A method, system and medium for predicting earthquake magnitude based on TLF-GV signal correlation

Abstract

This invention provides a method, system, and medium for predicting earthquake magnitude by correlating TLF-GV signals, belonging to the field of earthquake prediction technology. It aims to address the problems of traditional magnitude prediction models lacking physical constraints, insufficient utilization of TLF-GV features, poor interpretability, and lack of synergy with previous technologies. The method includes: extracting multi-source feature vectors from TLF-GV signals containing the absolute amplitude of gravity components, the relative amplitude of vibration components, and signal phase; and constructing a physical feature log10(Gravity-Relativity) model that shows a positive correlation between TLF-GV signal anomaly intensity and magnitude based on the Gründensky-Ründensky law. GVCI peak ), BF×T anomaly The method involves weighting features based on Bayesian factor (BF) values; using a physical empirical formula derived from historical sample fitting showing a positive correlation between the peak logarithm of the GVCI and magnitude as a physical constraint, and combining this physical constraint with the mean squared error of the magnitude prediction task to construct a loss function; and then constructing an XGBoost model, which is trained on the training set to obtain a prediction model for earthquake magnitude. This method, based on a multi-source feature model of the TLF-GV signal combined with a physically constrained XGBoost regression model, achieves quantitative prediction of earthquake magnitude.

Description

Technical Field

[0001] This invention relates to the field of earthquake early warning technology, specifically to a method, system, and medium for predicting earthquake magnitude using TLF-GV signal correlation. Background Technology

[0002] Earthquake magnitude prediction is one of the core elements of a short-term earthquake prediction system, and its accuracy directly determines the effectiveness of disaster impact assessment and emergency response. Current magnitude prediction technology has gradually evolved from traditional experience-driven approaches to a data- and model-driven approach, but it still focuses on traditional precursor signals (such as groundwater levels, geoelectric fields, GNSS crustal deformation, and microseismic activity sequences), and has not yet developed a technical system specifically for new precursor signals with unique physical properties.

[0003] In terms of technological development, early magnitude prediction relied on empirical formulas based on statistical laws. For example, the Gutenberg-Richard law (GR law) was used to infer the upper limit of magnitude from the frequency of regional earthquakes, or the Omori law was used to fit the attenuation characteristics of aftershock sequences to predict aftershock magnitudes. Although these methods have clear physical meaning, they rely on the strict assumption of "stationarity of seismic activity." When faced with complex geological structures (such as areas where multiple faults converge) or non-stationary precursor evolution (such as sudden micro-ruptures before an earthquake), the prediction error often exceeds magnitude 1.0, making it difficult to meet the needs of short-term forecasting. With the rise of machine learning technology, magnitude prediction has gradually shifted to a data-driven approach, with models such as Support Vector Machine (SVM), Random Forest, Long Short-Term Memory Network (LSTM), and ordinary XGBoost being widely used. SVM processes the nonlinear correlation between precursor signals (such as geoelectric field strength) and magnitude through kernel function mapping, achieving a mean absolute error (MAE) within 0.6 in the prediction of aftershock magnitudes of small and medium-sized earthquakes. LSTM captures the long-term evolution of GNSS deformation sequences with its temporal memory capability, and is used for medium- to long-term (months to years) regional magnitude upper limit prediction. Ordinary XGBoost, due to its feature importance assessment capability, is used for short-term magnitude prediction by integrating multiple traditional precursors such as groundwater level and total geomagnetic intensity.

[0004] However, current earthquake magnitude prediction methods based on traditional precursor signals, while effective in specific scenarios (such as aftershocks and medium- to long-term predictions), still have significant shortcomings in feature utilization, physical coupling, precursor coordination, and interpretability when faced with novel precursor signals that possess "coordination of two physical quantities and specific frequency domain response". Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a method for predicting earthquake magnitude based on TLF-GV signal correlation. This method is based on the multi-source characteristics of TLF-GV signals combined with a physically constrained XGBoost regression model to achieve quantitative prediction of earthquake magnitude.

[0006] To achieve the above objectives, the present invention provides the following technical solution.

