A joint probabilistic prediction method and system for three earthquake elements associated with TLF-GV signals

Abstract

This invention provides a method and system for joint probabilistic prediction of three earthquake elements associated with TLF-GV signals, belonging to the field of earthquake prediction technology. It aims to solve the problems of traditional three-element prediction lacking joint fusion, difficulty in handling evidence conflicts, and fixed weights. The method includes: extracting the physical characteristics of the TLF-GV signal, taking the prediction results of the previous three elements and model confidence parameters, concatenating them into a feature vector, inputting it into a dynamic weighted prediction model constructed by a shallow MLP, and outputting time, magnitude, and location weights; dividing the three elements into 3×3×3=27 "time-magnitude-location" propositions, generating a basic probability assignment (BPA) for each proposition based on the weights, and quantifying evidence contradictions through a conflict coefficient K; obtaining the joint BPA using the Dempster combination rule, calculating the mean of the confidence function and likelihood function of each proposition as the comprehensive probability, and extracting the proposition with the highest probability as the coupling result. This method significantly improves the scientific rigor and practicality of short-term strong earthquake prediction.

Description

Technical Field

[0001] This invention relates to the field of earthquake early warning technology, specifically to a method and system for joint probabilistic prediction of three earthquake elements associated with TLF-GV signals. Background Technology

[0002] Accurate prediction of the three elements of an earthquake (time, location, and magnitude) remains a major scientific challenge in the field of Earth sciences. Earthquakes originate in the extremely complex dynamic environment deep within the Earth, resulting from the long-term superposition and repeated action of multiple force sources within the Earth's crust. However, current observation technologies are unable to directly probe the state of rock masses at depths of 10,000 to 20,000 kilometers below the surface, making it impossible to accurately determine whether the underground rock mass is in a critical state of fracturing. Currently, we mainly rely on surface observation stations to obtain indirect signals through seismometry, topographic deformation, underground fluid analysis, and geomagnetism for inference. However, these signals are easily affected by environmental interference, and the recurrence cycle of strong earthquakes often far exceeds the time span of modern observation records, resulting in inherent defects in the usable research data.

[0003] Current earthquake probabilistic prediction methods are mainly divided into two major systems: physical models and data-driven approaches, but both have significant technical limitations. Physical models, based on fracture mechanics and crustal dynamics theories, construct numerical simulation frameworks. While they can describe fault kinematic processes, they struggle to fully characterize the nonlinear properties of geological bodies and complex fault interactions. They often simplify the coordinated activities of multi-fault systems, leading to insufficient accuracy in simulating the spatiotemporal distribution of earthquakes in actual predictions. Statistical models extrapolate by mining statistical patterns from historical earthquake data. Traditional Poisson models assume independent earthquake events and cannot explain the clustering characteristics of earthquakes. Even improved infectious models like ETAS, which can capture aftershock sequence patterns, still rely on ample historical data. Prediction reliability drops significantly in areas with short earthquake records or sparse activity, and they are ill-suited to handling low-probability extreme events.

[0004] With the development of artificial intelligence technology, machine learning methods have begun to be applied to earthquake prediction. However, while improving the ability to process multi-source data, they also bring new problems. Although deep learning models can integrate multi-dimensional data such as seismic waves and GPS deformation, the "black box effect" makes the prediction logic difficult to interpret, and the model performance is highly dependent on the quality of training data. Existing monitoring data often suffers from "contamination" due to factors such as station relocation and equipment replacement, and the lack of standardized quality control processes directly affects the robustness of the models. At the same time, the lack of understanding of the physical mechanisms of the models contrasts sharply with the experience-based reliance of data-driven models. The former cannot effectively utilize real-time observation data to optimize predictions, while the latter struggles to incorporate prior knowledge such as geological structures, resulting in a long-standing disconnect between the two approaches.

[0005] The imperfections in multi-source information fusion mechanisms further restrict the improvement of prediction accuracy. Existing methods generally suffer from unreasonable weight allocation when integrating data from different observation methods. More importantly, existing technologies often focus on single-factor predictions, neglecting the intrinsic correlation between time, location, and magnitude. Furthermore, the long-term, medium-term, and short-term forecasts are loosely connected and lack dynamic update mechanisms, making it difficult to provide timely and reliable probability assessments during critical stages of seismic activity and failing to meet the actual needs of disaster prevention and mitigation. Summary of the Invention

[0006] To address the aforementioned issues, this invention provides a joint probabilistic prediction method for three earthquake elements associated with TLF-GV signals. This method systematically solves the problems existing in the prior art by unifying the probability dimension, dynamically adapting weights, actively resolving conflicts, and providing interpretable output, thus significantly improving the scientific rigor and practicality of short-term strong earthquake prediction.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A joint probabilistic prediction method for three earthquake elements associated with TLF-GV signals includes the following steps:

[0009] Physical features 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 ;

[0010] A three-element prediction model is constructed, including a time probability prediction model, a magnitude probability prediction model, and a location probability prediction model. The reliability of each prediction model is determined, and the model is concatenated with physical features to form a feature vector for weight allocation. The feature vector is then input into a dynamic weight prediction model based on a fuzzy gated network to obtain the allocation results of time weight, magnitude weight, and location weight.

