Earthquake prediction method, system, device and storage medium
By employing a dedicated kernel function and a multi-kernel learning method to perform high-dimensional mapping and weight allocation on multidimensional heterogeneous earthquake precursor data, the problem of difficulty in fusing heterogeneous features is solved, thereby improving the accuracy and reliability of earthquake prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XI AN JIAOTONG UNIV
- Filing Date
- 2026-04-07
- Publication Date
- 2026-06-26
AI Technical Summary
Existing earthquake prediction systems suffer from low accuracy due to the difficulty in directly fusing multi-source earthquake precursor data, which is characterized by heterogeneity.
A dedicated kernel function is used to perform high-dimensional mapping of multidimensional heterogeneous features, and feature fusion is achieved through attention weight network and multi-kernel learning method. Combined with weight allocation of multi-source earthquake precursor data and sub-model weighting, linear fusion of heterogeneous features is realized.
It improves the accuracy of earthquake prediction, ensures that features can be directly used for linear calculations, dynamically adapts to the differences in importance of multi-source features, and provides more valuable prediction input.
Smart Images
Figure CN122283808A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of earthquake prediction technology, and in particular to an earthquake prediction method, system, device and storage medium. Background Technology
[0002] Earthquake prediction, or earthquake forecasting, involves accurately predicting the location, time, and magnitude of an earthquake before it occurs—the three elements of an earthquake. This greatly helps prevent casualties caused by large earthquakes occurring unexpectedly. Earthquake prediction remains a major scientific and technological challenge that the world has yet to solve. Earthquakes account for over 85% of natural disaster losses, urgently requiring groundbreaking theories and technological methods. Seismologists both domestically and internationally are dedicated to studying geological structures and crustal movements, hoping to reveal the causes of earthquakes and, with the help of advances in information science, developing advanced observation equipment. Based on this, a multidisciplinary observation system has been formed, encompassing seismometry, crustal deformation, electromagnetism, subsurface fluid dynamics, and auxiliary observations. Currently, earthquake monitoring and prediction mainly employ three methods: seismological methods, seismic statistical methods, and earthquake precursor methods.
[0003] Seismic geology methods predict potential earthquake zones by analyzing geological structures. Seismic statistics methods assess the risk of future earthquakes based on historical earthquake data. Precursor methods, on the other hand, focus on identifying macroscopic and microscopic signs preceding earthquakes, such as ground noise, unusual light patterns, seismic clouds, changes in groundwater levels, and fluctuations in geophysical fields. While these methods have made some progress in earthquake research, they largely rely on empirical inferences and are highly subjective.
[0004] Currently, existing earthquake prediction systems extract features from precursory signals such as crustal deformation, electromagnetic anomalies, and underground fluid anomalies. These extracted features are then simply stitched together and input into the prediction model to predict earthquakes. However, multiple precursory signals are of different types, resulting in heterogeneous features that are difficult to fuse directly, leading to low accuracy in earthquake prediction results from existing systems. Summary of the Invention
[0005] In view of the deficiencies of the existing technology, the present invention provides an earthquake prediction method, system, device and storage medium, which solves the existing problems.
[0006] The present invention adopts the following technical solution: In a first aspect, the present invention provides an earthquake prediction method, comprising the following steps: Real-time acquisition of multi-source earthquake precursor data for the target area; Feature extraction is performed on multi-source earthquake precursor data to obtain corresponding multi-dimensional heterogeneous features; and specific kernel functions are set for different types of heterogeneous features. Based on the proportion of different heterogeneous feature types in the current multidimensional heterogeneous features, a corresponding dedicated kernel function is determined. The current multidimensional heterogeneous features are then mapped to a high dimension using the dedicated kernel function to obtain a mapping vector. Based on the correlation between different multidimensional heterogeneous features and earthquake prediction targets, different weights are assigned to different earthquake precursor data. Multiple mapping vectors are linearly fused with multiple weights to obtain a preliminary fused vector. The preliminary fused vector is then dimensionality-reduced to obtain a fused feature vector. The fused feature vectors are input into the sub-models corresponding to different earthquake prediction targets to obtain the probability values of the corresponding prediction results. The probability values of different prediction results are weighted to obtain the final prediction result and the corresponding confidence level. Different earthquake prediction targets include the time, magnitude, and location of the earthquake. Preferably, the multi-source earthquake precursor data includes ultra-low frequency dynamic gravity anomaly disturbance data, groundwater level data, geomagnetic field change data, topographic deformation data, underground fluid chemical data, electromagnetic anomaly data, and seismic activity data.
[0007] Preferably, the heterogeneous feature types include frequency domain features, time-series features, and statistical features; The specific kernel function for the frequency domain features is as follows: K1(a,b)=exp(-||ab||² / (2σ²)); The specific kernel function for the time-series features is as follows: K2(a,b)=(a・b+c) d ; The specific kernel function for the statistical features is as follows: K3(a,b) = a・b + c; In the formula, K1 is the kernel function specific to the frequency domain features, K2 is the kernel function specific to the time series features, K3 is the kernel function specific to the statistical features, a and b are different feature vectors of the same domain feature type, σ is the kernel width hyperparameter of K1, c is the offset hyperparameter of K2 and K3, and d is the order hyperparameter of K2.
