A seismic pre-start mechanism monitoring system

Abstract

The application discloses a kind of earthquake pre-enabling mechanism monitoring systems, it is related to earthquake monitoring technical field.The application includes multimodal sensor array module, edge computing node, intelligent analysis center and early warning release unit;Multimodal sensor array unit includes MEMS accelerometer cluster, backbone network and infrasound detection station;Edge computing node is equipped with seismic signal processing chip, specifically including preprocessing module and feature extraction module;Intelligent analysis center weights the reliability of each sensor by attention mechanism, outputs P / S wave confidence, generates three-level cascading decision unit;Early warning release unit adjusts alarm range dynamically according to estimated intensity by multi-channel broadcast protocol.The application detects P wave characteristics by each sensor continuous sampling, edge node executes real-time FFT edge, and starts regional cooperation when any node detects P wave characteristics, and generates three-level cascading decision unit by CNN network parallel processing and Transformer time series analysis, reduces earthquake early warning response delay, reduces earthquake false alarm.

Description

Technical Field

[0001] This invention belongs to the field of earthquake monitoring technology, and in particular relates to an earthquake pre-activation mechanism monitoring system. Background Technology

[0002] Traditional earthquake early warning systems generally suffer from three major technical bottlenecks: (1) High response latency: The existing system relies on centralized data processing (such as the Japan Meteorological Agency system, which takes an average of 3.5 seconds), which cannot meet the millisecond-level response requirements of scenarios such as high-speed rail and nuclear power plants; (2) High false alarm and false negative rates: The false alarm rate of a single sensor architecture for non-seismic vibrations (such as blasting and mechanical operations) is 12% to 18%; (3) Insufficient coverage: Conventional communication protocols (such as Cell Broadcast) have a coverage rate of less than 60% in mountainous and marine areas.

[0003] To address the aforementioned issues, this application provides an earthquake pre-activation mechanism monitoring system. Summary of the Invention

[0004] The purpose of this invention is to provide an earthquake early warning mechanism monitoring system. Through continuous sampling by various sensors, edge nodes perform real-time FFT edge detection. When any node detects P-wave characteristics, regional collaboration is initiated. Parallel processing of CNN networks and Transformer time series analysis are performed to generate a three-level cascaded decision unit, which solves the problems of large response delay, high false alarm rate and insufficient coverage of existing earthquake early warning systems.

[0005] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: This invention is an earthquake pre-activation mechanism monitoring system, comprising a multimodal sensor array module, edge computing nodes, an intelligent analysis center, and an early warning release unit; The multimodal sensing array unit includes a MEMS accelerometer cluster, a backbone network, and an infrasound detection station. The MEMS accelerometer cluster is evenly spaced on the building base to capture high-frequency vibration signals and identify the first arrival time and amplitude characteristics of the P-wave. The backbone network consists of optical fibers laid along the main fault zone for real-time inversion of the strain state along the fiber. The infrasound detection station is an infrasound array arranged at intervals of 10–50 km to analyze the source depth through three propagation paths. The edge computing node is equipped with a seismic signal processing chip, specifically including a preprocessing module and a feature extraction module; the preprocessing module is used to perform real-time preprocessing of data uploaded by the multimodal sensor array unit; the feature extraction module is used for parallel processing of CNN network and Transformer time series analysis; The intelligent analysis center uses an attention mechanism to weight the credibility of each sensor, outputs P / S wave confidence, and generates a three-level cascaded decision unit. The early warning issuing unit dynamically adjusts the alarm range according to the estimated intensity through a multi-channel broadcast protocol.

[0006] As a preferred technical solution, the MEMS accelerometer cluster provides raw waveform data to edge computing nodes to realize microscale vibration sensing, mainly capturing tectonic seismic waves in the 0.1-20Hz frequency band; The backbone network can invert the strain state along the optical fiber in real time through laser pulse reflection signals, forming 1,000 virtual measuring points per kilometer, and using phase-sensitive optical time-domain reflectometry to detect ultra-low frequency geological activities. The infrasound detection station uses an array of infrasound sensors to detect atmospheric pressure fluctuations of 0.01 to 10 Hz and analyzes the source depth through a triple propagation path.

