Tlf-gv signal correlation earthquake time prediction method and system

Abstract

The application discloses a TLF-GV signal correlation earthquake time prediction method and system, relates to the cross technical field of earthquake short-term and impending prediction and machine learning, and comprises the following steps: acquiring a GVCI time sequence of a TLF-GV signal, inputting the GVCI time sequence into a pre-trained Informer time sequence prediction model, and obtaining a GVCI sequence in a future period of time; constructing a Weibull statistical prior distribution representing an interval between an abnormal peak value of the TLF-GV and an earthquake time based on historical earthquake events; taking the future GVCI sequence as observation evidence, dynamically updating parameters of the Weibull statistical prior distribution through Bayes theorem, and obtaining a posterior Weibull distribution; and calculating an earthquake condition probability in different time windows in the future according to the posterior Weibull distribution; the method can capture long-term dependence and dynamic statistical calibration of a TLF-GV signal derivative sequence, and the reliability of earthquake time prediction is improved.

Description

Technical Field

[0001] This invention relates to the field of short-term earthquake prediction and machine learning, specifically to a method and system for predicting earthquake occurrence time by correlating TLF-GV signals. Background Technology

[0002] Earthquake timing prediction is a crucial component of short-term earthquake forecasting systems, and its accuracy directly impacts the timeliness of disaster warnings and the scientific rigor of emergency responses. Traditional earthquake timing prediction methods primarily rely on statistical models based on seismic activity (such as recurrence cycle models and ETAS models) or empirical formulas based on single precursor observations (such as groundwater, geoelectricity, and geomagnetism). However, these methods generally suffer from insufficient ability to capture long-period, non-stationary precursor signals, lack of dynamic calibration mechanisms, and weak correlations in physical meaning, making it difficult to meet the demands of high-precision short-term forecasting.

[0003] With the development of artificial intelligence technology, machine learning methods have been gradually introduced into the field of earthquake prediction. Early studies mostly used traditional time series models such as ARIMA and exponential smoothing to fit precursor data, but these methods require stable data and are difficult to adapt to the non-stationary evolution of precursor features. Subsequently, deep learning models such as recurrent neural networks (RNN) and long short-term memory networks (LSTM) were used to model the earthquake timing of electromagnetic signals and microseismic activity sequences because they can capture time series dependencies. Some studies have used LSTM memory units to learn the precursor changes several days before the earthquake, reducing the prediction error to a certain range. In recent years, although the Transformer model has improved the ability to extract local features through the self-attention mechanism, its computational complexity increases quadratically with the sequence length, making it difficult to efficiently process long-period (such as several days to several weeks) precursor time series data.

[0004] There is an existing "Method for Predicting Major Earthquakes Based on Atmospheric Tidal Gravity Precursors" constructed for ULF-DG signals. This scheme attempts to provide a quantitative time anchor for short-term prediction through statistical modeling. Its specific implementation path is as follows: First, a micro-galaxy level disturbance signal is collected using a dynamic atmospheric tidal gravimeter. After removing background noise through a basic filtering algorithm, a paired sample library of "ULF-DG precursor signal-earthquake event" is established. Then, a decision tree algorithm is used to fit the data of "time difference from the appearance of the precursor to the earthquake" in the sample library to generate a cumulative probability distribution curve. For example, the curve shape can be used to intuitively show the "probability of earthquake occurring within N days after the appearance of the precursor". The core logic is "single gravity signal acquisition + decision tree statistical modeling", which attempts to transform empirical judgment into probability-driven decision-making. This scheme has the following defects: (1) The atmospheric tidal / solid tidal gravimeter on which this scheme depends has a design target frequency band that is different from the core frequency domain (1-30) of the low-frequency gravity-vibration signal TLF-GV signal. (1) The mismatch of mHz) leads to the lack of effective input basis for the model; (2) The scheme has a strong constraint that “an earthquake will inevitably occur within 20 days after the occurrence of an anomaly” in the statistical model, which forces the cumulative probability curve to converge to 100% after 20 days. The model has a mechanical time limit constraint, which violates the nonlinear law of crustal stress evolution; (3) The traditional gravimeter and simple filtering method used in this scheme lack an effective suppression mechanism for non-seismic interference (such as sudden changes in air pressure and instrument drift). The lack of anti-interference mechanism and low signal credibility lead to the distortion of the sample library and model.

[0005] In summary, existing methods struggle to capture the long-term dependence of TLF-GV signal-derived sequences and lack dynamic statistical calibration, leading to fundamental errors, loss of reliability, insufficient scientific rigor, and limited practicality in predicting earthquake occurrence time. Consequently, the prediction results are almost impossible to apply to actual short-term forecasting operations. Summary of the Invention

[0006] To address the shortcomings of existing technologies, such as difficulty in capturing the long-term dependence of TLF-GV signal-derived sequences and lack of dynamic statistical calibration, which leads to fundamental errors, loss of reliability, insufficient scientific rigor, and limited practicality in predicting earthquake occurrence time, this invention proposes a TLF-GV signal-correlated earthquake occurrence time prediction method and system, thereby solving the problems existing in the prior art.