[0008] Preferably, different weights are assigned to different earthquake precursor data based on the correlation between different multidimensional heterogeneous characteristics and earthquake prediction objectives, specifically including the following steps: Different multidimensional heterogeneous features are standardized to obtain the corresponding feature vectors; Unify the dimensions of multiple feature vectors; The unified feature vectors are input into the attention weight network to obtain the corresponding weights. The weights are then optimized using a grid search method. The attention weight network employs a multilayer perceptron.
[0009] Preferably, the preliminary fusion vector is as follows: F1=w1・φ1+w2・φ2+w3・φ3; In the formula, F1 is the initial fusion vector, w1, w2 and w3 are the weights corresponding to different multidimensional heterogeneous features, and φ1, φ2 and φ3 are the mapping vectors corresponding to different multidimensional heterogeneous features.
[0010] Preferably, the sub-model is either a single model or a composite model, and different single models are combined to obtain the composite model; Weighting the probability values of different forecast results involves the following steps: Determine the AUC value for each sub-model; The weights are normalized by weighting multiple AUC values to obtain the corresponding weights; Multiply the weights by their corresponding probability values and sum them to obtain the overall confidence level of the final forecast result.
[0011] Preferably, the real-time acquisition of multi-source earthquake precursor data for the target area includes the following steps: Real-time acquisition of multi-source earthquake precursor data from different regions through APIs on earthquake monitoring websites and institutions; Multi-source earthquake precursor data from different regions are partitioned using a locality-sensitive hash function to obtain multiple subsets. A similarity metric is used to evaluate the data in the subsets, and based on the evaluation results, multi-source earthquake precursor data from the same region are stored in the same Omen. Adaptive grouping of multi-source earthquake precursor data from different sources yields a multidimensional heterogeneous database; Real-time acquisition of multi-source earthquake precursor data for the target area from a multi-dimensional heterogeneous database.
[0012] Secondly, the present invention provides an earthquake prediction system, comprising: The acquisition module is used to acquire multi-source earthquake precursor data of the target area in real time; The extraction module is used to extract features from multi-source earthquake precursor data to obtain corresponding multi-dimensional heterogeneous features; and to set specific kernel functions for different heterogeneous feature types. The fusion module is used to determine the corresponding dedicated kernel function based on the primary and secondary status of different heterogeneous feature types in the current multidimensional heterogeneous features. The dedicated kernel function is used to perform high-dimensional mapping of the current multidimensional heterogeneous features to obtain a mapping vector. Based on the correlation between different multidimensional heterogeneous features and earthquake prediction targets, different weights are assigned to different earthquake precursor data. Multiple mapping vectors are linearly fused with multiple weights to obtain a preliminary fusion vector. The preliminary fusion vector is then dimensionality-reduced to obtain a fusion feature vector. The prediction module is used to input the fused feature vector into the sub-models corresponding to different earthquake prediction targets to obtain the probability values of the corresponding prediction results; the probability values of different prediction results are weighted to obtain the final prediction result and the corresponding confidence level; different earthquake prediction targets include the time, magnitude and location of the earthquake.
[0013] Thirdly, 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, wherein the processor executes the program to implement the earthquake prediction method described above.
[0014] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described earthquake prediction method.
[0015] Compared with the prior art, the above-mentioned at least one technical solution adopted by the present invention can achieve the following beneficial effects: This invention maps different heterogeneous feature types to a high-dimensional regenerative kernel Hilbert space using a dedicated kernel function, ensuring that features can be directly linearly processed. Simultaneously, it quantifies the importance of different earthquake precursor data, dynamically adapting to the differences in importance among multi-source features. Finally, the mapped vector is weighted with its corresponding weights, achieving linear fusion of heterogeneous features. This solves the problem of directly integrating heterogeneous features and provides more valuable input for subsequent earthquake prediction, thus improving the accuracy of earthquake prediction. Attached Figure Description
[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0017] Figure 1 This is a flowchart of an earthquake prediction method according to the present invention; Figure 2 This is a diagram illustrating the data storage strategy based on locality-sensitive hash function-adaptive grouping of the present invention. Figure 3 This is a schematic diagram of the strong earthquake prediction model framework based on ULF-DG anomaly disturbances of the present invention. Detailed Implementation
[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0019] To improve the accuracy of short-term strong earthquake prediction, this invention proposes an earthquake prediction method. Based on ULF-DG precursor data and integrating multi-source data and artificial intelligence technology, a higher-precision strong earthquake prediction model is established. This leads to the creation of an intelligent strong earthquake prediction platform based on cross-modal multi-source data, comprising the following steps: Step 1: Real-time acquisition of multi-source earthquake precursor data for the target area Multi-source earthquake data includes ultra-extremely low frequency dynamic gravity (ULF-DG) precursor data, other earthquake precursor data, and seismic activity data.