[0007] As a preferred technical solution, the edge computing node preprocessing process is as follows: Step YU01, Adaptive Kalman Filtering: Establish an ARIMA model of traffic vibration noise as the input of the process noise covariance matrix Q. Dynamically adjust the measurement noise covariance matrix R according to the real-time signal-to-noise ratio. When the signal-to-noise ratio is <15dB, increase the value of R to suppress high-frequency interference. Combine the signal edge features extracted by wavelet transform to constrain the update amplitude of the state transition matrix F. Step YU02, Wavelet Packet Transform Processing: A 5-level decomposition is performed using the db4 wavelet basis function to obtain 16 sub-bands ranging from 0.5 to 100 Hz. The decomposed... The frequency band (4–32 Hz) was recombined into the characteristic frequency band of the P-wave. The frequency band (32-100Hz) was used as a candidate frequency band for S-wave. The coefficients of each node were normalized using the L2 norm to eliminate differences in sensor sensitivity. Step YU03, Dynamic Baseline Correction: The sensor zero-point drift is identified by the sliding window T test (window length 1s). Based on the data from the built-in temperature sensor, the baseline offset is corrected by applying a quadratic polynomial compensation model. The baseline reference value is updated every 10ms and smoothed by exponential weighted moving average.

[0008] As a preferred technical solution, when the feature extraction module performs parallel processing using a CNN network, it needs to identify the first arrival point of the P-wave. The specific process is as follows: Step P01: Time-domain synchronization of the data acquired by the multimodal sensor array unit is performed by using a sliding window to divide the continuous waveform and eliminate clock drift between sensors; Step P02: Set the time to 0.3 to 0.5 seconds for short windows (STA) and 30 to 60 seconds for long windows (LTA); Step P03: Calculate the average value of the signal energy over a period of time to determine the long-term trend of the signal; Step P04: Compare it with the average signal value within the next short window to determine if the signal has increased; the specific calculation formula is as follows: ; ; ; In the formula, This represents the average energy value over a short time window. This represents the average energy over a long time window. For waveform data, They are and Window length; Step P05: Calculate the AIC value for each candidate time window length, and select the minimum AIC value as the mutation point, which is the first arrival of the P wave.

[0009] As a preferred technical solution, in step P05, the formula for calculating AIC is as follows: ; In the formula, and These represent the variances of the data within the two time periods, respectively. This represents the total number of sampling points in the data window. This is the current earthquake record; When calculating AIC, the first step is to calculate the data for several consecutive seconds within the window. If the result does not meet the P-wave triggering condition, the window slides forward in 1-second increments to continue calculating the current window data segment. It then determines whether the P-wave triggering condition is met, until a position that meets the condition appears. This point that meets the condition is the initial arrival of the P-wave.

[0010] As a preferred technical solution, the CNN network also needs to extract the polarization features of the S-wave during parallel processing. The specific extraction process is as follows: Step S01: Input triaxial seismic waveform data (EW / NS / Up-Down components), and determine the S-wave search window by the P-wave arrival time (P-wave arrival time + ΔT, ΔT is calculated based on the epicentral distance). Step S02: Construct the covariance matrix of the three-axis data within the sliding window, perform eigenvalue decomposition, obtain the three principal axis directions of the polarization ellipsoid, and extract the polarization parameters as: degree of polarization (0-1 interval), principal positive polarization direction (direction angle / elevation angle), and polarization plane stability index; Step S03: Input the triaxial waveform and polarization parameter matrix into the CNN network for feature enhancement; the CNN network structure is as follows: the first layer is a 3-convolution kernel (5×5×3) to extract spatiotemporal features, the second layer is a deformable convolution to adapt to different polarization morphologies; the output layer is a Softmax classification (S-wave / non-S-wave). Step S04: Establish a baseline model, construct a normal infrasound template library through K-means clustering, and generate energy distribution baselines for each frequency band; Step S05: Perform anomaly feature extraction and multi-site infrasound array consistency analysis. Time-domain anomalies are identified by long-period fluctuations lasting more than 30 seconds; frequency-domain anomalies are detected by abnormal enhancements in the range of 0.1 to 0.5 Hz. Step S06: The CNN recognition network processes the time-domain branch and the frequency-domain branch respectively, and outputs the anomaly probability score; the time-domain branch is a 1D CNN processing the original waveform, and the frequency-domain branch is a 2D CNN processing the STFT spectrum.

