Earth surface anomaly monitoring data fusion system based on vibration signal processing

Abstract

The invention provides an earth surface anomaly monitoring data fusion system based on vibration signal processing, and relates to the technical field of earthquake precursor monitoring, and the system comprises a data collection module which is used for carrying out the real-time monitoring of the earth surface through a distributed multi-parameter sensor network, and obtaining the time sequence monitoring data related to the earthquake precursor; the signal processing module is used for processing the time sequence monitoring data by adopting a two-parameter joint optimization algorithm and outputting a separation signal component containing precursor information; the data fusion and state inversion module is used for carrying out fusion processing on the separated signal components by adopting a dual-time scale analysis framework and outputting a multi-scale precursor state estimation result and an uncertainty quantitative index; and the monitoring and early warning module is used for outputting earthquake precursor grading early warning information through a three-layer anomaly recognition architecture and probabilistic risk assessment. According to the invention, high-precision separation and identification of earthquake precursor related signals can be realized, and the false alarm phenomenon is reduced through an intelligent early warning mechanism.

Description

Technical Field

[0001] The present invention relates to the technical field of earthquake precursor monitoring, and particularly to a surface anomaly monitoring data fusion system based on vibration signal processing. Background Art

[0002] Earthquake precursor observation is an important basis for earthquake prediction research. By monitoring the abnormal changes of the geophysical field before an earthquake, possible earthquake precursor information can be identified. Earthquake precursor phenomena involve changes in various physical parameters, including seismic activity, crustal deformation, groundwater dynamics, geomagnetism, geoelectric field, etc. These parameters often exhibit complex spatio-temporal evolution characteristics during the earthquake gestation process. Establishing a high-precision, multi-parameter comprehensive earthquake precursor observation system is of great significance for improving earthquake prediction ability and reducing earthquake disasters.

[0003] Currently, the technologies for earthquake precursor observation mainly include traditional single-parameter monitoring methods such as seismic monitoring networks, geomagnetic observation, geoelectric observation, and deformation observation, as well as multi-parameter comprehensive observation technologies developed in recent years. However, the existing technologies generally have the following problems: First, single-parameter observation is easily affected by environmental interference and it is difficult to accurately identify true earthquake precursor signals; Second, there is a lack of an effective fusion mechanism between multi-parameter observation data, and the complementary information between different parameters cannot be fully utilized; Third, the existing methods have limited ability to separate different physical sources in mixed signals, resulting in low accuracy of precursor signal identification; Finally, traditional methods lack an adaptive mechanism and it is difficult to perform dynamic optimization according to changes in the actual observation environment.

[0004] The Chinese invention patent application No. CN202111504790.3 discloses a method for screening and analyzing earthquake precursor characteristics. This method obtains earthquake electromagnetic disturbance signals by deploying a monitoring network, and uses filtering processing, feature extraction, and anomaly detection technologies to identify earthquake precursor characteristics, and combines the sliding quartile method and the isolation forest algorithm for anomaly detection. Although this method has made certain progress in earthquake electromagnetic disturbance monitoring, this method mainly processes single-type electromagnetic signals and lacks the comprehensive separation ability for mixed signals of multiple physical parameters; its anomaly detection mechanism is relatively simple and it cannot effectively handle complex signal mixing and multi-source interference problems. Summary of the Invention

[0005] In view of this, the present invention provides a surface anomaly monitoring data fusion system based on vibration signal processing, which realizes high-precision separation and identification of earthquake precursor-related signals, improves the accuracy and reliability of multi-scale precursor state estimation, and reduces false alarm phenomena through an intelligent early warning mechanism.

[0006] The technical solution of the present invention is realized as follows: The present invention provides a surface anomaly monitoring data fusion system based on vibration signal processing, including: A data acquisition module, which is used to monitor the surface in real time by using a distributed multi-parameter sensor network, and output time-series monitoring data related to earthquake precursors based on a dual-frequency sampling strategy; A signal processing module, which is used to process the time-series monitoring data by using a dual-parameter joint optimization algorithm, realize adaptive wavefield separation by simultaneously optimizing the demixing matrix and the parameters of the seismic structure model, and output separated signal components containing precursor information; A data fusion and state inversion module, which is used to perform fusion processing on the separated signal components by using a dual-time scale analysis framework, and output multi-scale precursor state estimation results and uncertainty quantification indicators; A monitoring and early warning module, which is used to analyze the multi-scale precursor state estimation results by using an adaptive threshold mechanism and an early warning fatigue suppression algorithm, and output earthquake precursor classification early warning information through a three-layer anomaly recognition architecture and probabilistic risk assessment.

