Geological disaster monitoring and early warning system based on deep learning
By simultaneously collecting and calibrating multiple monitoring data, constructing multi-dimensional collaborative features, and dynamically adjusting benchmark thresholds, the data deviation and early warning deviation problems of existing earthquake early warning systems have been solved, thereby improving the accuracy and adaptability of earthquake early warning.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HENAN PROVINCIAL SECOND INST OF RESOURCE & ENVIRONMENTAL INVESTIGATION CO LTD
- Filing Date
- 2026-04-30
- Publication Date
- 2026-06-19
AI Technical Summary
Existing earthquake monitoring and early warning systems do not simultaneously access data from multiple monitoring devices and lack data calibration steps, resulting in acquisition delays and data deviations that affect the accuracy of feature extraction and judgment results. They also lack dynamic benchmark thresholds and weighting coefficients, leading to early warning deviations. Furthermore, they lack parameter iterative optimization mechanisms, which affect the timeliness and adaptability of early warning responses.
The system synchronously calls upon base stations, basic stations, and meteorological and geological monitoring equipment to collect various monitoring data and calibrate them according to timestamps. Abnormal data is removed, multi-dimensional collaborative feature parameters are constructed, and comprehensive feature values are calculated by combining dynamic weight coefficients. The baseline threshold is dynamically adjusted, and the risk level is confirmed by combining data from adjacent monitoring points. The priority and delay of early warning push are dynamically adjusted, and the parameters are iteratively optimized by comparing data after the early warning.
Through multi-source data collaborative processing and dynamic threshold adjustment, the earthquake early warning system has achieved precision, real-time capability, and adaptability, improving the accuracy of early warning judgment and the precision of push notifications, and ensuring the adaptability and timeliness of early warning information.
Smart Images
Figure CN122245034A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of earthquake early warning technology, and in particular to a geological disaster monitoring and early warning system based on deep learning. Background Technology
[0002] The field of earthquake early warning technology encompasses related technologies such as seismic wave propagation monitoring, earthquake event identification, earthquake intensity determination, and early warning information generation and dissemination. Based on earthquake monitoring data, this field achieves rapid response after an earthquake through data acquisition, analysis, and event determination. The core content includes acquiring earthquake monitoring signals, identifying initial earthquake disturbances, calculating earthquake parameters, and pushing early warning information to affected areas. The overall technology field covers the complete technical aspects of ground monitoring station deployment, monitoring signal acquisition and transmission, earthquake event feature extraction, early warning threshold determination, and early warning command issuance.
[0003] Among them, the geological disaster monitoring and early warning system based on deep learning refers to a system that uses a deep learning model to process earthquake monitoring data and realize earthquake early warning. This patent topic addresses technical matters such as earthquake monitoring waveform data processing, earthquake P-wave identification, earthquake magnitude estimation, and earthquake impact range determination. It extracts earthquake monitoring waveform features by constructing a convolutional neural network model, analyzes the temporal features of earthquake data using a recurrent neural network model, completes model training and parameter updates based on a labeled earthquake monitoring dataset, classifies and determines the real-time collected earthquake monitoring data using the trained deep learning model, and generates early warning-related information by combining preset earthquake parameter determination rules.
[0004] Existing technologies only process earthquake monitoring waveform data, without simultaneously calling multiple monitoring devices to collect various types of data. The lack of data calibration steps can easily lead to acquisition delays and data deviations. The lack of data classification and preprocessing means that abnormal data can affect the accuracy of feature extraction and judgment results. The absence of dynamic benchmark thresholds and weighting coefficients makes it difficult to accurately determine risk levels. The lack of parameter iteration and optimization mechanisms means that adjustments cannot be made according to actual disaster conditions, which can easily lead to early warning deviations and affect the timeliness and adaptability of early warning responses. Summary of the Invention
[0005] The main objective of this invention is to provide a geological disaster monitoring and early warning system based on deep learning, which can effectively solve the problems mentioned in the background art.