[0020] ULF-DG precursor data represents minute perturbations in the Earth's gravitational field that occur before earthquakes. Closely related to the earthquake gestation and occurrence process, it holds promise as an effective earthquake precursor, providing a novel approach for short-term earthquake prediction. This signal typically operates in the millihertz (mHz) frequency range, possessing extremely long wavelengths (ranging from tens to thousands of kilometers) and strong penetrating power (reaching deep into the Earth's interior and seawater). It is suitable for research in exploring the Earth's internal structure, monitoring changes in the gravitational field, and earthquake prediction.
[0021] The application of ULF-DG earthquake precursor signals in earthquake prediction is still in its early stages, and related research is rarely reported in academic literature. This invention aims to utilize this signal to develop a new earthquake prediction method, with the goal of improving prediction accuracy.
[0022] Other precursory earthquake data include, but are not limited to, groundwater level data, geomagnetic field variation data, topographic deformation data, underground fluid chemical data, and electromagnetic anomaly data.
[0023] Groundwater level data includes real-time water level values, water level change rates, and cumulative changes; geomagnetic field change data includes geomagnetic field component data, component change rates, and monitoring environment auxiliary data; topographic deformation data includes ground rise and fall, horizontal displacement, and tilt changes (acquired through GPS, leveling, and inclinometers); underground fluid chemistry data includes changes in groundwater radon (Rn), mercury (Hg), and other trace elements, as well as water chemical ion concentrations; electromagnetic anomaly data includes abnormal fluctuations in the earth's electric field strength and polarization (supplementary electromagnetic precursors in addition to geomagnetic field changes).
[0024] Seismic activity data includes historical earthquake baseline data, specifically including the occurrence time, magnitude, epicenter latitude and longitude, and focal depth of recorded earthquakes within the global or target region.
[0025] This invention collects multi-source earthquake precursor data through an automated data collection process. The main purpose of this automated process is to ensure the real-time updating and integrity of the earthquake precursor data, reduce the tediousness and error rate of manual operations, and improve the efficiency and reliability of data collection. It includes:
[0026] (1) Development of Automated Scripts: Automated scripts were developed using the Python programming language, relying on its rich libraries for network requests and data processing, such as requests, BeautifulSoup, and Scrapy. These scripts were designed to automatically identify and extract earthquake precursor data from publicly available earthquake monitoring websites and databases, such as ultra-low frequency dynamic gravity anomalies, groundwater levels, and geomagnetic field changes. In this way, the latest earthquake precursor information can be continuously collected from multiple data sources.
[0027] (2) Data collection using APIs (Application Programming Interfaces): Many earthquake monitoring and research institutions provide data APIs, through which standardized data can be directly accessed and collected. Develop clients based on RESTful or SOAP (Simple Object Access Protocol) protocols to request and receive data according to the API documentation. Using APIs enables access to real-time data, ensuring the timeliness and accuracy of data collection.
[0028] (3) Logging and Error Handling: In the automated data collection process, logging and error handling mechanisms are introduced to ensure the stability and reliability of data collection. These mechanisms can capture and record abnormal situations in a timely manner during the data collection process, which facilitates problem tracking and timely repair, and ensures the continuity and high efficiency of data collection work.
[0029] This invention also constructs a database that supports multi-source heterogeneity and high real-time performance, aiming to provide a unified, efficient, and scalable storage and management platform for earthquake precursor data.
[0030] NoSQL (Not Only Structured Query Language) databases (such as MongoDB) or distributed database systems (such as Hadoop HBase) are used to store structured and unstructured data. This design allows data to be efficiently organized, indexed, and queried, supporting complex data analysis and processing needs. Seismic data can be clearly divided into structured and unstructured data based on data models, storage formats, and parsing difficulty.
[0031] Due to the sheer volume and complexity of the data, including structured and unstructured data stored in different system environments, a multi-source heterogeneous, real-time database with a hybrid architecture of Hadoop and Storm was designed to achieve data fusion.
[0032] The collected data is multi-source, heterogeneous, and real-time, requiring the data processing system to be diverse and real-time. Using the ArcPy interface, a data service bus algorithm based on multi-source heterogeneous data is designed to integrate web text data, ULF-DG data, and other precursory data stored in a unified Windows environment. This is used to build a unified platform database, enabling information sharing and interaction between different systems, breaking down communication barriers between heterogeneous data, improving system real-time performance, and completing the fusion of multi-source heterogeneous data within the system.
[0033] Reference Figure 2 When the aforementioned data is stored in a distributed manner, it is required that data from the same region be distributed within the same partition block to ensure data integrity and load balancing of the system data distribution. This invention proposes a novel data storage strategy: adaptive grouping using a locality-sensitive hash function (LSH). First, an LSH is proposed for data partitioning, projecting data objects with study area numbers into the same subset of data, resulting in data partitions on the OMen (Optical Method Interface). A similarity metric function is designed to evaluate the data partitioning results, ensuring that data from the same region resides within the same OMen, thus guaranteeing regional data integrity. Next, an adaptive grouping strategy is designed to ensure uniform distribution of the OMen stages, resulting in a Key-Value table including raster labels, standard power, and normalized amplitude. This minimizes the overall task runtime, ensuring data balance within the partitions, thereby improving cluster efficiency and system real-time performance. Based on this, a multidimensional heterogeneous database based on a hybrid Hadoop and Storm architecture is constructed.