[0007] Preferably, the dual-parameter joint optimization algorithm includes: A1. Preprocess and whiten the time-series monitoring data to obtain a precursor signal mixing matrix; A2. Initialize the demixing matrix and the parameters of the seismic structure model; A3. Construct a joint objective function containing an independence measure and seismic physical constraints based on the precursor signal mixing matrix; A4. Calculate the gradients of the joint objective function with respect to the demixing matrix and the parameters of the seismic structure model; A5. Update the demixing matrix and the parameters of the seismic structure model simultaneously by using an adaptive learning rate; A6. Determine whether the convergence criterion is satisfied. If it is satisfied, end the iteration; if not, return to step A4 to continue the iteration; A7. Output the optimized demixing matrix to separate the precursor signal mixing matrix to obtain separated signal components containing precursor information.

[0008] Preferably, the joint objective function is: , where is the joint objective function, is the precursor signal after separation, W is the demixing matrix, is the parameter vector of the seismic structure model, is the theoretical mixing matrix based on the seismic structure model, is the independence measure function, is the adaptive regularization function, is the Frobenius norm.

[0009] Preferably, the calculation formula of the adaptive threshold mechanism is: , wherein, is the adaptive threshold at time t, is the basic threshold, is the uncertainty adjustment coefficient, is the variance of the precursor state estimation, is the time-varying adjustment factor.

[0010] Preferably, the early warning fatigue suppression algorithm includes: B1. Obtain the current precursor early warning judgment result and historical early warning records as inputs; B2. Set a statistical time window to collect historical precursor early warning data within the window; B3. Count the number of precursor early warnings and the corresponding verification results within the window; B4. Calculate the recent precursor early warning accuracy rate and early warning frequency indicators; B5. Calculate the precursor early warning suppression factor based on the ratio of the early warning frequency to the threshold and the accuracy deviation; B6. Multiply the original precursor early warning result by the early warning suppression factor to obtain the final suppressed early warning output.

[0011] Preferably, the calculation formula of the early warning suppression factor is: , wherein, is the early warning suppression factor, is the recent number of precursor early warnings, is the early warning number threshold, is the recent precursor early warning accuracy rate, is the target accuracy rate.

[0012] Preferably, the dual-frequency sampling strategy includes: under normal conditions, sampling is performed at a standard frequency of 100 - 200 Hz, and under precursor abnormal conditions, high-frequency sampling is performed at 500 - 1000 Hz. When the change rate of the precursor signal amplitude exceeds a preset threshold, it automatically switches to the high-frequency sampling mode.

[0013] Preferably, the dual-time-scale analysis framework includes: in the short time scale, a 1 - 10 minute time window is used to capture the rapid change process of the precursor signal, and in the long time scale, a 1 - 24 hour time window is used to identify the trend change of the precursor signal. The fusion weight is dynamically adjusted according to the quality and uncertainty of the precursor signal.

[0014] Preferably, the three-layer anomaly recognition architecture includes: the signal layer identifies the type of precursor anomaly event through template matching, the state layer monitors the change of the physical properties of earthquake precursors based on statistical process control, and the system layer identifies the precursor anomaly propagation mode through spatial correlation analysis.

[0015] Preferably, the convergence criteria of the dual-parameter joint optimization algorithm include: (1) the relative change of the objective function is less than the set threshold; (2) the change amplitude of the parameter vector is less than the convergence tolerance; (3) the number of iterations reaches the maximum limit value.