[0011] The time prediction results in the three-element prediction model are divided into multiple core intervals, the magnitude prediction results are divided into magnitude intervals, and the high-probability areas in the location prediction results are merged into multiple core intervals, generating multiple propositions of time-magnitude-location.

[0012] Based on the weighting results, a weighted Basic Probability Assignment (BPA) is generated for each proposition, and the conflict coefficient of each proposition is determined. The degree of conflict is judged based on the conflict coefficient, and BPAs exceeding a preset threshold are weighted and discounted. Then, the Dempster combination rule is used to obtain the joint BPA, and the confidence function and likelihood function of each proposition are extracted.

[0013] The mean of the confidence function and likelihood function for each proposition is taken as the comprehensive probability for that proposition, and the proposition with the highest comprehensive probability is taken as the output item.

[0014] Preferably, the construction of the three-element prediction model includes a time probability prediction model, a magnitude probability prediction model, and a location probability prediction model, and the determination of the reliability of each prediction model includes the following steps:

[0015] The gravity-vibration contribution index (GVCI) time series of the TLF-GV signal was extracted, and the 7-day time probability was predicted using the Informer time series prediction model. Specifically, based on historical earthquake events and their associated TLF-GV signal anomaly data, a Weibull statistical prior distribution characterizing the time interval from the TLF-GV anomaly peak to the earthquake occurrence was constructed. The future GVCI sequence predicted by the Informer time series prediction model was used as observational evidence. The parameters of the Weibull statistical prior distribution were dynamically updated using Bayes' theorem to obtain the posterior Weibull distribution. A likelihood function P(E|λ,k) was constructed to evaluate the probability of the observational evidence E in the Weibull distribution. According to Bayes' theorem, the Weibull prior distribution and the likelihood function were combined, and the posterior distribution π(λ,k|E) of the parameter λ was calculated using the Markov chain Monte Carlo (MCMC) sampling method. The mean of the posterior distribution was then used as the updated parameter λ. post According to parameter λ post Determine the reliability of the time model:

[0016] ;

[0017] Multi-source feature vectors and physical feature vectors of the TLF-GV signal are extracted, and the predicted magnitude is obtained through a magnitude probability prediction model. The reliability of the magnitude model is determined based on the physical constraint deviation of the predicted magnitude.

[0018] ;

[0019] Gravity gradient grids from TLF-GV signals are extracted, and location probability heatmaps are obtained through location probability prediction models. The reliability of the location model is determined based on the high-probability regions in the probability heatmap and their distance from the fault. .

[0020] Preferably, the generation of the multiple propositions of time-magnitude-location includes the following steps:

[0021] The 7-day time probability is divided into 3 core intervals as time intervals, including T1 (24~48h), T2 (48~72h) and T3 (72~96h), and the cumulative probability of each time interval is calculated;

[0022] Based on the range of uncertainty, three magnitude intervals are fixedly divided, including M1 (( -0.2~ M2 ~ +0.2)) and M3 ( +0.2~ +0.4), to ensure matching with the time and location interval dimensions;

[0023] High-probability areas with a probability ≥ 0.6 are extracted from the heat map and merged into 3 core regions as location intervals, including the main high-probability area L1, the second high-probability area L2, and the potential area L3. The average probability of each location region is calculated.

[0024] Construct a time-magnitude-location proposition and an empty set based on the time interval, magnitude interval, and location interval.

[0025] Preferably, the step of generating the weighted basic probability assignment (BPA) for each proposition based on the weight allocation result includes the following steps:

[0026] Based on the cumulative probability P(T) of each proposition over the time interval i ), Magnitude Probability P(M) j ) and the average probability P(L) of the location region k After weighting and multiplying, extend to the entire domain θ to transform into the BPA of each proposition:

[0027] m(θ ijk )=m t (T i )×m m (M j )×m l (L k ), m(θ)=1-Σm(θ) ijk );

[0028] Where, m t (T i )=w t ×P(T i () represents time BPA, m m (M j )=w m ×P(M j ) represents the magnitude BPA, m l (L k )=w l ×P(L k ) represents location BPA; w t w m and w l The weights are the weights of the three factors.

[0029] Preferably, the step of determining the degree of conflict based on the conflict coefficient and weighting the BPA exceeding a preset threshold by discounting specifically includes:

[0030] The degree of conflict K in obtaining the three elements of evidence:

[0031] ;

[0032] If K < 0.3, Dempster's combination rule is applied directly; if K ≥ 0.3, an improved rule is applied, with weighted discounts applied to conflicting evidence.

[0033] ;

[0034] Where α is the conflict allocation coefficient.