[0011] As a preferred technical solution, the specific process of the Transformer timing analysis is as follows: Step T01: Receive the P-wave first arrival time series from N sensors and establish a unified time reference coordinate; Step T02: Construct the feature matrix, which specifically includes time-domain features, frequency-domain features, and spatial features; Step T03: Calculate the P-wave propagation time difference between each pair of sensors using a sliding window (window length 1-3 seconds), dynamically remove outliers, and generate a propagation time difference matrix; Step T04: Add virtual reference nodes (for calibrating system errors), apply the Floyd-Warshall algorithm to complete the indirect node time difference, and generate a weighted directed graph where the nodes are sensors and the edge weights are time differences; The Transformer network structure is shown below: The encoder includes: an input layer, positional encoding, and a multi-head attention mechanism; The output layer is a concatenation of the time difference matrix and frequency domain features; the position encoding is a 3D position code generated using the sensor's spatial coordinates; the multi-head attention mechanism focuses on learning the spatiotemporal correlation of the sensors. Step T05: Extract the time-frequency features of each sensor signal through a 1D convolutional layer, and spatially weight the features using attention weights; Step T06: Global average pooling generates a 128-dimensional feature vector, containing three sub-features: source orientation features (20% dimension), propagation path features (50% dimension), and geological structure features (30% dimension). The missing data is compensated by linear interpolation using a GPS-synchronized clock. Step T07: Construct the feature matrix, which specifically includes time-domain features, frequency-domain features, and spatial features; among which, the time-domain features include the P / S wave arrival time difference of each sensor (N×N matrix), the frequency-domain features include the main frequency / bandwidth of each node signal (N×3 matrix), and the spatial features include the three-dimensional coordinates of the sensor nodes (N×3 matrix).

[0012] As a preferred technical solution, the intelligent analysis center generates a three-level cascaded decision unit, including the first level: single-node rapid triggering, the second level: regional multi-source data fusion, and the third level: deep learning magnitude prediction. When a single node is triggered quickly (<0.5s), a hardware-accelerated localized rapid response mechanism is used to process MEMS sensor data streams in real time through an FPGA chip. An improved STA / LTA algorithm is used to identify the first arrival characteristics of P waves. Combined with an adaptive threshold dynamic adjustment strategy, millisecond-level abnormal vibration detection is achieved. After the local node is triggered, a primary alarm is broadcast immediately, and the regional collaborative process is started simultaneously to ensure that the blind zone radius is ≤15km and the first report delay is ≤0.5s. When multi-source data is fused in the region (1-2s), the heterogeneous data (MEMS acceleration / DAS strain / infrasound) of the edge node cluster are integrated, and the spatiotemporal reference is aligned through GPS time synchronization and three-dimensional wave field propagation model. The multi-source features are weighted by attention mechanism and combined with Bayesian network to process sensor uncertainty, generating a regional seismic situation map covering a radius of 50km, and outputting the epicenter location and intensity prediction (error <1km). When using deep learning for magnitude prediction (error ±0.3), the residual spatiotemporal convolutional network (RSTCN) is used to fuse the outputs of the first two stages with the historical earthquake database. The Transformer encoder is used to model the long-range wave propagation dependency. The multi-task learning framework simultaneously outputs the magnitude, depth, and intensity distribution predictions. Combined with adversarial examples, the robustness of the model is improved, and finally, accurate early warning instructions (magnitude error ≤ ±0.3) are generated to guide emergency response.

[0013] As a preferred technical solution, the early warning release unit adopts a multi-protocol redundant communication architecture. The 5G broadcast channel utilizes MBSFN single-frequency network technology to achieve millisecond-level regional coverage. The LoRaWAN gateway ensures connectivity in remote areas (transmission distance ≥ 3km) through the 470-510MHz frequency band. Satellite relay relies on BeiDou short message service to solve the communication blind spot problem. The hierarchical strategy dynamically divides the early warning area based on the real-time intensity prediction model: areas with intensity VII or above trigger full-protocol broadcast (including audible and visual alarms), areas with intensity V to VI use 5G+LoRaWAN combination push, and areas with intensity IV or below only receive lightweight notifications via APP.

