Strong earthquake prediction method and device based on earthquake precursor information, medium and equipment

Abstract

The invention discloses a strong earthquake prediction method and device based on earthquake precursor information, a medium and equipment, and relates to the technical field of earthquake prediction. When it is determined that the atmospheric tidal gravity signal is an atmospheric tidal gravity abnormal signal, inputting the atmospheric tidal gravity abnormal signal into a pre-constructed gradient elevator model, and performing spatial partition simulation on the research area to obtain a simulated magnitude of each partition; compensating the energy loss of the signal in the complex medium propagation by using an attenuation function to obtain the actual magnitude of each partition; further determining the distribution condition of a seismic source risk area in combination with regional seismic spatial distribution characteristics, and outputting a strong seismic prediction result of the research area according to the distribution condition; on the basis, a characteristic contribution degree analysis method is introduced, physical interpretability test is carried out on model output, and an influence mechanism of key characteristics of abnormal signals on a prediction result is clarified. According to the method, the accuracy of strong earthquake quantitative prediction is improved.

Description

Technical Field

[0001] This invention relates to the field of earthquake prediction technology, and in particular to a method, apparatus, medium, and equipment for predicting strong earthquakes based on earthquake precursor information. Background Technology

[0002] With the explosive growth of data in scientific research and all aspects of production and daily life, humanity has entered the era of big data. How to fully utilize massive amounts of data and extract its value has become a focus of attention in various fields. Currently, machine learning has become a major technical means to solve many application problems.

[0003] Over the past few decades, scientists have conducted extensive research in the field of earthquake prediction. Although earthquake prediction has been found to be "far more difficult than imagined," a wealth of observational evidence suggests that earthquakes are not entirely unpredictable. Strong earthquakes often exhibit certain patterns and precursory phenomena, making earthquake prediction theoretically possible. In the area of ​​short-term strong earthquake prediction, interdisciplinary research combining earthquake precursor information with artificial intelligence methods is gradually increasing. Researchers are attempting to use complex mathematical models and machine learning algorithms to predict earthquakes, but overall, qualitative assessments remain dominant, and quantitative predictions of strong earthquakes are still not feasible. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, device, medium, and equipment for strong earthquake prediction based on earthquake precursor information to address the above-mentioned technical problems. This method improves the accuracy of quantitative prediction of strong earthquakes.

[0005] The present invention adopts the following technical solution: This invention provides a method for predicting strong earthquakes based on earthquake precursor information, including: Atmospheric tidal gravity signals are collected in real time using an atmospheric tidal gravimeter located at the observation point. When the atmospheric tidal gravity signal is determined to be an atmospheric tidal gravity anomaly signal, the atmospheric tidal gravity anomaly signal is input into the pre-constructed gradient lift machine model. The gradient lift machine model is used to simulate the spatial partitioning of the study area to obtain the simulated magnitude of each partition in the study area. The study area is the area centered on the observation point and within a preset radius threshold. The simulated magnitude of each zone is converted into the actual magnitude using an attenuation function; the attenuation function is used to compensate for energy loss of the signal during propagation through complex geological structures. The spatial distribution characteristics of earthquakes in the study area were analyzed to determine the distribution of earthquake source risk zones in the study area; Based on the actual magnitude and focal risk zone distribution of each zone, the prediction results for strong earthquakes in the study area are determined.

[0006] Optionally, the method further includes: After collecting atmospheric tidal gravity signals, the signals are input into a classification model to determine the signal classification results. The signal classification results include whether the atmospheric tidal gravity signal is an anomalous atmospheric tidal gravity signal or a non-annomous atmospheric tidal gravity signal. The classification model is trained on a dataset composed of positive and negative balanced samples, consisting of anomalous atmospheric tidal gravity signals monitored before a strong earthquake and normal atmospheric tidal gravity signals monitored before a strong earthquake.

[0007] Optionally, the attenuation function is: ; in, To simulate magnitude, This represents the actual magnitude of the strong earthquake. The attenuation coefficient is... This represents the distance from the location of the strong earthquake to the observation point.

[0008] Optionally, the training process of the gradient boosting machine model includes: Multiple samples of atmospheric tidal gravity signals were acquired, and based on the characteristic correlation of the atmospheric tidal gravity signals, several relatively independent features in each sample of atmospheric tidal gravity signals were screened. For any sample atmospheric tidal gravity signal, the sample distance between the strong earthquake and the observation point is determined based on the observation point of the sample atmospheric tidal gravity signal and the corresponding strong earthquake location. The sample distance and the corresponding sample magnitude are substituted into the attenuation function to obtain the sample simulated magnitude. A sample set is constructed based on the atmospheric tidal gravity signals of the samples after feature screening and the corresponding simulated magnitudes of the samples; The gradient booster model is obtained by training the sample set.