[0007] A method for predicting earthquake occurrence time by correlating TLF-GV signals includes the following steps:

[0008] Obtain the gravity-vibration contribution index (GVCI) time series from the potential source region to the low-frequency gravity-vibration TLF-GV signal;

[0009] The GVCI time series is input into the pre-trained Informer time series prediction model to predict the GVCI sequence in the future period, and the potential earthquake time window is initially determined based on the predicted sequence.

[0010] Based on historical earthquake events and their associated TLF-GV signal anomaly data, a Weibull statistical prior distribution is constructed to characterize the interval from the peak of the TLF-GV signal anomaly to the earthquake occurrence time. Using the future GVCI sequence predicted by the Informer time series prediction model as observational evidence, the parameters of the Weibull statistical prior distribution are dynamically updated through Bayes' theorem to obtain the posterior Weibull distribution.

[0011] Based on the posterior Weibull distribution, the conditional probability of earthquake occurrence within different future time windows is calculated; the earthquake occurrence time is predicted based on the conditional probability.

[0012] Furthermore, after obtaining the gravity-vibration contribution index (GVCI) time series from the potential seismic source region to the low-frequency gravity-vibration TLF-GV signal, the GVCI time series is preprocessed, specifically including the following steps:

[0013] The 3σ criterion was used to remove outliers in the GVCI time series whose GVCI values ​​exceeded [μ-3σ, μ+3σ]; where μ and σ are the mean and standard deviation of GVCI during the TLF-GV quiet period.

[0014] Missing values ​​in the GVCI time series after outlier removal are filled and repaired using linear interpolation or Kalman interpolation.

[0015] The repaired GVCI time series was subjected to min-max normalization to map the GVCI values ​​to the [0,1] interval;

[0016] The normalized long sequence is sliced ​​according to a fixed time window and step size to obtain the preprocessed GVCI time series.

[0017] Furthermore, the step of inputting the GVCI time series into the pre-trained Informer time series prediction model to predict the GVCI sequence for a future period specifically includes the following steps:

[0018] Extract the GVCI time series according to the preset historical time window length;

[0019] The truncated GVCI time series is input into the encoder of the pre-trained Informer time series prediction model. The ProbSparse self-attention mechanism is used to process the input truncated GVCI time series to extract the long-term evolution pattern features of the TLF-GV signal, which are "gradual rise during the calm period - peak during the abnormal period - sudden drop before the oscillation". The sequence feature vector that can reflect the co-evolution trend of the TLF-GV signal is generated.

[0020] Using a sequence feature vector reflecting the co-evolution trend of TLF-GV signals and an initialized target sequence as input to the decoder, GVCI sequences for a future period are gradually generated in parallel through a masked multi-head attention mechanism and a cross-attention mechanism with the encoder output. pred (t).

[0021] Furthermore, the preliminary determination of the potential earthquake occurrence time window based on the predicted sequence specifically includes the following steps:

[0022] GVCI based on decoder output pred (t), based on the evolution law of TLF-GV signal, a sudden drop judgment threshold is set, if GVCI pred (t+1)-GVCI pred (t)<-0.2×GVCI pred,peak GVCI pred,peak To predict the TLF-GV anomaly peak in the sequence, and if this condition is met for three consecutive time steps, then t is determined to be the potential earthquake start time, and the corresponding time window is [t, t+24] hours.

[0023] Furthermore, the historical earthquake events and their associated TLF-GV signal anomaly data meet the following screening criteria: magnitude M≥5.0, GVCI peak value≥1.5, and TLF-GV signal anomaly duration≥12 hours.

[0024] Furthermore, the probability density function of the Weibull prior distribution is:

[0025] ;

[0026] Where, Δ t λ represents the interval from the abnormal peak value of the TLF-GV signal to the time of oscillation, and λ represents the abnormal characteristic lifetime Δ of the TLF-GV. t the median of k For shape parameters.

[0027] Furthermore, the step of using the future GVCI sequence predicted by the Informer time series prediction model as observational evidence, and dynamically updating the parameters of the Weibull statistical prior distribution through Bayes' theorem to obtain the posterior Weibull distribution, specifically includes the following steps:

[0028] GVCI predicted from the Informer model pred (t) extracts observational evidence feature E; the observational evidence feature includes at least: the ratio R of the predicted GVCI peak to the historical average peak. peak The predicted duration of the TLF-GV anomaly, D anomaly The steep slope S of the GVCI sequence drop ;

[0029] Construct a likelihood function P(E|λ,k) to evaluate the probability of the observed evidence feature E occurring given the Weibull distribution parameters;

[0030] Based on Bayes' theorem, the Weibull prior distribution is combined with the likelihood function. The posterior distribution π(λ,k|E) of parameters λ and k is calculated using the Markov chain Monte Carlo (MCMC) sampling method. The mean of the posterior distribution is then used as the updated parameter λ. post k post .

[0031] Furthermore, the conditional probability of earthquake occurrence within different time windows in the future is calculated based on the posterior Weibull distribution, specifically by using the updated λ. post k post Obtain the conditional probability of the final earthquake occurrence time. The process is represented as follows:

[0032] ;

[0033] in, T future For the future.