[0007] A method for predicting earthquake magnitude using TLF-GV signal correlation includes the following steps:

[0008] Multi-source feature vectors were extracted from the low-frequency gravity-vibration signal TLF-GV signal, including the peak GVCI of the gravity-vibration contribution index GVCI time series. peak Bayesian factor BF, TLF-GV abnormal duration T anomaly The frequency entropy and spatial cluster size are used to predict the probability of vibration based on the GVCI time series; wherein, the TLF-GV signal contains the absolute amplitude of the gravity component, the relative amplitude of the vibration component, and the signal phase;

[0009] Based on the GR law, a physical characteristic of positive correlation between the anomaly intensity and magnitude of the TLF-GV signal is constructed according to the peak value of the GVCI time series. The logarithm of the GVCI peak value is log10( GVCI peak According to BF and T anomaly Constructing the physical characteristics BF×T that couple the dominant physical processes of TLF-GV with the duration of anomalies anomaly ;

[0010] Based on the Bayesian factor (BF) value, the dominant physical process of TLF-GV is determined. When gravity dominates the TLF-GV signal, the GVCI is... peak Weighting, when vibration dominates the TLF-GV signal, for T anomaly Empowerment;

[0011] The logarithm of the peak value of GVCI was obtained by fitting historical samples (log10). GVCI peak The physical empirical formulas that are positively correlated with magnitude are used as physical constraints. The physical constraints are used as regularization terms and combined with the mean square error of the magnitude prediction task to construct a loss function.

[0012] An XGBoost model is constructed, and a prediction model for earthquake magnitude prediction is obtained by training a training set based on a loss function and weighted multi-source feature vectors, two types of physical feature vectors, and corresponding actual magnitude labels.

[0013] Preferably, the prediction standard deviation σ on the validation set is determined based on the prediction model. model , where σ model To validate the sliding window standard deviation of the prediction error within a preset number of months;

[0014] Predicted standard deviation σ model Calculate the predicted magnitude by combining the reliability of the characteristics of the sample to be predicted. Uncertainty range: .

[0015] Preferably, the step of obtaining the GVCI peak logarithm log10 based on historical samples is... GVCI peak The physical empirical formulas that are positively correlated with magnitude serve as physical constraints, and include the following steps:

[0016] Fit physical empirical formulas based on historical samples:

[0017] ;

[0018] In the formula, a , b The fitting coefficients are determined based on historical earthquake data from the region.

[0019] Preferably, the step of constructing a loss function by combining physical constraints as a regularization term with the mean square error of the magnitude prediction task includes the following steps:

[0020] A regularization term is constructed based on empirical physical formulas to penalize deviations between predicted magnitude and empirical physical formulas:

[0021] ;

[0022] The loss function is constructed by combining physical constraints as a regularization term with the mean square error of the magnitude prediction task:

[0023] ;

[0024] In the formula, To predict the magnitude Compared with the actual magnitude The mean square error; This is the physical regularization intensity coefficient.

[0025] Preferably, the step of determining the dominant physical process of TLF-GV based on the Bayesian factor BF value specifically includes:

[0026] If BF > 1, then it is a gravity-dominated TLF-GV signal, corresponding to significant mass migration in the source region. Weight × 1.2;

[0027] If BF < 1, then the vibration-dominated TLF-GV signal corresponds to active fault micro-fractures, and T... anomaly Weight × 1.1.

[0028] Preferably, the method further includes feature screening of multi-source feature vectors and two types of physical feature vectors, using Pearson correlation coefficient to screen features with a correlation ≥0.6 with magnitude, and eliminating weakly correlated features.

[0029] Preferably, the method further includes data cleaning and feature standardization after extracting the multi-source feature vectors of the TLF-GV signal, including the following steps:

[0030] Calculate the feature missing rate of the multi-source feature vector data. If the feature missing rate is ≤3%, fill it with the TLF-GV feature mean of similar earthquake cases. If the missing rate is >3%, remove the sample and use the Z-score criterion to remove extreme abnormal features after missing value processing.

[0031] X is standardized using Z-score norm =(X-μ) / σ handles all features; μ and σ in the formula are the mean and standard deviation of the features of historical TLF-GV earthquakes.

[0032] This invention also provides a TLF-GV signal-correlated earthquake magnitude prediction system, the system comprising:

[0033] processor;

[0034] A memory on which computer programs that can run on the processor are stored;

[0035] The computer program, when executed by the processor, implements the steps of the TLF-GV signal-correlated earthquake magnitude prediction method.

[0036] The present invention also provides a computer-readable storage medium storing a data processing program, which, when executed by a processor, implements the steps of the TLF-GV signal-correlated earthquake magnitude prediction method.