[0034] This invention achieves real-time updates of earthquake precursor data through the development of automated scripts and APIs, and integrates acquired ULF-DG information in real time, focusing on solving the problems of efficiency and accuracy in data collection. Based on this, an innovative data service bus algorithm is designed to break down data barriers between different systems and achieve efficient communication of massive heterogeneous data; a locality-sensitive hashing (LSH) storage strategy is used to achieve efficient storage; and a multi-source, heterogeneous, real-time database based on a hybrid Hadoop and Storm architecture is constructed to support the storage, processing, and analysis of multi-source data.
[0035] Step 2: Extract features from multi-source earthquake precursor data to obtain corresponding multidimensional heterogeneous features. Data preprocessing and feature extraction are key steps in improving the accuracy and efficiency of earthquake prediction models. They can extract key information from raw data, reduce noise interference, and improve the model's learning efficiency.
[0036] Data preprocessing includes data cleaning and standardization. The purpose of data cleaning is to identify and handle outliers, missing values, and erroneous data to ensure data quality and consistency. By applying various data cleaning techniques, such as missing value imputation, outlier detection, and replacement strategies, the data is transformed into a uniform format. Simultaneously, data standardization ensures that data from different sources and with different dimensions can be appropriately scaled, eliminating the influence between data dimensions and improving the stability and predictive ability of the model.
[0037] Feature extraction aims to extract representative features from raw data, which are crucial for training forecasting models. Traditional feature engineering methods use techniques such as signal processing and statistical analysis to extract valuable features from earthquake precursor data, such as frequency domain features and time series statistical features. Feature selection strategies, such as random forests and principal component analysis (PCA), further filter out the subset of features that have the greatest impact on model forecasts, thereby reducing data dimensionality and improving model training efficiency and forecast accuracy. Although feature extraction based on deep learning technology has achieved great success, processing data using traditional feature engineering during the data processing stage remains important for improving the model's generalization ability and robustness, and can significantly enhance the performance of ensemble models.
[0038] Specifically, for any precursor data, the time-domain data is mapped to the frequency domain using frequency domain analysis techniques (such as Fourier transform, short-time Fourier transform, and wavelet transform), and frequency domain features used to characterize the frequency distribution of the data are extracted. These frequency domain features include power spectral density (PSD, unit: V² / Hz), dominant frequency components (frequency values where data energy is concentrated), harmonic distribution (number of frequency components that are integer multiples of the dominant frequency), frequency band energy ratio (percentage of energy in a certain frequency band to the total energy), and frequency offset (deviation between the measured dominant frequency and the historical average).
[0039] Meanwhile, time series analysis techniques (such as autoregressive models, autoregressive moving average models, autoregressive difference moving average models, Bayesian mutation point detection, and sliding window statistics) are used to extract time series features to capture the dynamic changes of data over time. These time series features include the slope of the time series trend (unit: data units / day), the intensity of periodicity (the proportion of seasonal variance to total variance), the autocorrelation coefficient (ACF value lagged by k days), the number of mutation points (the number of mutations within a certain time period), the magnitude of mutation (the difference between the data before and after the mutation point), and the stationarity index (ADF test statistic).
[0040] For single precursor data, in addition to frequency domain features and time series features, statistical features can also be extracted through statistical quantification algorithms. Statistical features include mean (the average level of data within a certain time period), kurtosis (the steepness of the data distribution; kurtosis > 3 indicates a peaked distribution, suggesting the existence of extreme values in the data), maximum amplitude (the difference between the maximum and minimum values of data within a certain time period), duration of extreme values (the continuous time for data to exceed a threshold), and rate standard deviation (the degree of fluctuation in the rate of rise / fall). For example, features such as "maximum fluctuation value of geomagnetic intensity on a single day" and "deviation of the mean geomagnetic value over three consecutive days" can be extracted from geomagnetic field change data.
[0041] Step 3: Fuse the features extracted from multi-source seismic data.
[0042] By integrating earthquake precursor data from different physical quantities, multi-source information can be used to improve the accuracy and robustness of forecasts.
[0043] A "two-layer fusion architecture" (feature-level fusion + decision-level fusion) is adopted to achieve efficient fusion of multi-source earthquake precursor data.
[0044] Feature-level fusion: At the feature level, feature vectors from different sources (such as ULF-DG data, geomagnetic field data, and groundwater level data) are spatially aligned, heterogeneously adapted, and weighted using technical means to form a unified high-dimensional feature matrix, thereby enhancing the model's understanding of the complex relationships of earthquake precursors.
[0045] Since different precursor data have different feature dimensions and units, the influence of units is first eliminated by feature standardization / normalization; then, feature vectors of different dimensions are adjusted to the same length by unifying feature dimensions and interpolation to ensure consistency of feature space.
[0046] This invention employs a strategy combining "attention-guided weighted fusion" and "Multiple Kernel Learning (MKL)" to achieve dynamic integration of features from multiple sources. Attention-guided weighted fusion: An attention weight network (i.e., a multilayer perceptron) is constructed, and dynamic weights are assigned to features from different sources based on the "correlation between features and earthquake prediction targets." To objectively determine the weights, a grid search method is used for parameter tuning. MKL: For heterogeneous features (such as numerical features and statistical features), different kernel functions are designed (e.g., Gaussian kernels for continuous features and polynomial kernels for discrete features). The kernel functions map the features to a high-dimensional reproducing kernel Hilbert space (RKHS) before linear fusion.