[0034] The present invention also includes a TLF-GV signal-correlated earthquake occurrence time prediction system, comprising:

[0035] The acquisition module is used to acquire the gravity-vibration contribution index (GVCI) time series from the potential source region to the low-frequency gravity-vibration TLF-GV signal;

[0036] The sequence prediction module is used to input the GVCI time series into the pre-trained Informer time series prediction model, predict the GVCI sequence in the future, and preliminarily determine the potential earthquake time window based on the predicted sequence.

[0037] The posterior Weibull distribution construction module is used to construct a Weibull statistical prior distribution representing the interval from the peak of the TLF-GV signal anomaly to the earthquake occurrence time based on historical earthquake events and their associated TLF-GV signal anomaly data. Using the future GVCI sequence predicted by the Informer time series prediction model as observational evidence, the parameters of the Weibull statistical prior distribution are dynamically updated through Bayes' theorem to obtain the posterior Weibull distribution.

[0038] The prediction module is used to calculate the conditional probability of earthquake occurrence within different time windows in the future based on the posterior Weibull distribution; and to predict the earthquake occurrence time based on the conditional probability.

[0039] This invention provides a method for predicting earthquake occurrence time by correlating TLF-GV signals, which has the following beneficial effects:

[0040] This invention learns the complete evolutionary pattern of the GVCI sequence from "TLF-GV calm period - TLF-GV anomaly period - pre-earthquake occurrence" using the Informer model, extending the long-term reliance on the capture length from 10 hours in traditional models to 30 days. Based on the Weibull distribution, it constructs a statistical prior for "TLF-GV signal anomaly peak to earthquake occurrence time" in historical earthquake cases, and uses Bayes' theorem to update statistical parameters with real-time GVCI sequences (reflecting the dynamics of TLF-GV signals), making the prediction results both data-driven dynamism and the reliability of TLF-GV related statistical regularities. It outputs the probabilistic earthquake occurrence time associated with TLF-GV, directly supporting the hierarchical decision-making of "monitoring-early warning-emergency response" and meeting the real-time requirements of short-term earthquake prediction. Attached Figure Description

[0041] Figure 1 This is a diagram illustrating the architecture of the TLF-GV signal-correlated earthquake occurrence time prediction method in this embodiment of the invention.

[0042] Figure 2 This is a flowchart illustrating the method for predicting earthquake occurrence time by correlating TLF-GV signals in an embodiment of the present invention. Detailed Implementation

[0043] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0044] This invention proposes a method for predicting earthquake occurrence time by correlating TLF-GV signals, aiming to accurately capture the long-term dependency of TLF-GV-correlated GVCI sequences. By utilizing the ProbSparse attention mechanism of the Informer model, it learns the complete evolutionary pattern of the GVCI sequence from "TLF-GV calm period - TLF-GV anomaly period - pre-earthquake occurrence," extending the long-term dependency capture length from 10 hours in traditional models to 30 days. It also achieves dynamic statistical calibration of TLF-GV correlation: based on the Weibull distribution, it constructs a statistical prior for "TLF-GV signal anomaly peak to earthquake occurrence time" from historical earthquake cases, and combines Bayes' theorem with real-time GVCI sequences (inverse) to... It dynamically updates statistical parameters to reflect the TLF-GV signal, ensuring that the prediction results combine the dynamic nature of data-driven forecasts with the reliability of TLF-GV-related statistical regularities; it outputs the probabilistic occurrence time associated with TLF-GV: for short-term forecasting needs, it outputs the daily conditional probability of occurrence for the next 7 days (e.g., "P=55% in the next 24 hours, P=78% in the next 48 hours"), supporting graded decision-making based on TLF-GV signal risk; it ensures synergy with preceding TLF-GV technologies: it strictly adapts to the GVCI index format (timestamp, numerical sequence, TLF-GV anomaly annotation) output from the TLF-GV signal processing stage, requiring no additional data preprocessing and directly inheriting the results of preceding technologies.

[0045] like Figure 1 , Figure 2 As shown, with "TLF-GV signal association with GVCI sequence evolution learning - TLF-GV correlation statistical prior construction - dynamic Bayesian calibration - probability output" as the core logic, this method specifically includes the following steps:

[0046] S1. Input Data Preprocessing: Receive the GVCI exponential time series output from the TLF-GV signal processing stage (including TLF-GV calm and abnormal period data, sampling frequency 1Hz, time span ≥72 hours), and perform data cleaning and normalization. Transform the GVCI exponential time series output from the TLF-GV signal processing stage into the input format of the adaptation model, eliminating noise and dimensional influences while preserving the cooperative characteristics of the TLF-GV signal. Among them, the low-frequency gravity-vibration TLF-GV signal includes the absolute amplitude of the gravity component, the relative amplitude of the vibration component, and the signal phase; it has three characteristics: ultra-low frequency (1~30mHz), micro-amplitude (1~10μGal), and short-term (appearing 1~20 days before the 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 by the team of this invention in a patent filed on the same day, "A Network Monitoring Method and System for Short-Term Earthquake Precursor Information".