[0037] The beneficial effects of this invention are:

[0038] This invention proposes a TLF-GV signal-correlated earthquake magnitude prediction method. Addressing the shortcomings of existing patents that rely solely on the single feature of "abnormal energy" and only cover strong earthquakes with M≥6.0, this method integrates multi-source features of the TLF-GV signal, covering moderate to strong earthquakes with M≥4.0. This expands the applicable magnitude range and significantly improves the prediction accuracy of moderate to strong earthquakes. Furthermore, this method injects TLF-GV physical constraints, such as the GR law correlation, into the loss function, reducing physical bias in strong earthquake prediction and effectively improving prediction accuracy. Leveraging the duration characteristics of TLF-GV anomalies, it effectively extends the warning time, increasing the emergency preparedness window several times compared to existing patents. Attached Figure Description

[0039] Figure 1 This is a flowchart of a method according to an embodiment of the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0041] Example 1

[0042] Currently, the industry's solutions for magnitude prediction mainly fall into two categories: one is a magnitude prediction scheme based on earthquake precursors without TLF-GV information, and the other is a magnitude prediction scheme based on ultra-low frequency dynamic gravity (ULF-DG) information. Both types of schemes have their own technical logic and limitations, as detailed below:

[0043] (a) Other earthquake precursor magnitude prediction schemes in the industry

[0044] Among the other popular earthquake precursor magnitude prediction schemes currently in use, the most widely applied is the post-earthquake magnitude estimation scheme based on seismic wave parameters. Its core is to use seismic wave signals such as P-waves and S-waves captured by seismic networks after an earthquake, measure the maximum amplitude and vibration period of the waves, and combine them with the epicentral distance between the station and the hypocenter, and substitute them into empirical formulas such as Richter's magnitude formula and body wave magnitude formula to calculate the magnitude. It is mostly used for rapid assessment of disaster within minutes after an earthquake. However, this scheme relies entirely on signals generated after the earthquake and cannot use pre-earthquake precursors to achieve short-term prediction. Moreover, in areas with sparse seismic network coverage, such as the ocean and remote mountainous areas, the amplitude measurement error is likely to exceed 25%, resulting in a magnitude estimation deviation that often exceeds 0.5, and the error is more obvious for small earthquakes (M<5.0). Another popular approach is to predict the magnitude range based on crustal deformation monitoring. This involves using surface displacement maps obtained through InSAR technology and long-term displacement time-series data from GNSS stations (displacement rates are mostly measured in millimeters per year). By combining this with the empirical relationship of "fault length-slip-magnitude," the intensity of fault activity can be inverted, thereby predicting the potential magnitude range for the next few to decades. For example, monitoring of plate boundary fault zones can provide a range of "M6.0~7.0." However, this approach relies on long-term accumulated deformation data, and the prediction timescale is in the year. It cannot capture short-term anomalies before earthquakes and can only output a vague magnitude range, without providing a specific magnitude value. It is completely ineffective for moderate to strong earthquakes with M < 6.0 due to the weak deformation signal. In addition, statistical forecasting based on underground fluid anomalies is also a common approach. This involves monitoring macroscopic anomalies such as sudden rises and falls in well water levels and changes in underground radon concentration, and combining historical data to statistically determine the correlation probability between anomalies and earthquake magnitude. However, this approach relies on manual judgment to determine whether anomalies are precursors to earthquakes, and is easily affected by non-earthquake factors such as rainfall and human pumping, resulting in a misjudgment rate of over 40%. Furthermore, it lacks a standardized mathematical model and cannot output quantitative magnitudes, thus limiting its engineering application value.

[0045] (II) Solutions based on dynamic gravity information

[0046] The solution based on ultra-low frequency dynamic gravity (ULF-DG) information (invention patent: a method, device, equipment and storage for strong earthquake magnitude prediction, 202410300161.6) is currently the industry's standardized magnitude prediction scheme based on dynamic gravity signals. Its core principle is as follows: A ULF-DG instrument collects dynamic gravity characteristic information with amplitude oscillations between 1.5mV and 28mV before a strong earthquake occurs. First, 25 potential anomaly features, such as amplitude, duration, and variance, are extracted from the signal. Then, correlation coefficient analysis is used to screen out the "anomaly energy" feature with the highest positive correlation coefficient to the magnitude. Next, the average energy value per 10,000 energy units is calculated, and a correlation scatter plot is plotted with the corresponding average magnitude. The Pearson correlation coefficient is fitted to verify the significant positive correlation between anomaly energy and magnitude. Then, an exponential regression model of anomaly energy and magnitude is established, and a standardized exponential relationship is obtained after logarithmic transformation of the magnitude. Finally, the effectiveness of the model is verified through mean square error, coefficient of determination, and residual analysis. In practical applications, the predicted magnitude can be output simply by inputting the anomaly energy. While this scheme achieves quantitative prediction of dynamic gravity characteristics, it has significant problems: First, its applicability is narrow, only applicable to strong earthquakes with M≥6.0. For moderate to strong earthquakes with M<6.0, the weak anomalous energy signals make feature extraction difficult, resulting in prediction errors exceeding 0.6 magnitude. Second, it has poor regional adaptability, failing to consider differences in geological conditions across regions (e.g., the dynamic gravity signal characteristics differ between plate boundaries and inland fault zones). Applying the same model parameters to different regions reduces prediction accuracy by more than 50%. Third, it lacks physical constraints. The model is a purely data-driven exponential regression, failing to incorporate seismic physics such as fault rupture critical conditions, deep mass migration, and magnitude correlation. It cannot explain the physical nature of the "exponential relationship between anomalous energy and magnitude," making it difficult to trace the cause of extreme deviations (e.g., misjudging M5.0 as M6.5), leading to significant challenges in engineering iteration and optimization.