[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The deep learning-based geological disaster monitoring and early warning system includes the following steps: Synchronously calling benchmark stations, basic stations, general stations, and meteorological and geological monitoring equipment to collect seismic motion parameters, environmental data, meteorological data, and regional geological parameters; aligning and calibrating the acquisition delay and data deviation according to timestamps to obtain synchronous calibration data; preprocessing the synchronous calibration data according to parameter type, removing abnormal data and supplementing and improving it to obtain valid data; calling the valid data to construct multi-dimensional collaborative feature parameters, and calculating a comprehensive feature value by combining dynamically adjusted weight coefficients; determining a dynamic benchmark threshold based on real-time anomaly-free data in the monitoring area, judging the risk level by comparing the comprehensive feature value with the benchmark threshold, and confirming the validity of the level by combining data from adjacent monitoring points; dynamically adjusting the early warning push priority and delay according to the confirmed risk level and the location parameters of monitoring points and user terminals, accurately pushing early warning information adapted to different terminals; calling real-time monitoring data after the early warning and the actual disaster situation, comparing the deviation between the early warning data and the actual data, adjusting parameters such as feature weights and benchmark thresholds, and iteratively optimizing.
[0007] Preferably, the acquisition of multi-source monitoring data specifically involves: calling the reference station's velocity meter, accelerometer, and environmental monitoring equipment to collect ground motion velocity, acceleration, and environmental monitoring data, and transmitting them via a dedicated fiber optic line with a communication bandwidth of ≥2M; calling the basic station's accelerometer and environmental monitoring equipment to collect ground motion acceleration and environmental monitoring data, and transmitting them via a dedicated fiber optic line with a communication bandwidth of ≥2M; calling the general station's intensity meter to collect ground motion acceleration data, and transmitting it via a wireless network; and simultaneously calling the hourly rainfall data from meteorological monitoring equipment and the stratigraphic hardness and fault distribution density data from regional geological monitoring equipment.
[0008] Preferably, the calibration of multi-source monitoring data is specifically as follows: align all collected data by timestamp, calculate the acquisition delay of each type of data at the same timestamp, and when the acquisition delay of ground motion parameters exceeds 0.5 seconds or the acquisition delay of environmental and meteorological parameters exceeds 1 second, re-collect the real-time data of the corresponding equipment; calculate the acquisition difference of the same monitoring parameter at three adjacent monitoring points, and when the difference exceeds 5% of the acquisition range of the parameter, call the equipment self-test data to judge the operating status. If the equipment is fault-free, take the average of the data from the three monitoring points as valid data; if the equipment is faulty, discard the faulty data.
[0009] Preferably, the data preprocessing and anomaly removal are as follows: Synchronous calibration data are processed according to parameter type. For ground motion velocity and acceleration data, the moving average value over the past 10 minutes is calculated. If the absolute value of the difference between the real-time acquired value and the moving average value is greater than 30% of the moving average value, it is judged as abnormal data, and the adjacent 5 minutes of normal data are called up and supplemented by linear interpolation. For environmental and meteorological data, the deviation between the acquired value and the average value over the past 1 hour is calculated. If the deviation exceeds 20% of the average value, the abnormal data is removed and data from adjacent monitoring points in the same area are called up to supplement it. It is confirmed that the time of acquisition of regional geological parameters is no more than 72 hours from the present. Otherwise, the data is re-acquired.
[0010] Preferably, the construction of collaborative feature parameters and the calculation of comprehensive feature values are specifically as follows: The ground motion velocity, acceleration, hourly rainfall, stratum hardness, and fault distribution density from the valid data are retrieved; the ratio of ground motion velocity to acceleration, the ratio of hourly rainfall to the average rainfall over the past 24 hours, the product of fault distribution density and stratum hardness, and the rate of change of ground motion acceleration are calculated; the corresponding weighting coefficients are retrieved, which are dynamically adjusted based on the geological type of the monitoring area and the rainfall over the past 24 hours; the sum of the products of each feature parameter and its corresponding weighting coefficient is calculated to obtain the comprehensive feature value.
[0011] Preferably, the determination of the dynamic benchmark threshold and the judgment of risk level are as follows: the benchmark threshold is determined by twice the average of the comprehensive characteristic values of the monitoring area with no abnormal monitoring data in the past hour; the ratio of the comprehensive characteristic value to the benchmark threshold is calculated, and a ratio ≥2.0 is judged as high risk, 1.5~2.0 as medium risk, 1.0~1.5 as low risk, and <1.0 as no risk; the average of the comprehensive characteristic values of the three adjacent monitoring points is calculated simultaneously, and the calculation is recalculated when the difference between the average and the current monitoring point's comprehensive characteristic value exceeds 15%; if two or more of the three adjacent monitoring points are consistent with the current risk level, the risk level is confirmed to be valid.