[0016] The present invention has the following beneficial effects compared with the prior art: (1) By constructing a complete surface anomaly monitoring data fusion system, the present invention realizes the full-process technology integration from data acquisition, signal processing, data fusion to monitoring and early warning, and effectively solves the technical problems such as low signal separation accuracy, poor multi-parameter fusion effect, and low early warning accuracy in traditional earthquake precursor observation; (2) By simultaneously optimizing the demixing matrix and the parameters of the seismic structure model and introducing seismic physical constraints as regularization terms, the algorithm overcomes the problem of unstable separation results caused by the lack of physical constraints in traditional blind source separation methods, and improves the accuracy and physical rationality of the separation of mixed precursor signals; (3) By setting two analysis windows with different time scales, short-term and long-term, and dynamically adjusting the fusion weights according to the signal quality, the framework can capture both the rapid changes and long-term trends of earthquake precursor signals, and solves the problem that important precursor information is easily missed in single-time-scale analysis; (4) The adaptive threshold mechanism dynamically adjusts the early warning threshold by considering the uncertainty and time-varying characteristics of precursor state estimation, avoids the problem of poor adaptability of the fixed threshold method in different observation environments, and reduces false alarms and missed alarms caused by environmental changes; (5) By calculating the suppression factor through statistical historical early warning frequency and accuracy, the early warning fatigue suppression algorithm effectively suppresses the early warning output with high frequency and low accuracy, and solves the problem of easy generation of early warning fatigue in traditional early warning systems; (6) The three-layer anomaly recognition architecture realizes multi-level analysis from single-point anomaly detection to spatial pattern recognition through the hierarchical recognition mechanism of the signal layer, state layer and system layer, and overcomes the limitation that single-level anomaly recognition is easily affected by local interference. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0018] Figure 1 It is a schematic diagram of the system structure of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0019] Next, in combination with the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0020] As Figure 1 shown, the present invention provides a surface anomaly monitoring data fusion system based on vibration signal processing, including: A data acquisition module for real-time monitoring of the surface using a distributed multi-parameter sensor network and outputting time-series monitoring data related to earthquake precursors based on a dual-frequency sampling strategy; A signal processing module for processing the time-series monitoring data using a dual-parameter joint optimization algorithm, realizing adaptive wavefield separation by simultaneously optimizing the demixing matrix and seismic structure model parameters, and outputting separated signal components containing precursor information; A data fusion and state inversion module for performing fusion processing on the separated signal components using a dual-time scale analysis framework and outputting multi-scale precursor state estimation results and uncertainty quantification indicators; A monitoring and warning module for analyzing the multi-scale precursor state estimation results using an adaptive threshold mechanism and a warning fatigue suppression algorithm, and outputting earthquake precursor classification warning information through a three-layer anomaly recognition architecture and probabilistic risk assessment.

[0021] Specifically, in an embodiment of the present invention, the data acquisition module adopts a distributed multi-parameter sensor network architecture, deploying multiple sensor nodes in the target monitoring area, and each node is configured with various sensor devices such as a three-axis accelerometer, an inclination angle sensor, and a pore water pressure gauge. The three-axis accelerometer is used to monitor surface micro-motions and vibration signals, and its sampling accuracy requirement reaches the microgravity level, capable of capturing weak surface motion changes caused by earthquake precursors. The inclination angle sensor is responsible for monitoring surface tilt changes, and its measurement accuracy needs to reach the arcsecond level for detecting surface tilt anomalies caused by crustal deformation. The pore water pressure gauge is used to monitor groundwater pressure changes, and its pressure measurement accuracy requirement reaches the Pascal level, capable of reflecting the change characteristics of groundwater dynamics parameters.

[0022] The layout of the sensor network follows the principle of multi-level spatial sampling, and the spatial configuration of sensor nodes is determined according to the geological structure characteristics and seismic activity characteristics of the monitoring area. The main sensor nodes are arranged in a regular grid pattern, and the node spacing is determined according to the target monitoring frequency and seismic wavelength characteristics, generally set to one-quarter to one-half of the wavelength corresponding to the main frequency band of earthquake precursor signals. For key monitoring areas, such as near active faults and areas with concentrated historical earthquakes, the sensor layout density needs to be increased by more than one time to form a dense observation network. The vertical layout of sensor nodes considers the terrain undulation and geological stratification characteristics. In mountainous and hilly areas, the sensors are distributed along the elevation gradient, and the elevation interval is determined according to the terrain slope and the complexity of the geological structure. In plain areas, the sensors mainly consider the characteristics of groundwater level changes and soft soil layer distribution, and the burial depth of the sensors is determined according to local geological exploration data, generally set in the soil layer above the stable bedrock surface. Auxiliary sensor nodes are set in the boundary area to monitor the propagation characteristics of earthquake precursor signals at the outer edge of the monitoring area. The spacing of the auxiliary nodes can be appropriately relaxed, but it is necessary to ensure the signal continuity with the main monitoring network. In special geological structure areas, such as karst development areas and soft formation distribution areas, it is necessary to adjust the sensor type and burial depth according to specific geological conditions.

[0023] The system adopts a dual-frequency sampling strategy to adaptively sample earthquake precursor-related signals. The core of this strategy is to dynamically adjust the sampling frequency according to the signal state. In the normal monitoring state, the system uses the standard sampling frequency for data acquisition, and its frequency range is set to 100 - 200 Hz, which can meet the sampling requirements of conventional earthquake precursor signals. When the system detects that the amplitude change rate of the precursor signal exceeds the preset threshold, it automatically switches to the high-frequency sampling mode, and at this time the sampling frequency is increased to the range of 500 - 1000 Hz. The judgment condition for the sampling mode switch is: , where is the amplitude of the currently sampled signal, represents the time, is the preset amplitude change rate threshold. The high-frequency sampling mode can capture the rapid change process of earthquake precursor signals and ensure that important transient abnormal information is not missed.