[0035] Preferably, the step of obtaining the joint BPA using Dempster's combination rule and extracting the confidence function and likelihood function for each proposition includes the following steps:

[0036] Perform Dempster combinatorial operations on the BPAs of each proposition to obtain the final joint BPA m. comb (θ ijk And calculate the reliability function Bel and the likelihood function Pl for each proposition:

[0037] ;

[0038] ;

[0039] The mean of the confidence function and likelihood function for each proposition is taken as the overall probability for that proposition:

[0040] .

[0041] Preferably, the dynamic weight prediction model is constructed from a shallow MLP, the training set consists of historical TLF-GV physical features and model credibility feature vectors, as well as the corresponding manually labeled optimal weights, and the mean squared error is used as the loss function.

[0042] This invention also provides a joint probabilistic prediction system for three earthquake elements associated with TLF-GV signals, the system comprising:

[0043] processor;

[0044] A memory on which computer programs that can run on the processor are stored;

[0045] The computer program, when executed by the processor, implements the steps of the TLF-GV signal-correlated seismic three-element joint probability prediction method.

[0046] 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 seismic three-element joint probability prediction method.

[0047] The beneficial effects of this invention are:

[0048] This invention proposes a joint probability prediction method for three earthquake elements associated with TLF-GV signals. This method transforms the probability distribution of earthquake occurrence time, the probability interval of magnitude, and the probability heatmap of location into a unified comprehensive probability confidence level. Based on a fuzzy gated network combined with the physical characteristics of TLF-GV signals and model confidence, the method dynamically allocates the weights of the three elements to avoid the limitations of fixed weights. It outputs a quantitative result of "time-magnitude-location" coupling and identifies conflicts in the prediction of the three elements through DS evidence theory. It uses a conflict coefficient and improved combination rules to resolve contradictions, ensuring the consistency of the comprehensive probability and effectively improving the accuracy of the comprehensive probability prediction. Attached Figure Description

[0049] Figure 1 This is a flowchart of a method according to an embodiment of the present invention. Detailed Implementation

[0050] 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.

[0051] Example 1

[0052] In the prior art, a satellite thermal infrared brightness temperature anomaly technology for short-term prediction of strong earthquakes (patent publication number CN101793972A) is a representative solution for short-term prediction of strong earthquakes (M≥6.0). This scheme uses satellite remote sensing technology to capture pre-earthquake thermal infrared brightness temperature anomalies and constructs a statistical prediction framework of "thermal anomaly spatiotemporal evolution - magnitude and energy correlation": First, it uses multi-source remote sensing data such as polar-orbiting satellites and geostationary satellites to extract brightness temperature anomaly areas within 10-20 days before the earthquake (e.g., warming area of ​​100,000-600,000 square kilometers); then, it fuses multi-source signals by using fixed empirical weights (thermal infrared 0.6, infrasound 0.2, topographic stress thermal line 0.2) to establish a statistical relationship of "warming area - magnitude" (e.g., M6.0 earthquake corresponds to a warming area of ​​400,000 square kilometers); finally, it outputs discrete three-element prediction results, such as "the probability of a strong earthquake of M6.5-7.0 occurring in the region of 34°±0.5′N, 108°±0.5′E on the eastern edge of the Qinghai-Tibet Plateau in the next 15 days is 25%".

[0053] While this scheme achieves joint prediction of the three elements of strong earthquakes, its technical design differs fundamentally from that of this invention, and it has significant limitations in prediction accuracy, dynamic adaptability, and interpretability. From the perspective of coupling in the prediction dimensions, the output of the three elements in this scheme is a discrete combination rather than a physically correlated joint probability: time prediction only provides a fixed period of 10-20 days, unable to be refined to a short window of 48 hours; location prediction relies on a 1°×1° grid division, without dynamically adjusting spatial accuracy based on fault activity; magnitude prediction is achieved through statistical extrapolation of warming area, without establishing a physical model of "energy release-fault rupture-magnitude" (such as a quantitative relationship between a specific fault length and magnitude). The root of this design lies in the fact that its core logic is based on statistical correlation of historical earthquakes, rather than modeling the physical mechanisms of strong earthquake gestation processes, resulting in prediction results that cannot output quantitative reliability of "time-magnitude-location" coupling as in this invention (e.g., "the comprehensive probability of an M6.5±0.3 earthquake occurring in the 34°15′40″N region in the next 48 hours is 72%").

[0054] Regarding dynamic weight adaptation, this scheme employs a completely fixed empirical weight allocation strategy, failing to consider the time-varying characteristics of precursor signals and the dynamic changes in model reliability. For example, when solar activity causes a sharp drop in the signal-to-noise ratio of satellite thermal infrared data (e.g., the accuracy of brightness temperature anomaly identification drops from 80% to 50% during geomagnetic storms), or when infrasound signals exhibit false high values ​​due to meteorological interference, the fixed weights still forcibly maintain a 0.6% proportion for thermal infrared, incorporating invalid interference signals into the fusion and amplifying prediction bias. The essence of this problem is that this scheme fails to construct a correlation mechanism between "signal physical characteristics - model reliability parameters - dynamic weights," resulting in the weights being unable to adjust in real time according to signal quality, making it difficult to adapt to the complex dynamic changes of strong earthquake precursors. This contrasts sharply with the dynamic weight allocation achieved by this invention through a fuzzy gating network.