[0012] Preferably, the dynamic adjustment of the warning level is specifically as follows: call the confirmed risk level and the longitude and latitude parameters of the monitoring point, calculate the distance between the monitoring point and the warning center; determine the push priority according to the risk level and distance, push to high-risk areas first, the push delay for high-risk areas shall not exceed 1 second, for medium-risk areas not exceed 2 seconds, and for low-risk areas not exceed 3 seconds.
[0013] Preferably, the precise push of early warning information is as follows: calling the user terminal type parameter, mobile terminals push text warnings containing risk level, comprehensive feature value, expected impact range and risk avoidance suggestions; TV and radio terminals push voice + text warnings; terminals in key industries such as railways, power grids and nuclear power push customized warnings containing specific emergency response instructions; after the push, the reach rate is counted, and if the reach rate is less than 95%, the push interface is called again for a second push.
[0014] Preferably, the closed-loop data correction and parameter iteration are as follows: Real-time monitoring data within 10 minutes of the warning being issued is retrieved, the comprehensive characteristic value and risk level are recalculated, and the deviation value is calculated by comparing it with the data at the time of the warning; if the deviation value exceeds 10% of the comprehensive characteristic value at the time of the warning, the characteristic parameter weight coefficient is adjusted and recalculated; if the adjusted deviation value is less than 10%, the new weight coefficient is stored; actual earthquake occurrence data during the warning period is retrieved, and the matching degree between the warning risk level and the actual damage level is calculated; if the matching degree is less than 80%, the benchmark threshold calculation coefficient is adjusted until the matching degree is ≥80%; the adjusted parameters are synchronously applied to subsequent warning processes.
[0015] Preferably, the deep learning model is specifically applied to comprehensive feature value calculation, dynamic benchmark threshold adjustment, weight coefficient iteration, and risk level judgment. The model is continuously trained by real-time monitoring data to improve the accuracy of parameter calculation and risk judgment.
[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention simultaneously collects data from multiple monitoring devices, aligns and calibrates data according to timestamps to reduce data deviation, classifies and preprocesses data to remove abnormal data to ensure data validity, constructs multi-dimensional collaborative features and calculates comprehensive feature values by combining dynamic weights, determines dynamic benchmark thresholds based on real-time abnormal data, confirms the validity of risk levels by combining data from adjacent monitoring points, dynamically adjusts the priority and delay of early warning push to achieve accurate push, and uses post-warning data comparison to adjust parameters for iterative optimization, thereby improving the accuracy of early warning judgment and push precision, ensuring the adaptability of early warning information, and achieving continuous improvement in early warning performance. Attached Figure Description
[0017] Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 This is a schematic diagram of the data processing and feature calculation process of the present invention; Figure 3 This is a schematic diagram of the risk assessment and early warning push process of the present invention; Figure 4 This is a schematic diagram of the closed-loop iterative optimization process of the system according to the present invention. Detailed Implementation
[0018] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.
[0019] Please see Figures 1 to 4This invention discloses a geological disaster monitoring and early warning system based on deep learning. Its core purpose is to solve the technical pain points of existing geological disaster monitoring and early warning systems, such as asynchronous data acquisition, low early warning accuracy, and fixed parameters that cannot be adaptively optimized. Through multi-source data collaborative processing, dynamic threshold adjustment, closed-loop iterative optimization, and deep application of deep learning models, the system achieves more accurate, real-time, and adaptive upgrades for geological disaster early warning. The following detailed description of the specific working principle, operation process, and calculation process of this system is provided in conjunction with specific embodiments.
