Slope early warning method and system based on deep learning

Abstract

The invention discloses a slope early warning method and system based on deep learning, and particularly relates to the technical field of slope early warning, and the method comprises the steps: S1, multi-source data collection, S2, dynamic graph construction, S3, meta-learning model initialization, S4, space-time fusion prediction, S5, dynamic risk assessment, and S6, graded early warning triggering. Through multi-modal data fusion, an innovative model architecture and an intelligent early-warning mechanism, the slope early-warning capability can be remarkably improved, multi-source data are fused, a cross-modal attention mechanism is utilized, the slope state is comprehensively and accurately reflected, the early-warning accuracy is improved, a dynamic graph structure is constructed to be combined with a meta-learning engine, different slopes are adapted, continuous optimization can be achieved, and the early-warning capability of the slope is improved. Meanwhile, a scientific grading early warning system is established, a historical case library and related equipment are linked, resources are efficiently allocated, life and property safety is guaranteed, and disaster losses are reduced.

Description

Technical Field

[0001] The present invention relates to the field of slope warning technology, and more specifically, to a slope warning method and system based on deep learning. Background Art

[0002] In geological disaster prevention and control, slope stability monitoring and early warning are crucial to protecting people's lives and property and the stability of infrastructure. With the development of deep learning technology, its application in slope early warning is increasing. For example, the reservoir slope deformation early warning method based on deep learning algorithm adopts the fusion of the main prediction model and several secondary prediction models, and the landslide displacement prediction method based on graph deep learning uses global navigation satellite system positioning data.

[0003] However, existing technologies generally have limitations. Most technologies only focus on a single type of data input, the model generalization ability is insufficient, the early warning mechanism has weak adaptability, and the real-time and automation levels are low, making it difficult to meet the complex and changeable slope monitoring and early warning needs. In view of this, the present invention provides a slope early warning method and system based on deep learning. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, the present invention provides a slope early warning method and system based on deep learning to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a slope early warning method based on deep learning, specifically comprising the following steps:

[0006] S1. Multi-source data acquisition: synchronous acquisition of satellite remote sensing images, GNSS displacement monitoring data, microseismic signals, groundwater level data, and rainfall forecast information of the slope area;

[0007] S2. Dynamic graph construction, using monitoring points as nodes, and calculating the maximum principal stress difference between nodes based on real-time stress field analysis. ,like >Preset threshold , then establish node connections and assign weights , generating a dynamic adjacency matrix, wherein the preset threshold It is pre-set based on the geological conditions of the slope and historical monitoring data to determine whether there is stress correlation between monitoring points;

[0008] S3. Initialize the meta-learning model by inputting the slope, lithology, and vegetation coverage characteristics of the target slope into the pre-trained meta-network and outputting the initialization parameters of the multimodal spatiotemporal graph neural network. ;

[0009] S4, spatiotemporal fusion prediction: the data from step S1 is input into the MSTGNN model after spatiotemporal registration, spatial features are extracted through the graph convolution layer, temporal features are extracted through the gated recurrent unit, and heterogeneous data are fused through the cross-modal attention mechanism to output the displacement prediction value for the next 1 hour. ;

[0010] S5. Dynamic risk assessment, according to the formula Calculating risk index ,in, To predict the displacement, The rainfall in the next 24 hours. is the microseismic anomaly index, is the microseismic signal weight matrix, 、 and is the slope type related weight, is the Sigmoid activation function;

[0011] S6, graded warning trigger, when ≥0.3, start the warning, press The value interval performs a graded response:

[0012] 0.3≤ <0.6: trigger manual inspection instructions;

[0013] 0.6≤ <0.8: Triggering equipment restriction and evacuation preparation instructions:

[0014] ≥0.8: Trigger emergency evacuation instructions and path planning.

[0015] Preferably, step S2 specifically includes:

[0016] The node attribute vector is defined as ;

[0017] Recalculate the slope stress field and update the adjacency matrix every 6 hours based on the latest sensor data;

[0018] Strengthened connections are added to nodes in the fault zone area, and the weight is increased to 1.5 times the normal value.

[0019] Preferably, in step S3, the pre-trained meta-network is pre-trained on a dataset containing 100+ historical slope cases:

[0020] Initialization parameters After fine-tuning the target slope data for 7 days, the loss function adopts the MAE+ displacement curve second-order derivative regularization term: ,in, is the actual displacement, To predict the displacement, =0.05, is the second-order derivative of the displacement curve.