[0055] The lack of conflict resolution capability is another key limitation of this scheme. When contradictory multi-source precursors appear (e.g., significant thermal infrared warming anomalies but no response from topographic stress thermal lines), this scheme merely masks the conflict by using a fixed-weighted average, failing to quantify the degree of conflict (e.g., conflict coefficient) or analyze the source of the conflict based on physical meaning (e.g., whether it is a false signal caused by atmospheric disturbance). For example, in a certain regional observation, thermal infrared showed a warming area of ​​700,000 square kilometers (corresponding to a magnitude of M7.0 or higher), but infrasound showed no abnormal response. This scheme still outputs a comprehensive anomaly index with fixed weights, failing to determine whether the "risk signal originates from a single unreliable precursor." In contrast, this invention quantifies the three elements to predict contradictions by defining a conflict coefficient and combines it with an improved DS combination rule for targeted resolution, increasing the conflict resolution rate to over 90%, significantly outperforming the passive fusion mode of traditional statistical frameworks.

[0056] Furthermore, this scheme lacks interpretability in its prediction results, only outputting the final probability values ​​without providing the basis for weight allocation or analysis of conflict sources. In strong earthquake early warning decision-making, decision-makers cannot trace key information such as "why the thermal infrared weight is higher than the infrasound weight" or "whether the probability results are affected by data conflicts." This contrasts sharply with the interpretability of this invention, which includes explanations for weighting (e.g., "the location model weight of 0.4 is due to high fault constraint satisfaction") and conflict analysis (e.g., "the 15% conflict degree stems from a slight discrepancy between time and magnitude"), significantly limiting its ability to support scientific decision-making. Therefore, this invention proposes a joint probabilistic prediction method for three earthquake elements associated with TLF-GV signals, such as... Figure 1 As shown, the main steps include:

[0057] S1: Extract physical features 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 .

[0058] S2: Construct a three-element prediction model, including a time probability prediction model, a magnitude probability prediction model, and a location probability prediction model. Determine the credibility of each prediction model and concatenate it with physical features to form a feature vector for weight allocation. Input the feature vector into a dynamic weight prediction model to obtain the time weight, magnitude weight, and location weight allocation results.

[0059] S3: Divide the time prediction results in the three-element prediction model into multiple core intervals, the magnitude prediction results into magnitude intervals, and merge the high-probability areas in the location prediction results into multiple core intervals to generate multiple propositions of time-magnitude-location.

[0060] S4: Generate the weighted basic probability assignment (BPA) for each proposition based on the weight allocation results, and determine the conflict coefficient for each proposition; determine the degree of conflict based on the conflict coefficient, and weight discount the BPA that exceeds the preset threshold, then use the Dempster combination rule to obtain the joint BPA, and extract the confidence function and likelihood function for each proposition.

[0061] S5: Take the mean of the confidence function and likelihood function for each proposition as the overall probability for that proposition, and take the proposition with the highest overall probability as the output item.

[0062] Specifically:

[0063] 1. Step 1: Input data preprocessing.

[0064] 1.1 Data Reception: Four types of core data are received, all of which are related to the TLF-GV signal and preceding prediction results:

[0065] The TLF-GV dataset of low-frequency gravity-vibration signals 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) in accordance with risk levels, combined with a dual-mode collaborative cold atom interferometer. The specific acquisition method is disclosed by the team of this invention in a patent application filed on the same day, a method for network monitoring of short-term earthquake precursor information.

[0066] ① Earthquake occurrence time probability data (from the time prediction stage): daily conditional probability distribution of earthquake occurrence over the next 7 days (e.g., P) t1 =46% (24h), P t2 =77% (48h), P t3 =89% (72h)), and Weibull parameter λ post k post (Reflecting the credibility of the time model). Specifically, a likelihood function P(E|λ,k) is constructed to evaluate the probability of observed evidence E in the Weibull distribution. Based on Bayes' theorem, the Weibull prior distribution is combined with the likelihood function. The posterior distribution π(λ,k|E) of the parameter λ 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 and k post .

[0067] Among them, the method of obtaining the earthquake occurrence probability based on GVCI time series prediction is disclosed in a patent filed on the same day: an earthquake occurrence time prediction method and system, specifically based on the GVCI time series of TLF-GV signals, 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 the potential earthquake occurrence time window; a Weibull statistical prior distribution representing the time interval from the TLF-GV anomaly peak to the earthquake occurrence is constructed based on historical earthquake events; using the future GVCI sequence as observational evidence, the parameters of the Weibull statistical prior distribution are dynamically updated through 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.

[0068] ② Magnitude probability data (from the magnitude prediction stage): Magnitude estimate =6.3 and uncertainty range 6.3±0.2, which are converted into probability intervals (e.g. M1:6.1~6.3 (probability 40%), M2:6.3~6.5 (probability 50%), M3:6.5~6.7 (probability 10%)) and physical constraint deviations (e.g. 0.1, reflecting the credibility of the magnitude model).