[0020] Example 1: This example primarily focuses on the acquisition and synchronous calibration of multi-source geological disaster monitoring data, constructing a unified time-series basic data set. Specifically, it includes: synchronously accessing monitoring equipment from reference stations, basic stations, general stations, and meteorological and geological monitoring equipment to collect seismic motion parameters, environmental data, meteorological data, and regional geological parameters. In the implementation process, the reference station is equipped with velocity sensors, accelerometers, and environmental monitoring equipment. The velocity sensor is a VS-100 model, the accelerometer is an AC-200 model, and the environmental monitoring equipment is an EN-300 model environmental monitor. It collects seismic velocity, seismic acceleration, and environmental monitoring data. Data transmission uses a 2.5M fiber optic leased line to ensure data transmission stability and real-time performance. The basic station is equipped with an AC-200 model accelerometer and an EN-300 model environmental monitor to collect seismic acceleration and environmental monitoring data. The same 2.5M fiber optic dedicated line is used for transmission; the general station is equipped with an LD-400 intensity meter to collect ground motion acceleration data, and uses a 4G wireless network for transmission, with a transmission rate stable at over 10Mbps; the QX-500 meteorological monitoring equipment is used simultaneously to collect hourly rainfall data, and the DZ-600 regional geological monitoring equipment is used to collect stratum hardness and fault distribution density data, with a stratum hardness acquisition accuracy of 0.1MPa and a fault distribution density acquisition accuracy of 0.01 faults / km². The sampling frequency of all acquisition equipment is uniformly set to 10Hz to ensure that all types of data can be acquired synchronously. All collected data were aligned using timestamps with a precision of 0.01 seconds. The acquisition delay of each parameter at the same timestamp was calculated. If the acquisition delay of the seismic motion parameter exceeded 0.5 seconds, the corresponding equipment was re-acquired to collect real-time data for supplementary acquisition. If the acquisition delay of the environmental and meteorological parameters exceeded 1 second, supplementary acquisition was performed. The difference between the collected data from three adjacent monitoring points for the same monitoring parameter was calculated. The seismic acceleration acquisition range is 0-1 m / s², and the 5% value of the range is 0.05 m / s². The seismic acceleration acquisition values of the three adjacent monitoring points were 0.3 m / s², 0.32 m / s², and 0.31 m / s², respectively. The difference between these values was less than 0.05 m / s². After retrieving the equipment self-test data to confirm that there were no faults, the average value of the three values, 0.31 m / s², was taken as the valid data for that timestamp. Finally, the synchronous calibration of all data was completed, and the synchronous calibration data was obtained.
[0021] Example 2, based on the synchronous calibration data in Example 1, further implements data classification preprocessing and abnormal data removal and supplementation to obtain standardized and effective data. Specifically, it includes: classifying and processing the synchronous calibration data according to parameter type; for ground motion velocity and acceleration data, calculating the sliding average value with a 10-minute sliding window. In the specific implementation, the 10-minute sliding average value of ground motion velocity is 0.6 m / s, and a certain real-time acquired value is 0.85 m / s. The absolute value of the difference between the two is 0.25 m / s. This value is greater than 30% of the sliding average value, i.e., 0.18 m / s, so the acquired value is determined to be abnormal data, and the monitoring data is retrieved. Within a 5-minute interval, the normally collected data were 0.62 m / s, 0.58 m / s, and 0.61 m / s. A valid data point of 0.6 m / s was obtained by linear interpolation. For environmental and meteorological data, the average value was calculated with a 1-hour period. If the deviation of a certain environmental data point from the average value exceeded 20%, the data was directly discarded, and data from adjacent monitoring points in the same area were retrieved to supplement it. The timeliness of the regional geological parameters was checked. The stratigraphic hardness data was collected 48 hours ago, and the fault distribution density data was collected 36 hours ago; neither exceeded the 72-hour timeliness requirement, so no re-collection was necessary. After the above processing, all data were valid.