[0024] The sensor network adopts a distributed synchronous acquisition mechanism, and all sensor nodes achieve time synchronization through the GPS clock synchronization or network clock synchronization protocol. The synchronization accuracy requirement reaches the microsecond level. Each sensor node is configured with a local data cache unit, which can store the original sampling data for a certain period of time. When network communication is abnormal, the local cache mechanism can prevent data loss. Data transmission uses a combination of wired and wireless methods. Wired transmission is used for data backhaul of core nodes, and wireless transmission is used for data acquisition of edge nodes.

[0025] Specifically, in one embodiment of the present invention, the signal processing module implements adaptive wavefield separation using a dual-parameter joint optimization algorithm. This algorithm extracts the separated signal components containing precursor information by simultaneously optimizing the demixing matrix and the seismic structure model parameters. The signal processing process is as follows: A1. Preprocess and whiten the time-series monitoring data to obtain the precursor signal mixing matrix , where represents the observed signals of sensor channels, where respectively represent the observed signals of the 1st, 2nd, denotes matrix transpose.

[0026] The whitening transformation eliminates the second-order statistical correlation between signals through decorrelation processing. The whitening matrix is obtained by performing eigenvalue decomposition on the covariance matrix: , where is the eigenvalue diagonal matrix, is the eigenvector matrix, denotes matrix transpose. The signal after whitening transformation has the unit covariance property, is the signal after whitening transformation, is the whitening matrix, is the precursor signal mixing matrix.

[0027] A2. Initialize the demixing matrix and the seismic structure model parameters.

[0028] In the initialization stage of the algorithm, it is necessary to set the initial values of the demixing matrix and the seismic structure model parameter vector . The demixing matrix is initialized as the identity matrix or the initial estimate obtained through principal component analysis. The seismic structure model parameters are initialized according to prior geological information, including physical parameters such as formation density, elastic modulus, and medium attenuation coefficient. The seismic structure model parameters are used to construct the theoretical mixing matrix , which describes the theoretical propagation characteristics of different physical source signals in the sensor array. The construction of the theoretical mixing matrix is based on the physical mechanism of seismic wave propagation, considering the effects of formation structure, medium parameters, and geometric attenuation on signal propagation.

[0029] A3. Construct a joint objective function containing independence measure and seismic physical constraints based on the precursor signal mixing matrix. The expression of the objective function is: , Among them, is the joint objective function, is the separated precursor signal, is the demixing matrix, is the seismic structure model parameter vector, is the theoretical mixing matrix based on the seismic structure model, is the independence measure function, is the adaptive regularization function, is the Frobenius norm. The independence measure function uses negative entropy or mutual information as the measurement criterion to evaluate the statistical independence of the separated signals. The adaptive regularization function is dynamically adjusted according to the deviation degree of the model parameters from the prior reference values, and its expression is: , where is the adaptive regularization function, is the basic regularization coefficient, is the adjustment parameter, is the reference geological model parameter, is the seismic structure model parameter vector, is the L2 norm. When the model parameters are close to the reference values, the regularization intensity is high; when the parameters deviate from the reference values, the regularization intensity gradually decreases, allowing the algorithm to adjust the model parameters according to the observed data.

[0030] A4. Calculate the gradients of the joint objective function with respect to the demixing matrix and the seismic structure model parameters.

[0031] Gradient calculation is the core step of the optimization algorithm, and the partial derivatives of the objective function with respect to the demixing matrix and the model parameters need to be calculated separately. The gradient with respect to the seismic structure model parameters is: , The gradient with respect to the demixing matrix is: , where represents the gradient operator with respect to the seismic structure model parameter , represents the gradient operator with respect to the demixing matrix , is the joint objective function, is the demixing matrix, is the vector of the originally observed mixed signals, is the independence measure function, is the theoretical mixing matrix based on the seismic structure model, represents the matrix transpose, is the adaptive regularization function, is the Frobenius norm. The derivative of the theoretical mixing matrix with respect to the model parameters is involved in the gradient calculation , which is obtained by numerical differentiation or analytical differentiation methods, and reflects the influence of the changes in geological model parameters on the signal propagation characteristics.

[0032] A5. Simultaneously update the unmixing matrix and seismic structure model parameters using an adaptive learning rate.