[0047] The GVCI exponent time series output by the TLF-GV signal processing stage includes the following steps: The original TLF-GV time series signal is decomposed using the CEEMDAN algorithm (Adaptive Noise Complete Set Empirical Mode Decomposition) to obtain several Intrinsic Mode Function (IMF) components; gravity-related IMF components and vibration-related IMF components in the corresponding frequency bands of the original TLF-GV time series signal are selected from the IMF components, and denoised and reconstructed respectively to obtain clean dynamic gravity and vibration signals; physical enhancement processing is performed on the reconstructed dynamic gravity and vibration signals; based on the enhanced dynamic gravity and vibration signals, the gravity-vibration contribution index (GVCI) and Bayesian factor (BF) are calculated; candidate outliers are marked by threshold filtering of the GVCI; feature vectors of each candidate outlier are extracted based on the Bayesian factor, and these feature vectors are used as input to an isolated forest to output an anomaly score for each candidate outlier; suspected outliers are marked by comparing each anomaly score with a threshold; and valid outliers are screened based on the physical mechanism of Coulomb stress change. The specific method of obtaining it is disclosed by the team of this invention in a patent filed on the same day: a method and system for detecting anomalies in earthquake prediction signals based on TLF-GV.

[0048] S1.1 Data Reception: Receive the GVCI index time series output by the TLF-GV signal processing stage. The format is JSON, containing the following fields: time_stamp (time stamp, accurate to the second), gvci_value (GVCI index value, quantifying the gravity-vibration coordination intensity of the TLF-GV signal), and tlf_gv_anomaly_flag (TLF-GV anomaly flag, 0 = TLF-GV calm period, 1 = TLF-GV anomaly period). The time span is ≥72 hours (including at least one complete TLF-GV anomaly cycle).

[0049] S1.2 Data Cleaning:

[0050] Extreme values ​​removal: Outliers with GVCI values ​​exceeding [μ-3σ, μ+3σ] are removed using the 3σ criterion (μ and σ are the mean and standard deviation of the TLF-GV GVCI during the quiet period, calculated from the quiet period data of the TLF-GV signal processing stage).

[0051] Missing value repair: If the proportion of missing points is ≤5%, linear interpolation method is used to fill in the missing points (missing length ≤10 sampling points) or Kalman interpolation method is used to fill in the missing points (missing length >10 sampling points) to ensure the continuity of time sequence and avoid the break of TLF-GV signal evolution information.

[0052] S1.3 Normalization: Min-max normalization is used to map the GVCI values ​​to the [0,1] interval, preserving the relative differences in the coordinated strength of TLF-GV signals. Formula:

[0053] ;

[0054] in, GVCI min , GVCI max These represent the minimum and maximum GVCI values ​​in the TLF-GV signal processing sequence, respectively.

[0055] S1.4 Sequence Slicing: The normalized GVCI sequence is sliced ​​according to "time window = 72 hours, step size = 1 hour" to construct the model input sample (each sample contains 72 hours of TLF-GV associated historical GVCI data, corresponding to 1 prediction target: the probability of earthquakes in the next 7 days).

[0056] S2. Informer Model Training and Prediction: Construct an Informer model adapted to the GVCI sequence (reflecting the coordinated characteristics of TLF-GV signals), learn the "gradual rise-sudden fall" evolution pattern of the GVCI sequence, and output the predicted GVCI index sequence for the next 7 days and the preliminary judgment of the time point of the earthquake (e.g., "the GVCI drops suddenly within the next 48 hours, corresponding to the abnormal weakening of the TLF-GV signal, which is judged as a potential earthquake window"), and capture the key turning point from the abnormality to the earthquake in the TLF-GV signal.

[0057] S2.1 Informer model structure design (adapted to TLF-GV associated GVCI sequences).

[0058] S2.1.1 Encoder:

[0059] ① Input layer: Receives a 72-hour × 1-dimensional GVCI normalized sequence (72 time steps, each step size corresponds to 1 hour of TLF-GV signal coordination strength).

[0060] ②Embedding: The 1-dimensional sequence is mapped to a 512-dimensional feature space, and sinusoidal position embedding is used to preserve the temporal information of the TLF-GV signal evolution.

[0061] ③ProbSparse attention layer: The ProbSparse attention mechanism is adopted (computational complexity O(L logL), where L is the sequence length). The number of attention heads is set to 8, the hidden layer dimension is 512, and the dropout is 0.1. It effectively captures the long-term dependency of the GVCI sequence "TLF-GV calm period gradual rise-TLF-GV abnormal period peak-early drop before the oscillation".

[0062] ④ Encoder output: Generates a 512-dimensional sequence feature vector, reflecting the trend of TLF-GV signal co-evolution.

[0063] S2.1.2, Decoder:

[0064] ① Input layer: Receives the initial prediction sequence of "7 days × 1 dimension" (initial value is set as the mean of GVCI during the TLF-GV calm period).

[0065] ② Masked Multi-Head Attention: Prevents the decoder from focusing on information from future time steps, ensuring the causality of TLF-GV signal evolution prediction.

[0066] ③ Cross Attention: Combines the TLF-GV sequence features output by the encoder to optimize the predicted sequence.

[0067] ④ Decoder output: GVCI index prediction sequence for the next 7 days (168 hours) pred (t) (reflects future changes in the cooperative strength of TLF-GV signals).