[0047] The high-dimensional feature matrix after fusion is reduced in dimensionality by feature reduction techniques (such as principal component analysis, t-distributed random neighborhood embedding, and autoencoder), redundant information is removed, core discriminative features are retained, and the final "fusion feature vector" is formed for subsequent model training.
[0048] This embodiment uses ULF-DG data (precursor 1), groundwater level data (precursor 2), and geomagnetic field change data (precursor 3) as examples.
[0049] ULF-DG data has 5-dimensional features, including power spectral density, dominant frequency component, time series trend slope, autocorrelation coefficient, and mean; groundwater level data has 5-dimensional features, including frequency band energy proportion, frequency offset, number of abrupt change points, maximum amplitude, and rate standard deviation; geomagnetic field variation data has 5-dimensional features, including harmonic distribution, stationarity index (ADF value), abrupt change amplitude, kurtosis, and duration of extreme values.
[0050] The fusion process follows the feature-level fusion logic in the "two-layer fusion architecture," as detailed below: (1) Standardization / Normalization to Eliminate Dimensions: Due to the differences in the characteristic dimensions of the three precursors (e.g., power spectral density in V² / Hz, time-series trend slope in mm / day, and mean in nT), Z-score standardization was used: the mean μ and standard deviation σ were calculated for each feature dimension, and the original feature value x was converted to x'=(x-μ) / σ. After processing, the 5-dimensional features of the three precursors were all converted into dimensionless data with a mean of 0 and a standard deviation of 1, ensuring data comparability.
[0051] (2) Dimensional unification and interpolation: By unifying the feature dimensions and interpolating, the feature vectors of different dimensions are adjusted to the same length. In this example, the three precursors are all 5-dimensional, and the dimensions are consistent, so no interpolation is needed. If there is a precursor with 3-dimensional features, it is necessary to complete it to 5-dimensional by linear interpolation or cubic spline interpolation (such as completing the missing dimensions based on the continuity of the feature time series) to ensure that the feature space structure is consistent.
[0052] (3) Weighted fusion guided by attention mechanism: A multilayer perceptron is constructed as the attention weight network. The input is the standardized feature vectors of three precursors, and the output is three weight values w1, w2, and w3 (satisfying w1+w2+w3=1). The weights are dynamically allocated based on the correlation between the features and the earthquake prediction target. Example output: Assuming that ULF-DG data has the highest correlation with strong earthquakes, the weight w1=0.45; geomagnetic field data is second, w3=0.35; and groundwater level data w2=0.2. The weights are optimized using a grid search method.
[0053] (4) Multi-kernel learning (MKL) for handling heterogeneous features: The 5-dimensional features of the three precursors are heterogeneous (precursor 1 is mainly frequency domain features, precursor 2 is mainly time series features, and precursor 3 is mainly statistical features). Dedicated kernel functions are designed for different types of features: Frequency domain features (such as power spectral density and frequency band energy ratio): use the Gaussian kernel function K1(a,b)=exp(-||ab||² / (2σ²)).
[0054] Time-series features (such as trend slope, number of abrupt change points): using the polynomial kernel function K2(a,b)=(a・b+c). d .
[0055] Statistical features (such as mean and kurtosis): use the linear kernel function K3(a,b)=a・b+c.
[0056] In the formula, K1 is the kernel function specific to the frequency domain features, K2 is the kernel function specific to the time series features, K3 is the kernel function specific to the statistical features, a and b are different feature vectors of the same domain feature type, σ is the kernel width hyperparameter of K1, c is the offset hyperparameter of K2 and K3, and d is the order hyperparameter of K2.
[0057] Each precursor's 5-dimensional feature vector is mapped to a high-dimensional regenerative kernel Hilbert space (RKHS) using a dedicated kernel function, resulting in mapped vectors φ1, φ2, and φ3 (with dimensions higher than 5, determined by the kernel function parameters).
[0058] (5) Linear fusion and dimensionality reduction: The attention weights and the mapped high-dimensional feature vectors are linearly combined (the weights determine the contribution of each source feature) to obtain the initial fusion vector: F1=w1・φ1+w2・φ2+w3・φ3. Then, the dimensionality of F1 is initially reduced by principal component analysis (PCA) to remove redundant information, and finally a "fusion feature vector F" with a dimension of 3~5 (if the first 3 principal components are retained, the cumulative variance contribution rate is ≥95%), which is used for subsequent model training.
[0059] Step 4: Input the fused feature vectors into the sub-models corresponding to different earthquake prediction targets to obtain the probability values of the corresponding prediction results; weight the probability values of different prediction results to obtain the final prediction result and the corresponding confidence level.
[0060] Input a single model: If the fused feature vectors emphasize a certain type of feature (e.g., primarily frequency domain features), then input a single model that excels at that feature for training, for example: Frequency-domain dominant fusion features → convolutional neural network, extracting local frequency domain correlation features; Temporally dominant fusion features → Long Short-Term Memory Network, capturing long-term dependencies.