[0022] Example 3, based on the effective data of Example 2, further constructs multi-dimensional collaborative feature parameters and completes the calculation of comprehensive feature values. Simultaneously, it determines dynamic benchmark thresholds and completes the geological hazard risk level assessment and validity verification. Specifically, it includes: calling up the effective data for ground motion velocity (0.6 m / s), ground motion acceleration (0.31 m / s²), hourly rainfall (5 mm), stratum hardness (5 MPa), and fault distribution density (0.3 faults / km²), and sequentially calculating the ratio of ground motion velocity to acceleration, the ratio of hourly rainfall to the average rainfall over the past 24 hours, the product of fault distribution density and stratum hardness, and the rate of change of ground motion acceleration. In the specific implementation process, the ratio of ground motion velocity to acceleration is 1.94, the average rainfall over the past 24 hours is 2.5 mm, the ratio of hourly rainfall to the mean is 2.0, the product of fault distribution density and stratum hardness is 1.5, and the rate of change of ground motion acceleration is 0.02 m / s³. Weighting coefficients were set according to the geological type of the monitoring area: the weight of the ground motion parameter ratio was 0.3, the weight of the rainfall parameter ratio was 0.2, the weight of the geological vulnerability parameter was 0.3, and the weight of the acceleration change rate was 0.2. Since the rainfall in the past 24 hours did not reach the adjustment threshold, the weighting coefficients remained unchanged. Each characteristic parameter was multiplied by its corresponding weighting coefficient and then summed: 1.94 × 0.3 + 2.0 × 0.2 + 1.5 × 0.3 + 0.02 × 0.2, resulting in a final value of 1.436. This value is the comprehensive feature value. The entire parameter calculation was performed using a deep learning model. The comprehensive feature values corresponding to the monitoring data with no anomalies in the past hour were extracted: 1.38, 1.45, 1.41, 1.39, 1.42, and 1.40. The mean was calculated to be 1.41, and twice this mean, 2.82, was set as the dynamic baseline threshold. The ratio of the comprehensive feature value of 1.436 to the dynamic baseline threshold is 0.51, which is less than 1.0, initially indicating no risk. The comprehensive feature values of three adjacent monitoring points (1.42, 1.45, and 1.44) are retrieved, and their average is 1.437. The difference between this average and the current monitoring point's comprehensive feature value is 0.001, which is less than 15% of the current comprehensive feature value (0.2154). Since all three adjacent monitoring points are also determined to be risk-free, the current risk level is confirmed to be valid. The risk level classification rule is: a ratio greater than or equal to 2.0 indicates high risk, 1.5 to 2.0 indicates medium risk, 1.0 to 1.5 indicates low risk, and less than 1.0 indicates no risk. All judgment processes are completed using a deep learning model.
[0023] Example 4, building upon the risk level confirmation in Example 3, further achieves dynamic adjustment of warning levels and precise delivery of warning information across multiple terminals. Specifically, it includes: retrieving risk level and monitoring point latitude and longitude parameters; calculating the spherical distance between the monitoring point and the warning center; and setting delivery priorities based on risk level and distance. The delivery delay for high-risk areas is no more than 1 second, for medium-risk areas no more than 2 seconds, and for low-risk areas no more than 3 seconds. The closer the monitoring point is to the warning center, the higher the delivery priority. In a specific implementation, a monitoring point has a longitude of 110.000° and a latitude of 30.000°, while the warning center has a longitude of 110.500° and a latitude of 30.500°. The calculated straight-line distance between them is 61.2 km. The system retrieves user terminal type parameters. Mobile terminals push text warnings including risk level, comprehensive characteristic value, expected impact range, and risk avoidance suggestions. The expected impact range is determined based on the monitoring point location and risk level: high-risk impact range is 5km around the monitoring point, medium-risk is 3km around the monitoring point, and low-risk is 1km around the monitoring point. Television and radio terminals push voice and text warnings, with voice broadcasts at 10-second intervals and text information continuously displayed until the warning is lifted. Terminals in key industries such as railways, power grids, and nuclear power push customized warnings corresponding to emergency response instructions. For example, railway terminals push "Immediately reduce speed to 20km / h and stop at the nearest station," and power grid terminals push "Cut off power lines in high-risk areas and activate backup power." After the push is completed, the reach rate of the warning information is statistically analyzed in real time. The calculated reach rate is 93%, which is lower than 95%. The push interface is called again to push the warnings to the unreached terminals, and the reach rate is improved to 98%, achieving accurate delivery of warning information.