[0068] S2.2 Model Training and Optimization

[0069] S2.2.1 Dataset Construction:

[0070] ① Training set: GVCI sequences of 100+ historical earthquakes (GVCI data of typical earthquakes from 2010 to 2023 were calculated by backtracking using the TLF-GV signal processing method), totaling 10,000+ samples, with each sample associated with a corresponding TLF-GV signal anomaly label.

[0071] ② Validation set / Test set: 20% and 10% of the historical samples are selected respectively. The validation set is used for hyperparameter tuning, and the test set is used for model generalization verification.

[0072] S2.2.2 Loss Function: A composite loss function of "mean squared error (MSE) + TLF-GV anomaly peak loss (APL)" is adopted to highlight the prediction accuracy of the TLF-GV anomaly peak and the sudden drop segment before the onset of the earthquake in the GVCI sequence.

[0073] ;

[0074] in, T peak T represents the time window of the TLF-GV anomalous peak in the GVCI sequence. drop The time window for the sudden drop in GVCI before the earthquake (corresponding to the abnormal weakening of the TLF-GV signal) is α=0.8 (weighting coefficient, determined through validation set optimization).

[0075] S2.2.3 Training parameters: The optimizer uses AdamW (learning rate = 1e-4, weight decay = 1e-5), the number of iterations = 100 rounds, the batch size = 32, and an early stopping strategy is adopted (training stops if the validation set loss does not decrease for 10 consecutive rounds).

[0076] S2.3 Preliminary judgment of earthquake occurrence time:

[0077] GVCI based on decoder output pred (t), based on the evolution law of TLF-GV signal, a "sudden drop judgment threshold" is set: if GVCI pred (t+1)-GVCI pred (t)<-0.2×GVCI pred,peak (GVCI pred,peak If the TLF-GV abnormal peak value is predicted in the sequence and the condition is met for three consecutive time steps, then t is determined to be the "potential time start point of the earthquake", and the corresponding time window is [t, t+24] hours (initially judged as a high-risk window, reflecting the high probability period of earthquake after the TLF-GV signal abnormality).

[0078] S3. Construction of Weibull Prior Distribution: Based on the data of "TLF-GV signal anomaly peak to earthquake time" from 100+ historical earthquake cases (determined by backtracking the peak value of GVCI index), Weibull distribution parameters are fitted to establish a statistical prior for earthquake time, providing TLF-GV related statistical basis for subsequent calibration.

[0079] S3.1 Historical Data Extraction:

[0080] Historical earthquake case selection criteria: M≥5.0, GVCI peak value≥1.5, TLF-GV anomaly duration≥12 hours. Earthquake cases with blurred TLF-GV signals (e.g., due to strong meteorological interference) were excluded to ensure a valid sample size≥100, avoiding invalid samples from affecting parameter fitting accuracy. From the GVCI sequences of 100+ historical earthquakes (calculated retrospectively using TLF-GV signal processing methods), the "GVCI anomaly peak occurrence time t" was extracted. peak (corresponding to the strongest anomaly moment of the TLF-GV signal) and "actual time of tremor t" earthquake Calculate the time difference Δt = t earthquake -t peak (Unit: hours, reflecting the interval from the TLF-GV anomaly to the occurrence of an earthquake).

[0081] S3.2, Weibull distribution fitting:

[0082] The Weibull distribution is suitable for describing the failure time distribution from "TLF-GV anomaly evolution to oscillation", and its probability density function is:

[0083] ;

[0084] Where λ is the median of the “TLF-GV anomaly characteristic lifetime” Δt, which is the time when there is a 50% probability of an earthquake, and k is the “shape parameter” (when k>1, Δt is positively skewed, which conforms to the rule that “the probability of an earthquake first increases and then decreases after the peak of the TLF-GV anomaly”).

[0085] S3.3 Parameter estimation: Maximum likelihood estimation (MLE) is used to fit λ and k. Based on 100+ historical Δt data (correlated with TLF-GV anomalies), λ=48 hours and k=1.8 are calculated (which can be dynamically updated according to the TLF-GV anomaly data of new earthquake cases).

[0086] S3.4, Prior probability distribution: The cumulative distribution function (CDF) of Δt generated by the fitted parameters, i.e., the TLF-GV associated prior probability of the earthquake occurrence time:

[0087] ;

[0088] For example, when T=24 hours, P prior =1-e-(24 / 48) ^1.8 ≈32% (reflecting the statistical probability of oscillation occurring 24 hours after a TLF-GV anomaly); at T=48 hours, P prior ≈63%.

[0089] S4. Weibull-Bayes Dynamic Calibration: Using the GVCI prediction sequence output by Informer (reflecting the future evolution trend of TLF-GV signals) as observational evidence, the Weibull distribution parameters are updated through Bayes' theorem to generate the daily conditional probability of earthquake occurrence for the next 7 days.

[0090] S4.1 Observational Evidence Extraction (Correlation of TLF-GV Signal Features):

[0091] GVCI output from Informer pred In (t), three key observation features are extracted (reflecting the similarity between the current GVCI sequence and historical TLF-GV anomaly features):

[0092] ①Feature 1: R peak =Predicted GVCI peak value / Historical TLF-GV anomaly average peak value (GVCI peak value average obtained from TLF-GV signal processing of 100+ earthquake cases).