[0033] The parameter update uses the gradient descent method with an adaptive learning rate to simultaneously update the unmixing matrix and seismic structure model parameters. The update formula is as follows: , , where is the vector of seismic structure model parameters after the (k + 1)-th iteration, is the vector of seismic structure model parameters after the k-th iteration, is the joint objective function with respect to the parameter , is the unmixing matrix, is the unmixing matrix after the (k + 1)-th iteration, is the unmixing matrix after the k-th iteration, is the joint objective function with respect to the unmixing matrix , and are the adaptive learning rates, which are dynamically adjusted according to the gradient magnitude and iteration history. The adaptive learning rate adopts a variant of the AdaGrad or Adam optimizer, which can automatically adjust the step size according to the historical information of parameter updates and improve the convergence stability of the algorithm. To ensure the numerical stability of the unmixing matrix, the unmixing matrix needs to be orthogonalized after each update to ensure its reversibility requirement.

[0034] A6. Judge whether the convergence criterion is satisfied. If it is satisfied, end the iteration; if not, return to step A4 to continue the iteration.

[0035] The convergence criterion includes three conditions: the relative change of the objective function is less than the set threshold, the change magnitude of the parameter vector is less than the convergence tolerance, and the number of iterations reaches the maximum limit value. The relative change criterion of the objective function is: , where is the relative change threshold, is the joint objective function after the (k + 1)-th iteration, is the joint objective function after the k-th iteration. The change magnitude criterion of the parameter vector is: and , where and is the convergence tolerance, is the vector of seismic structure model parameters after the (k + 1)-th iteration, is the vector of seismic structure model parameters after the k-th iteration, is the demixing matrix after the (k + 1)-th iteration, is the demixing matrix after the k-th iteration, is the L2 norm, is the Frobenius norm. When any convergence condition is satisfied, the algorithm terminates the iteration and outputs the optimized demixing matrix.

[0036] A7. Output the optimized demixing matrix to separate the precursor signal mixing matrix to obtain the separated signal components containing precursor information.

[0037] In the algorithm output stage, the optimized demixing matrix is used to separate the precursor signal mixing matrix to obtain the separated signal components containing precursor information . Each separated signal component corresponds to a specific physical process, such as surface deformation caused by earthquake precursors, groundwater dynamics changes, rock mass stress adjustment, etc. The separated signals are screened through physical interpretation and quality assessment to select the signal components highly related to earthquake precursors. The algorithm also outputs the estimation results of the seismic structure model parameters, and these parameters reflect the geological structure characteristics and physical properties of the monitoring area.

[0038] Specifically, in an embodiment of the present invention, the data fusion and state inversion module uses a dual-time scale analysis framework to fuse the separated signal components output by the signal processing module. This framework realizes multi-scale precursor state estimation by setting two different time scale analysis windows, short-term and long-term. The short-term scale uses a time window of 1 - 10 minutes, mainly used to capture the rapid change process of earthquake precursor signals, and can identify sudden precursor abnormal events and transient signal characteristics. The long-term scale uses a time window of 1 - 24 hours, used to identify the trend change of earthquake precursor signals, and can capture the slowly accumulating precursor process and long-term evolution pattern. The analysis windows of the two time scales adopt a sliding window mechanism to ensure the continuity and real-time nature of state estimation.

[0039] The short-term scale analysis extracts features and estimates the state of the separated signal components through the sliding window statistical method. For each short-term time window for the signal components within , represents the time instant, represents the value of the k-th separated signal component at time instant t, is the short-term time window length, is the time window for short-term scale analysis. Calculate the statistical feature vector , including statistical quantities such as the mean, variance, skewness, and kurtosis of the signal amplitude. State estimation is achieved through the Bayesian filtering framework, and the state transition equation is: , where is the short-term state vector at time t, is the state transition matrix, is the short-term state vector at time t - 1, is the process noise at time t. The observation equation is: , where is the statistical feature vector at time t, is the observation matrix, is the short-term state vector at time t, is the observation noise at time t. The state estimate and covariance matrix are recursively updated through the Kalman filter algorithm to obtain the short-term precursor state estimate and its uncertainty quantification .

[0040] Long-term scale analysis uses trend analysis and periodic decomposition methods to identify the long-term evolution characteristics of precursor signals. For the signal components within the long-term time window , is the length of the long-term time window, represents the time window for long-term scale analysis, represents the time. Empirical mode decomposition or wavelet transform is used to extract the trend component and periodic component. Trend analysis fits the long-term trend of the signal through linear regression or non-parametric regression methods, and the trend estimation model is: , where, is the value of the trend component at time t, is the trend parameter, is the residual term. Periodic analysis identifies the periodic components in the signal through spectral analysis methods and extracts the main periodic frequencies and corresponding amplitude information. The long-term state estimate combines the trend and periodic analysis results to construct the long-term state vector , and the long-term state estimate and its uncertainty are updated through a similar Bayesian filtering framework.