[0024] Example 5, building upon the early warning push completed in Example 4, further implements closed-loop data correction, iterative optimization of core system parameters, and full-process application of deep learning models. Specifically, it includes: retrieving real-time monitoring data within 10 minutes of the early warning release, recalculating the comprehensive feature value to 1.46, comparing it with the comprehensive feature value of 1.436 at the time of the early warning, and finding a deviation of 0.024. This value is less than 10% of the comprehensive feature value at the time of the early warning, i.e., 0.1436, so no adjustment of the feature parameter weight coefficient is needed. If an actual geological disaster occurs during the early warning period, the matching degree between the early warning risk level and the actual degree of damage is calculated. When the matching degree is 75% but less than 80%, the baseline threshold calculation coefficient is adjusted, changing the original 2x coefficient to 1.9x, recalculating the baseline threshold to 2.679, and recalculating the ratio of the comprehensive feature value to the threshold. The coefficient is continuously adjusted until the matching degree reaches above 80%. The finally adjusted weight coefficient and baseline threshold calculation coefficient are then synchronously applied to subsequent early warning processes to complete parameter iterative optimization. The deep learning model employs a fusion architecture of convolutional neural networks and recurrent neural networks. The model takes into account multi-source collected data, synchronously calibrated data, and preprocessed valid data, and outputs comprehensive feature values, dynamic baseline thresholds, weight coefficients, and risk levels. Each model training iteration uses 1000 sets of valid monitoring data, with 500 iterations and a learning rate of 0.001. After training, the model's parameter calculation accuracy reaches 99.2%, and its risk assessment accuracy reaches 98.5%. The model incorporates comprehensive feature value calculation, dynamic threshold adjustment, weight coefficient iteration, and risk level determination throughout the entire process. It is continuously trained and updated using real-time collected monitoring data, constantly optimizing the computational logic and judgment rules to enhance the adaptive capabilities of the geological disaster monitoring and early warning system, replacing the traditional fixed threshold and experience-based judgment methods.
[0025] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended technical solutions and their equivalents.
Claims
1. A geological disaster monitoring and early warning system based on deep learning, characterized in that, Includes the following steps: The system synchronously calls upon reference stations, basic stations, general stations, and meteorological and geological monitoring equipment to collect seismic motion parameters, environmental data, meteorological data, and regional geological parameters. It aligns and calibrates the acquisition delay and data deviation according to the timestamp to obtain synchronous calibration data. The synchronous calibration data is preprocessed according to parameter type, abnormal data is removed and supplemented to obtain effective data. The effective data is used to construct multi-dimensional collaborative feature parameters, and the comprehensive feature value is calculated by combining dynamically adjusted weight coefficients. A dynamic baseline threshold is determined based on real-time anomaly-free data of the monitoring area. The risk level is judged by comparing the comprehensive feature value with the baseline threshold, and the validity of the level is confirmed by combining data from adjacent monitoring points. Based on the confirmed risk level and the location parameters of monitoring points and user terminals, the priority and delay of early warning push are dynamically adjusted to accurately push early warning information adapted to different terminals. After the early warning is invoked, real-time monitoring data and actual disaster conditions are performed. The deviation between the early warning data and the actual data is compared, and parameters such as feature weights and benchmark thresholds are adjusted and iteratively optimized.
2. The geological disaster monitoring and early warning system based on deep learning according to claim 1, characterized in that, The acquisition of multi-source monitoring data specifically involves: calling the base station velocity meter, accelerometer and environmental monitoring equipment to collect ground motion velocity, acceleration and environmental monitoring data, and transmitting them using a dedicated fiber optic line with a communication bandwidth of ≥2M; The basic station accelerometer and environmental monitoring equipment are used to collect ground motion acceleration and environmental monitoring data, which are transmitted using a dedicated fiber optic line with a communication bandwidth of ≥2M. The seismic acceleration data was collected by using a general-purpose intensity meter and transmitted via wireless network. Simultaneously access hourly rainfall data from meteorological monitoring equipment and stratigraphic hardness and fault distribution density data from regional geological monitoring equipment.
3. The geological disaster monitoring and early warning system based on deep learning according to claim 1, characterized in that, The calibration of multi-source monitoring data is as follows: align all collected data by timestamp, calculate the acquisition delay of each type of data at the same timestamp, and when the acquisition delay of ground motion parameters exceeds 0.5 seconds or the acquisition delay of environmental and meteorological parameters exceeds 1 second, re-collect the real-time data of the corresponding equipment; calculate the acquisition difference of the same monitoring parameter at three adjacent monitoring points, and when the difference exceeds 5% of the acquisition range of the parameter, call the equipment self-test data to judge the operating status. If the equipment is fault-free, take the average of the data from the three monitoring points as the valid data; if the equipment is faulty, discard the faulty data.