[0021] Preferably, the cross-modal attention mechanism of step S4 specifically includes:

[0022] Calculate the crack density index using DeepLabV3+ network on satellite images ;

[0023] The energy of the 5-30Hz abnormal frequency band is extracted from the microseismic signal through 1D-CNN and wavelet packet transform. ;

[0024] Attention weight calculation formula: ,in, 、 、 are optical, earthquake, and rainfall feature vectors, respectively. is the attention weight matrix, which is used to determine the attention weights of data of different modalities and realize heterogeneous data fusion.

[0025] Preferably, step S5 further includes:

[0026] S5.1. Automatically adjust the risk threshold based on the prediction error. If the prediction error is less than 5% for three consecutive times, the risk threshold will be lowered by 10%. If the false alarm rate is greater than 20%, the risk threshold will be raised by 15%.

[0027] S5.2. The weight setting rule is: the rock slope weight is set to =0.6, =0.3, = 0.1, then the soil slope is set to =0.4, =0.5, =0.1.

[0028] Preferably, step S6 further includes:

[0029] S6.1. Based on the historical disposal case library, use case-based reasoning algorithms to match the optimal disposal solution for the current scenario;

[0030] S6.2, when When ≥0.6, the drainage pump station and traffic signal equipment in the slope area will be controlled in a linked manner.

[0031] The present invention also provides a system for implementing the above-mentioned slope early warning method based on deep learning, comprising:

[0032] Multi-source sensing module, integrating SAR satellite receiver, GNSS displacement sensor, microseismic accelerometer, groundwater level meter and rainfall sensor;

[0033] a dynamic graph processor configured to calculate the slope stress field in real time and construct and update the node adjacency matrix;

[0034] A meta-learning engine with a built-in pre-trained meta-network that generates initialization parameters for a multimodal spatiotemporal graph neural network based on input slope features;

[0035] The spatiotemporal prediction core deploys a multimodal spatiotemporal graph neural network, including graph convolutional layers, gated recurrent units, and cross-modal attention modules;

[0036] Risk decision unit, execution risk index Calculation and classification of warning rules;

[0037] The feedback optimization module implements incremental learning through an elastic weight solidification algorithm and updates model parameters every 24 hours.

[0038] Preferably, the feedback optimization module specifically includes a short-term learning cycle: incremental training in a sliding window manner, constraining the update amplitude of important parameters: <0.1, long-term evolution cycle: retrain the meta-network every quarter and optimize the risk formula weights to reduce the false positive rate by 5-10%.

[0039] Preferably, in the multi-source perception module, the satellite image resolution is ≥0.5m, and the deformation monitoring accuracy reaches millimeter level;

[0040] When data is missing, the GAN generator is used to fill it in, and the generated data is correlated with the measured data. >0.92.

[0041] Preferably, it also includes:

[0042] Edge computing unit: NVIDIA Jetson Orin module is used to deploy the MSTGNN model, with response latency less than 3s;

[0043] LoRa emergency communication unit: transmits warning instructions to terminal devices, including risk area maps, evacuation routes and equipment control signals.

[0044] Technical effects and advantages of the present invention:

[0045] 1. The present invention achieves multimodal data fusion by synchronously collecting satellite remote sensing images, GNSS displacement monitoring data, microseismic signals, groundwater level data, rainfall forecast information, and video surveillance data. Utilizing technologies such as the cross-modal attention mechanism, it fully explores the correlations between different types of data. Compared with traditional slope early warning technologies that rely on a single or a few types of data, this technology can more comprehensively reflect the slope status and effectively avoid information omissions, thereby significantly improving the accuracy of slope risk warnings. For example, under complex geological conditions, the combination of microseismic signals and video surveillance data can promptly detect potential hazards such as internal rock fractures and surface rockfalls in the slope, providing early warnings.

[0046] 2. This invention constructs a dynamic graph structure based on real-time stress field analysis and generates customized model parameters for different slopes through a meta-learning engine. It can continuously optimize the model using short-term incremental learning and long-term model evolution. This enables the system to not only adapt to dynamic changes in slope conditions but also quickly adapt to different types of slopes (such as rock slopes and soil slopes) without requiring large amounts of data retraining. This solves the problems of poor generalization and adaptability to new scenarios faced by traditional models and reduces model deployment and maintenance costs.

