representing a sensor to aggregate feature information from its neighboring nodes, thereby updating each node's own feature representation. The output of GCN is an updated node feature matrix that contains local spatial structure information. GCN captures complex spatiotemporal relationships between data by iteratively updating the feature state of each node. The feature state update formula is:

[0046] in, For the first The node feature matrix output by the layer; For the first The node feature matrix output by the layer; Given an adjacency matrix with self-loops, , The graph adjacency matrix of the spatiotemporal graph. It is the identity matrix; for The corresponding degree matrix (diagonal matrix) satisfies That is, the degree of each node is equal to its own degree plus the number of all its neighbors; For the first Layer-learnable weight matrix; This is the activation function.

[0047] After the L-layer iterations of the GCN described above, the final node feature matrix is ​​obtained. This is the required set of node feature vectors. Node feature vectors can effectively aggregate information from their neighborhood, thereby identifying potential security risks.

[0048] To obtain a global feature vector characterizing the macroscopic state of the entire slope or a specific area, after completing the L-layer iteration of the GCN described above, the final feature matrix of all graph nodes (representing sensors) can also be processed. Perform a global pooling operation; specifically, global average pooling can be used to calculate the final feature matrix of all graph nodes. The sums are then averaged to obtain a fixed-length one-dimensional global feature vector. , , where N is the total number of nodes in the spatiotemporal graph. Each dimension represents a certain average or comprehensive response pattern of the entire slope monitoring system, which can be used for macroscopic stability assessment. The global feature vector is only used for rapid macroscopic assessment of the overall slope stability at a single moment and is not included in subsequent time-series modeling.

[0049] Next, the node feature matrix output by GCN, which contains spatial information, is obtained. Then, it is expanded step-by-step and input into a Long Short-Term Memory (LSTM) network. The node feature matrix is ​​a two-dimensional matrix (dimension N×F, where N is the number of nodes in the spatiotemporal graph and F is the feature dimension) output by the GCN and retains all spatial features of the sensor nodes. Specifically, the node feature matrices of T consecutive time steps can be stacked along the time axis to form an N×F×T spatiotemporal feature sequence (where T is the length of the time window), which is then expanded step-by-step and input into the LSTM. The purpose is to capture the temporal evolution pattern of each monitoring point while retaining the spatial granularity of the nodes, providing a foundation for subsequent point-by-point hazard localization and spatiotemporal correlation analysis. The core mechanisms of LSTM (such as forget gate, input gate, output gate, etc.) can effectively capture long-term dependencies in the time series; the LSTM model uses the node features at each time step... As input, after calculation, the output is the timing state at the current moment. This time series state integrates the evolutionary pattern of slope state over time; the entire calculation process of the LSTM model can be described by the following formula:

[0050] in, It is a set of node features (including node feature matrices at multiple time points output by GCN). It is a time series.

[0051] GCN output and the time-series state set output by LSTM (Including the temporal states at multiple moments) combined to obtain comprehensive features. , is represented as:

[0052] in, , These are the weight matrices corresponding to the node feature set and the temporal state set, respectively.

[0053] After obtaining the comprehensive feature Z that integrates spatiotemporal characteristics, it is input into a multi-level risk perception model suitable for slope safety monitoring scenarios; this multi-level risk perception model specifically consists of one or more fully connected layers and a final... The output layer is constructed as follows: the prediction process of this multi-level risk perception model is as follows: a) The comprehensive feature Z is processed through the first fully connected layer, and its dimensionality is reduced to intermediate features describing the overall stability of the slope; b) The intermediate features are input into the second fully connected layer, where neurons correspond to different hazard patterns such as "abnormal displacement rate," "stress mutation," and "excessive environmental factors," and pattern matching is performed; c) The pattern matching results are input into... Output layer, which outputs a probability distribution (i.e., the probability that the slope state belongs to a predefined category); d) The multi-level risk perception model finally takes the category with the highest probability as the predicted hazard category label Y.

[0054] The mathematical form of the entire prediction process of the multi-level risk perception model can be expressed as:

[0055] in, and Weights and biases for the classification model (i.e., the multi-level risk perception model).

[0056] In the entire process of identifying slope safety hazards based on comprehensive feature Z, the classification model is the final output layer of the entire multi-scale spatiotemporal graph model. This classification model can use a temporal convolutional network or a fully connected network suitable for time series data. The output Y of this classification model is the health status label of the slope within the current time window.

[0057] In summary, GCN is responsible for extracting "spatial" features, LSTM is responsible for extracting "temporal" features, and the fusion of the two constitutes the core of the multi-scale spatiotemporal model. The final classifier (i.e., the above-mentioned classification model, also known as the above-mentioned multi-level risk perception model) is the output module of the multi-scale spatiotemporal model.

[0058] As one possible implementation, the aforementioned preset physical constraint information may include preset physical constraint equations. Based on this, the step S106 performed by each edge computing node (i.e., based on the preset physical constraint information and the corresponding preprocessed data, performing security hazard detection on the corresponding dynamic spatiotemporal correlation features and performing corresponding risk classification and early warning) may include: performing security hazard detection on the corresponding dynamic spatiotemporal correlation features based on the preset physical constraint equations and the corresponding preprocessed data; performing corresponding risk level assessment based on the corresponding security hazard detection results; and triggering corresponding early warning based on the corresponding risk level assessment results.