[0069] Among them, the multi-source feature vector and physical feature vector of the TLF-GV signal are extracted, and the predicted magnitude is obtained by using a magnitude probability prediction model. The above technical means are disclosed in the patent filed on the same day: a method, system and medium for predicting earthquake magnitude by TLF-GV signal association.

[0070] ③ Location probability data (from the location prediction stage): Heat map of the source location probability (5km×5km grid, such as Pl1=75% (N30.6° area), Pl2=50% (N30.7° area)), and the distance between the high probability area and the fault (such as 2.5km, reflecting the credibility of the location model).

[0071] Among them, the gravity gradient grid of the TLF-GV signal is extracted, and the location probability heatmap is obtained by predicting the location probability using a location probability prediction model based on the U-Net++ model. The above technical means are disclosed in the patent filed on the same day: a method, system and medium for predicting the location of earthquake source associated with TLF-GV signal.

[0072] ④ TLF-GV physical characteristics (from signal processing stage): GVCI peak value (e.g., 20), BF value (e.g., 0.8, vibration dominant), and abnormal duration (e.g., 35 hours), used for dynamic weight allocation.

[0073] 1.2 Data standardization and format unification:

[0074] ① Time probability discretization: Divide the 7-day time probability into 3 core intervals: T1 (24~48h), T2 (48~72h), and T3 (72~96h), and calculate the cumulative probability of each interval (e.g., P(T1)=77%, (P(T2)=89%)).

[0075] ② Magnitude probability intervalization: Based on the range of uncertainty, three magnitude intervals are fixedly divided: M1 (( -0.2~ M2 ~ +0.2), M3 ( +0.2~ +0.4), to ensure matching with the time and location interval dimensions;

[0076] ③ Location probability focusing: Extract high probability areas with a probability ≥ 0.6 from the heat map and merge them into 3 core regions: L1 (main high probability area, such as N30.6°), L2 (secondary high probability area, such as N30.7°), and L3 (potential area, such as N30.5°), and calculate the average probability of each region;

[0077] ④ Model credibility quantization: Convert the model credibility parameters into quantized values ​​of 0 to 1.

[0078] Reliability of the time model ( The closer it is to the historical average of 48, the higher the reliability.

[0079] Credibility of magnitude model (When the deviation is level 0.1, C) m =0.9);

[0080] Location model credibility (At a distance of 2.5km, C) l =0.75).

[0081] 1.3 Feature Vector Construction: Concatenate "TLF-GV physical features + model credibility" into a weighted input vector F=[GVCI]. peak , BF, T anomaly C t C m C l (e.g., [20, 0.8, 35, 0.85, 0.9, 0.75]), with 6 dimensions.

[0082] 2. Step 2: Dynamic weight allocation for fuzzy gating network

[0083] (1) Objective: Based on the physical characteristics and model credibility of TLF-GV, dynamically allocate the weights of the three-factor prediction model to avoid the limitations of fixed weights. The core is to construct a fuzzy gated network to learn the nonlinear relationship between "features and weights".

[0084] (2) Operational details:

[0085] 2.1 Fuzzy Gated Network Structure Design

[0086] The network is a shallow MLP (suitable for training with small samples to avoid overfitting), and its structure is as follows:

[0087] ① Input layer: 6-dimensional feature vector F (TLF-GV physical features + model credibility);

[0088] ② Hidden layers: 2 layers, the first layer has 12 neurons (activation function ReLU), the second layer has 6 neurons (activation function ReLU), and overfitting is suppressed by Dropout (probability 0.1);

[0089] ③ Output layer: 3 neurons (activation function Softmax), outputting the dynamic weights of the three elements (W=[w t , w m ,w l ](w t +w m +w l =1), which correspond to the model weights of time, magnitude, and location, respectively.

[0090] 2.2 Network Training and Optimization

[0091] 2.2.1 Construction of training dataset:

[0092] ① Sample source: Feature vectors F from 100+ historical earthquakes + manually labeled optimal weights W true The rules for manually labeled weights include: GVCI. peak When the magnitude is greater than 25 km, the magnitude weight is ≥0.4; when the distance between the high-probability zone and the fault is <2.5 km, the location weight is ≥0.4; and the probability of earthquake P... earthquake When the time weight is greater than 0.8, it should be ≥0.35 to ensure that the subjective error of annotation is ≤10%. true The model is labeled by domain experts based on "TLF-GV feature importance + model credibility" (e.g., "a higher GVCI peak value increases the magnitude weight, and a closer fault distance increases the location weight").

[0093] ② Loss function: The mean squared error (MSE) is used to minimize the deviation between the predicted weights and the true weights.

[0094] ;

[0095] 2.2.2 Training parameters: Optimizer Adam (learning rate 1e-4, weight decay 1e-5), number of iterations 50, batch size=8, early stopping strategy (stop if the validation set loss does not decrease for 8 consecutive rounds).