[0047] 3. The present invention has established a scientific graded early warning mechanism, which accurately triggers different levels of early warning responses according to the risk index, from manual inspection to emergency evacuation, to ensure that the response measures match the risk level. At the same time, based on the historical disposal case library, the optimal disposal plan is generated through case reasoning algorithm, and the linkage control of drainage pumping stations, traffic signal equipment, etc. is realized. Compared with the traditional simple early warning method, the present invention can organize resources for risk disposal more efficiently, maximize the protection of human life and property safety, and reduce disaster losses. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is the overall flow chart of the present invention.

[0049] Figure 2 It is a system diagram of the present invention. DETAILED DESCRIPTION

[0050] Example 1

[0051] As attached Figure 1 As shown, an embodiment of the present invention provides a slope warning method based on deep learning, which specifically includes the following steps:

[0052] S1. Multi-source data acquisition: In the slope area, data is collected synchronously through multiple devices and systems, and the collected data will be transmitted to the data processing center for subsequent processing and analysis. Specifically:

[0053] Satellite remote sensing image acquisition uses satellites equipped with synthetic aperture radar to obtain image data of the slope area every 6 hours, with an image resolution of no less than 0.5m. At the same time, multi-temporal interferometry measurement technology is used to monitor the surface deformation rate with an accuracy of up to millimeter level;

[0054] GNSS displacement monitoring data collection: high-precision GNSS displacement sensors are deployed at key nodes on the slope. The sensor accuracy is ±0.5mm. Displacement data of each monitoring point is collected at a fixed frequency, such as once per minute, to obtain deformation information of the slope in the spatial dimension.

[0055] Microseismic signal acquisition: Install a microseismic accelerometer with a frequency response range of 0.1-100 Hz to collect microseismic signals generated by rock fracture and sliding inside the slope in real time, capturing the mechanical changes inside the slope;

[0056] Collect groundwater level data, arrange groundwater level meters, measure groundwater level once an hour, record dynamic changes in water level, and provide groundwater-related data for slope stability analysis;

[0057] Obtain rainfall forecast information: access the meteorological department's rainfall forecast system to obtain rainfall forecast data for the next 24 hours, divide it into four time periods, and calculate the equivalent infiltration volume of rainfall in each time period;

[0058] S2. Dynamic graph construction:

[0059] Node definition and attribute setting, each monitoring point on the slope is defined as a node in the graph structure, and the node attribute vector is set as ,The displacement data of each node comes from the GNSS displacement sensor, the vibration frequency is obtained by ,spectral analysis of the microseismic signal, and the groundwater level data is measured by a ,groundwater level meter;

[0060] Edge connection and weight calculation, based on real-time stress field analysis, calculate the maximum principal stress difference between each node The stress field analysis can be performed by using the finite element analysis method, combined with the geological structure of the slope, rock and soil mechanics parameters, etc. >Preset threshold (This threshold is pre-set based on the geological conditions of the slope, historical monitoring data and engineering experience. For example, for a certain type of rock slope, =1.5mpa), then a connection is established between the corresponding two nodes and a weight is assigned ,In this way, a dynamic adjacency matrix reflecting the stress correlation between ,the monitoring points on the slope is constructed;

[0061] Adjacency matrix updates: Every 6 hours, based on the latest sensor data, the slope stress field is recalculated, and the connection relationships and weights between nodes are updated. This allows the dynamic adjacency matrix to reflect changes in the mechanical state of the slope in real time. For nodes in areas with complex geological conditions and poor stability, such as fault zones, reinforced connections are added, and their connection weights are increased to 1.5 times the normal value to highlight the importance of these areas in slope stability analysis.

[0062] S3. Meta-learning model initialization:

[0063] Meta-network pre-training: Collect historical data on more than 100 different types of slopes (including rock slopes and soil slopes). This data includes information such as slope gradient, lithology, vegetation coverage, displacement monitoring data, and microseismic signals. Use this data to pre-train the meta-network. During the training process, use appropriate optimization algorithms (such as stochastic gradient descent) and loss functions to enable the meta-network to learn the mapping relationship between different slope characteristics and appropriate model parameters.