[0009] Optionally, a sample set is constructed based on the atmospheric tidal gravity signals of the samples after feature filtering and the corresponding simulated magnitudes of the samples, including: The atmospheric tidal gravity signals of the samples after feature screening are standardized. Based on the standardized atmospheric tidal gravity signals and the corresponding simulated magnitudes of the samples, a sample set is obtained through Bootstrap resampling expansion.

[0010] Optionally, the spatial distribution characteristics of earthquakes include the locations of historical strong earthquakes and their corresponding historical magnitudes; the spatial distribution characteristics of earthquakes in the study area are analyzed to determine the distribution of focal risk zones in the study area, including: Based on the locations and corresponding magnitudes of historical strong earthquakes in the study area, the distribution of seismic source risk zones in the study area is determined using kernel density estimation.

[0011] This invention provides a strong earthquake prediction device based on earthquake precursor information, comprising: The acquisition module is used to acquire atmospheric tidal gravity signals in real time through an atmospheric tidal gravimeter located at the observation point; The prediction module is used to input the atmospheric tidal gravity anomaly signal into a pre-built gradient lift model when the atmospheric tidal gravity signal is determined to be an atmospheric tidal gravity anomaly signal. The gradient lift model is used to perform spatial partitioning simulation of the study area to obtain the simulated magnitude of each partition in the study area. The study area is the area centered on the observation point and within a preset radius threshold. The correction module is used to convert the simulated magnitude of each zone into the actual magnitude using an attenuation function; the attenuation function is used to compensate for the energy loss of the signal propagating through complex geological structures. The analysis module is used to analyze the spatial distribution characteristics of earthquakes in the study area and determine the distribution of seismic source risk zones in the study area. The determination module is used to determine the strong earthquake prediction results for the study area based on the actual magnitude and source risk zone distribution of each partition.

[0012] The present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described strong earthquake prediction method based on earthquake precursor information.

[0013] The present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the above-mentioned strong earthquake prediction method based on earthquake precursor information.

[0014] The above-mentioned at least one technical solution adopted in this invention can achieve the following beneficial effects: In this invention, the simulated magnitude is predicted using atmospheric tidal gravity anomaly signals instead of directly using the original atmospheric tidal gravity signals. This is equivalent to filtering noise in advance and avoiding interference from normal signals in the prediction. Subsequently, the simulated magnitude is compensated in reverse using an attenuation function to avoid errors caused by propagation loss. Finally, the distribution of seismic source risk zones is identified based on the spatial distribution characteristics of earthquakes. This ensures that the strong earthquake prediction results include not only the actual magnitude but also the spatial information of high-risk locations, further improving the accuracy of strong earthquake prediction compared to single-dimensional magnitude prediction. Attached Figure Description

[0015] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0016] Figure 1 This is a schematic diagram of a strong earthquake prediction method based on earthquake precursor information provided by the present invention; Figure 2 This is a schematic diagram of ATG signal feature correlation analysis provided by the present invention; Figure 3 A SHAP value distribution diagram of various characteristic parameters of an ATG signal provided by the present invention; Figure 4 A SHAP overall feature diagram provided by the present invention; Figure 5 A schematic diagram of a computer device for implementing a strong earthquake prediction method based on earthquake precursor information, provided by the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0018] In the field of earthquake prediction, there is a growing trend of interdisciplinary research combining earthquake precursor information with artificial intelligence methods. Researchers are attempting to use complex mathematical models and machine learning algorithms to predict earthquakes. While machine learning-based methods have made significant progress in long- and medium-term earthquake prediction, and have also begun to show breakthroughs in short-term strong earthquake prediction, the field as a whole remains in the exploratory stage.

[0019] Since the beginning of the 21st century, numerous studies have utilized artificial intelligence methods for earthquake prediction, achieving significant results. For example, researchers proposed a strong earthquake prediction method based on seismic activity indicators, defining eight seismic activity indicators based on the temporal distribution of earthquakes to characterize regional potential seismic activity. They conducted comparative experiments in Southern California and the San Francisco Bay Area using models such as Recurrent Neural Networks (RNNs), Back Propagation Neural Networks (BPNNs), and Radial Basis Function Networks (RBFs). The results showed that RNNs performed best. Subsequently, the model was replaced with a Probabilistic Neural Network (PNN) for prediction in the same region, showing that PNNs performed better for earthquakes of magnitude 4.5–6.0. ​​Researchers have also conducted earthquake prediction studies using time-series earthquake magnitude data and seismic electric signals (SES), employing various neural network architectures to evaluate their effectiveness in predicting earthquakes in Greece. Some models achieved prediction accuracy exceeding 80% for small-magnitude earthquakes. Researchers further constructed a dataset containing 60 seismic activity indicators and trained a hybrid model of Support Vector Machine (SVM) and Hierarchical Neural Network (SVM-HNN) to predict earthquakes of magnitude 5.0 and above in regions such as Southern California, achieving a significant improvement in accuracy compared to previous methods. Another research team developed a deep learning model trained using aftershock data from the past decade to predict the distribution of aftershocks up to a year after a major earthquake, achieving prediction accuracy exceeding traditional empirical models. Subsequently, researchers also constructed a Focal Mechanism Neural Network (FMNet) model, using the full waveform of an earthquake to achieve real-time estimation of focal mechanism parameters, achieving a breakthrough in the field of rapid seismic parameter inversion.