[0059] Continuing from the previous example, the physical constraint equations are a set of equations used to verify whether the model output (i.e., the comprehensive feature Z of the final output of the multi-scale spatiotemporal map model and its corresponding preliminary hazard category label Y) conforms to the basic principles of geotechnical mechanics. The core of the physical constraint equations includes the following three equations: 1) Linear elastic constitutive equation: describes the stress-strain relationship of a material in the elastic stage and is the basis of all constraints; the expression of this equation is: ,in, For stress; The elastic modulus of a material represents its stiffness. Strain represents the degree of deformation of a material; This is the amount of deformation; The original length of the material; 2) Strength constraint equation: Used to determine whether the material has yielded and failed. The Mohr-Coulomb strength criterion is used as the strength constraint for the slope soil and rock mass; the expression of this equation is: ,in, For shear stress, It is normal stress. The internal friction angle is denoted by ; this equation constrains the stress state represented by the model output to not exceed the shear strength of the material. 3) Deformation Compatibility Constraint Equation: Used to check whether the deformation between different monitoring points satisfies physical continuity; the expression of this equation is: ,in, For strain tensor, and The equation constrains that the strain field calculated from measurements by different displacement gauges should be continuous and without abrupt changes in space.

[0060] Of the three equations mentioned above, the linear elastic constitutive equation is the physical foundation, while the strength constraint equation and the deformation compatibility equation are specific applications and derivation criteria of the linear elastic constitutive equation in different analytical dimensions.

[0061] By analyzing the real-time monitoring data (i.e., the data obtained after preprocessing and caching of the data from each sensor at the current moment through the edge computing node), the current stress and strain values ​​can be obtained. Then, the corresponding real-time monitoring data is substituted into the physical constraint model (i.e., the three equations mentioned above) for anomaly detection.

[0062] The entire process of analyzing real-time monitoring data and inputting it into a physical constraint model for anomaly detection is as follows: First, features directly related to physical quantities are extracted from the output Z of the multi-scale spatiotemporal graphical model (e.g., the deformation ΔL is extracted from the feature vector of the corresponding displacement gauge node). Then, the extracted features (e.g., the deformation ΔL) are substituted into the deformation compatibility constraint equation to calculate the theoretical strain. Simultaneously, the measured stress is directly obtained from the eigenvectors of the corresponding stress nodes. ; to adapt theory Substituting into the linear elastic constitutive equation, the theoretical stress is calculated. ;Measured stress With theoretical stress Perform a response-based comparison, and simultaneously, the measured stress Substitute the stress into the strength constraint equation and determine whether the stress state exceeds the shear strength under the Mohr-Coulomb criterion; if the measured stress Much greater than theoretical stress (Poor deformation compatibility), or measured stress If the yield surface defined by the strength constraint equation is close to or exceeds the yield surface, it is determined that the physical constraint is not satisfied, and the real-time monitoring data at that moment is marked as a physical anomaly (i.e., there is a potential safety hazard).

[0063] In the above-mentioned anomaly detection process based on the physical constraint model, in order to accurately identify slope safety hazards, a safety threshold can be set according to the yield strength of the material, the expression of which is:

[0064] in, This is a safety threshold; The allowable yield strength of the material; For safety margin; The model output (i.e., the comprehensive feature Z output by the multi-scale spatiotemporal graph model and its corresponding hazard category label Y) is compared with the safety threshold to identify potential safety hazards.

[0065] Accurate identification of slope safety hazards is a process driven by models and verified against physical constraints. First, the model outputs a preliminary hazard identification result Y based on the learned spatiotemporal patterns. Then, it calls the real-time monitoring data of the abnormal nodes representing the corresponding sensors, as indicated by the hazard identification result Y, and substitutes them into the physical constraint equations for secondary verification. Finally, the preliminary hazard identification result Y and the secondary verification result are fused, thereby achieving accurate identification of slope safety hazards. The fusion process can be as follows: if the model determines the slope is abnormal and the secondary verification result shows that the physical constraints are not met, then the slope is confirmed to have a safety hazard, and the subsequent risk level assessment process (i.e., quantifying and calculating the corresponding risk level based on the slope safety hazard identification result) will proceed. If the model determines the slope is abnormal but the secondary verification result shows that the physical constraints are met, it may be a model false alarm or sensor noise, and the warning level will be lowered.