[0047] Current earthquake magnitude prediction methods based on traditional precursor signals, while effective in specific scenarios (such as aftershocks and medium- to long-term predictions), have significant shortcomings in feature utilization, physical coupling, precursor coordination, and interpretability. More importantly, there is no magnitude prediction technology for novel precursor signals with "gravity-vibration co-anomaly and specific frequency domain response of 1-30mHz", such as the low-frequency gravity-vibration signal TLF-GV. The TLF-GV signal dataset contains data such as the absolute amplitude of gravity components, the relative amplitude of vibration components, and signal phase. It has three characteristics: ultra-low frequency (1~30mHz), micro-amplitude (1~10μGal), and short-term (occurring 1~20 days before an earthquake). It is collected by deploying atmospheric tidal gravimeters (main layer, relative gravity, wide-area capture) and solid tidal gravimeters (secondary layer, relative gravity, gravity component supplement) according to risk level, combined with a dual-mode collaborative cold atom interferometer. The specific acquisition method is disclosed in the patent application filed on the same day by the invention team: a network monitoring method for short-term earthquake precursor information.

[0048] To address this gap, this invention proposes a "physically constrained XGBoost regression model," which for the first time achieves deep integration of multi-source characteristics of TLF-GV signals with geophysical laws, constructing a magnitude prediction system adapted to this new type of signal. The main steps include:

[0049] S1: Extract the multi-source feature vector of the TLF-GV signal, including the peak GVCI of the gravity-vibration contribution index (GVCI) time series. peak Bayesian factor BF, TLF-GV abnormal duration T anomaly The frequency entropy and spatial cluster size are used to predict the probability of vibration based on the GVCI time series; wherein, the TLF-GV signal contains the absolute amplitude of the gravity component, the relative amplitude of the vibration component and the signal phase.

[0050] S2: Based on the GR law, a physical characteristic of positive correlation between the anomaly intensity and magnitude of the TLF-GV signal is constructed according to the peak value of the GVCI time series. The logarithm of the GVCI peak value is log10( GVCI peak According to BF and T anomaly Constructing the physical characteristics BF×T that couple the dominant physical processes of TLF-GV with the duration of anomalies anomaly .

[0051] S3: Determine the dominant physical process of TLF-GV based on the Bayesian factor (BF) value. When gravity dominates the TLF-GV signal, the GVCI... peak Weighting, when vibration dominates the TLF-GV signal, for T anomaly Empowerment.

[0052] S4: Obtain the logarithm of the peak value of GVCI (log10) based on historical sample fitting. GVCI peak The physical empirical formulas that are positively correlated with magnitude are used as physical constraints. These physical constraints are then combined with the mean square error of the magnitude prediction task to construct a loss function, which is also used as a regularization term.

[0053] S5: Construct an XGBoost model, and train it using a training set based on the loss function and weighted multi-source feature vectors, two types of physical feature vectors, and corresponding actual magnitude labels to obtain a prediction model for earthquake magnitude prediction.

[0054] Detailed steps are as follows Figure 1 As shown, it specifically includes:

[0055] 1. Step 1: Input Data Preprocessing

[0056] 1.1 Data Reception: Receives two types of core data, both from the TLF-GV pre-processing technology stages:

[0057] ① The feature_vector (multi-source feature vector) output from the low-frequency gravity-vibration signal TLF-GV signal processing stage: including GVCI peak value GVCI peak (Quantitative TLF-GV gravity-vibration co-intensity), Bayesian factor BF value (reflecting the dominant physical process of TLF-GV signal), and anomaly duration T anomaly (Duration of TLF-GV anomaly from startup to peak), frequency entropy E freq (TLF-GV signal frequency domain distribution uniformity), spatial cluster size S cluster The five core features are: the number of monitoring points with TLF-GV anomaly coverage, and the number of monitoring points with TLF-GV anomalies.