[0020] Overall, although existing models have shown some effectiveness in experiments, most are limited to laboratory simulations or historical data retrospective predictions, and still fall far short of practical needs in short-term earthquake prediction. As a famous paper points out, the assertion that "earthquakes are unpredictable" reflects the enormous obstacles facing short-term earthquake prediction: the earthquake gestation process is extremely complex, small earthquakes can cascade into large ones, and there is a lack of objective definitions of precursory anomalies and sufficient statistical evidence to link them to impending earthquakes. Therefore, many scholars hold a cautious or pessimistic attitude towards short-term earthquake prediction. However, a series of devastating earthquakes in recent decades (such as the 2008 Wenchuan earthquake (magnitude 8.0), the 2011 Japan east coast earthquake (magnitude 9.0), and the 2023 Turkey earthquake (magnitude 7.8)) failed to be effectively predicted, causing enormous casualties and economic losses. This, in turn, motivates researchers to continue exploring new approaches, attempting to find breakthroughs in the seemingly elusive field of short-term earthquake prediction.

[0021] Among numerous precursory phenomena of earthquakes, changes in the gravity field are considered an important physical quantity for studying the gestation and occurrence of earthquakes because they directly reflect the migration and density changes of underground materials. As early as the 1960s, scholars began to explore the patterns of earthquake precursors through gravity observations. For example, researchers established connections between crustal deformation and gravity changes, reported pre-earthquake gravity change signals in North China, and used superconducting gravimeters to monitor gravity disturbances related to tidal strain. Some scholars also used the GRACE satellite to observe coseismic and post-earthquake gravity effects of large earthquakes. These studies showed that measurable gravity anomalies may indeed accompany earthquake gestation, but the observation methods at that time (such as satellite gravity, absolute or relative gravimeters) still had problems with insufficient sensitivity and limited resolution in capturing weak anomalies in the short pre-earthquake phase. With the continuous development of high-precision gravimeters, capturing pre-earthquake physical anomalies has become possible. The "atmospheric tidal gravity anomaly," proposed in recent years, refers to gravity disturbances in the 1–30 mHz frequency band with amplitudes of approximately 1–10 μGal, which are considered closely related to mass migration and stress accumulation during the earthquake preparation stage. Existing field studies have shown that this type of anomaly has a high occurrence rate before strong earthquakes, demonstrating good physical correlation and monitoring continuity, and is therefore considered a potential short-term earthquake precursor. However, most existing research focuses on anomaly identification or empirical verification of individual cases, lacking a systematic method for quantitatively predicting the three elements (time, epicenter size, and magnitude) of strong earthquakes. Therefore, developing a quantitative prediction system based on atmospheric tidal gravity anomalies has become a key direction for overcoming the bottleneck in short-term earthquake prediction.

[0022] Current work on short-term applications of atmospheric tidal gravity (ATG) precursors largely remains focused on identifying anomalies and comparing individual cases. It lacks verifiable and transferable quantitative mappings that directly map observational characteristics to magnitude and source distance. Regarding path effects and physical normalization, many studies directly input the raw amplitudes into the model, rarely explicitly addressing the attenuation and loading transmission of ATG signals through the "distance-medium" relationship. This leads to far-field overestimation and near-field underestimation, disrupting stable mappings to magnitude and source distance. High-precision gravity studies show that atmospheric mass redistribution and loading have significant local / regional / global zonal contributions. Zonal modeling or the introduction of empirical attenuation / band normalization are necessary to project observations under different distances and media conditions onto comparable scales. Regarding modeling objectives and validation paradigms, existing earthquake studies related to ATG / atmospheric-ionospheric anomalies mostly output classification probabilities of "earthquake presence / absence" or "whether an earthquake occurred within a certain time window," with joint regressions directly addressing magnitude and focal distance being uncommon. Furthermore, many methods lack prospective testing and standardized evaluation, making it difficult to prove cross-regional generalization. The earthquake predictability community has established a prospective testing framework and a complete set of statistical tests for consistency, spatiality, quantity, and magnitude; however, many models have not undergone similar rigorous field validation, posing risks of insufficient generalization and transferability. There are also issues with small sample sizes and evaluation bias in machine learning for earthquake prediction. In terms of feature engineering and interpretability, ATG spectral features often exhibit strong collinearity (e.g., adjacent dominant frequency amplitudes are similar), which, without systematic band-dependent correction and decorrelation processing, can lead to unstable model parameters and overfitting. Existing research shows that using frequency-dependent pressure equivalent / loading corrections can significantly reduce residuals and improve spectral consistency, indicating that band processing is a necessary step; however, the implementation of this in existing precursor modeling is inconsistent. On the other hand, many models still lack a unified interpretable framework to quantify the positive and negative contributions of each feature to the output, limiting the formulation of engineering thresholds and business rules. Finally, regarding academic and engineering attempts to use atmospheric tidal gravity for pre-earthquake applications, there have been reviews and case reports in China, but overall, they still mainly focus on anomaly occurrence rates and typical earthquake case reviews. Systematic quantitative models and unified evaluations of anomalies and earthquakes are still insufficient, which is one of the direct bottlenecks that makes it difficult to scale up the methods.