[0061] The trained single model serves as a "basic support module," providing core algorithm units for the composite model.
[0062] Direct input to the composite model: If the fused feature vector contains multiple types of features such as frequency domain, time series, and statistics (such as the "dual-layer fusion" result in the patent), then it can be directly input into the composite model, leveraging the complementary advantages of multiple models. For example: The model is trained by fusing feature vectors into a CNN module (extracting local features in the frequency domain), a BiLSTM module (capturing temporal dependencies), and an Attention layer (enhancing key features). This process is then used as an "optimized and upgraded model".
[0063] Final destination: The trained single model / composite model serves as the core architecture of the "time prediction sub-model", "magnitude prediction sub-model" and "location prediction sub-model" respectively, and is used for subsequent decision-making integration (such as stacking and weighted averaging) to finally output the three-element earthquake prediction results.
[0064] In the development of a strong earthquake intelligent prediction system, model training and the application of ensemble learning methods are key to improving prediction accuracy. The goal of this stage is to improve the overall prediction accuracy and robustness by combining the training of a single precursor model with ensemble strategies, integrating the advantages of different models in handling specific types of data.
[0065] This method employs ensemble learning techniques (such as random forests and gradient boosting trees) to synthesize forecasts from different models, improving the accuracy and reliability of the final forecast through voting or weighted averaging. This results in a final earthquake forecast for the three key elements (time, magnitude, and location), fundamentally addressing the problem of insufficient generalization ability and large prediction bias inherent in single models.
[0066] Since different models have different prediction output formats (such as some models outputting the probability distribution of magnitude, and some models outputting the point prediction value of magnitude), the results are first standardized to unify them into "probability vectors" or "quantitative indicators": for example, the point prediction value of magnitude is converted into "probability distribution of magnitude range (such as 4-5, 5-6)," and the latitude and longitude of the location prediction is converted into "probability heat map of spatial grid."
[0067] The principles of single and composite models are shown in Table 1.
[0068] Table 1. Principles of Single and Composite Models The three-element prediction model is a collection of three specific sub-models: "time, magnitude, and location." Each sub-model is built upon a single model or a composite model, and the specific relationships are as follows: (1) Time prediction sub-model: Based on the construction of "single model (LSTM, GRU, Transformer)," the core logic is to capture the temporal variation pattern of precursor data through LSTM / GRU, strengthen the correlation between "time difference and earthquake probability" through the self-attention mechanism of Transformer, and finally output the "model of the change of cumulative probability of earthquake occurrence with time difference".
[0069] (2) Magnitude prediction sub-model: Based on the construction of "composite model (CNN) + single model (SVR-ER)," the frequency domain features (such as power spectral density and frequency band energy) of ULF-DG data are first extracted by CNN, and then the small sample features of "signal duration-amplitude" are captured by support vector regression (SVR). The quantitative relationship of "feature-energy-magnitude" is established by combining the exponential regression model (ER), and finally the magnitude interval prediction is output.
[0070] (3) Location prediction sub-model: Based on the "composite model (Transformer+Attention)", the "subtle feature differences" (such as main frequency offset and amplitude fluctuation) of ULF-DG data obtained from observation stations in different geographical locations are analyzed by Transformer. The correlation between "feature differences-epicenter location / source depth" is quantified by combining attention mechanism, and finally the predicted values of epicenter latitude and longitude and source depth are output.
[0071] Calculate the AUC (Area Under Curve) value (model performance index, ranging from 0 to 1, the closer to 1 the better the performance) of each sub-model: Time sub-model AUC1=0.9 (predicting "probability of occurrence in the next 7 days 0.8"); Magnitude sub-model AUC2=0.8 (predicting "probability of magnitude 5-6 0.7"); Location sub-model AUC3=0.85 (predicting "probability of occurrence at 30°N, 105°E 0.9").
[0072] Weight calculation formula: c i =AUC i / (AUC1+AUC2+AUC3); Substituting the values, we get: c1 = 0.9 / (0.9 + 0.8 + 0.85) = 0.9 / 2.55 ≈ 0.346; c2 = 0.8 / 2.55 ≈ 0.308; c3 = 0.85 / 2.55 ≈ 0.346.
[0073] The three elements work together to produce the final forecast result, which is not a single value, but a comprehensive probability combination of "time-magnitude-location". The format is: "Next 7 days (probability 0.8×0.346≈0.277 weight contribution), magnitude 5-6 (probability 0.7×0.308≈0.216 weight contribution), epicenter located at 30°±0.5°N, 105°±0.5°E (probability 0.9×0.346≈0.311 weight contribution)", and the comprehensive confidence score is output as 0.277+0.216+0.311≈0.804.
[0074] For different earthquake precursors, models suitable for the characteristics of earthquake data are selected and customized to capture complex patterns in the precursor data. Earthquake prediction models aim to analyze the Earth's physical processes and predict the time, location, and intensity of earthquakes by identifying patterns and signals that may foreshadow them. With the support of big data and high-performance computing technologies, significant progress has been made in the field of deep learning models for earthquake prediction. Applicable deep learning models such as CNN, LSTM, GRU, and Transformer have demonstrated excellent performance in various aspects. Selectable single models include Convolutional Neural Networks (CNN), Long Short-Term Memory Networks (LSTM), Gated Recurrent Units (GRU), Bidirectional Long Short-Term Memory Networks (BiLSTM), Transformer, and Multi-Task Learning (MTL).