[0058] The formula for calculating the Gravity-Vibration Contribution Index (GVCI) is as follows:

[0059] ;

[0060] In the formula: This represents the instantaneous value of the enhanced dynamic gravity signal. The instantaneous value of the enhanced vibration signal; σg The standard deviation of the gravity signal during the quiescent period; σa This represents the standard deviation of the vibration signal during the quiet period.

[0061] The formula for calculating the Bayesian factor (BF) is:

[0062] ;

[0063] In the formula: D This is an earthquake event; Indicates earthquakeD When it occurs, the likelihood probability of the enhanced dynamic gravity signal; Indicates earthquake D When it occurs, the likelihood probability of the enhanced vibration signal.

[0064] ② Auxiliary parameters output by the earthquake occurrence time prediction stage: Earthquake occurrence probability (P) earthquake (e.g., the probability of an earthquake occurring in the next 72 hours) is used to enhance the correlation between features and magnitude (TLF-GV features corresponding to a high probability of an earthquake are more reliable).

[0065] Specifically, the probability of earthquake occurrence is predicted based on GVCI time series data. This invention team disclosed in a patent application filed on the same day: "A Method and System for Predicting Earthquake Occurrence Time." Specifically, it predicts earthquake occurrence time based on the GVCI time series of the TLF-GV signal. The GVCI time series is input into a pre-trained Informer time series prediction model to obtain the GVCI sequence for a future period and preliminarily determine potential earthquake occurrence time windows. A Weibull statistical prior distribution characterizing the time interval between TLF-GV anomaly peaks and earthquake occurrence is constructed based on historical earthquake events. Using future GVCI sequences as observational evidence, the parameters of the Weibull statistical prior distribution are dynamically updated using Bayes' theorem combined with Markov chain Monte Carlo sampling to obtain the posterior Weibull distribution. Based on the posterior Weibull distribution, the conditional probability of earthquake occurrence within different future time windows is calculated.

[0066] 1.2 Data Cleaning:

[0067] ① Missing value handling: If the feature missing rate is ≤3%, fill it with the mean of TLF-GV features from similar earthquake cases; if the missing rate is >3%, remove the sample (to avoid introducing invalid information);

[0068] ② Outlier Removal: The Z-score criterion (|Z|>3) is used to remove extreme outliers (such as GVCI). peak >10, far exceeding the normal abnormal range of TLF-GV, to avoid affecting model training.

[0069] 1.3 Feature Standardization and Physical Feature Construction:

[0070] ① Standardization: Z-score standardization is adopted (X norm =(X-μ) / σ) processes all features, where μ and σ are the mean and standard deviation of features from 100+ historical TLF-GV seismic cases, ensuring consistent feature magnitudes;

[0071] ②Construction of physical features: Based on the physical meaning of TLF-GV signals, two additional derivative physical features are added:

[0072] i.log10(GVCI peakBecause the intensity of the TLF-GV signal anomaly is logarithmically positively correlated with the magnitude (in accordance with the GR law), this characteristic is directly related to the physical laws of magnitude.

[0073] ii.BF×T anomaly : Reflects the coupling between the dominant physical process (gravity / vibration) of TLF-GV and the duration of the anomaly. For example, if vibration dominates (BF<1) and the duration is long, it often corresponds to a larger magnitude.

[0074] 1.4 Feature Selection: Features with a correlation ≥ 0.6 with magnitude were selected using the Pearson correlation coefficient (weakly correlated features were removed; features with a correlation < 0.4 were also removed), ultimately retaining 6-7 core features (such as log10(GVCI)). peak ), BF×T anomaly S cluster P earthquake .

[0075] 2. Step 2: Construction of the Physically Constrained XGBoost Model

[0076] 2.1 XGBoost Core Regressor Design

[0077] 2.1.1 Model Structure:

[0078] ① Basic parameters: Number of decision trees (n_estimators) = 100 (balancing accuracy and efficiency), tree depth (max_depth) = 6 (avoiding overfitting), learning rate (learning_rate) = 0.1, subsample = 0.8 (sample sampling rate), colsample_bytree = 0.8 (feature sampling rate);

[0079] ② Objective function: The mean squared error (MSE) of the regression task is used by default, and a physical regularization term is injected later to form a composite loss function;

[0080] ③ Optimizer: The Adam optimizer (weight decay = 1e-5) is used to accelerate model convergence and avoid gradient oscillation.

[0081] 2.1.2 Training Dataset Construction:

[0082] ① Training set: TLF-GV correlation features of 100+ historical earthquakes (calculated back through pre-sequence techniques), each sample contains "screened feature vector + actual magnitude label";

[0083] ② Validation set / Test set: Select 20% and 10% of the historical samples respectively. The validation set is used for hyperparameter tuning (such as adjusting tree depth and learning rate), and the test set is used to evaluate the model's generalization ability (such as MAE and R²).