[0066] The risk level assessment process described above mainly includes the following steps: First, the risk value is calculated. The risk value R is a comprehensive indicator, not solely determined by stress, but rather integrating the dynamic spatiotemporal correlation characteristics output by the multi-scale spatiotemporal graph model and the verification results of multiple physical constraint equations. The formula for calculating the risk value R is as follows:

[0067] in: The stress ratio, based on the strength constraint equation, is used to measure how close the measured stress is to the shear failure threshold of the soil and rock mass. , , All are determined by calculation using strength constraint equations. It reflects the degree to which the stress approaches the failure threshold; The strain rate anomaly index is based on the deformation compatibility equation. , This reflects whether the deformation conforms to physical laws; This is an anomaly index of correlation degree based on dynamic spatiotemporal correlation characteristics. The difference between the predicted and actual relationships between nodes can be extracted from the comprehensive features Z output by the multi-scale spatiotemporal graph model. The probability of a hazard (i.e., the probability that the slope condition belongs to the category corresponding to the hazard category label Y) directly output by the model is given by the above classification model (i.e., the above multi-level risk perception model). , , , The weights are adjusted according to the actual situation of the slope (such as stress control or displacement control). This formula unifies multi-source monitoring data (such as stress, displacement, etc.), dynamic spatiotemporal correlation characteristics, and physical constraint equations into a quantitative framework, making the risk level assessment more scientific and comprehensive. Allowable stress refers to the maximum safe stress value calculated based on the strength constraint equation and safety factor of the rock and soil mass under long-term service conditions. This represents the stress value actually collected by on-site monitoring equipment; Ultimate stress refers to the stress value at which a rock or soil mass reaches the critical state of shear failure under the strength constraint equation, that is, the upper limit of stress that the material can withstand. This represents the strain value actually collected by the on-site monitoring equipment; This represents the ideal strain value that satisfies deformation continuity, calculated based on the deformation compatibility equation. This represents the permissible critical value used to determine whether strain deviation is abnormal.

[0068] Secondly, based on the calculated risk value Potential safety hazards on slopes are classified into different risk levels. The risk level classification method is as follows: like If so, it is considered to be of low risk (i.e., the risk level is judged as low risk), which is usually within an acceptable range; like If it is, it is considered to be of medium risk (i.e., the risk level is determined to be medium risk). like If the risk level is high, it is considered to be high-risk and urgent measures must be taken to prevent potential safety hazards.

[0069] After completing the above risk level assessment process, a multi-level early warning mechanism is automatically triggered. This mechanism is implemented by pre-establishing early warning level classification standards, which primarily include three states: Normal (corresponding to low risk), Attention (corresponding to medium risk), and Danger (corresponding to high risk). A risk level assessment model can be pre-established to calculate risk values ​​and determine the risk level based on these values, thereby outputting the current early warning instruction: If... If the risk level is low and the status is normal, no additional measures are required. In this case, a current warning instruction representing the normal status is output. If If the risk level is determined to be medium risk, an alert status is automatically triggered, a warning is sent to the monitoring personnel, prompting them to conduct regular checks and strengthen observation. In this case, a current warning command representing this alert status is output; if If the risk level is determined to be high, a dangerous state will be automatically triggered, requiring emergency response measures (including initiating evacuation procedures, notifying the emergency team, etc.) and taking necessary protective measures. In this case, a current warning instruction representing the dangerous state will be output to ensure the safety of slope monitoring.

[0070] In practical applications, the early warning strategy can be dynamically adjusted based on actual abnormal situations, which can be specifically expressed as follows:

[0071] in, Maintain current strategy This indicates that the current strategy will be maintained (or the existing monitoring plan will remain unchanged); This indicates an increase in monitoring frequency (or a higher frequency of monitoring); This indicates the deployment of additional resources (or the allocation of emergency reinforcements).

[0072] In the process of automatically triggering a multi-level early warning mechanism based on the risk level assessment results, an early warning strategy feedback mechanism can be further introduced. This mechanism records and provides feedback on the adjusted early warning strategy. The adjustment method of the early warning strategy can be specifically expressed as follows:

[0073] in, This is the current early warning strategy; Adjustments to the early warning strategy.

[0074] To facilitate understanding, the implementation process of the above-mentioned slope safety hazard identification and early warning method is described in the following example using a specific application.

[0075] See Figure 2 As shown, the overall process of the above-mentioned slope safety hazard identification and early warning method mainly includes the following steps: Step S1: Deploy a multi-source sensor network and edge computing nodes in the slope area. The multi-source sensor network collects multi-source monitoring data in real time and transmits it to the edge computing nodes, thereby realizing the real-time collection and efficient transmission of multi-source monitoring data of the slope.

[0076] In step S1 above, the multi-source sensor network includes, but is not limited to, displacement sensor networks, stress sensor networks, environmental sensor networks, vibration sensor networks, etc.

[0077] In step S1 above, the functions of the edge computing node include, but are not limited to: receiving, preprocessing, fusing and analyzing multi-source monitoring data of slopes collected by the multi-source sensor network in the corresponding monitoring area in real time; completing data noise reduction, anomaly identification, feature extraction and local lightweight model inference; and realizing real-time judgment and graded early warning of slope safety status. At the same time, it performs real-time data storage and historical data association to ensure that the integrity and reliability of the data can still be maintained in the event of network interruption or other abnormal situations.

[0078] Step S2: The edge computing node performs data preprocessing on the collected multi-source monitoring data.