[0014] The present invention has the following beneficial effects: (1) This invention continuously samples through each sensor, performs real-time FFT edge on edge nodes, and initiates regional collaboration when any node detects P-wave features. It also performs parallel processing of CNN network and Transformer time series analysis to generate a three-level cascaded decision unit, thereby reducing earthquake early warning response delay and earthquake false alarms.

[0015] (2) The hierarchical architecture of the intelligent analysis center in this invention draws on the collaborative logic of "monitoring-forecasting-response" in the disaster prevention system. Through the progressive processing of hardware-accelerated perception (level 1), data fusion analysis (level 2), and intelligent decision-making (level 3), the accuracy is improved step by step while ensuring the response speed. The system supports automatic isolation of abnormal nodes and hot switching of models to ensure reliability under extreme conditions.

[0016] (3) This invention uses a multi-source heterogeneous sensor fusion and hierarchical intelligent decision-making mechanism to achieve a lower false alarm rate and higher reliability compared to traditional systems, while gradually improving decision-making accuracy while ensuring response speed.

[0017] (4) In this invention, CNN is responsible for spatial feature extraction and Transformer is used to process time series correlation. The two achieve information fusion through feature cross-layer. Each edge node can independently complete the preliminary judgment of earthquake events in its region, which significantly reduces the dependence on the intelligent analysis center.

[0018] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1This is a schematic diagram of the structure of an earthquake pre-activation mechanism monitoring system according to the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and 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.

[0022] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0023] To make the purpose, technical solution, and advantages of this application clearer, the following description is provided in conjunction with the appendix. Figure 1 The present application will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the application.

[0024] Please see Figure 1 As shown, the present invention is an earthquake pre-activation mechanism monitoring system, including a multimodal sensor array module, an edge computing node, an intelligent analysis center, and an early warning release unit; The multimodal sensing array unit includes a MEMS accelerometer cluster, a backbone network, and infrasound detection stations. The MEMS accelerometer cluster is evenly spaced on the building base to capture high-frequency vibration signals and identify the first arrival time and amplitude characteristics of P waves. The backbone network consists of optical fibers laid along the main fault zone, used for real-time inversion of the strain state along the fiber. The infrasound detection stations are infrasound arrays arranged at intervals of 10–50 km, used to analyze the source depth through three propagation paths. The edge computing node is equipped with a seismic signal processing chip, specifically including a preprocessing module and a feature extraction module; the preprocessing module is used to perform real-time preprocessing of data uploaded by the multimodal sensor array unit; the feature extraction module is used for parallel processing of CNN network and Transformer time series analysis. The intelligent analysis center uses an attention mechanism to weight the credibility of each sensor, outputs P / S wave confidence scores, and generates a three-level cascaded decision unit. The early warning issuing unit dynamically adjusts the alarm range according to the estimated intensity through a multi-channel broadcast protocol.

[0025] MEMS accelerometer clusters provide raw waveform data to edge computing nodes and capture high-frequency vibration signals through sampling at ≥500Hz. This is used to identify the arrival time and amplitude characteristics of P waves, enabling microscale vibration sensing. The main function is to capture tectonic seismic waves in the 0.1–20Hz frequency band. The backbone network can invert the strain state along the optical fiber in real time through laser pulse reflection signals, forming 1,000 virtual measuring points per kilometer, and using phase-sensitive optical time-domain reflectometry to detect ultra-low frequency geological activity; The array of infrasound sensors deployed at the infrasound detection station detects atmospheric pressure fluctuations in the range of 0.01 to 10 Hz. By analyzing the source depth through a triple propagation path, it supplements the information on underground-surface coupled vibrations, and has a particular advantage in early identification of deep-source earthquakes. The MEMS accelerometer cluster and the backbone network are triggered synchronously. The MEMS accelerometer cluster captures high-frequency vibration details, while the backbone network records the propagation characteristics of long-period wave fields.