[0084] 2.2 Physical Constraint Injection

[0085] A physical regularization term is added to the XGBoost loss function to force the model's predictions to conform to the geophysical laws of the TLF-GV signal. The composite loss function is defined as follows:

[0086] ;

[0087] in:

[0088] ① Predicting the magnitude Compared with the actual magnitude The mean square error ensures data-driven accuracy;

[0089] ② Physical regularization intensity coefficient (obtained through validation set tuning) =0.5), controlling the degree of influence of physical constraints;

[0090] ③ Physical empirical formula based on TLF-GV signal (obtained by fitting 100+ historical samples) a =2.5、 b =3.0), reflecting a positive correlation between the peak logarithm of the GVCI and the magnitude (consistent with the GR law, i.e., the larger the magnitude, the stronger the TLF-GV co-anomaly). The regional adaptation parameter table for the physical empirical formula is as follows, with parameters determined based on fitting historical earthquake cases in the region:

[0091] Table 1. Fitting parameters for historical earthquake cases in the region;

[0092]

[0093] ④ Regularization term : Penalize the deviation between the predicted magnitude and the physical empirical formula, and avoid extreme values ​​in the model output that violate the physical meaning of the TLF-GV signal (e.g. =8.0 but =0.5, which clearly does not conform to empirical relationships.

[0094] Furthermore, a second-order constraint is added to address the TLF-GV-dominated physical processes reflected by the BF value:

[0095] ① If BF > 1 (gravity dominates TLF-GV signal, corresponding to significant mass migration in the source region), then increase the weight in the features (e.g., (Weight × 1.2)

[0096] ②If BF < 1 (vibration dominates TLF-GV signal, corresponding to active fault micro-fracture), then T anomaly The weight is multiplied by 1.1 to ensure that the impact of different physical processes on the magnitude is reasonably represented.

[0097] 2.3 Model Training and Optimization

[0098] 2.3.1 Training Process:

[0099] ① Initialize model parameters (such as number of trees, depth);

[0100] ② Iterative training: In each iteration, the composite loss function Obj is calculated, and the model parameters (such as the split threshold of the decision tree and the weight of the leaf nodes) are updated through gradient descent.

[0101] ③ Early stopping strategy: If the validation set loss does not decrease for 10 consecutive rounds, stop training to avoid overfitting;

[0102] ④ Model Evaluation: The evaluation metrics on the test set must meet the following requirements: MAE ≤ 0.3 and R² ≥ 0.85. Otherwise, the hyperparameters should be readjusted (e.g., by increasing the regularization strength). β ).

[0103] 2.3.2 Hyperparameter Tuning: A grid search was used to traverse the parameter combinations, and the optimal parameters are as follows:

[0104] Table 2 Optimal combination of hyperparameters;

[0105]

[0106] 3. Step 3: Model Interpretation and Feature Importance Analysis

[0107] 3.1 Feature Importance Calculation: Using the `feature_importances` attribute of XGBoost, the contribution of each feature is calculated (based on the information gain of the feature in the decision tree split), and the feature importance ranking is output; Example ranking (based on historical samples):

[0108] ①log10(GVCI peak (Contribution ≈ 45%): The intensity of the TLF-GV co-anomaly is the core influencing factor on the magnitude;

[0109] ②BF×T anomaly (Contribution ≈ 25%): The coupling between the dominant physical process and the duration of the anomaly is the second most important factor.

[0110] ③S cluster (Contribution ≈ 15%): The spatial coverage of TLF-GV anomalies reflects the magnitude of the earthquake source;

[0111] ④P earthquake (Contribution ≈ 10%): Earthquake occurrence probability assists in verifying the reliability of the feature;

[0112] ⑤ Other characteristics (contribution ≈ 5%).

[0113] 3.2 Interpretation of Physical Meaning: Explain the prediction logic by considering the importance of features, for example:

[0114] If a sample predicts an earthquake magnitude M=6.5, the core basis is "log10(GVCI)". peak =1.4 (corresponding to GVCI) peak =25, TLF-GV strong cohesive anomaly), BF×T anomaly =30) (Vibration is dominant and the anomaly lasts for 30 hours), which is consistent with the physical law that "strong TLF-GV anomalies correspond to large magnitude earthquakes";

[0115] If the prediction deviation (e.g.) =7.0 but the actual M=6.0), the problem can be located by feature importance (e.g., log10(GVCI)). peak Measurement errors lead to excessively high weights.