[0079] In step S2 above, data preprocessing mainly includes filtering, smoothing and normalization. The core of data preprocessing is to filter out noise interference and outliers in the data through filtering.

[0080] Step S3: Based on the preprocessed multi-source monitoring data, construct a multi-scale spatiotemporal graph model.

[0081] In step S3 above, the multi-scale spatiotemporal map model is used to extract the dynamic spatiotemporal correlation features of multi-source monitoring data of the slope, effectively identifying potential slope safety hazards and abnormal evolution trends. The core output of the multi-scale spatiotemporal map model is a node spatiotemporal feature matrix with dimensions N×F×T (where N is the number of nodes in the spatiotemporal map, F is the feature dimension, and T is the time window length), and a trend vector used to characterize the abnormal evolution trend of the overall or local area of ​​the slope. This matrix and this vector quantify the state of each monitoring point of the slope (i.e., the node representing the sensor) at different historical times and the mutual influence between the state monitored by the monitoring points and the states monitored by adjacent monitoring points. It includes specific information such as "node feature vector of the slope top displacement acceleration zone" and "characteristics of abnormal increase in lower soil pressure caused by rainfall infiltration", which serve as the core basic information for safety hazard detection and risk quantification in the subsequent step S5.

[0082] Step S4 involves training and parameter aggregation of the constructed multi-scale spatiotemporal model using a federated learning framework, balancing data privacy protection with model training effectiveness, and improving the model's adaptability to slope monitoring data and prediction accuracy.

[0083] In step S4 above, the process of training the multi-scale spatiotemporal graph model and aggregating parameters mainly includes: First, federated learning is a distributed learning approach that improves overall model performance by aggregating model parameters from participating nodes without sharing local data, thereby updating the model. Each edge computing node trains its own multi-scale spatiotemporal graph model (i.e., a local model), and the training optimization objective of this model is to minimize the global loss function.

[0084] in, The global loss function; This refers to the number of nodes participating in the learning process (i.e., the number of edge computing nodes). For the first Loss function for each edge computing node; The actual value; These are the model's predicted values; Each edge computing node updates the model parameters locally and calculates the amount of model parameter update.

[0085] Then, the edge computing nodes upload the calculated local update of the model parameters to the central server, ensuring that each edge computing node only uploads the updated model parameters to the central server and does not transmit the original monitoring data. Specifically, the edge computing nodes calculate the model parameter update using the following formula:

[0086] in, For the first The amount of model parameter updates for each edge computing node; For the first Updated model parameters for each edge computing node; For the first The original model parameters of each edge computing node; This process effectively protects data privacy and reduces data transmission costs. After receiving the model parameter updates uploaded by all edge computing nodes, the central server performs an aggregation operation. The aggregation process uses a weighted average strategy to reasonably reflect the contribution of each edge computing node. The aggregation process can be represented as follows:

[0087] in, These are global model parameters; For the first Number of samples per edge computing node; This represents the total number of samples; This is an assignment operation, meaning that the calculation result on the right is assigned to the variable on the left.

[0088] Edge computing nodes are the "physical clients" that perform federated learning; graph nodes are the "data feature objects" being learned. On each edge computing node (the client of the federated learning framework), a multi-scale spatiotemporal graph model is trained using local monitoring data within its jurisdiction.

[0089] By weighted aggregation of the model parameter updates across all nodes, a more accurate global model can be obtained. After parameter aggregation, the central server sends the updated global model to all participating edge computing nodes, enabling them to continue the next round of local model training.

[0090] Step S5 involves combining the physical constraint equations to detect potential safety hazards in the model output, accurately identifying potential safety hazards on the slope, and quantitatively calculating the corresponding risk level based on the safety hazard detection results, providing a scientific basis for early warning decisions.

[0091] Step S6: Based on the safety hazard detection results and risk level calculation results, automatically trigger a multi-level early warning mechanism, and dynamically adjust the early warning strategy according to the actual abnormal situation to ensure the timeliness and pertinence of the early warning response.

[0092] Step S7: After implementing the early warning strategy, the edge-center collaborative online model update mechanism is activated: Each edge computing node independently performs local incremental updates of model parameters using newly collected monitoring data and early warning response results (such as manually verified false / missed reports); at the same time, the edge computing nodes periodically report the changes in model parameters generated by this update, after de-identification processing, to the central server. After aggregating the changes in model parameters from multiple edge computing nodes, the central server performs a global asynchronous aggregation to optimize the global model.

[0093] This approach of "local real-time iteration and global asynchronous optimization" can fully integrate newly collected monitoring data and early warning response feedback results, continuously optimize model performance, and improve the ability to identify and warn of slope disaster hazards.