[0075] In complex problems like earthquake prediction, it's necessary to comprehensively consider multiple models to overcome the limitations of each individual model in processing sequential data. Suitable composite models for this study include CNN-LSTM, CNN-GRU, CNN-BiLSTM, CNN-BiGRU, CNN-LSTM-Attention, CNN-GRU-Attention, CNN-BiLSTM-Attention, CNN-BiGRU-Attention, TCN-Attention, and ARIMA-CNN-Transformer. The selection of composite models requires subsequent model evaluation to find the optimal model combination. Taking the CNN-BiLSTM-Attention composite model as an example, this model combines a convolutional neural network (CNN) and a long short-term memory network (LSTM), offering advantages such as local and global feature capture, spatiotemporal relationship modeling, hierarchical feature extraction, and self-attention mechanisms.
[0076] For example, prediction models based on ULF-DG anomalies: By utilizing high-precision ULF-DG anomaly data and fully leveraging machine learning and deep learning technologies, a three-element prediction model for strong earthquakes based on ULF-DG anomaly information was created, such as... Figure 3 As shown. Regarding time prediction, LSTM, GRU, and Transformer were used to study and reveal the signal variation law and time cumulative probability of earthquake occurrence as the time difference increases, developing a time difference-based earthquake occurrence probability (P) prediction model. Regarding magnitude prediction, a quantitative relationship between "ULF-DG anomalous signal characteristics and their energy" and magnitude was established based on CNN research. Features include duration and amplitude within the interval. A support vector regression (SVR) model was used to capture small-sample features such as duration and amplitude within the interval in the ULF-DG signal. Combined with a complex exponential regression (ER) model between magnitude and energy, a magnitude prediction model based on ULF-DG anomalous signal characteristics and their energy was established. Regarding epicenter location prediction, through in-depth analysis of the subtle differences in the high-precision ULF-DG signals captured by ULF-DG acquisition instruments at different geographical locations, a quantitative relationship between epicenter location, focal depth, and subtle differences in ULF-DG signal characteristics was discovered and revealed based on Transformer, thereby constructing a strong earthquake epicenter location prediction model. A swarm intelligence optimization algorithm is used to adaptively optimize the hidden layer nodes, network parameters, and learning rate in various models, thereby improving the timeliness, robustness, and generalization ability of model predictions.
[0077] Voting and Stacking: A simple voting mechanism and a stacking ensemble method are used to combine the forecasts of different models. The voting mechanism makes the final decision based on the forecasts of the majority of models, while the stacking ensemble learns how to optimally combine the forecasts of various models by training a new model.
[0078] Model fusion techniques: These techniques (such as Boosting and Bagging) are used to improve the performance of a single model and reduce forecast bias and variance by combining multiple models.
[0079] By comprehensively applying ensemble learning and model fusion technologies, the strong earthquake intelligent prediction system can effectively improve the accuracy and reliability of predictions, providing strong technical support for earthquake early warning.
[0080] Step 5: Properly splitting the training and test datasets is crucial for evaluating the model's generalization ability on unseen data. This helps avoid overfitting and ensures the model's effectiveness in real-world applications. Random partitioning is typically used to ensure dataset representativeness. A common practice is to divide the dataset into 70% training data and 30% test data, or adjust the ratio based on specific circumstances. Furthermore, stratified sampling can be used to maintain a consistent proportion of data across classes in both the training and test sets, thereby improving the model's generalization ability.
[0081] Cross-validation is a more robust model evaluation method that aims to reduce the impact of the training and test set splitting on the model evaluation results, thereby obtaining more reliable model performance estimates. The most commonly used cross-validation method is k-fold cross-validation, where the dataset is divided into k subsets. Each time, one subset is set as the test set, and the rest are used as the training set. This process is repeated k times, each time selecting a different test set, and the average of the k test results is calculated as the model performance estimate. This method is particularly suitable for situations where the amount of data is not particularly large.
[0082] Model performance evaluation aims to quantify a model's predictive ability using specific metrics. This is crucial for comparing the performance of different models, selecting the best model, and guiding further model optimization. Performance evaluation metrics include accuracy, recall, F1 score, and area under the ROC curve (AUC). The choice of metric depends on the specific requirements of the forecasting task.
[0083] Model optimization and parameter tuning aim to further improve the model's predictive ability and efficiency, ensuring optimal performance in practical applications. For traditional machine learning algorithms, optimization and tuning often employ methods such as grid search and random search to systematically traverse the model parameter space and find the optimal model configuration. For deep learning-based artificial intelligence models, training often uses stochastic gradient descent or other gradient-based optimizers, such as Adagrad and Adam, to optimize model parameters. Simultaneously, cross-validation techniques are used to evaluate the model's performance on different subsets of data, ensuring the model's generalization ability.
[0084] At this stage, the goal is to deploy the trained model to the production environment and establish mechanisms to support continuous optimization and iteration of the model based on new data.