[0023] To address the shortcomings of existing technologies, this invention provides a method for predicting strong earthquakes based on earthquake precursor information. This method acquires and corrects atmospheric tidal gravity anomaly signals, extracts and filters features, introduces an attenuation function for physical normalization, and constructs a joint prediction framework by combining a gradient boosting model and kernel density estimation to achieve quantitative prediction of the magnitude and focal range of strong earthquakes.

[0024] The technical solutions provided by the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0025] Figure 1 This is a schematic diagram of a strong earthquake prediction method based on earthquake precursor information according to the present invention, which specifically includes the following steps: S101 collects atmospheric tidal gravity signals in real time using an atmospheric tidal gravimeter located at the observation point.

[0026] For example, the observation point can be located at an observation base that can effectively respond to strong earthquakes within a radius of approximately 7,000 kilometers.

[0027] S102, when the atmospheric tidal gravity signal is determined to be an atmospheric tidal gravity anomaly signal, the atmospheric tidal gravity anomaly signal is input into the pre-constructed gradient lift machine model. The gradient lift machine model is used to simulate the spatial partitioning of the study area to obtain the simulated magnitude of each partition in the study area. The study area is the area centered on the observation point and within the preset radius threshold.

[0028] In one embodiment, after the atmospheric tidal gravity signal is acquired, it is input into a classification model to determine the signal classification result of the atmospheric tidal gravity signal. The signal classification result includes whether the atmospheric tidal gravity signal is an anomalous atmospheric tidal gravity signal or a non-anomalous atmospheric tidal gravity signal. The classification model is trained on a dataset composed of positive and negative balanced samples, which consists of anomalous atmospheric tidal gravity signals monitored before a strong earthquake and normal atmospheric tidal gravity signals monitored before a strong earthquake.

[0029] In one embodiment, the training process of the gradient booster model includes: acquiring multiple sample atmospheric tidal gravity signals, and filtering out several relatively independent features from each sample atmospheric tidal gravity signal based on the characteristic correlation of the atmospheric tidal gravity signals; for any sample atmospheric tidal gravity signal, determining the sample distance between the strong earthquake and the observation point based on the observation point of the sample atmospheric tidal gravity signal and the corresponding strong earthquake location, substituting the sample distance and the corresponding sample magnitude into the attenuation function to obtain the sample simulated magnitude; constructing a sample set based on the sample atmospheric tidal gravity signals after feature filtering and the corresponding sample simulated magnitude; and training the gradient booster model using the sample set. During the training process, a feature contribution analysis method is introduced to perform a physical interpretability test on the model output, clarifying the influence mechanism of key features of anomalous signals on the prediction results.

[0030] Specifically, the data sources for the datasets of classification models and gradient boosting machine models include: Data Sources and Study Scope: The basic data used consists of two parts: first, atmospheric tidal disturbance signals collected by the Atmospheric Tidal Dynamic Gravimeter between September 20, 2021 and April 27, 2024; second, earthquake data from the National Earthquake Science Data Center (China Earthquake Networks Center Rapid Earthquake Catalog), including catalog information (containing key information such as magnitude, focal depth, and epicenter geographical coordinates) of earthquakes with a magnitude ≥ 6.0 occurring within the study area and within the same time frame as mentioned above. The study area is limited to the Eurasian Plate and its surrounding areas (defined as longitude 30°E-80°E, latitude 20°S-90°N).

[0031] Data Characteristics and Dataset Construction: From September 20, 2021 to April 27, 2024, a total of 77 strong earthquakes with a magnitude greater than 6.0 occurred in the study area. Analysis of the temporal and spatial correspondence between ATG signals and these earthquake events revealed that significant ATG anomalous signals were detected before 63 of these 77 strong earthquakes, exhibiting specific spatiotemporal evolution characteristics and demonstrating their potential as earthquake precursors. To construct the model training dataset, "non-excited state" signal samples (i.e., normal signals without anomalous disturbances, with a target magnitude set to 0) were introduced, forming a positive-negative balanced dataset together with anomalous "excited state" signals (ATG anomalous signals). This strategy of including negative samples can improve the model's generalization ability, avoid learning only pre-earthquake anomalies while ignoring signal characteristics under normal conditions, and thus reduce the risk of overfitting. Specifically, a two-stage modeling approach is employed: first, a classification model is trained to determine whether a given signal is in an excited state (indicating the presence of pre-seismic anomalies); if it is determined to be in an excited state, a Gradient Boosting Machine (GBM) model is used to predict the corresponding earthquake magnitude. This classification-decision two-step method avoids noise interference when directly predicting magnitude for all signals. The final dataset contains features extracted from ATG signals and information on corresponding seismic events, totaling 126 sample records.