[0026] The preprocessing process for edge computing nodes is as follows: Step YU01, Adaptive Kalman Filtering: Establish an ARIMA model of traffic vibration noise as the input of the process noise covariance matrix Q. Dynamically adjust the measurement noise covariance matrix R according to the real-time signal-to-noise ratio. When the signal-to-noise ratio is <15dB, increase the value of R to suppress high-frequency interference. Combine the signal edge features extracted by wavelet transform to constrain the update amplitude of the state transition matrix F. Step YU02, Wavelet Packet Transform Processing: A 5-level decomposition is performed using the db4 wavelet basis function to obtain 16 sub-bands ranging from 0.5 to 100 Hz. The decomposed... The frequency band (4–32 Hz) was recombined into the characteristic frequency band of the P-wave. The frequency band (32-100Hz) was used as the candidate frequency band for S-wave. The coefficients of each node were normalized using the L2 norm to eliminate the differences in sensor sensitivity. Step YU03, Dynamic Baseline Correction: The sensor zero-point drift is identified by the sliding window T test (window length 1s). Based on the data from the built-in temperature sensor, the baseline offset is corrected by applying a quadratic polynomial compensation model. The baseline reference value is updated every 10ms and smoothed by exponential weighted moving average. Through the cascaded processing of these three technologies, a clean seismic signal with a signal-to-noise ratio improvement of ≥20dB is finally output, providing high-quality input for subsequent feature extraction.

[0027] When the feature extraction module performs parallel processing in a CNN network, it needs to identify the first arrival point of the P wave. The specific process is as follows: Step P01: Time-domain synchronization of the data acquired by the multimodal sensor array unit is performed by using a sliding window to divide the continuous waveform and eliminate clock drift between sensors; Step P02: Set the time to 0.3 to 0.5 seconds for short windows (STA) and 30 to 60 seconds for long windows (LTA); Step P03: Calculate the average value of the signal energy over a period of time to determine the long-term trend of the signal; Step P04: Compare it with the average signal value within the next short window to determine if the signal has increased; the specific calculation formula is as follows: ; ; ; In the formula, This represents the average energy value over a short time window. This represents the average energy over a long time window. For waveform data, They are and Window length; Step P05: Calculate the AIC value for each candidate time window length, and select the minimum AIC value as the mutation point, which is the first arrival of the P wave.

[0028] In step P05, the formula for calculating AIC is as follows: ; In the formula, and These represent the variances of the data within the two time periods, respectively. This represents the total number of sampling points in the data window. This is the current earthquake record; When performing AIC calculations, the first step is to calculate the data for several consecutive seconds within the window. If the result does not meet the P-wave triggering condition, the window slides forward in 1-second increments to continue calculating the current window data segment. It then determines whether the P-wave triggering condition is met, until a position that meets the condition appears. This point that meets the condition is the initial arrival of the P-wave.

[0029] When processing data in parallel using a CNN network, it is also necessary to extract the polarization features of the S-wave. The specific extraction process is as follows: Step S01: Input triaxial seismic waveform data (EW / NS / Up-Down components), and determine the S-wave search window by the P-wave arrival time (P-wave arrival time + ΔT, ΔT is calculated based on the epicentral distance). Step S02: Construct the covariance matrix of the three-axis data within the sliding window, perform eigenvalue decomposition, obtain the three principal axis directions of the polarization ellipsoid, and extract the polarization parameters as: degree of polarization (0-1 interval), principal positive polarization direction (direction angle / elevation angle), and polarization plane stability index; Step S03: Input the triaxial waveform and polarization parameter matrix into the CNN network for feature enhancement; the CNN network structure is as follows: the first layer is a 3-convolution kernel (5×5×3) to extract spatiotemporal features, the second layer is a deformable convolution to adapt to different polarization morphologies; the output layer is a Softmax classification (S-wave / non-S-wave). Step S04: Establish a baseline model, construct a normal infrasound template library through K-means clustering, and generate energy distribution baselines for each frequency band; Step S05: Perform anomaly feature extraction and multi-site infrasound array consistency analysis. Time-domain anomalies are identified by long-period fluctuations lasting more than 30 seconds; frequency-domain anomalies are detected by abnormal enhancements in the range of 0.1 to 0.5 Hz. Step S06: The CNN recognition network processes the time-domain branch and the frequency-domain branch respectively, and outputs anomaly probability scores; the time-domain branch is a 1D CNN processing the original waveform, and the frequency-domain branch is a 2D CNN processing the STFT spectrum. Cross attention is used to output anomaly probability scores (0-1 points).