[0041] The fusion weights are dynamically adjusted according to the quality and uncertainty of the precursor signals, and the weight calculation is based on the signal quality index and the confidence of the state estimate. The fusion weights for short-term and long-term state estimates are respectively: , , where, is the fusion weight of the short-term state estimate at time t, is the fusion weight of the long-term state estimate at time t, and are the short-term and long-term signal quality indicators at time t, respectively, is the matrix trace operation, and are the uncertainty covariance matrices of the short-term state estimate and the long-term state estimate at time t, respectively. The signal quality indicator is calculated by comprehensively considering factors such as signal amplitude, signal-to-noise ratio, spectral purity, and time-domain stability. When the signal quality is high and the uncertainty is low, the corresponding weight increases; conversely, the weight decreases.

[0042] The multi-scale precursor state estimate is obtained by weighted fusion of the short-term and long-term state estimates: , is the fused precursor state estimate vector at time t, is the fusion weight of the short-term state estimate at time t, is the fusion weight of the long-term state estimate at time t, is the short-term precursor state estimate vector at time t, is the long-term precursor state estimate vector at time t.

[0043] The uncertainty quantification of the fused state estimate takes into account the uncertainties and correlations of the two time-scale estimates: , where, is the uncertainty covariance matrix of the fused state estimate at time t, is the fusion weight of the short-term state estimate at time t, is the fusion weight of the long-term state estimate at time t, and are the uncertainty covariance matrices of the short-term state estimate and the long-term state estimate at time t, respectively, is the covariance matrix between the short-term and long-term state estimates. The uncertainty quantification indicators include the standard deviation, confidence interval, and prediction variance of the state estimate.

[0044] The module also implements a state inversion function to inversely infer the possible physical source parameters from the observed precursor state estimate. The state inversion is based on the physical mechanism model of earthquake precursors to establish a mapping relationship between the state estimate and the physical source parameters. The inversion process uses an optimization algorithm to solve the non-linear inversion problem: , wherein is the physical source parameter vector, is the fusion precursor state estimation vector at time t, is the forward physical model, is the regularization term, is the regularization coefficient. The inversion result provides a physical explanation of earthquake precursors, including key information such as possible epicenter location, intensity, occurrence time, etc.

[0045] The output results include the multi-scale precursor state estimation vector, the corresponding uncertainty quantification index, the fusion weight sequence, and the state inversion parameters.

[0046] Specifically, in an embodiment of the present invention, the monitoring and early warning module uses an adaptive threshold mechanism to analyze and process the multi-scale precursor state estimation results output by the data fusion and state inversion module. This mechanism adapts to the change characteristics of precursor signals under different time and environmental conditions by dynamically adjusting the warning threshold. The calculation formula of the adaptive threshold is: , wherein, is the adaptive threshold at time t, is the basic threshold, is the uncertainty adjustment coefficient, is the variance of the precursor state estimation, is the time-varying adjustment factor.

[0047] The time-varying adjustment factor is obtained through historical pattern learning, and its expression is: , where t represents the time, is the time-varying adjustment factor, is the periodic basis function, including time patterns such as annual cycle, seasonal cycle, daily cycle, etc., is the total number of periodic basis functions; is the corresponding weight coefficient, which is estimated by fitting the historical abnormal event data through the least squares method: , where is the basis function matrix, is the historical abnormal event vector, represents the matrix transpose; the sum of the weights is 1. To avoid system instability caused by frequent threshold adjustment, an exponentially weighted moving average is used to smoothly update the threshold: ; wherein is the smoothing parameter, is the original calculated threshold for the (k + 1)-th update, that is, the current newly calculated value, is the adaptive threshold after the (k + 1)-th update, is the adaptive threshold for the k-th update, that is, the historical smoothed value.

[0048] The early warning fatigue suppression algorithm dynamically adjusts the early warning sensitivity by statistically analyzing the historical early warning frequency and accuracy rate, effectively solving the problem of early warning fatigue caused by frequent false alarms. The early warning fatigue suppression algorithm includes: B1. Obtain the current precursor early warning judgment result and historical early warning records as inputs; B2. Set a statistical time window to collect historical precursor early warning data within the window; B3. Statistically count the number of precursor early warnings and corresponding verification results within the window; B4. Calculate the recent precursor early warning accuracy rate and early warning frequency indicators; B5. Calculate the precursor early warning suppression factor based on the ratio of the early warning frequency to the threshold and the accuracy deviation; B6. Multiply the original precursor early warning result by the early warning suppression factor to obtain the final suppressed early warning output.