[0094] In step S7 above, the continuous optimization process of model performance mainly includes: 1) Dynamic update of model parameters: New monitoring data is input to obtain a new monitoring dataset. This operation is specifically represented as follows:

[0095] in, For input features, The characteristic matrix; For the corresponding output results, This is the output vector; This represents the total number of samples in the dataset; Set the model parameters that need to be updated as follows The objective of updating the model parameters is to minimize the new loss function, which is defined as follows:

[0096] in, Given model parameters and new monitoring data The comprehensive loss function (i.e., the new loss function); For the purposes of 1 to Summing all samples; The loss function itself is typically used to measure the difference between predicted and actual values. For the model to the first Sample The predicted output depends on the sample features and the current model parameters. ; For corresponding The true label or target value is usually the result that the model hopes to predict. The gradient descent method is used to dynamically update the model parameters. This operation is specifically represented as follows:

[0097] in, The learning rate; For the comprehensive loss function Regarding parameters The gradient.

[0098] 2) Adjusting model parameters based on early warning response results: Incorporate early warning response results into the update process; the modified parameters can be updated using weighted coefficients. Represented as:

[0099] in, These are the adjusted model parameters; The change in model parameters; This operation enables the adjustment of model parameters, allowing the model to respond to detected alarms.

[0100] 3) Model Evaluation: After each model parameter update, the model performance is evaluated. This operation can be performed using a validation dataset. The expression is as follows:

[0101] in, For the validation set Normalize the number of samples in the sample. It is a verification set The total number of samples in the dataset is used to calculate the average performance of the model. For the validation set Each sample in Perform summation; To use the adjusted model parameters For input features The prediction results; To calculate the prediction results With real labels The loss between the two sets of data is the primary metric used to evaluate the model's performance on that sample. P represents the model's average performance metric, which is the average loss value of the model on the validation set.

[0102] As a specific example, the technical approach of the above-mentioned slope safety hazard identification and early warning method mainly includes the following parts: (I) Deployment of Multi-Source Sensor Networks and Edge Computing Nodes: A multi-source sensor network is deployed in the slope area using a 5m×5m monitoring grid. Specifically: the displacement sensor network uses crack gauges and deep displacement gauges to monitor surface cracks and deep soil displacement; the stress sensor network uses earth pressure gauges and steel stress gauges to monitor soil pressure within the slope and the stress state of the support anchors; the environmental sensor network uses rain gauges, temperature and humidity sensors, and wind speed sensors to monitor rainfall, slope temperature and humidity, and surrounding wind speed; and the vibration sensor network uses vibration sensors to monitor vibration signals generated by vehicle movement and rock weathering. Meanwhile, three edge computing nodes are deployed in the slope monitoring area to receive, preprocess, and infer local models of multi-source monitoring data from the upper, middle, and lower sections of the slope, respectively. The edge computing nodes have data storage capabilities to ensure that multi-source monitoring data is not lost when the network is interrupted.

[0103] (ii) Edge computing nodes preprocess the collected multi-source monitoring data: First, the edge computing nodes use a weighted moving average method for filtering to remove noise and outliers from the data. Then, the edge computing nodes are smoothed using the exponential smoothing method; Furthermore, edge computing nodes reduce random fluctuations in signals through smoothing processes, ensuring the reliability of subsequent analysis; After smoothing is complete, the edge computing nodes continue to perform normalization processing to unify data from different sources to the same standard range, ensuring data consistency.

[0104] After data preprocessing is completed, the edge computing nodes can transmit the preprocessed data results to the central server for further analysis. This data transmission process first establishes a data transmission channel, then compresses the data using a specific encoding scheme to reduce network transmission time, and finally outputs cleaned monitoring data for advanced analysis and decision support.

[0105] (III) Based on the preprocessed multi-source monitoring data, construct a multi-scale spatiotemporal model: Using the monitoring data of each sensor as nodes of the spatiotemporal graph, and constructing the edges of the spatiotemporal graph based on the spatial distance between sensors and the temporal correlation of data, an adjacency matrix is ​​generated, and a multi-scale spatiotemporal model is constructed using the adjacency matrix of the spatiotemporal graph. After constructing the multi-scale spatiotemporal graph model, the spatiotemporal graph structure is further optimized using GCN. Specifically, GCN captures complex spatiotemporal relationships between data by iteratively updating the node feature states. The specific update formula for the node feature states is as follows:

[0106] in, For activation functions; For the first Layer spatiotemporal graph nodes The node feature matrix; For spatiotemporal graph nodes The adjacent nodes; For the first Layer-learnable weight matrix; For the first Layer bias terms; After multiple iterations, the feature vectors of spatiotemporal graph nodes can effectively aggregate information from their neighborhoods, thereby identifying potential structural changes and security risks; To extract dynamic spatiotemporal correlation features from the constructed multi-scale spatiotemporal graph model, feature aggregation is first performed. Specifically, the global feature representation of each node can be calculated based on the relationship between the node and its neighboring nodes. These features (i.e., global feature representations) can include the clustering coefficient, betweenness centrality, etc., of the node, which can reflect the importance of the node in the entire spatiotemporal structure.