[0096] 2.2.3 Weighted Output Example: Input feature F=[20,0.8,35,0.85,0.9,0.75], network output W=[0.3,0.35,0.35] (time weight 0.3, magnitude weight 0.35, location weight 0.35), which conforms to the logic that "the higher the credibility of the magnitude model (0.9), the higher the weight".

[0097] 2.3 Verification of Weight Reasonableness

[0098] The output weights are considered reasonable if they meet the following conditions:

[0099] ① When the confidence level of a certain model is ≤0.5, its weight is ≤0.2 (to avoid low-confidence models dominating the fusion).

[0100] ②GVCI in TLF-GV characteristics peak When the magnitude is >25 (strong anomaly), the magnitude weight is ≥0.4 (strong anomalies correspond to more reliable magnitude predictions); if this condition is not met, the network should be retrained or the feature vectors adjusted.

[0101] 3. Step 3: Integration of DS Evidence Theories

[0102] (1) Purpose: to transform the weighted probability of the three elements into evidence, handle conflicts and integrate them into a comprehensive probability reliability. The core is to define the identification framework and improve the conflict combination rules.

[0103] (2) Operational details:

[0104] 3.1 Defining the Recognition Framework and Basic Probability Assignment (BPA)

[0105] 3.1.1 Identification Framework θ: Based on the three-element interval, a coupled identification framework is constructed, containing all combinations of "time-magnitude-location", totaling 3×3×3=27 propositions (such as θ). 111 =T1∩M1∩L1, θ112=T1∩M1∩L2), and “empty set Ø”;

[0106] 3.1.2 Basic Probability Assignment (BPA) Calculation: The weighted probabilities of the three factors are converted into the BPA (assignment of confidence in evidence) for each proposition:

[0107] ①Time BPA: m t (T i )=w t ×P(T i (e.g., m) t (T1) = 0.3 × 77% = 23.1%), m t (θ t )=w t ×(1-ΣP(T i ))(θ t (This refers to the entire time domain, reflecting the uncertainty of temporal evidence).

[0108] ② Magnitude BPA: m m (M j )=w m ×P(M j (e.g., m) m (M2) = 0.35 × 50% = 17.5%), (m m (θ m )=wm ×(1-ΣP(M j ));

[0109] ③Location BPA: m l (L k )=w l ×P(L k (e.g., m) l (L1) = 0.35 × 75% = 26.25%), m l (θ l )=w l ×(1-ΣP(L k ));

[0110] ④ Joint BPA: Extended to the entire θ domain through the product of independent evidence, such as m(θ) ijk )=m t (T i )×m m (M j )×m l (L k ), m(θ)=1-Σm(θ) ijk (Global trust reflects overall uncertainty).

[0111] 3.2 Handling and Combining Conflicts of Evidence

[0112] 3.2.1 Calculation of Conflict Coefficient: First, calculate the degree of conflict K among the three elements of evidence (the larger K is, the more severe the conflict): .

[0113] If K < 0.3 (low conflict), Dempster's combination rule is applied directly; if K ≥ 0.3 (high conflict), an improved rule (weighted discounting of conflicting evidence) is applied.

[0114] ;

[0115] Where α=0.5 (conflict allocation coefficient, optimized through the validation set), some conflicts are allocated to the whole domain to avoid contradictory results.

[0116] 3.2.2 Application of Combination Rules: Perform Dempster combination on the processed BPA to obtain the final combined BPA m. comb (θ ijk And calculate the confidence function (Bel, minimum confidence) and likelihood function (Pl, maximum confidence) for each proposition:

[0117] ;

[0118] ;

[0119] The overall probability is the average of Bel and Pl: .

[0120] 3.3 Comprehensive Probability Extraction

[0121] Select P comb (θ ijk The largest proposition is taken as the core result, for example:

[0122] Maximum proposition θ 221 =T2(48~72h)∩M2(6.3~6.5)∩L1(N30.6°), Bel=65%, Pl=80%, then the comprehensive probability Pcomb=72.5%), which can be simplified to "the comprehensive probability of an M6.3~6.5 earthquake occurring in the N30.6° area in the next 48~72 hours is 72%".

[0123] 4. Step 4: Output of Comprehensive Probability Report

[0124] (1) Purpose: To generate an intuitive and interpretable comprehensive report to support early warning decision-making, including comprehensive probability, weight basis, and conflict analysis.

[0125] (2) Report content (in JSON+PDF format, including visualization charts):

[0126] ① Enter basic information:

[0127] TLF-GV physical characteristics (GVCI peak 20, BF=0.8);

[0128] Three-factor independent prediction results (time probability, magnitude range, location heat map);

[0129] ② Weight allocation results:

[0130] The dynamic weights are W = [0.3, 0.35, 0.35].

[0131] Weighting criteria (e.g., "a magnitude weight of 0.35 has high credibility because the physical constraint deviation is only 0.1").