[0064] Target slope parameter generation and fine-tuning: For target slopes that require early warning monitoring, obtain characteristic information such as slope, lithology, and vegetation coverage. Input these characteristic vectors into the pre-trained meta-network, which outputs the initialization parameters of the multimodal spatiotemporal graph neural network. Then, using the 7-day monitoring data of the target slope (sampling interval is 2h), Fine-tune. During the fine-tuning process, MAE + displacement curve second-order derivative regularization term is used as the loss function: ,in, =0.05, by continuously adjusting the parameters, the prediction error of the model on the target slope is minimized, thereby improving the adaptability and prediction accuracy of the model;

[0065] S4. Spatiotemporal fusion prediction:

[0066] Data spatiotemporal registration: The satellite remote sensing images, GNSS displacement monitoring data, microseismic signals, groundwater level data, and rainfall forecast information collected in step S1 are uniformly projected into the same coordinate system. Bilinear interpolation is used to achieve alignment in time and space to ensure the consistency of each data in time and space, with an error control within ±0.1 pixel.

[0067] Feature extraction and fusion, the registered data is input into the MSTGNN model. In the model, the graph structure data is first processed by the graph convolution layer GCN to extract the features in the spatial dimension. The formula is ,in, is the dynamic adjacency matrix, is the node feature matrix, and then the gated recurrent unit GRU is used to process the time series data to capture the time-dependent features. The formula is For satellite images, the crack density index is calculated using the DeepLabV3+ network. , for microseismic signals, 1D-CNN and wavelet packet transform are used to extract the energy of the 5-30Hz abnormal frequency band Finally, the cross-modal attention mechanism is used to achieve the fusion of heterogeneous data. The attention weight calculation formula is: , input the fused features into the multi-layer perceptron MLP, and output the displacement prediction value for the next 1 hour ;

[0068] S5. Dynamic risk assessment:

[0069] Risk index calculation, according to the formula Calculating risk index ,in, To predict the displacement, The rainfall in the next 24 hours. is the microseismic anomaly index (which can be obtained by analyzing the characteristics of microseismic signals, such as the intensity and frequency of abnormal signals), is the microseismic signal weight matrix, 、 and is the slope type related weight. For rock slope, =0.6, =0.3, =0.1, for soil slope, set =0.4, =0.5, =0.1, is the Sigmoid activation function;

[0070] Risk threshold adjustment: automatically adjust the risk threshold based on the prediction error. If the prediction error is less than 5% for three consecutive times, it means that the model prediction accuracy is high, and the risk threshold can be lowered by 10% to improve the sensitivity of the warning; if the false alarm rate is greater than 20%, the risk threshold will be increased by 15% to reduce the occurrence of false alarms;

[0071] S6. Graded warning trigger:

[0072] Early warning activation and graded response, when the calculated risk index ≥0.3, the early warning mechanism is activated. Intervals of values ​​perform different graded responses:

[0073] When 0.3≤ When the value is less than 0.6, the manual inspection instruction is triggered and the relevant personnel are notified to conduct manual inspection of the slope to check for potential risks; when the value is less than 0.6, the manual inspection instruction is triggered and the relevant personnel are notified to conduct manual inspection of the slope to check for potential risks; When the value is less than 0.8, the equipment restriction and evacuation preparation instructions are triggered to restrict the operation of equipment near the slope and make preparations for the evacuation of personnel; when When the value is ≥0.8, emergency evacuation instructions and path planning are triggered, personnel are immediately organized to evacuate, and the system generates the optimal evacuation path to guide personnel to evacuate safely;

[0074] The treatment plan is generated and linked with the equipment. Based on the historical treatment case library, the optimal treatment plan for the current scenario is matched through case reasoning algorithm. When the value is ≥0.6, the drainage pump station and traffic signal equipment in the slope area will be controlled in a coordinated manner, such as starting the drainage pump station to drain water and adjusting the traffic signal to control vehicle traffic to ensure the safety of personnel and equipment.

[0075] Example 2

[0076] As attached Figure 2 As shown, an embodiment of the present invention further provides a system for implementing the above-mentioned slope early warning method based on deep learning, comprising:

[0077] The multi-source sensing module integrates a SAR satellite receiver, a GNSS displacement sensor, a microseismic accelerometer, a groundwater level meter, and a rainfall sensor. The SAR satellite receiver receives satellite remote sensing images at set time intervals. The GNSS displacement sensor, microseismic accelerometer, groundwater level meter, and rainfall sensor collect data in real time at their respective sampling frequencies. The collected data is transmitted to the data processing center via wired or wireless communication.