[0116] 4. Step 4: Magnitude Prediction and Uncertainty Output

[0117] 4.1 Magnitude Prediction: Input the preprocessed TLF-GV feature vector to be predicted into the trained model, and output the predicted magnitude value. Example: Input feature "log10(GVCI) peak =1.3, BF×T anomaly =28、S cluster =8、P earthquake =0.85”, model output =6.3).

[0118] 4.2 Uncertainty Analysis: Based on the model's prediction standard deviation (σ) on the validation set model ≈0.3 level, σ model The calculation method is the 'sliding window standard deviation of the prediction error of the validation set over the past 3 months', with a window size of 10 earthquake cases. σ is updated once for each new earthquake case. model This ensures that the standard deviation reflects the latest model error level. It also considers the reliability of the features of the sample to be predicted (e.g., P0). earthquake The higher the value, the stronger the reliability. The range of uncertainty in magnitude calculation is as follows: Example: P earthquake When P = 0.85, the uncertainty range is 6.3 ± (0.5 + 0.5 × 0.15) × 0.3 = 6.3 ± 0.17, which simplifies to M6.3 ± 0.2; if P earthquake =0.5 (low reliability), the range is expanded to 6.3±0.3.

[0119] 4.3 Structured Report Output: The report includes the following core content (in JSON / PDF format):

[0120] Input basic information: TLF-GV feature vector (e.g., GVCIpeak=20, BF=0.8), and the probability of seismicity;

[0121] Magnitude Result: Predicted Magnitude Uncertainty range (e.g., M6.3±0.2);

[0122] Feature Importance: Ranking and Physical Interpretation of the Contribution of Each TLF-GV Feature;

[0123] Physical constraint satisfaction: the deviation between the predicted magnitude and the physical empirical formula (e.g., "deviation = magnitude 0.1, physical constraints are satisfied").

[0124] In this embodiment, compared with the closest prior art: a method, device, equipment and storage for predicting strong earthquake magnitude (202410300161.6), it has the following advantages:

[0125] 1. Improved prediction accuracy and applicability: Existing patents rely solely on the single feature of "abnormal energy" and only cover strong earthquakes with M≥6.0; this invention integrates 6~7-dimensional multi-source features of TLF-GV signals (such as log...). 10 (GVCI peak ), BF×T anomaly The test set MAE is reduced to below 0.3, covering moderate and strong earthquakes with M≥4.0, expanding the applicable magnitude range by 40% compared to existing patents, and improving the prediction accuracy of moderate and strong earthquakes by 50%.

[0126] 2. Breakthrough in physical rationality and short-term early warning capability: Existing patents are purely data-driven models, which are prone to physical contradictions such as "M5.0 being misjudged as M6.5", and the early warning time is only a few hours. This invention injects TLF-GV physical constraints (such as GR law correlation) into the loss function, and the physical deviation of strong earthquake prediction is ≤0.2. Relying on the "anomaly 1~20 days before earthquake" characteristic of TLF-GV, the early warning time is extended to 1~20 days, and the emergency preparation window is increased several times compared with existing patents.

[0127] 3. Optimization of interpretability and problem location efficiency: Existing patents use "black box" index models that cannot explain the "logic between energy and magnitude," making it difficult to trace the source of deviations; this invention, through XGBoost feature importance analysis, clarifies the log... 10 (GVCI peak (Contribution ≈ 45%), BF×T anomaly With core features such as (contribution rate ≈ 25%), the efficiency of problem localization is improved by 60% compared with existing patents, supporting rapid model iteration.

[0128] The above is one embodiment of the TLF-GV signal-correlated earthquake magnitude prediction method provided in this example. Based on the same idea, this example also provides a corresponding TLF-GV signal-correlated earthquake magnitude prediction system. Specific limitations of the TLF-GV signal-correlated earthquake magnitude prediction system can be found in the limitations of the TLF-GV signal-correlated earthquake magnitude prediction method described above, and will not be repeated here. Each module in the above TLF-GV signal-correlated earthquake magnitude prediction system can be implemented entirely or partially through software, hardware, or a combination thereof. Each module can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0129] This embodiment also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 The provided method for predicting earthquake magnitude by associating TLF-GV signals.