[0030] The specific process of Transformer timing analysis is as follows: Step T01: Receive the P-wave first arrival time series from N sensors and establish a unified time reference coordinate; Step T02: Construct the feature matrix, which specifically includes time-domain features, frequency-domain features, and spatial features; Step T03: Calculate the P-wave propagation time difference between each pair of sensors using a sliding window (window length 1-3 seconds), dynamically remove outliers, and generate a propagation time difference matrix; Step T04: Add virtual reference nodes (for calibrating system errors), apply the Floyd-Warshall algorithm to complete the indirect node time difference, and generate a weighted directed graph where the nodes are sensors and the edge weights are time differences; The Transformer network structure is shown below: The encoder includes: an input layer, positional encoding, and a multi-head attention mechanism; The output layer is a concatenation of the time difference matrix and frequency domain features; the position encoding is a 3D position code generated using the sensor's spatial coordinates; the multi-head attention mechanism focuses on learning the spatiotemporal correlation of the sensors. Step T05: Extract the time-frequency features of each sensor signal through a 1D convolutional layer, and spatially weight the features using attention weights; Step T06: Global average pooling generates a 128-dimensional feature vector, containing three sub-features: source orientation features (20% dimension), propagation path features (50% dimension), and geological structure features (30% dimension). The missing data is compensated by linear interpolation using a GPS-synchronized clock. Step T07: Construct the feature matrix, which specifically includes time-domain features, frequency-domain features, and spatial features; among which, the time-domain features include the P / S wave arrival time difference of each sensor (N×N matrix), the frequency-domain features include the main frequency / bandwidth of each node signal (N×3 matrix), and the spatial features include the three-dimensional coordinates of the sensor nodes (N×3 matrix).

[0031] The intelligent analysis center generates a three-level cascaded decision-making unit, including the first level: single-node rapid triggering, the second level: regional multi-source data fusion, and the third level: deep learning magnitude prediction. When a single node is triggered quickly (<0.5s), a hardware-accelerated localized rapid response mechanism is used to process MEMS sensor data streams in real time through an FPGA chip. An improved STA / LTA algorithm is used to identify the first arrival characteristics of P waves. Combined with an adaptive threshold dynamic adjustment strategy, millisecond-level abnormal vibration detection is achieved. After the local node is triggered, a primary alarm is broadcast immediately, and the regional collaborative process is started simultaneously to ensure that the blind zone radius is ≤15km and the first report delay is ≤0.5s. When multi-source data is fused in the region (1-2s), the heterogeneous data (MEMS acceleration / DAS strain / infrasound) of the edge node cluster are integrated, and the spatiotemporal reference is aligned through GPS time synchronization and three-dimensional wave field propagation model. The multi-source features are weighted by attention mechanism and combined with Bayesian network to process sensor uncertainty, generating a regional seismic situation map covering a radius of 50km, and outputting the epicenter location and intensity prediction (error <1km). When using deep learning for magnitude prediction (error ±0.3), the residual spatiotemporal convolutional network (RSTCN) is used to fuse the outputs of the first two stages with the historical earthquake database. The Transformer encoder is used to model the long-range wave propagation dependency. The multi-task learning framework simultaneously outputs the magnitude, depth, and intensity distribution predictions. Combined with adversarial examples, the robustness of the model is improved, and finally, accurate early warning instructions (magnitude error ≤ ±0.3) are generated to guide emergency response.

[0032] The early warning release unit adopts a multi-protocol redundant communication architecture. The 5G broadcast channel uses MBSFN single-frequency network technology to achieve millisecond-level regional coverage. The LoRaWAN gateway ensures connectivity in remote areas (transmission distance ≥3km8) through the 470-510MHz frequency band. Satellite relay relies on BeiDou short message service to solve the communication blind spot problem. The graded strategy dynamically divides the early warning area based on the real-time intensity prediction model: the intensity zone of VII and above triggers full protocol broadcast (including sound and light alarm), the intensity zone of V to VI uses 5G+LoRaWAN combination push, and the intensity zone of IV and below is only notified by the APP.