[0107] The clustering coefficient of a node is calculated by the ratio of the actual number of connections between adjacent nodes to the theoretical maximum number of connections, and is used to characterize the degree of association within the node's local neighborhood. The formula for calculating the clustering coefficient of a node is as follows:

[0108] In the formula, For nodes Clustering coefficient, For nodes The total number of neighboring nodes (i.e., the degree of the node). For nodes The actual number of edges between all neighbors. This means The maximum number of possible edges between neighbors; The betweenness centrality of a node is calculated by statistically analyzing the proportion of all nodes that pass through it in the shortest path. It characterizes the degree of control a node has over the propagation of global associations. The formula for calculating the betweenness centrality of a node is:

[0109] In the formula, For nodes betweenness centrality, For the node To the node The total number of all shortest paths, For the node To the node All shortest paths passing through nodes The number of entries.

[0110] After feature extraction (i.e., the global feature representation calculation process described above), potential safety hazards are identified based on the dynamic features output by the model. Specifically, different thresholds can be set to determine whether the output of the constructed model exceeds the normal range, thereby identifying potential safety hazards. This process requires real-time analysis of the model output. The model output analysis process is as follows: First, the extracted dynamic spatiotemporal features are compared with the preset normal range; then, by determining the threshold, it is determined whether there are significant hidden dangers; finally, the judgment results provide basic information for subsequent analysis (such as whether to initiate a further risk level assessment process).

[0111] (iv) Training and parameter aggregation of the constructed multi-scale spatiotemporal graph model using a federated learning framework: First, multi-scale spatiotemporal graph models are trained on the edge computing nodes of the upper, middle and lower sections of the slope. The training optimization objective of this model is to minimize the global loss function. Each edge computing node updates the model parameters locally and calculates the parameter update amount; Then, the edge computing nodes upload the calculated local update of the model parameters to the central server; After receiving the model parameter updates uploaded by all edge computing nodes, the central server uses a weighted average strategy to aggregate the data, so as to reasonably reflect the contribution of each edge computing node. By weighted aggregation of the model parameter updates of all nodes, a more accurate global model can be obtained. Finally, after parameter aggregation, the central server sends the updated global model to all participating edge computing nodes so that the edge computing nodes can continue to train their local models in the next round. The models of each edge computing node are iteratively trained using the corresponding update algorithm, and finally the optimized multi-scale spatiotemporal graph models of each edge computing node are output.

[0112] This federated learning framework enables efficient training and updating of multi-scale spatiotemporal graph models, preserving data privacy of each edge computing node as much as possible. Throughout the implementation process, the edge computing nodes and the central server establish a connection through a secure encrypted communication protocol to ensure the security of model parameter updates and aggregation.

[0113] (v) Based on the physical constraint equations, perform safety hazard detection on the model output results, accurately identify potential safety hazards, and calculate the corresponding risk levels: First, a physical constraint model is established using the constitutive equation of linear elastic materials; by analyzing real-time monitoring data, the current stress and strain values ​​can be obtained. Then, the corresponding real-time monitoring data is substituted into the physical constraint model for anomaly detection, and a safety threshold is set to determine whether there are any safety hazards. When the detected stress exceeds the set safety threshold, the data point is marked as abnormal, and its risk value R is calculated. Then, based on the calculated risk level... The value determines the risk level of potential safety hazards. It was determined to be low risk at that time. It was determined to be of medium risk at the time. (It was determined to be high risk at the time).

[0114] (vi) Based on the results of safety hazard detection and risk level assessment, trigger a multi-level early warning mechanism and automatically adjust the early warning strategy: After edge computing nodes complete the detection of security risks through physical constraint equations, they compare the calculated risk level with the preset safety threshold to determine the trigger warning level; based on the warning level, they generate corresponding alarm information and transmit the alarm information to the terminal device of the monitoring operator, thereby achieving the effect of dynamic monitoring and timely response; In implementing the early warning mechanism, the early warning strategy can be dynamically adjusted based on historical monitoring data and current hazard detection results; this process ensures that the model is automatically adjusted after each data collection to maintain a rapid response capability to changes in slope condition.

[0115] (vii) Dynamic updates of model parameters after the implementation of the early warning strategy to reflect new monitoring data and early warning response results: First, an incremental learning algorithm is used to update the model parameters based on new monitoring data. The basic principle is to adjust the previous model using newly collected data, thereby reflecting environmental changes and new monitoring information. Then, the model parameters are adjusted based on the early warning response results so that the model can respond to the detected alarms; After each model parameter update, use the validation dataset. Evaluate the model performance; After the model parameters are updated and the model performance is evaluated, the new model parameters will be used to regenerate the multi-scale spatiotemporal graph model. Then, by calculating the node states and connections in the network, the adjacency matrix of the graph will be updated, and the model will be optimized through a graph convolutional network to better extract new dynamic spatiotemporal correlation features.

[0116] The model parameter update process between edge computing nodes and the central server is executed in the following steps: First, new monitoring data is collected, then data processing is performed to remove abnormal data, and finally, incremental learning algorithms are used to update the model parameters to reflect the new monitoring data and early warning response results. This process ensures that the model parameters are dynamically adjusted as the data changes, maintaining the edge computing nodes' ability to respond quickly to potential risks.