[0032] It should be noted that, to improve model efficiency, the extracted signal features can be screened to remove redundant information. Pearson correlation coefficient analysis, combined with a visualization heatmap, was used to assess the correlation between features and remove highly correlated features, retaining only a subset of features with independent information. The Pearson correlation coefficient r reflects the degree of linear correlation between two variables: generally, |r| < 0.4 is considered low correlation, 0.4 ≤ |r| < 0.7 is moderate correlation, and |r| ≥ 0.7 is high correlation. Heatmap observation revealed that some features exhibited extremely high correlation—for example, the amplitude values ​​of the first five dominant frequencies of the signal were almost identical, indicating high correlation. In other words, the spectral energy of the ATG signal is mainly concentrated in the dominant frequency band of approximately 1 Hz, with other frequency components having extremely small amplitudes, resulting in high redundancy among these dominant frequency amplitude features. This phenomenon suggests that the signal source may be related to a slowly changing physical process, with the vast majority of signal energy accumulating near the dominant frequency, and the contribution of other frequency bands being negligible. To avoid the adverse effects of such multicollinearity on the model, 18 relatively independent features were ultimately retained for modeling analysis, such as… Figure 2 As shown, Figure 2 This is a schematic diagram of the correlation analysis of ATG signal characteristics.

[0033] In one embodiment, a sample set is constructed based on the atmospheric tidal gravity signals of the selected samples and the corresponding simulated magnitudes. This includes: standardizing the atmospheric tidal gravity signals of the selected samples; and expanding the sample set through Bootstrap resampling based on the standardized atmospheric tidal gravity signals of the samples and the corresponding simulated magnitudes. Specifically, before inputting the data into the model, the features can be standardized. This involves subtracting the mean of each feature from its value in the training set and then dividing by the standard deviation, thereby converting the features into a standard normal distribution with zero mean and unit variance. This eliminates the influence of differences in units and numerical ranges on model training, ensuring that each feature contributes to the same order of magnitude to model training, which helps improve model performance. Given the limited number of samples, a Bootstrap sampling method is introduced to expand the data and evaluate model robustness. Bootstrap sampling constructs multiple new datasets of the same size as the original dataset by randomly drawing samples with replacement from the original dataset. Samples not selected in each sampling can be used as the corresponding test set. By performing multiple bootstrap iterations, multiple training-test splits are generated for model training and cross-validation. This strategy makes full use of limited data; multiple resampling not only improves the model's generalization ability but also effectively mitigates the risk of overfitting. The combination of standardization and bootstrap resampling makes model training more stable and reliable.

[0034] Furthermore, ATG signals possess unique advantages in strong earthquake monitoring due to their ability to penetrate media such as seawater and the Earth's crust over long distances. However, it must be considered that signal amplitude weakens with increasing propagation distance, and complex media environments may also reduce signal clarity. Therefore, attenuation correction is needed for observed signals at different distances during modeling and prediction. This invention employs an attenuation function that incorporates the distance between the epicenter and the observation point to correct the observed signal amplitude and generate a "simulated magnitude." As the core target variable for model prediction. In short, it is assumed that the initial earthquake magnitude is... M The propagation attenuation coefficient is α The distance from the epicenter to the observation point is d (Calculating distances in kilometers using the Haversian formula on the Earth's surface), then the corresponding distances can be calculated using empirical formulas. .this This represents the equivalent magnitude the earthquake would have occurred near the observation point. This attenuation correction compensates for signal propagation losses in complex geological structures, ensuring data from sensors at different distances are comparable on the same scale, thus improving model stability and prediction accuracy.

[0035] Optionally, the attenuation function is: (1); in, To simulate magnitude, This represents the actual magnitude of the strong earthquake. The attenuation coefficient is... This represents the distance from the location of the strong earthquake to the observation point.

[0036] The datasets constructed through feature selection, sample expansion, and magnitude attenuation were used for training to obtain classification models and GBM models, respectively.

[0037] Based on the constructed GBM model, atmospheric tidal gravity anomaly signals are input into the GBM model. The GBM model then performs spatial partitioning simulation of the study area, obtaining simulated magnitudes for each partition. Specifically, a study area with a radius of approximately 7000 km is defined, centered on the observation point, and this area is divided into several irregular spatial partitions. Each partition serves as an independent prediction unit, and the GBM model outputs the corresponding simulated magnitude value. This spatial discretization strategy avoids relying solely on the radial symmetry assumption, more accurately reflecting the differences in signal intensity under different directions and distances. As the distance from the observation point increases, the simulated magnitude typically shows a gradual attenuation trend, thus reflecting the energy dissipation characteristics of atmospheric tidal gravity signals during propagation. Through this method, the GBM model fully incorporates the spatial variation law of signal intensity, providing magnitude estimation results that conform to physical attenuation characteristics.