[0132] ③ Overall probability results:

[0133] Core propositions (future 48~72h, M6.3~6.5, N30.6°) and overall probability 72%;

[0134] Uncertainty range (Bel=65%~Pl=80%)

[0135] ④ Conflict Analysis:

[0136] Conflict coefficient K=12% (low conflict);

[0137] Sources of conflict (e.g., "a slight contradiction exists between time probability and position probability, which has been resolved through combination rules").

[0138] ⑤ Risk advice:

[0139] Overall probability ≥ 70%: Activate high-risk warning and focus on monitoring core areas;

[0140] Follow-up actions: Continue to monitor the TLF-GV signal. If the GVCI peak continues to rise, the overall probability needs to be updated.

[0141] Compared to the existing technology CN101793972A (Short-term Prediction Satellite Thermal Infrared Brightness Temperature Anomaly Technology for Strong Earthquakes), this embodiment addresses its key shortcomings through a core architecture of "fuzzy gating network + improved DS evidence theory." The technical effects and advantages are as follows:

[0142] First, it breaks through the limitations of discrete prediction and achieves accurate prediction through the coupling of three elements. Existing technologies can only output discrete results such as "an M6.5-7.0 earthquake will occur in a 1°×1° area in the next 15 days", with an epicenter area of ​​over 100,000 square kilometers and a minimum time resolution of 10 days. This invention, by unifying the probability dimension, outputs a coupled result such as "an M6.5±0.3 earthquake will occur in the 34°15′40″N area in the next 48 hours (comprehensive probability 72%)", improving short-term temporal accuracy by more than 40% compared to existing technologies.

[0143] Second, it dynamically adapts to signal changes, overcoming the rigidity of fixed weights. Existing technologies use fixed weights of "0.6 for thermal infrared and 0.2 for infrasound," which easily amplify errors when the signal is interfered with (such as when clouds obscure thermal infrared data). This invention uses a fuzzy gated network to correlate TLF-GV signal features with model credibility in real time, improving prediction stability by 35% in signal fluctuation scenarios and making it more suitable for the nonlinear signal characteristics of the imminent stage of strong earthquakes.

[0144] Third, it proactively resolves evidence conflicts and improves the reliability of data fusion. Existing technologies rely solely on weighting to mask data contradictions (such as abnormal thermal infrared but no infrasound response). This invention quantifies contradictions through conflict coefficients and resolves them using improved DS rules combined with physical meaning. The conflict resolution rate exceeds 90%, and the reliability of the fusion results is improved by 25% compared to existing technologies, reducing false alarms.

[0145] Fourth, it enhances the interpretability of decisions and moves away from "black box" outputs. Existing technologies only provide probability values ​​and cannot trace the prediction logic; this invention includes weighting rationale (such as "location weight 0.4 is due to high fault constraint satisfaction") and conflict analysis (such as "conflict degree 15% stems from unstable time model parameters"), improving the efficiency of disaster prevention decision adjustment by 50%.

[0146] The above is an embodiment of the TLF-GV signal-correlated earthquake three-element joint probability prediction method provided in this example. Based on the same idea, this embodiment also provides a corresponding TLF-GV signal-correlated earthquake three-element joint probability prediction system. Specific limitations of the TLF-GV signal-correlated earthquake three-element joint probability prediction system can be found in the limitations of the TLF-GV signal-correlated earthquake three-element joint probability prediction method above, and will not be repeated here. Each module in the above TLF-GV signal-correlated earthquake three-element joint probability 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.

[0147] 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 is a joint probabilistic prediction method for three earthquake elements associated with TLF-GV signals.

[0148] 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.