[0033] It is worth noting that the various units included in the above system embodiments are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy differentiation and are not used to limit the scope of protection of the present invention.

[0034] Furthermore, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware, and the corresponding program can be stored in a computer-readable storage medium.

[0035] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. An earthquake pre-activation mechanism monitoring system, comprising a multimodal sensor array module, edge computing nodes, an intelligent analysis center, and an early warning release unit, characterized in that: The multimodal sensing array unit includes a MEMS accelerometer cluster, a backbone network, and an infrasound detection station. The MEMS accelerometer cluster is evenly spaced on the building base to capture high-frequency vibration signals and identify the first arrival time and amplitude characteristics of the P-wave. The backbone network consists of optical fibers laid along the main fault zone for real-time inversion of the strain state along the fiber. The infrasound detection station is an infrasound array arranged at intervals of 10–50 km to analyze the source depth through three propagation paths. The edge computing node is equipped with a seismic signal processing chip, specifically including a preprocessing module and a feature extraction module; The preprocessing module is used to perform real-time preprocessing of data uploaded by the multimodal sensor array unit; the feature extraction module is used for parallel processing of the CNN network and time-series analysis of the Transformer. The intelligent analysis center uses an attention mechanism to weight the credibility of each sensor, outputs P / S wave confidence, and generates a three-level cascaded decision unit. The early warning issuing unit dynamically adjusts the alarm range according to the estimated intensity through a multi-channel broadcast protocol.

2. The earthquake pre-activation mechanism monitoring system according to claim 1, characterized in that, The MEMS accelerometer cluster provides raw waveform data for edge computing nodes; the backbone network can invert the strain state along the optical fiber in real time through laser pulse reflection signals, forming 1,000 virtual measuring points per kilometer, and uses phase-sensitive optical time-domain reflectometry to detect ultra-low frequency geological activity; the infrasound detection station array-deployed infrasound sensors detect atmospheric pressure fluctuations of 0.01 to 10 Hz, and analyzes the source depth through triple propagation paths.

3. The earthquake pre-activation mechanism monitoring system according to claim 1, characterized in that, The edge computing node preprocessing process is as follows: Step YU01, Adaptive Kalman Filtering: Establish an ARIMA model of traffic vibration noise as the input of the process noise covariance matrix Q, dynamically adjust the measurement noise covariance matrix R according to the real-time signal-to-noise ratio, and combine the signal edge features extracted by wavelet transform to constrain the update amplitude of the state transition matrix F. Step YU02, Wavelet Packet Transform Processing: A 5-level decomposition is performed using the db4 wavelet basis function to obtain 16 sub-bands ranging from 0.5 to 100 Hz. The decomposed... The frequency band was reorganized into a P-wave characteristic frequency band. The frequency band is used as a candidate frequency band for S-wave, and the coefficients of each node are normalized using the L2 norm. Step YU03, Dynamic Baseline Correction: The zero-point drift of the sensor is identified by the sliding window T test. Based on the data from the built-in temperature sensor, the baseline offset is corrected by applying a quadratic polynomial compensation model. The baseline reference value is updated every 10ms and smoothed by exponential weighted moving average.

4. The earthquake pre-activation mechanism monitoring system according to claim 1, characterized in that, When the feature extraction module performs parallel processing using the CNN network, it needs to identify the first arrival point of the P wave. The specific process is as follows: Step P01: Time-domain synchronization of the data acquired by the multimodal sensor array unit is performed by using a sliding window to segment the continuous waveform; Step P02: Set the time for short windows to 0.3 to 0.5 seconds, and for long windows to 30 to 60 seconds; Step P03: Calculate the average value of the signal energy over a period of time to determine the long-term trend of the signal; Step P04: Compare it with the average signal value within the next short window to determine if the signal has increased; Step P05: Calculate the AIC value for each candidate time window length, and select the minimum AIC value as the mutation point, which is the first arrival of the P wave.

5. The earthquake pre-activation mechanism monitoring system according to claim 1, characterized in that, As stated above.