[0049] The calculation formula for the early warning suppression factor is: , where is the early warning suppression factor, is the recent number of precursor early warnings, is the early warning number threshold, is the recent precursor early warning accuracy rate, calculated by statistically analyzing the verification results of early warning events, is the target accuracy rate, serving as a benchmark for evaluating the early warning quality.

[0050] The final early warning output is obtained by multiplying the original precursor early warning result by the early warning suppression factor: , is the minimum early warning threshold, is the final early warning output, is the original precursor early warning result, is the early warning suppression factor, represents the maximum value function; the early warning suppression factor is set with a lower limit to ensure that the system still maintains the basic early warning function when the suppression mechanism is activated.

[0051] The three - layer anomaly recognition architecture realizes multi - level anomaly detection and recognition from the signal layer to the system layer. The signal layer identifies precursor anomaly event types through the template matching method. This layer maintains a template library of various typical precursor signals, including signal feature templates of different types such as earthquake precursors, artificial interference, and environmental noise. Template matching calculates the similarity between the observed signal and the template signal using the correlation coefficient or the normalized cross - correlation function. Signals with a similarity exceeding the set threshold are identified as anomaly events of the corresponding type. The state layer monitors the changes in the physical properties of earthquake precursors based on the statistical process control method. This layer conducts statistical analysis on the results of precursor state estimation, calculates control chart statistics such as mean, variance, trend slope, etc. Parameters. When the statistic exceeds the control limit, it is determined as a state anomaly. The setting of the control limit takes into account the statistical distribution characteristics of the parameters under normal conditions and the false alarm rate requirements of the system. The system layer identifies the propagation mode of precursor anomalies through spatial correlation analysis. This layer analyzes the spatial correlation characteristics of precursor signals among multiple sensor nodes and identifies the spatial propagation path and influence range of anomaly signals through the correlation matrix and clustering analysis methods.

[0052] The probabilistic risk assessment model establishes a hierarchical risk assessment system and calculates the comprehensive risk level through a multi - index weighted scoring method. The risk assessment is based on four core indicators: The signal amplitude change rate reflects the degree of change in the intensity of the precursor signal; the frequency - domain feature deviation measures the deviation of the signal spectrum feature from the normal state; the spatial consistency index evaluates the spatial consistency of precursor signals at multiple monitoring points; the historical pattern matching degree assesses the danger level of the current anomaly through the pattern similarity with historical earthquake precursor events. The comprehensive risk index is calculated through a linear weighted combination: , where, is the comprehensive risk index, is the signal amplitude change rate, is the frequency - domain feature deviation, is the spatial consistency index, is the historical pattern matching degree, are the weight coefficients of each index. The weight coefficients are determined based on historical data statistics and expert experience and can be optimized and adjusted through machine learning methods. The sum of the weights is 1. The risk level is classified according to the comprehensive risk index and is set to four levels: low risk, medium risk, high risk, and extremely high risk. Each level corresponds to different early - warning response strategies and disposal measures.

[0053] The system implements a hierarchical progressive early warning strategy, including three levels: status early warning, event early warning, and trend early warning. Status early warning targets abnormal fluctuations in precursor state parameters and is triggered when the state estimate exceeds the normal range, mainly used to remind monitoring personnel to pay attention to changes in the system state. Event early warning targets detected abnormal events and is triggered when the three-layer abnormal identification architecture identifies clear precursor abnormal events. The early warning information includes key information such as event type, occurrence time, and affected range. Trend early warning targets long-term trend changes and is triggered when the system detects a persistent abnormal trend and conforms to the earthquake precursor evolution pattern. The early warning information includes the trend development direction, possible evolution results, and recommended countermeasures. The early warning output adopts a standardized format, containing complete information such as timestamp, early warning level, affected area, and recommended measures.

[0054] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection scope of the present invention.

Claims

1. A surface anomaly monitoring data fusion system based on vibration signal processing, characterized in that, Including: A data acquisition module, which is used to monitor the ground surface in real time by using a distributed multi-parameter sensor network, and output time-series monitoring data related to earthquake precursors based on a dual-frequency sampling strategy; A signal processing module, which is used to process the time-series monitoring data by using a dual-parameter joint optimization algorithm, realize adaptive wavefield separation by simultaneously optimizing the demixing matrix and earthquake structure model parameters, and output separated signal components containing precursor information; A data fusion and state inversion module, which is used to perform fusion processing on the separated signal components by using a dual-time scale analysis framework, and output multi-scale precursor state estimation results and uncertainty quantification indicators; A monitoring and early warning module, which is used to analyze the multi-scale precursor state estimation results by using an adaptive threshold mechanism and an early warning fatigue suppression algorithm, and output earthquake precursor classification early warning information through a three-layer anomaly recognition architecture and probabilistic risk assessment.