[0149] 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 joint probabilistic prediction method for three earthquake elements associated with TLF-GV signals, characterized in that, Includes the following steps: Physical features of the extracted low-frequency gravity-vibration signal TLF-GV, including the peak value GVCI of the gravity-vibration contribution index GVCI time series peak , the Bayesian factor BF, the TLF-GV anomaly duration T anomaly ; A three-element prediction model is constructed, including a time probability prediction model, a magnitude probability prediction model, and a location probability prediction model. The reliability of each prediction model is determined, and the model is concatenated with physical features to form a feature vector for weight allocation. The feature vector is then input into a dynamic weight prediction model based on a fuzzy gated network to obtain the allocation results of time weight, magnitude weight, and location weight. The time prediction results in the three-element prediction model are divided into multiple core intervals, the magnitude prediction results are divided into magnitude intervals, and the high-probability areas in the location prediction results are merged into multiple core intervals, generating multiple propositions of time-magnitude-location. Based on the weighting results, a weighted basic probability assignment (BPA) is generated for each proposition, and the conflict coefficient of each proposition is determined. The degree of conflict is determined based on the conflict coefficient. BPAs exceeding the preset threshold are weighted and discounted. Then, the Dempster combination rule is used to obtain the joint BPA, and the reliability function and likelihood function of each proposition are extracted. The mean of the confidence function and likelihood function for each proposition is taken as the comprehensive probability for that proposition, and the proposition with the highest comprehensive probability is taken as the output item. The construction of the three-element prediction model, including a time probability prediction model, a magnitude probability prediction model, and a location probability prediction model, and the determination of the reliability of each prediction model, includes the following steps: The gravity-vibration contribution index (GVCI) time series of the TLF-GV signal was extracted, and the 7-day time probability was predicted using the Informer time series prediction model. Specifically, based on historical earthquake events and their associated TLF-GV signal anomaly data, a Weibull statistical prior distribution characterizing the time interval from the TLF-GV anomaly peak to the earthquake occurrence was constructed. The future GVCI sequence predicted by the Informer time series prediction model was used as observational evidence. The parameters of the Weibull statistical prior distribution were dynamically updated using Bayes' theorem to obtain the posterior Weibull distribution. A likelihood function P(E|λ,k) was constructed to evaluate the probability of the observational evidence E in the Weibull distribution. According to Bayes' theorem, the Weibull prior distribution and the likelihood function were combined, and the posterior distribution π(λ,k|E) of the parameter λ was calculated using the Markov chain Monte Carlo (MCMC) sampling method. The mean of the posterior distribution was then used as the updated parameter λ. post According to parameter λ post Determine the reliability of the time model: ； Multi-source feature vectors and physical feature vectors of the TLF-GV signal are extracted, and the predicted magnitude is obtained through a magnitude probability prediction model. The reliability of the magnitude model is determined based on the physical constraint deviation of the predicted magnitude. ； Gravity gradient grids from TLF-GV signals are extracted, and location probability heatmaps are obtained through location probability prediction models. The reliability of the location model is determined based on the high-probability regions in the probability heatmap and their distance from the fault. 。 2. The joint probabilistic prediction method for three earthquake elements associated with TLF-GV signals according to claim 1, characterized in that, The generation of the multiple propositions concerning time-magnitude-location includes the following steps: The 7-day time probability is divided into 3 core intervals as time intervals, including T1 (24~48h), T2 (48~72h) and T3 (72~96h), and the cumulative probability of each time interval is calculated; Based on the range of uncertainty, three magnitude intervals are fixedly divided, including M1 (( -0.2~ M2 ~ +0.2)) and M3 ( +0.2~ +0.4), to ensure matching with the time and location interval dimensions; High-probability areas with a probability ≥ 0.6 are extracted from the heat map and merged into 3 core regions as location intervals, including the main high-probability area L1, the second high-probability area L2, and the potential area L3. The average probability of each location region is calculated. Construct a time-magnitude-location proposition and an empty set based on the time interval, magnitude interval, and location interval.

3. The joint probabilistic prediction method for three earthquake elements associated with TLF-GV signals according to claim 2, characterized in that, The process of generating the weighted basic probability assignment (BPA) for each proposition based on the weight allocation result includes the following steps: Based on the cumulative probability P(T) of each proposition over the time interval i ), Magnitude Probability P(M) j ) and the average probability P(L) of the location region k After weighting and multiplying, extend to the entire domain θ to transform into the BPA of each proposition: m(θ ijk )=m t (T i )×m m (M j )×m l (L k ),m(θ)=1-Σm(θ ijk ); Where, m t (T i )=w t ×P(T i () represents time BPA, m m (M j )=w m ×P(M j ) represents the magnitude BPA, m l (L k )=w l ×P(L k ) represents location BPA; w t w m and w l The weights are the weights of the three factors.

4. The joint probabilistic prediction method for three earthquake elements associated with TLF-GV signals according to claim 3, characterized in that, The step of determining the degree of conflict based on the conflict coefficient and weighting the BPA exceeding the preset threshold includes: The degree of conflict K in obtaining the three elements of evidence: ； If K < 0.3, Dempster's combination rule is applied directly; if K ≥ 0.3, an improved rule is applied, with weighted discounts applied to conflicting evidence. ； Where α is the conflict allocation coefficient.

5. The joint probabilistic prediction method for three earthquake elements associated with TLF-GV signals according to claim 4, characterized in that, The method of obtaining the joint BPA using Dempster's combination rule and extracting the confidence function and likelihood function for each proposition includes the following steps: Perform Dempster combinatorial operations on the BPAs of each proposition to obtain the final joint BPA m. comb (θ ijk And calculate the reliability function Bel and the likelihood function Pl for each proposition: ； ； The mean of the confidence function and likelihood function for each proposition is taken as the overall probability for that proposition: 。 6. The joint probabilistic prediction method for three earthquake elements associated with TLF-GV signals according to claim 1, characterized in that, The dynamic weight prediction model is constructed from a shallow MLP. The training set consists of historical TLF-GV physical features and model credibility feature vectors, as well as the corresponding manually labeled optimal weights. Mean squared error is used as the loss function.

7. A joint probabilistic prediction system for three earthquake elements associated with TLF-GV signals, 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 seismic three-element joint probability prediction method as described in any one of claims 1 to 6.

8. 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 seismic three-element joint probability prediction method as described in any one of claims 1 to 6.

Images