6. The earthquake pre-activation mechanism monitoring system according to claim 5, characterized in that, In step P05, the formula for calculating AIC is as follows: ； In the formula, and These represent the variances of the data within the two time periods, respectively. This represents the total number of sampling points in the data window. This is the current earthquake record; When calculating AIC, the first step is to calculate the data for several consecutive seconds within the window. If the result does not meet the P-wave triggering condition, the window slides forward in 1-second increments to continue calculating the current window data segment. It then determines whether the P-wave triggering condition is met, until a position that meets the condition appears. This point that meets the condition is the initial arrival of the P-wave.

7. The earthquake pre-activation mechanism monitoring system according to claim 4, characterized in that, When the CNN network processes data in parallel, it is also necessary to extract the polarization features of the S-wave. The specific extraction process is as follows: Step S01: Input triaxial seismic waveform data and determine the S-wave search window by the P-wave first arrival time; Step S02: Construct the covariance matrix of the three-axis data within the sliding window, perform eigenvalue decomposition, and obtain the three principal axis directions of the polarization ellipsoid; Step S03: Input the triaxial waveform and polarization parameter matrix into the CNN network for feature enhancement; Step S04: Establish a baseline model, construct a normal infrasound template library through K-means clustering, and generate energy distribution baselines for each frequency band; Step S05: Extract abnormal features and perform consistency analysis of multi-site infrasound arrays; Step S06: The CNN recognition network processes the time-domain branch and the frequency-domain branch respectively, and outputs anomaly probability scores.

8. The earthquake pre-activation mechanism monitoring system according to claim 1, characterized in that, The specific process of Transformer timing analysis is as follows: Step T01: Receive the P-wave first arrival time series from N sensors and establish a unified time reference coordinate system; Step T02: Construct the feature matrix, which specifically includes time-domain features, frequency-domain features, and spatial features; Step T03: Calculate the P-wave propagation time difference between each pair of sensors using a sliding window, and generate a propagation time difference matrix; Step T04: Add virtual reference nodes, apply the Floyd-Warshall algorithm to complete the time difference of indirect nodes, and generate a weighted directed graph; Step T05: Extract the time-frequency features of each sensor signal through a 1D convolutional layer, and spatially weight the features using attention weights; Step T06: Global average pooling generates 128-dimensional feature vectors.

9. The earthquake pre-activation mechanism monitoring system according to claim 1, characterized in that, The intelligent analysis center generates a three-level cascaded decision-making unit, including the first level: single-node rapid triggering, the second level: regional multi-source data fusion, and the third level: deep learning magnitude prediction. When a single node is triggered quickly, a hardware-accelerated localized rapid response mechanism is used to process MEMS sensor data streams in real time through an FPGA chip. An improved STA / LTA algorithm is used to identify the first arrival characteristics of P waves. Combined with an adaptive threshold dynamic adjustment strategy, millisecond-level abnormal vibration detection is achieved. After the local node is triggered, a primary alarm is broadcast immediately, and the regional collaborative process is started simultaneously. When multi-source data is fused in a region, heterogeneous data from edge node clusters are integrated, and spatiotemporal references are aligned through GPS time synchronization and a three-dimensional wave field propagation model. Attention mechanism is used to weight multi-source features, and Bayesian network is used to process sensor uncertainties to generate a regional seismic situation map covering a radius of 50km, outputting the epicenter location and intensity prediction. When using deep learning for magnitude prediction, the outputs of the first two stages are fused with the historical earthquake database based on a residual spatiotemporal convolutional network. The long-range wave propagation dependency is modeled through a Transformer encoder. The multi-task learning framework simultaneously outputs magnitude, depth, and intensity distribution predictions. Combined with adversarial examples, the robustness of the model is improved, and finally, accurate early warning instructions are generated to guide emergency response.

10. The earthquake pre-activation mechanism monitoring system according to claim 1, characterized in that, The early warning release unit adopts a multi-protocol redundant communication architecture, and the hierarchical strategy dynamically divides the early warning area based on the real-time intensity prediction model: the intensity area of ​​VII and above triggers full protocol broadcast, the intensity areas of V to VI use 5G+LoRaWAN combination push, and the intensity areas of IV and below are only notified through the APP lightweight notification.

Images