2. The surface anomaly monitoring data fusion system based on vibration signal processing according to claim 1, wherein The dual-parameter joint optimization algorithm includes: A1. Preprocess and whiten the time-series monitoring data to obtain a precursor signal mixing matrix; A2. Initialize the demixing matrix and earthquake structure model parameters; A3. Construct a joint objective function containing independence measure and earthquake physical constraints based on the precursor signal mixing matrix; A4. Calculate the gradients of the joint objective function with respect to the demixing matrix and earthquake structure model parameters; A5. Update the demixing matrix and earthquake structure model parameters simultaneously by using an adaptive learning rate; A6. Judge whether the convergence criterion is satisfied. If it is satisfied, end the iteration. If it is not satisfied, return to step A4 to continue the iteration; A7. Output the optimized demixing matrix to separate the precursor signal mixing matrix to obtain separated signal components containing precursor information.

3. The surface anomaly monitoring data fusion system based on vibration signal processing according to claim 2, characterized in that, The joint objective function is: ， Among them, is the combined objective function, is the separated precursor signal, W is the demixing matrix, is the seismic structure model parameter vector, is the theoretical mixing matrix based on the seismic structure model, is the independence measure function, is the adaptive regularization function, is the Frobenius norm.

4. The surface anomaly monitoring data fusion system based on vibration signal processing according to claim 1, characterized in that The calculation formula of the adaptive threshold mechanism is: ， wherein, is the adaptive threshold at time t, is the basic threshold, is the uncertainty adjustment coefficient, is the variance of the precursor state estimation, is the time-varying adjustment factor.

5. The surface anomaly monitoring data fusion system based on vibration signal processing according to claim 1, characterized in that, The early warning fatigue suppression algorithm includes: B1. Obtain the current precursor early warning judgment result and historical early warning records as inputs; B2. Set a statistical time window to collect historical precursor early warning data within the window; B3. Count the number of precursor early warnings within the window and the corresponding verification results; B4. Calculate the recent precursor early warning accuracy rate and early warning frequency indicators; B5. Calculate the precursor early warning suppression factor based on the ratio of the early warning frequency to the threshold and the accuracy deviation; B6. Multiply the original precursor early warning result by the early warning suppression factor to obtain the suppressed final early warning output.

6. The surface anomaly monitoring data fusion system based on vibration signal processing according to claim 5, characterized in that The calculation formula of the early warning suppression factor is: ， Among them, is the warning suppression factor, is the number of recent precursor warnings, is the warning number threshold, is the accuracy rate of recent precursor warnings, is the target accuracy rate.

7. The surface anomaly monitoring data fusion system based on vibration signal processing according to claim 1, characterized in that The dual-frequency sampling strategy includes: Under normal conditions, sampling is performed at a standard frequency of 100 - 200 Hz. Under precursor abnormal conditions, high-frequency sampling is performed at 500 - 1000 Hz. When the amplitude change rate of the precursor signal exceeds a preset threshold, it automatically switches to the high-frequency sampling mode.

8. The surface anomaly monitoring data fusion system based on vibration signal processing according to claim 1, characterized in that The dual-time scale analysis framework includes: In the short time scale, a 1 - 10 minute time window is used to capture the rapid change process of the precursor signal. In the long time scale, a 1 - 24 hour time window is used to identify the trend change of the precursor signal. The fusion weight is dynamically adjusted according to the quality and uncertainty of the precursor signal.

9. The ground surface anomaly monitoring data fusion system based on vibration signal processing according to claim 1, characterized in that The three-layer anomaly recognition architecture includes: At the signal layer, precursor abnormal event types are identified through template matching. At the state layer, the physical property changes of earthquake precursors are monitored based on statistical process control. At the system layer, the precursor abnormal propagation mode is identified through spatial correlation analysis.

10. The surface anomaly monitoring data fusion system based on vibration signal processing according to claim 2, wherein The convergence criteria of the two-parameter joint optimization algorithm include: (1) the relative change of the objective function is less than the set threshold; (2) the change amplitude of the parameter vector is less than the convergence tolerance; (3) the number of iterations reaches the maximum limit value.

Images