[0038] S103 converts the simulated magnitude of each zone into the actual magnitude using an attenuation function; the attenuation function is used to compensate for energy loss of the signal during propagation through complex geological structures.

[0039] By performing the inverse operation of the aforementioned attenuation function, the simulated magnitudes of each region are converted into the corresponding actual earthquake magnitudes. In other words, based on the simulated magnitudes estimated from observed atmospheric tidal gravity anomalies, and considering distance attenuation, the possible actual magnitude at the epicenter is calculated.

[0040] S104 analyzes the spatial distribution characteristics of earthquakes in the study area and determines the distribution of earthquake source risk zones in the study area.

[0041] In one embodiment, the spatial distribution characteristics of earthquakes include the locations of historical strong earthquakes and their corresponding historical magnitudes; analyzing the spatial distribution characteristics of earthquakes in the study area to determine the distribution of source risk zones in the study area includes: determining the distribution of source risk zones in the study area using kernel density estimation based on the locations of historical strong earthquakes and their corresponding historical magnitudes in the study area.

[0042] Specifically, to further improve the accuracy of epicenter location prediction, the kernel density estimation (KDE) method is introduced to analyze the spatial distribution characteristics of earthquakes within the study area. First, KDE is used to calculate the spatial probability density distribution of earthquake events. KDE analysis is equivalent to superimposing the influence range on the geographical location of earthquakes, representing areas with frequent historical strong earthquakes with high probability density. The magnitude of historical earthquakes and the level of recent tectonic activity are also considered, weighting the density distribution so that historical earthquakes with larger magnitudes and areas with more frequent recent activity contribute more to the density estimation. Through this kernel density modeling, the distribution of potential high-risk seismic source areas within the study area is obtained.

[0043] S105. Based on the actual magnitude and source risk zone distribution of each zone, the prediction results of strong earthquakes in the study area are determined.

[0044] The actual magnitude and focal risk zone distribution of each zone are used to determine the strong earthquake prediction results for the study area.

[0045] In one embodiment, this invention fuses the magnitude prediction output of the GBM model with the high-probability source regions obtained from KDE to construct a multi-dimensional strong earthquake prediction framework. By combining the signal intensity attenuation with distance and information on historical seismic activity zones, and considering the weighting of factors such as recent tectonic activity, several high-probability source regions may be output within a certain range of the observation point. Within these regions, the potential magnitude is then estimated using the GBM results.

[0046] Model Training and Performance: The relationship between atmospheric tidal gravity signals and seismic events can be highly complex and nonlinear. GBM, as a decision tree ensemble method, can approximate complex functional relationships by progressively stacking multiple weak learning trees, thereby capturing potential nonlinear interactions between features. Furthermore, GBM is robust to noise and redundant features, facilitating the training of robust models on data containing multidimensional signal features. During model training, Mean Squared Error (MSE) is used as the loss function, fitting the residuals of the previous iteration in each iteration to gradually reduce the prediction error. A learning rate is introduced to control the contribution of each newly added tree to the model, balancing training speed and preventing overfitting. The model is validated using K-fold cross-validation, employing metrics such as Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), MSE, and Coefficient of Determination (R²) to comprehensively evaluate the model's predictive performance and generalization ability. Cross-validation ensures that the model performs well not only on the training set but also exhibits stable performance on unseen data. Training results show that the GBM model has a high fit on this dataset. The MAE on the training set is approximately 0.077, and the RMSE is approximately 0.112; the MAE on the test set is approximately 0.135, and the RMSE is approximately 0.219, with an R² of approximately 0.914. Overall, the errors in both training and testing are low, the prediction accuracy is high, and there is no obvious overfitting, indicating that the GBM model has good applicability and reliability for the data.

[0047] Model Result Interpretation (SHapley Additive Explained, SHAP): SHAP is a method for interpreting the output of a machine learning model. It demonstrates how the model makes predictions by showing the contribution of each feature to the prediction result. The SHAP analysis results based on the GBM model are as follows: The distribution of SHAP values ​​for each feature parameter is shown below. Figure 3As shown in the figure, this graph details the distribution of SHAP values ​​for each feature parameter in the dataset and reflects their specific impact on model predictions. The x-axis represents the magnitude of the feature's SHAP value, i.e., the strength of the feature's contribution to the prediction result; the y-axis represents the importance ranking of the features, from top to bottom, indicating the features that contribute the most to the model prediction. Each point represents the SHAP value of a sample in the dataset, with its color gradient from blue to red, representing the change from low to high feature value. Positive SHAP values ​​indicate a positive driving force of the feature on the model's prediction, i.e., the higher the feature value, the larger the predicted value; negative SHAP values ​​indicate a negative correlation between the feature and the prediction result, i.e., the higher the feature value, the smaller the predicted value. For example, when the red dots (high feature values) are concentrated in the positive SHAP value region, it indicates that the increase in feature value significantly drives the increase in the model's prediction result. Figure 4 A quantitative analysis of the SHAP distribution plot highlighted the feature parameters that contribute most to the GBM model's prediction results. It was found that "number of anomalies" and "duration (seconds)" are the two features with the most significant impact on the model output. This indicates that the number and duration of anomalous fluctuations in the signal significantly affect the model's judgment when predicting whether atmospheric tidal anomalies are associated with seismic events. In summary, SHAP analysis improves the interpretability of the model results, enabling a deeper understanding of the role of ATG anomaly signals in strong earthquake prediction at the feature level.