[0093] ②Feature 2: D anomaly =Predict the duration of TLF-GV anomalies (the time from GVCI>1.5 to the peak, reflecting the sustained intensity of TLF-GV anomalies).

[0094] ③ Feature 3: S drop =GVCI drop slope (GVCI) pred,peak -GVCI pred,peak +6) / 6 (Amplitude decrease rate over 6 hours, reflecting the rate of TLF-GV anomalous attenuation).

[0095] S4.2, Likelihood Function Construction (Associating TLF-GV Anomaly Confidence)

[0096] Define the likelihood probability P(E|λ,k) (where E is the observed feature): If the observed feature E and the historical TLF-GV anomaly feature (such as R) are similar, then the likelihood probability P(E|λ,k) is defined as follows: peak >1.2、D anomaly >24 hours, S drop The higher the matching degree (>0.1), the greater the likelihood probability, indicating that the current GVCI sequence is more likely to lead to oscillation. The Weibull parameter needs to be adjusted to "decreasing λ and increasing k" (the oscillation time after the TLF-GV anomaly is earlier and the probability is concentrated).

[0097] S4.3, Bayesian parameter update

[0098] According to Bayes' theorem, the posterior parameter distribution π(λ,k|E) is proportional to the product of the prior distribution and the likelihood function (integrating TLF-GV historical statistics and real-time evolution):

[0099] ;

[0100] ① Prior distribution π(λ,k): The Gamma distribution (λ~Gamma(a1, b1), k~Gamma(a2, b2)) is adopted. Based on the historical TLF-GV anomaly fitting (λ=48, k=1.8), the hyperparameters a1=48, b1=1, a2=18, b2=10 are set.

[0101] ② Posterior sampling: Using the Markov chain Monte Carlo (MCMC) method (such as the NUTS algorithm), π(λ,k|E) is sampled 1000 times, and the mean is taken as the updated parameter λ. post k post The MCMC sampling convergence judgment adopts the latent scaling factor (R-hat) criterion: when R-hat < 1.01 for 500 consecutive samplings, it is judged as convergence, the total number of samplings is ≥ 1000, to ensure the statistical reliability of posterior parameters λpost and kpost, and the parameter estimation error is ≤ 5%.

[0102] Example: If the observed feature E is "R peak =1.5、D anomaly =36 hours, S drop =0.15" (highly matches the historical TLF-GV strong anomaly), then the updated λ post =40 hours, k post =2.0 (Temperature of shock after TLF-GV anomaly is advanced); if E is "R peak =0.8、D anomaly =12 hours, S drop =0.05” (matches a weak anomaly with historical TLF-GV), then λ post =60 hours, k post =1.5 (time delay after TLF-GV anomaly).

[0103] S4.4 Calculation of posterior probability of earthquake occurrence:

[0104] From the updated λ post k post Generate the posterior cumulative distribution function, i.e., the conditional probability of the final earthquake occurrence time (correlated with TLF-GV real-time evolution):

[0105] ;

[0106] in, T future For future times (e.g., t=24 hours, t=48 hours).

[0107] Example: λ post =40、k post When P = 2.0, post (t=24)=1-e-(24 / 40)^2≈46%) (Probability of an earthquake occurring 24 hours after a TLF-GV anomaly), P post (t=48)=1-e-(48 / 40)^2≈77%) (Probability of an earthquake occurring 48 hours after a TLF-GV anomaly).

[0108] S5. Probability Result Output: Convert the conditional probability of earthquake occurrence time into a structured report and transmit it to the subsequent magnitude and location prediction stages to support the joint prediction of the three elements of earthquake.

[0109] Report content:

[0110] ① Input basic information: the time range of the GVCI sequence output by the TLF-GV signal processing stage, the TLF-GV anomaly label (quiet period / abnormal period), and the GVCI peak time predicted by the Informer model (corresponding to the TLF-GV predicted peak).

[0111] ② Dynamic probability results: The daily conditional probability of earthquake occurrence within the next 7 days (e.g., "P=46% in the next 24 hours, P=77% in the next 48 hours, P=89% in the next 72 hours, P=92% in the next 96 hours"), with probability change curves (marking the TLF-GV anomaly evolution stages).

[0112] ③ Parameter description: The updated Weibull parameter λ post k post Observational features R peak D anomaly S drop The specific value (corresponding to the TLF-GV anomaly intensity).

[0113] ④ Risk classification recommendations: Divide risk levels based on probability values ​​(P≥80% is "high risk, it is recommended to activate the early warning"; 50%≤P<80% is "medium risk, strengthen TLF-GV signal monitoring"; P<50% is "low risk, continue to observe TLF-GV signal").

[0114] ⑤ Output format: JSON (for subsequent magnitude and location prediction interface integration) and PDF (for manual review).