[0078] The dynamic graph processor, equipped with a dedicated computing unit, is used to calculate the slope stress field in real time. It receives monitoring data from the multi-source perception module, performs calculations based on a preset stress field calculation model (such as a finite element model), and constructs and updates the node adjacency matrix based on the calculation results. This processor has efficient data processing capabilities and can complete the dynamic graph construction and update tasks within the specified time, providing accurate data support for subsequent analysis.

[0079] The meta-learning engine has a built-in pre-trained meta-network that stores model parameters trained using a large amount of historical slope data. Upon receiving the characteristic information of the target slope, the meta-network quickly generates the initialization parameters for the multimodal spatiotemporal graph neural network. The engine is also responsible for fine-tuning these initialization parameters, interacting with the data processing center to obtain monitoring data of the target slope, and optimizing the parameters according to the set fine-tuning algorithm and loss function to ensure the applicability of the model on different slopes.

[0080] The spatiotemporal prediction core deploys the MSTGNN model, including graph convolutional layers, gated recurrent units, and cross-modal attention modules. It receives multi-source data after spatiotemporal registration, sequentially performs spatial feature extraction, temporal feature extraction, and cross-modal data fusion operations, and ultimately outputs displacement prediction values. This core uses high-performance computing chips (such as the NVIDIA Jetson Orin module) to ensure that the model can run quickly and complete prediction tasks in a short time, meeting the needs of real-time warning.

[0081] The risk decision unit receives the displacement prediction value from the spatiotemporal prediction core and other relevant data (such as rainfall, microseismic signals, etc.), and calculates the risk index according to the preset risk index calculation formula and graded warning rules. And determine whether to trigger an early warning. When an early warning is triggered, follow The value interval generates corresponding warning instructions and disposal suggestions, and transmits the warning information to the LoRa emergency communication unit;

[0082] Feedback optimization module:

[0083] In the short-term learning cycle, the feedback optimization module collects new monitoring data in a sliding window manner and performs incremental training on the model. During the training process, the elastic weight solidification algorithm is used to constrain the update range of important parameters. <0.1, preventing the model from forgetting previously learned knowledge when learning new data, ensuring the stability and accuracy of the model;

[0084] A long-term evolutionary cycle uses quarterly feedback from the optimization module to retrain the meta-network, incorporating newly accumulated slope case data into the training set. This optimizes the meta-network parameters to better adapt to the characteristics of different slopes. At the same time, the weights in the risk formula are optimized and adjusted. Through extensive verification and analysis of actual data, the false alarm rate is reduced by 5-10%, continuously improving system performance.

[0085] The edge computing unit uses the NVIDIA Jetson Orin module and deploys the MSTGNN model. This unit has powerful computing capabilities and can quickly process data locally, enabling real-time inference and prediction of the model with a response delay of less than 3 seconds, meeting the real-time requirements of slope warning.

[0086] The LoRa emergency communication unit is responsible for transmitting early warning instructions to the terminal device. The instruction content includes information such as risk area map, evacuation route and equipment control signals.

[0087] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A slope early warning method based on deep learning, characterized by: The specific steps include: S1. Multi-source data acquisition: synchronous acquisition of satellite remote sensing images, GNSS displacement monitoring data, microseismic signals, groundwater level data, and rainfall forecast information of the slope area; S2. Dynamic graph construction, using monitoring points as nodes, and calculating the maximum principal stress difference between nodes based on real-time stress field analysis. ,like >Preset threshold , then establish node connections and assign weights , generating a dynamic adjacency matrix, wherein the preset threshold It is pre-set based on the geological conditions of the slope and historical monitoring data to determine whether there is stress correlation between monitoring points; S3. Initialize the meta-learning model by inputting the slope, lithology, and vegetation coverage characteristics of the target slope into the pre-trained meta-network and outputting the initialization parameters of the multimodal spatiotemporal graph neural network. ; S4, spatiotemporal fusion prediction: the data from step S1 is input into the MSTGNN model after spatiotemporal registration, spatial features are extracted through the graph convolution layer, temporal features are extracted through the gated recurrent unit, and heterogeneous data are fused through the cross-modal attention mechanism to output the displacement prediction value for the next 1 hour. ; S5. Dynamic risk assessment, according to the formula Calculating risk index ,in, To predict the displacement, The rainfall in the next 24 hours. is the microseismic anomaly index, is the microseismic signal weight matrix, 、 and is the slope type related weight, is the Sigmoid activation function; S6, graded warning trigger, when ≥0.3, start the warning, press The value interval performs a graded response: 0.3≤ <0.6: trigger manual inspection instructions; 0.6≤ <0.8: Triggering equipment restriction and evacuation preparation instructions: ≥0.8: Trigger emergency evacuation instructions and path planning.