[0117] In summary, the above-mentioned methods for identifying and warning of slope safety hazards have the following beneficial effects: 1. Multi-dimensional perception and more comprehensive feature extraction: By deploying a multi-source sensor network, multi-dimensional monitoring data of slopes can be collected. Combined with the spatiotemporal correlation features of the data fused by the multi-scale spatiotemporal map model, the dynamic evolution trend of slope instability can be effectively captured, and the accuracy of slope safety hazard identification can be greatly improved.

[0118] 2. Edge intelligent processing for more real-time response: Data preprocessing and local lightweight model inference are deployed on edge computing nodes to achieve real-time judgment of slope safety status, effectively reducing data transmission latency and solving the problem of data lag in cloud processing in traditional slope safety monitoring methods, thus meeting the need for rapid response in the event of a sudden slope disaster.

[0119] 3. Federated learning training for more reliable privacy protection: The model is trained in a distributed manner by relying on the federated learning framework. Each edge computing node only uploads the model parameter update to the central server and does not transmit the original monitoring data to the central server. This not only realizes the coordinated optimization of the model, but also effectively protects the data privacy of each monitoring point, while reducing the data transmission cost.

[0120] 4. Combining physical constraints makes slope safety hazard detection more scientific: The output results of the multi-scale spatiotemporal diagram model are combined with the physical constraint equations of the slope soil and rock mass. Hazard detection is completed through material constitutive relations and yield strength thresholds, avoiding the overfitting problem of pure data-driven models, and making the identification of slope safety hazards more consistent with the actual failure mechanism of slope soil and rock mass.

[0121] 5. Quantitative risk classification for more targeted early warning: The slope risk level is calculated by a quantitative formula, and the risk is divided into three levels: low, medium and high. A corresponding multi-level early warning mechanism is established, and the early warning strategy is dynamically adjusted according to the risk level to effectively avoid over-warning and delayed warning, thereby improving the targeting and effectiveness of the early warning response.

[0122] 6. Dynamic model optimization and continuous performance improvement: The model parameters are dynamically updated by integrating newly collected monitoring data and early warning response feedback results, and the model performance is continuously evaluated using validation datasets to achieve iterative optimization of the model, ensuring that the model maintains high accuracy in identifying and warning of slope safety hazards as the slope environment changes.

[0123] Based on the above-mentioned method for identifying and warning of slope safety hazards, this invention also provides a device for identifying and warning of slope safety hazards. (See attached image.) Figure 3 As shown, the device may include the following modules: The preprocessing module 302 is used to preprocess multi-source monitoring data of the slope area through different edge computing nodes.

[0124] The extraction module 304 is used to extract dynamic spatiotemporal correlation features from the corresponding preprocessed data through the local model of each edge computing node; wherein, the local model is obtained by training the multi-scale spatiotemporal graph model of the corresponding edge computing node through federated learning by the central server.

[0125] The detection and early warning module 306 is used to detect potential safety hazards based on preset physical constraint information and corresponding preprocessed data of each edge computing node, and to conduct corresponding risk classification and early warning.

[0126] The slope safety hazard identification and early warning device provided in this embodiment of the invention has the same implementation principle and technical effect as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the device embodiment can be referred to the corresponding content in the aforementioned method embodiment.

[0127] This invention also provides an electronic device, such as... Figure 4 The diagram shows the structure of the electronic device, which includes a processor 41 and a memory 40. The memory 40 stores computer-executable instructions that can be executed by the processor 41. The processor 41 executes the computer-executable instructions to implement the above-mentioned slope safety hazard identification and early warning method.

[0128] exist Figure 4 In the illustrated embodiment, the electronic device further includes a bus 42 and a communication interface 43, wherein the processor 41, the communication interface 43, and the memory 40 are connected via the bus 42.

[0129] The memory 40 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 43 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 42 may be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. The bus 42 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 4 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0130] The processor 41 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the aforementioned slope safety hazard identification and early warning method can be completed through the integrated logic circuits in the processor 41 or through software instructions. The processor 41 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the slope safety hazard identification and early warning method disclosed in the embodiments of this invention can be directly manifested as execution by a hardware decoding processor, or as execution by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the field. The storage medium is located in the memory. The processor 41 reads the information in the memory and, in conjunction with its hardware, completes the steps of the slope safety hazard identification and early warning method of the aforementioned embodiment.

[0131] Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention.