[0115] Technical effects achieved by this invention (compared to existing technologies: a method for predicting major earthquakes based on atmospheric tidal gravity precursors):

[0116] 1. Significantly improved TLF-GV signal adaptability and data quality, solidifying the foundation for prediction input: Existing solutions suffer from instrument frequency band limitations (insufficient coverage of the TLF-GV core frequency domain 1-30mHz) and weak anti-interference capabilities, resulting in invalid input signals and a false positive rate exceeding 40% in the sample library. This invention relies on a TLF-GV signal processing link adapted to the full 1-30mHz frequency band, combined with CEEMDAN adaptive decomposition + Kalman filtering to purify GVCI sequences. At the same time, through the ProbSparse attention mechanism of the Informer model, the long-term dependency capture length of GVCI sequences is extended from 10 hours to 30 days, fully learning the evolution pattern of "quiet period - abnormal period - pre-earthquake", completely solving the problem of "invalid input foundation" and providing high-quality correlation data for prediction.

[0117] 2. The model closely follows crustal patterns, significantly reducing prediction errors: Existing schemes presuppose a mechanical constraint that "an earthquake will inevitably occur 20 days after an anomaly," which violates the nonlinear evolution of crustal stress, resulting in high false alarm and false negative rates. The Weibull-Bayes module of this invention has no time limit. It first constructs a statistical prior of "TLF-GV peak to earthquake occurrence time" based on 100+ historical earthquake cases, and then dynamically updates the distribution parameters using the real-time GVCI sequence output by Informer. This makes the prediction both statistically reliable and dynamically adaptable. In practice, it can reduce the false alarm rate by more than 30% and the false negative rate by more than 40%, perfectly matching the uncertainty characteristics of geological processes.

[0118] 3. Supporting practical short-term decision-making and achieving seamless collaboration with preceding technologies: Existing solutions only output a single curve of "probability of earthquake occurrence within N days," which cannot meet the needs of hierarchical decision-making and requires additional preprocessing for connection with other stages. This invention, for short-term forecasting scenarios, outputs the conditional probability of earthquake occurrence for each day within the next 7 days (e.g., "24-hour P=55%, 48-hour P=78%)," directly supporting hierarchical decision-making for "monitoring-early warning-emergency response." At the same time, it strictly adapts to the GVCI format (timestamp, numerical sequence, etc.) output by the preceding TLF-GV technology, accepting data without additional processing, improving the efficiency of the "collection-processing-prediction" link by 50%, solving the problems of "unintuitive output and cumbersome connection" in the preceding technology, and meeting the needs of actual business applications.

[0119] 4. Strong engineering practicality: The model is implemented based on the PyTorch framework. The trained Informer model takes ≤10 minutes to predict 7-day GVCI sequences (correlated with TLF-GV) and ≤2 minutes to calibrate with Weibull-Bayes, meeting the real-time requirements of short-term earthquake prediction (≤1 hour).

[0120] Based on the same inventive concept, this invention also proposes a TLF-GV signal-correlated earthquake occurrence time prediction system, comprising:

[0121] The acquisition module is used to acquire the gravity-vibration contribution index (GVCI) time series of the low-frequency gravity-vibration TLF-GV signal from the potential source region.

[0122] The sequence prediction module is used to input the GVCI time series into the pre-trained Informer time series prediction model, predict the GVCI sequence in the future, and preliminarily determine the potential earthquake time window based on the predicted sequence.

[0123] The posterior Weibull distribution construction module is used to construct a Weibull statistical prior distribution representing the interval from the peak of the TLF-GV signal anomaly to the earthquake occurrence time based on historical earthquake events and their associated TLF-GV signal anomaly data. Using the future GVCI sequence predicted by the Informer time series prediction model as observational evidence, the parameters of the Weibull statistical prior distribution are dynamically updated through Bayes' theorem to obtain the posterior Weibull distribution.

[0124] The prediction module is used to calculate the conditional probability of earthquake occurrence within different time windows in the future based on the posterior Weibull distribution; and to predict the earthquake occurrence time based on the conditional probability.

[0125] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting earthquake occurrence time using TLF-GV signal correlation, characterized in that, Includes the following steps: Obtain the gravity-vibration contribution index (GVCI) time series from the potential source region to the low-frequency gravity-vibration TLF-GV signal; The GVCI time series is input into the pre-trained Informer time series prediction model to predict the GVCI sequence in the future, and the potential earthquake time window is initially determined based on the predicted GVCI sequence. Based on historical earthquake events and their associated TLF-GV signal anomaly data, a Weibull statistical prior distribution characterizing the time interval from the peak of the TLF-GV signal anomaly to the earthquake occurrence is constructed. Using the future GVCI sequence predicted by the Informer time series prediction model as observational evidence, the parameters of the Weibull statistical prior distribution are dynamically updated through Bayes' theorem 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; the earthquake occurrence time is predicted based on the conditional probability.

2. The method for predicting earthquake occurrence time by correlating TLF-GV signals according to claim 1, characterized in that, After obtaining the Gravity-Vibration Contribution Index (GVCI) time series from the potential seismic source region to the low-frequency gravity-vibration TLF-GV signal, the GVCI time series is preprocessed, specifically including the following steps: The 3σ criterion was used to remove outliers in the GVCI time series whose GVCI values ​​exceeded [μ-3σ, μ+3σ]; where μ and σ are the mean and standard deviation of GVCI during the TLF-GV quiet period. Missing values ​​in the GVCI time series after outlier removal are filled and repaired using linear interpolation or Kalman interpolation. The repaired GVCI time series was subjected to min-max normalization to map the GVCI values ​​to the [0,1] interval; The normalized long sequence is sliced ​​according to a fixed time window and step size to obtain the preprocessed GVCI time series.