4. The geological disaster monitoring and early warning system based on deep learning according to claim 1, characterized in that, The data preprocessing and anomaly removal are as follows: Synchronous calibration data are processed according to parameter type. For ground motion velocity and acceleration data, the moving average value over the past 10 minutes is calculated. If the absolute value of the difference between the real-time acquired value and the moving average value is greater than 30% of the moving average value, it is judged as abnormal data, and the adjacent 5 minutes of normal data are called up and supplemented by linear interpolation. For environmental and meteorological data, the deviation between the acquired value and the average value over the past 1 hour is calculated. If the deviation exceeds 20% of the average value, the abnormal data is removed and data from adjacent monitoring points in the same area are called up to supplement it. It is confirmed that the geological parameter acquisition time in the area is no more than 72 hours from the present. Otherwise, the data is re-acquired.
5. The geological disaster monitoring and early warning system based on deep learning according to claim 1, characterized in that, The collaborative feature parameter construction and comprehensive feature value calculation are as follows: The ground motion velocity, acceleration, hourly rainfall, stratum hardness, and fault distribution density from the valid data are retrieved; the ratio of ground motion velocity to acceleration, the ratio of hourly rainfall to the average rainfall over the past 24 hours, the product of fault distribution density and stratum hardness, and the rate of change of ground motion acceleration are calculated; the corresponding weighting coefficients are retrieved, which are dynamically adjusted based on the geological type of the monitoring area and the rainfall over the past 24 hours; the sum of the products of each feature parameter and its corresponding weighting coefficient is calculated to obtain the comprehensive feature value.
6. The geological disaster monitoring and early warning system based on deep learning according to claim 1, characterized in that, The determination of the dynamic baseline threshold and the judgment of risk level are as follows: The baseline threshold is determined by twice the average of the comprehensive characteristic values of the monitoring area with no abnormal monitoring data in the past hour; the ratio of the comprehensive characteristic value to the baseline threshold is calculated, and a ratio ≥2.0 is judged as high risk, 1.5~2.0 as medium risk, 1.0~1.5 as low risk, and <1.0 as no risk; the average of the comprehensive characteristic values of the three adjacent monitoring points is calculated simultaneously, and the average is recalculated when the difference between the average and the current monitoring point's comprehensive characteristic value exceeds 15%; if two or more of the three adjacent monitoring points are consistent with the current risk level, it is confirmed as valid.
7. The geological disaster monitoring and early warning system based on deep learning according to claim 1, characterized in that, The dynamic adjustment of the warning level is as follows: after calling the confirmed risk level and the longitude and latitude parameters of the monitoring point, the distance between the monitoring point and the warning center is calculated; the priority of push is determined according to the risk level and distance, with priority given to high-risk areas, and the push delay for high-risk areas not exceeding 1 second, medium-risk areas not exceeding 2 seconds, and low-risk areas not exceeding 3 seconds.
8. The geological disaster monitoring and early warning system based on deep learning according to claim 1, characterized in that, The precise push of early warning information is as follows: by calling the user terminal type parameter, mobile terminals push text warnings containing risk level, comprehensive feature value, expected impact range and risk avoidance suggestions; TV and radio terminals push voice + text warnings; terminals in key industries such as railways, power grids and nuclear power push customized warnings containing specific emergency response instructions; after the push is pushed, the reach rate is counted, and if the reach rate is less than 95%, the push interface is called again for a second push.
9. The geological disaster monitoring and early warning system based on deep learning according to claim 1, characterized in that, The closed-loop data correction and parameter iteration process is as follows: Real-time monitoring data within 10 minutes of the warning issuance is retrieved, the comprehensive characteristic value and risk level are recalculated, and the deviation value is calculated by comparing it with the data at the time of the warning. If the deviation value exceeds 10% of the comprehensive characteristic value at the time of the warning, the weight coefficient of the characteristic parameter is adjusted and recalculated. If the adjusted deviation value is less than 10%, the new weight coefficient is stored. Actual earthquake data during the warning period is retrieved, and the matching degree between the warning risk level and the actual damage level is calculated. If the matching degree is less than 80%, the baseline threshold calculation coefficient is adjusted until the matching degree is ≥80%. The adjusted parameters are then synchronously applied to subsequent warning processes.
10. The geological disaster monitoring and early warning system based on deep learning according to claim 1, characterized in that, Deep learning models are specifically applied to comprehensive feature value calculation, dynamic benchmark threshold adjustment, weight coefficient iteration, and risk level judgment. By continuously training the model through real-time monitoring data, the accuracy of parameter calculation and risk judgment is improved.