2. The deep learning-based slope early warning method according to claim 1, characterized in that: Step S2 specifically includes: The node attribute vector is defined as ; Recalculate the slope stress field and update the adjacency matrix every 6 hours based on the latest sensor data; Strengthened connections are added to nodes in the fault zone area, and the weight is increased to 1.5 times the normal value.

3. The deep learning-based slope early warning method according to claim 2, characterized in that: In step S3, the pre-trained meta-network is pre-trained on a dataset containing 100+ historical slope cases; Initialization parameters After fine-tuning the target slope data for 7 days, the loss function adopts the MAE+ displacement curve second-order derivative regularization term: ,in, is the actual displacement, To predict the displacement, =0.05, is the second-order derivative of the displacement curve.

4. The deep learning-based slope early warning method according to claim 3, characterized in that: The cross-modal attention mechanism of step S4 specifically includes: Calculate the crack density index using DeepLabV3+ network on satellite images ; The energy of the 5-30Hz abnormal frequency band is extracted from the microseismic signal through 1D-CNN and wavelet packet transform. ; Attention weight calculation formula: ,in, 、 、 are optical, earthquake, and rainfall feature vectors, respectively. is the attention weight matrix, which is used to determine the attention weights of data of different modalities and realize heterogeneous data fusion.

5. The deep learning-based slope early warning method according to claim 4, characterized in that: Step S5 also includes: S5.

1. Automatically adjust the risk threshold based on the prediction error. If the prediction error is less than 5% for three consecutive times, the risk threshold will be lowered by 10%. If the false alarm rate is greater than 20%, the risk threshold will be raised by 15%. S5.

2. The weight setting rule is: the rock slope weight is set to =0.6, =0.3, = 0.1, then the soil slope is set to =0.4, =0.5, =0.

1.

6. The deep learning-based slope early warning method according to claim 5, characterized in that: Step S6 also includes: S6.

1. Based on the historical disposal case library, use case-based reasoning algorithms to match the optimal disposal solution for the current scenario; S6.2, when When ≥0.6, the drainage pump station and traffic signal equipment in the slope area will be controlled in a linked manner.

7. A system for implementing the deep learning-based slope early warning method according to claim 6, characterized in that: include: Multi-source sensing module, integrating SAR satellite receiver, GNSS displacement sensor, microseismic accelerometer, groundwater level meter and rainfall sensor; a dynamic graph processor configured to calculate the slope stress field in real time and construct and update the node adjacency matrix; A meta-learning engine with a built-in pre-trained meta-network that generates initialization parameters for a multimodal spatiotemporal graph neural network based on input slope features; The spatiotemporal prediction core deploys a multimodal spatiotemporal graph neural network, including graph convolutional layers, gated recurrent units, and cross-modal attention modules; Risk decision unit, execution risk index Calculation and classification of warning rules; The feedback optimization module implements incremental learning through an elastic weight solidification algorithm and updates model parameters every 24 hours.

8. The system according to claim 7, characterized in that: The feedback optimization module specifically includes a short-term learning cycle: incremental training in a sliding window manner, constraining the update amplitude of important parameters: <0.1, long-term evolution cycle: retrain the meta-network every quarter and optimize the risk formula weights to reduce the false positive rate by 5-10%.

9. The system according to claim 8, characterized in that: In the multi-source perception module, the satellite image resolution is ≥0.5m, and the deformation monitoring accuracy reaches millimeter level; When data is missing, the GAN generator is used to fill it in, and the generated data is correlated with the measured data. >0.

92.

10. The system according to claim 9, characterized in that: Also includes: Edge computing unit: NVIDIA Jetson Orin module is used to deploy the MSTGNN model, with response latency less than 3s; LoRa emergency communication unit: transmits warning instructions to terminal devices, including risk area maps, evacuation routes and equipment control signals.

Images