3. The method for predicting earthquake occurrence time by correlating TLF-GV signals according to claim 1, characterized in that, The process of inputting the GVCI time series into a pre-trained Informer time series prediction model to predict the GVCI sequence for a future period specifically includes the following steps: Extract the GVCI time series according to the preset historical time window length; The truncated GVCI time series is input into the encoder of the pre-trained Informer time series prediction model. The ProbSparse self-attention mechanism is used to process the input truncated GVCI time series to extract the long-term evolution pattern features of the TLF-GV signal, which reflect the "gradual rise during the calm period - peak during the abnormal period - sudden drop before the oscillation". The sequence feature vector that can reflect the co-evolution trend of the TLF-GV signal is generated. Using a sequence feature vector reflecting the co-evolution trend of TLF-GV signals and an initialized target sequence as input to the decoder, GVCI sequences for a future period are gradually generated in parallel through a masked multi-head attention mechanism and a cross-attention mechanism with the encoder output. pred (t).

4. The method for predicting earthquake occurrence time using TLF-GV signal correlation according to claim 1, characterized in that, The preliminary determination of potential earthquake occurrence time windows based on the predicted GVCI sequence specifically includes the following steps: GVCI based on decoder output pred (t), based on the evolution law of TLF-GV signal, a sudden drop judgment threshold is set, if GVCI pred (t+1)-GVCI pred (t)<-0.2×GVCI pred,peak GVCI pred,peak To predict the TLF-GV anomaly peak in the sequence, and if this condition is met for three consecutive time steps, then t is determined to be the potential earthquake start time, and the corresponding time window is [t, t+24] hours.

5. The method for predicting earthquake occurrence time using TLF-GV signal correlation according to claim 1, characterized in that, The historical earthquake events and their associated TLF-GV signal anomaly data meet the following screening criteria: magnitude M≥5.0, GVCI peak value≥1.5, and TLF-GV signal anomaly duration≥12 hours.

6. The method for predicting earthquake occurrence time by correlating TLF-GV signals according to claim 3, characterized in that, The probability density function of the Weibull statistical prior distribution is: ； Where, Δ t The interval from the abnormal peak value of the TLF-GV signal to the time of oscillation. λ The abnormal characteristic lifetime Δ of TLF-GV t the median of k For shape parameters.

7. The method for predicting earthquake occurrence time by correlating TLF-GV signals according to claim 6, characterized in that, The method of using the future GVCI sequence predicted by the Informer time series prediction model as observational evidence, and dynamically updating the parameters of the Weibull statistical prior distribution through Bayes' theorem to obtain the posterior Weibull distribution, specifically includes the following steps: GVCI predicted from the Informer model pred (t) extracts observational evidence feature E; the observational evidence feature includes at least: the ratio R of the predicted GVCI peak to the historical average peak. peak The predicted duration of the TLF-GV anomaly, D anomaly The steep slope S of the GVCI sequence drop ; Construct the likelihood function P(E| λ , k ), used to assess the probability of the observed evidence feature E occurring given the Weibull distribution parameters; Based on Bayes' theorem, the Weibull prior distribution is combined with the likelihood function, and the parameters are calculated using the Markov chain Monte Carlo (MCMC) sampling method. λ and k The posterior distribution π( λ , k |E), and use the mean of the posterior distribution as the updated parameter. λ post , k post .

8. The method for predicting earthquake occurrence time by correlating TLF-GV signals according to claim 7, characterized in that, The process involves calculating the conditional probability of earthquake occurrence within different time windows based on the posterior Weibull distribution, specifically through the updated... λ post , k post Obtain the conditional probability of the final earthquake occurrence time. The process is represented as follows: ； in, T future For the future.

9. A TLF-GV signal-correlated earthquake occurrence time prediction system, characterized in that, include: The acquisition module is used to acquire the gravity-vibration contribution index (GVCI) time series from the potential source region to the low-frequency gravity-vibration TLF-GV signal; The sequence prediction module is used to input the GVCI time series into the pre-trained Informer time series prediction model, predict the GVCI sequence in the future, and preliminarily determine the potential earthquake time window based on the predicted GVCI sequence. The posterior Weibull distribution construction module is used to construct a Weibull statistical prior distribution representing the interval from the peak of the TLF-GV signal anomaly to the earthquake occurrence time based on historical earthquake events and their associated TLF-GV signal anomaly data. Using the future GVCI sequence predicted by the Informer time series prediction model as observational evidence, the parameters of the Weibull statistical prior distribution are dynamically updated through Bayes' theorem to obtain the posterior Weibull distribution. The prediction module is used to calculate the conditional probability of earthquake occurrence within different time windows in the future based on the posterior Weibull distribution; and to predict the earthquake occurrence time based on the conditional probability.

Images