[0130] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0131] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting earthquake magnitude using TLF-GV signal correlation, characterized in that, Includes the following steps: Multi-source feature vectors were extracted from the low-frequency gravity-vibration signal TLF-GV, including the peak GVCI of the gravity-vibration contribution index GVCI time series. peak Bayesian factor BF, TLF-GV abnormal duration T anomaly The frequency entropy and spatial cluster size are used to predict the probability of vibration based on the GVCI time series; wherein, the TLF-GV signal contains the absolute amplitude of the gravity component, the relative amplitude of the vibration component, and the signal phase; Based on the GR law, a physical characteristic of positive correlation between the anomaly intensity and magnitude of the TLF-GV signal is constructed according to the peak value of the GVCI time series. The logarithm of the GVCI peak value is log10( GVCI peak According to BF and T anomaly Constructing the physical characteristics BF×T that couple the dominant physical processes of TLF-GV with the duration of anomalies anomaly ; Based on the Bayesian factor (BF) value, the dominant physical process of TLF-GV is determined. When gravity dominates the TLF-GV signal, the GVCI is... peak Weighting, when vibration dominates the TLF-GV signal, for T anomaly Empowerment; The logarithm of the peak value of GVCI was obtained by fitting historical samples (log10). GVCI peak The physical empirical formulas that are positively correlated with magnitude are used as physical constraints. The physical constraints are used as regularization terms and combined with the mean square error of the magnitude prediction task to construct a loss function. An XGBoost model is constructed, and a prediction model for earthquake magnitude prediction is obtained by training a training set based on a loss function and weighted multi-source feature vectors, two types of physical feature vectors, and corresponding actual magnitude labels.

2. The TLF-GV signal-correlated earthquake magnitude prediction method according to claim 1, characterized in that, This also includes determining the range of magnitude uncertainty based on the prediction results of the prediction model, specifically including: The prediction standard deviation σ on the validation set is determined based on the prediction model. model , where σ model To validate the sliding window standard deviation of the prediction error within a preset number of months; Predicted standard deviation σ model Calculate the predicted magnitude by combining the reliability of the characteristics of the sample to be predicted. Uncertainty range: , P earthquake This represents the probability of an earthquake.

3. The TLF-GV signal-correlated earthquake magnitude prediction method according to claim 1, characterized in that, The GVCI peak logarithm log10 was obtained by fitting historical samples. GVCI peak The physical empirical formulas that are positively correlated with magnitude serve as physical constraints, and include the following steps: Fit physical empirical formulas based on historical samples: ； In the formula, a , b The fitting coefficients are determined based on historical earthquake data from the region.

4. The TLF-GV signal-correlated earthquake magnitude prediction method according to claim 3, characterized in that, The method of constructing a loss function by combining physical constraints as a regularization term with the mean square error of the magnitude prediction task includes the following steps: A regularization term is constructed based on empirical physical formulas to penalize deviations between predicted magnitude and empirical physical formulas: ； The loss function is constructed by combining physical constraints as a regularization term with the mean square error of the magnitude prediction task: ； In the formula, To predict the magnitude Compared with the actual magnitude The mean square error; This is the physical regularization intensity coefficient.

5. The TLF-GV signal-correlated earthquake magnitude prediction method according to claim 1, characterized in that, The determination of the dominant physical process of TLF-GV based on the Bayesian factor BF value specifically includes: If BF > 1, then it is a gravity-dominated TLF-GV signal, corresponding to significant mass migration in the source region. Weight × 1.2; If BF < 1, then the vibration-dominated TLF-GV signal corresponds to active fault micro-fractures, and T... anomaly Weight × 1.

1.

6. The TLF-GV signal-correlated earthquake magnitude prediction method according to claim 1, characterized in that, It also includes feature filtering of multi-source feature vectors and two types of physical feature vectors, using Pearson correlation coefficient to filter features with a correlation of ≥0.6 with magnitude, and eliminating weakly correlated features.

7. The TLF-GV signal-correlated earthquake magnitude prediction method according to claim 1, characterized in that, It also includes data cleaning and feature standardization after extracting the multi-source feature vectors of the TLF-GV signal, including the following steps: Calculate the feature missing rate of the multi-source feature vector dataset. If the feature missing rate is ≤3%, fill it with the TLF-GV feature mean of similar earthquake cases. If the missing rate is >3%, remove the sample and use the Z-score criterion to remove extreme abnormal features after missing value processing. X is standardized using Z-score norm =(X-μ) / σ handles all features; μ and σ in the formula are the mean and standard deviation of the features of historical TLF-GV earthquakes.

8. A TLF-GV signal-correlated earthquake magnitude prediction system, characterized in that, The system includes: processor; A memory on which computer programs that can run on the processor are stored; When the computer program is executed by the processor, it implements the steps of the TLF-GV signal-correlated earthquake magnitude prediction method as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a data processing program, which, when executed by a processor, implements the steps of the TLF-GV signal-correlated earthquake magnitude prediction method as described in any one of claims 1 to 7.

Images