[0132] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the 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 to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the slope safety hazard identification and early warning method described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0133] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0134] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for identifying and warning of slope safety hazards, characterized in that, include: Preprocessing of multi-source monitoring data for slope areas using different edge computing nodes; Dynamic spatiotemporal correlation features are extracted from the preprocessed data using the local model of each edge computing node; wherein, the local model is obtained by training the multi-scale spatiotemporal graph model of the corresponding edge computing node using federated learning on the central server; Based on preset physical constraints and corresponding preprocessed data, each edge computing node detects potential security risks in the corresponding dynamic spatiotemporal correlation features and performs corresponding risk classification and early warning. Also includes: Multi-source monitoring data is collected by different multi-source sensor networks deployed in the slope area, and the collected multi-source monitoring data is sent to different edge computing nodes; wherein each edge computing node corresponds to at least one multi-source sensor network deployed in the corresponding monitoring area of ​​the slope area. The multi-scale spatiotemporal graph model adopts a preset graph structure corresponding to the monitoring area. In the preset graph structure, nodes represent sensors and edges between nodes represent the spatiotemporal correlation between data collected by different sensors. Dynamic spatiotemporal correlation features are extracted from the preprocessed data through the local model of each edge computing node, including: Each local model generates a corresponding node feature matrix for the preprocessed data based on the corresponding preset graph structure. The temporal state of the corresponding node feature matrix is ​​estimated by each local model according to a preset time step; By fusing the corresponding node feature matrix with its estimated temporal state in each local model, the corresponding comprehensive features are obtained. The corresponding comprehensive features are classified by each local model to obtain the corresponding dynamic spatiotemporal correlation features; The preset physical constraint information includes preset physical constraint equations; based on the preset physical constraint information and the corresponding preprocessed data, security hazard detection is performed on the corresponding dynamic spatiotemporal correlation features, and corresponding risk classification and early warning are carried out, including: Based on the preset physical constraint equations and the corresponding preprocessed data, safety hazard detection is performed on the corresponding dynamic spatiotemporal correlation features. Based on the results of the safety hazard detection, a corresponding risk level assessment is conducted, and a corresponding early warning is triggered based on the results of the risk level assessment.

2. The method for identifying and warning of slope safety hazards according to claim 1, characterized in that, Preprocessing of multi-source monitoring data for slope areas using different edge computing nodes includes: Each edge computing node performs filtering, smoothing, and normalization processing on the multi-source monitoring data corresponding to the relevant monitoring area.

3. The method for identifying and warning of slope safety hazards according to claim 1, characterized in that, The preset graph structure includes an adjacency matrix representing the corresponding spatiotemporal relationships, and the local model includes a corresponding graph convolutional network model; each local model generates a corresponding node feature matrix for the corresponding preprocessed data based on the corresponding preset graph structure, including: For each local model, the corresponding graph convolutional network model performs iterative calculations on the preprocessed data based on the corresponding node feature matrix and adjacency matrix to generate the corresponding node feature matrix.

4. The method for identifying and warning of slope safety hazards according to claim 1, characterized in that, The local model includes a corresponding long short-term memory network model; based on a preset time step, temporal state estimation is performed on the corresponding node feature matrix using each local model, including: For each local model, based on a preset time step, corresponding temporal input data is generated for the corresponding node feature matrix, and the corresponding temporal state is estimated by the corresponding long short-term memory network model based on the corresponding temporal input data.

5. The method for identifying and warning of slope safety hazards according to claim 1, characterized in that, The local model includes a corresponding multi-level risk perception model; each local model classifies the corresponding comprehensive features to obtain the corresponding dynamic spatiotemporal correlation features, including: For each local model, the corresponding multi-level risk perception model predicts the corresponding hidden danger category based on the corresponding comprehensive characteristics, which serves as the corresponding dynamic spatiotemporal correlation feature.

6. A slope safety hazard identification and early warning device, characterized in that, include: The preprocessing module is used to preprocess multi-source monitoring data of the slope area through different edge computing nodes; The extraction module is used to dynamically extract spatiotemporal correlation features from the corresponding preprocessed data through the local model of each edge computing node; wherein, the local model is obtained by training the multi-scale spatiotemporal graph model of the corresponding edge computing node through federated learning by the central server; The detection and early warning module is used to detect potential security risks based on preset physical constraint information and corresponding preprocessed data of each edge computing node, and to conduct corresponding risk classification and early warning. It also includes a sending module, which is used for: Multi-source monitoring data is collected by different multi-source sensor networks deployed in the slope area, and the collected multi-source monitoring data is sent to different edge computing nodes; wherein each edge computing node corresponds to at least one multi-source sensor network deployed in the corresponding monitoring area of ​​the slope area. The multi-scale spatiotemporal graph model adopts a preset graph structure corresponding to the monitoring area. In the preset graph structure, nodes represent sensors and edges between nodes represent the spatiotemporal correlation between data collected by different sensors. The extraction module includes: Each local model generates a corresponding node feature matrix for the preprocessed data based on the corresponding preset graph structure. The temporal state of the corresponding node feature matrix is ​​estimated by each local model according to a preset time step; By fusing the corresponding node feature matrix with its estimated temporal state in each local model, the corresponding comprehensive features are obtained. The corresponding comprehensive features are classified by each local model to obtain the corresponding dynamic spatiotemporal correlation features; The preset physical constraint information includes preset physical constraint equations; the detection and early warning module includes: Based on the preset physical constraint equations and the corresponding preprocessed data, safety hazard detection is performed on the corresponding dynamic spatiotemporal correlation features. Based on the results of the safety hazard detection, a corresponding risk level assessment is conducted, and a corresponding early warning is triggered based on the results of the risk level assessment.

7. An electronic device, characterized in that, The method includes a processor and a memory, the memory storing computer-executable instructions that can be executed by the processor, the processor executing the computer-executable instructions to implement the slope safety hazard identification and early warning method according to any one of claims 1 to 5.

Images