[0048] This invention models the relationship between observed atmospheric tidal gravity anomalies and the occurrence of strong earthquakes, enabling quantitative prediction of earthquake magnitude and focal distance. Specifically, it utilizes the spatiotemporal evolution characteristics of gravity anomalies, combining a gradient boosting model with kernel density estimation to jointly predict the focal range and magnitude. By introducing a signal attenuation function to normalize path and medium effects, and employing a feature selection mechanism to eliminate high collinearity and redundancy, the robustness and generalization ability of the model are improved under limited sample conditions. Furthermore, the SHAP interpretation tool clarifies the contribution of key features to the prediction results, enhancing the physical interpretability of the predictions. This invention aims to provide quantitative prediction results within a short, temporary period of several days to twenty days before an earthquake, overcoming the shortcomings of existing technologies that primarily rely on qualitative assessments, and providing actionable data support for earthquake risk assessment and emergency response.

[0049] This invention models the correspondence between atmospheric tidal gravity anomalies and strong earthquake events, enabling quantitative prediction of magnitude and focal distance in the short pre-earthquake phase. The model exhibits high stability and accuracy during training and testing, demonstrating its effectiveness in capturing the intrinsic link between atmospheric tidal gravity anomalies and strong earthquake occurrence. Regarding focal range prediction, a risk area highly consistent with historical high-frequency seismic zones is identified using kernel density estimation. By introducing a signal attenuation function to correct observations at different epicentral distances, the prediction results are more consistent with physical laws, effectively reducing errors caused by differences in far-field and near-field responses. Furthermore, the SHAP analysis tool is used to interpret the model results, clarifying the dominant role of features such as "anomaly frequency" and "duration" in the prediction, making the prediction process traceable and physically interpretable, and enhancing the model's credibility in scientific research and engineering applications. Overall, this invention can provide quantitative prediction results and data support in a short pre-earthquake phase of several days to twenty days, offering new insights for earthquake risk management and emergency preparedness.

[0050] In terms of technical advantages, this invention represents a significant improvement over traditional methods that rely on empirical judgment. The prediction results represent a shift from qualitative assessment to quantitative output. The method integrates gradient boosting models and kernel density estimation, creating complementarity between magnitude prediction and focal range identification, thus improving the completeness of the prediction. By utilizing feature selection, bootstrap resampling, and signal attenuation functions, it enhances robustness under small sample conditions and the model's generalization ability. Furthermore, the result interpretation within the SHAP framework provides transparent model information for subsequent researchers and practical users, avoiding the risk of "black box" predictions.

[0051] When applying the strong earthquake prediction method based on earthquake precursor information provided by this invention, it is not necessary to rely on... Figure 1 The steps shown are executed in sequence. The specific execution order of each step can be determined as needed, and this invention does not impose any restrictions on it.

[0052] The above describes a strong earthquake prediction method based on earthquake precursor information provided by one or more embodiments of the present invention. Based on the same idea, the present invention also provides a corresponding strong earthquake prediction device based on earthquake precursor information, the device comprising: The acquisition module is used to acquire atmospheric tidal gravity signals in real time through an atmospheric tidal gravimeter located at the observation point; The prediction module is used to input the atmospheric tidal gravity anomaly signal into a pre-built gradient lift model when the atmospheric tidal gravity signal is determined to be an atmospheric tidal gravity anomaly signal. The gradient lift model is used to perform spatial partitioning simulation of the study area to obtain the simulated magnitude of each partition in the study area. The study area is the area centered on the observation point and within a preset radius threshold. The correction module is used to convert the simulated magnitude of each zone into the actual magnitude using an attenuation function; the attenuation function is used to compensate for the energy loss of the signal propagating through complex geological structures. The analysis module is used to analyze the spatial distribution characteristics of earthquakes in the study area and determine the distribution of seismic source risk zones in the study area. The determination module is used to determine the strong earthquake prediction results for the study area based on the actual magnitude and source risk zone distribution of each partition.

[0053] Specific limitations regarding strong earthquake prediction devices based on earthquake precursor information can be found in the above section on limitations of strong earthquake prediction methods based on earthquake precursor information, and will not be repeated here. Each module in the aforementioned strong earthquake prediction device based on earthquake precursor information can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0054] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 A method for predicting strong earthquakes based on earthquake precursor information is provided.

[0055] The present invention also provides Figure 5 The schematic diagram of the computer device shown is as follows: Figure 5 As shown, at the hardware level, this computer device includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then executes it to achieve the above. Figure 1 A method for predicting strong earthquakes based on earthquake precursor information is provided.

[0056] 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 methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention 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.

[0057] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this invention.

Claims

1. A method for predicting a strong earthquake based on earthquake precursor information, characterized by, The method comprises the following steps: collecting an atmospheric tide gravity signal in real time by an atmospheric tide gravimeter located at an observation point; when it is determined that the atmospheric tide gravity signal is an atmospheric tide gravity anomaly signal, inputting the atmospheric tide gravity anomaly signal into a gradient boosting machine model constructed in advance, simulating a spatial partition of a research area by the gradient boosting machine model, and obtaining a simulated magnitude of each partition in the research area; the research area is an area within a preset radius threshold centered on the observation point; converting the simulated magnitude of each partition into an actual magnitude by an attenuation function; the attenuation function is used to compensate for the energy loss of the signal in the propagation of complex geological structures; analyzing the spatial distribution characteristics of earthquakes in the research area to determine the distribution of the source risk area of the research area; determining the strong earthquake prediction result of the research area according to the actual magnitude of each partition and the distribution of the source risk area.

2. The method of claim 1, wherein, The method further comprises: after collecting the atmospheric tide gravity signal, inputting the atmospheric tide gravity signal into a classification model to determine the signal classification result of the atmospheric tide gravity signal; the signal classification result includes that the atmospheric tide gravity signal is an atmospheric tide gravity anomaly signal or a non-atmospheric tide gravity anomaly signal; the classification model is obtained by training a data set composed of positive and negative balanced samples based on sample atmospheric tide gravity anomaly signals monitored before a strong earthquake occurs and sample atmospheric tide gravity normal signals without a strong earthquake.

3. The method of claim 1, wherein, The attenuation function is: ； wherein, M is the magnitude of the earthquake, M0 is the actual magnitude of the earthquake, K is the attenuation coefficient, R is the distance from the location of the earthquake to the observation point.

4. The method of claim 3, wherein, The training process of the gradient boosting machine model comprises: obtaining a plurality of sample atmospheric tide gravity signals, and selecting a plurality of relatively independent features in each sample atmospheric tide gravity signal according to the feature correlation of the atmospheric tide gravity signal; for any sample atmospheric tide gravity signal, determining a sample distance between the occurrence of a strong earthquake and the observation point according to the observation point of the sample atmospheric tide gravity signal and the corresponding strong earthquake position, and inputting the sample distance and the corresponding sample magnitude into the attenuation function to obtain a sample simulated magnitude; constructing a sample set according to the sample atmospheric tide gravity signal after the feature selection and the corresponding sample simulated magnitude; training the gradient boosting machine model by using the sample set.

5. The method of claim 4, wherein, Constructing a sample set according to the sample atmospheric tide gravity signal after the feature selection and the corresponding sample simulated magnitude comprises: standardizing the sample atmospheric tide gravity signal after the feature selection; expanding the sample set by Bootstrap resampling according to the standardized sample atmospheric tide gravity signal and the corresponding sample simulated magnitude.

6. The method of claim 1, wherein, The spatial distribution characteristics of earthquakes include the positions of historical strong earthquakes and the corresponding historical magnitudes; analyzing the spatial distribution characteristics of earthquakes in the research area to determine the distribution of the source risk area of the research area comprises: determining the distribution of the source risk area of the research area by using kernel density estimation according to the positions of historical strong earthquakes in the research area and the corresponding historical magnitudes.

7. A strong earthquake prediction device based on earthquake precursor information, characterized in that, The method comprises the following steps: a collection module for collecting an atmospheric tide gravity signal in real time by an atmospheric tide gravimeter located at an observation point; The prediction module is configured to input the atmospheric tidal gravity anomaly signal to a pre-constructed gradient boosting machine model when determining that the atmospheric tidal gravity signal is the atmospheric tidal gravity anomaly signal, simulate the research region by the gradient boosting machine model in a spatial zoning manner, and obtain simulated magnitudes of each zone in the research region; the research region is a region within a preset radius threshold and centered on the observation point; The correction module is configured to convert the simulated magnitudes of each zone into actual magnitudes by using an attenuation function; The attenuation function is configured to compensate for energy loss of the signal in propagation in a complex geological structure; The analysis module is configured to analyze spatial distribution characteristics of earthquakes in the research region, and determine a distribution of a seismic source risk area of the research region; The determination module is configured to determine a strong earthquake prediction result of the research region according to the actual magnitudes of each zone and the distribution of the seismic source risk area.

8. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, and the computer program is executed by the processor to implement the method in any one of claims 1-6.

9. A computer device, comprising: The computer program is stored in the memory and executable on the processor, and the processor implements the method in any one of claims 1-6 when executing the program.

Images