[0085] Model deployment.
[0086] The model can be deployed on local servers or cloud platforms. Containerization technologies such as Docker can improve deployment flexibility and portability. To ensure high availability and scalability, container orchestration tools such as Kubernetes can be used to manage containerized model services. Simultaneously, appropriate interfaces, such as REST APIs, need to be established so that applications can easily interact with the model.
[0087] Real-time data monitoring and model updates.
[0088] Monitoring real-time data and updating the model based on new data is crucial for ensuring the long-term effectiveness and adaptability of the model. Implementing a data monitoring mechanism to periodically evaluate model performance or based on specific trigger conditions, and automatically triggering model retraining when performance deteriorates, requires establishing an effective data pipeline to ensure that newly collected data can be used for model training in a timely manner. Furthermore, adopting a microservice architecture can facilitate seamless model updates and replacements.
[0089] This invention addresses technical problems at three levels: the data layer, the algorithm layer, and the application layer. First, at the data level: based on ULF-DG anomaly information, and on the basis of ULF-DG anomaly information, fully integrate other modal earthquake precursor information to build and form a multi-source, heterogeneous, real-time updated, and integrated earthquake precursor big data database.
[0090] Second, at the algorithm level: Based on ultra-low frequency dynamic gravity anomaly data, combined with multi-source and cross-modal precursor information, multi-modal features are extracted and dynamically fused based on deep learning technology to overcome the inherent long-tail distribution characteristics of earthquake precursor databases. By combining the advantages of traditional machine learning models and deep learning models through ensemble learning, a high-precision prediction probability model for the occurrence time, magnitude, and epicenter location of strong earthquakes is established.
[0091] Thirdly, at the application level: how to simulate and improve extreme situations such as missing source data based on the integrated earthquake precursor big data and cross-modal earthquake prediction model, test and optimize indicators such as model calibration and uncertainty estimation, and establish a hybrid intelligent optimization strong earthquake short-term prediction system.
[0092] Based on the same concept, the present invention also provides an earthquake prediction system, including an acquisition module, an extraction module, a fusion module and a prediction module.
[0093] The acquisition module is used to acquire multi-source earthquake precursor data of the target area in real time.
[0094] The extraction module is used to extract features from multi-source earthquake precursor data to obtain corresponding multi-dimensional heterogeneous features; and sets exclusive kernel functions for different heterogeneous feature types.
[0095] The fusion module is used to determine the corresponding dedicated kernel function based on the primary and secondary status of different heterogeneous feature types in the current multidimensional heterogeneous features. The dedicated kernel function is used to perform high-dimensional mapping of the current multidimensional heterogeneous features to obtain a mapping vector. Based on the correlation between different multidimensional heterogeneous features and earthquake prediction targets, different weights are assigned to different earthquake precursor data. Multiple mapping vectors are linearly fused with multiple weights to obtain a preliminary fusion vector. The preliminary fusion vector is then dimensionality-reduced to obtain a fusion feature vector.
[0096] The prediction module is used to input the fused feature vector into the sub-models corresponding to different earthquake prediction targets to obtain the probability values of the corresponding prediction results; the probability values of different prediction results are weighted to obtain the final prediction result and the corresponding confidence level; different earthquake prediction targets include the time, magnitude and location of the earthquake.
[0097] Based on the same concept, the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described earthquake prediction method.
[0098] Based on the same concept, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described earthquake prediction method.
[0099] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.
[0100] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.
Claims
1. An earthquake prediction method, characterized in that, Includes the following steps: Real-time acquisition of multi-source earthquake precursor data for the target area; Feature extraction is performed on each earthquake precursor data to obtain corresponding multidimensional heterogeneous features; And specific kernel functions are set for different heterogeneous feature types; Based on the proportion of different heterogeneous feature types in the current multidimensional heterogeneous features, a corresponding dedicated kernel function is determined. The current multidimensional heterogeneous features are then mapped to a high dimension using the dedicated kernel function to obtain a mapping vector. Based on the correlation between different multidimensional heterogeneous features and earthquake prediction targets, different weights are assigned to different earthquake precursor data. Multiple mapping vectors are linearly fused with multiple weights to obtain a preliminary fused vector. The preliminary fused vector is then dimensionality-reduced to obtain a fused feature vector. The fused feature vectors are input into the sub-models corresponding to different earthquake prediction targets to obtain the probability values of the corresponding prediction results; the probability values of different prediction results are weighted to obtain the final prediction result and the corresponding confidence level. Different earthquake prediction targets include the time, magnitude, and location of the earthquake.
2. The earthquake prediction method as described in claim 1, characterized in that, The multi-source earthquake precursor data includes ultra-low frequency dynamic gravity anomaly disturbance data, groundwater level data, geomagnetic field change data, topographic deformation data, underground fluid chemical data, electromagnetic anomaly data, and seismic activity data.
3. The earthquake prediction method as described in claim 1, characterized in that, The heterogeneous feature types include frequency domain features, time series features, and statistical features; The specific kernel function for the frequency domain features is as follows: K1(a,b)=exp(-||ab||² / (2σ²)); The specific kernel function for the time-series features is as follows: K2(a,b)=(a・b+c) d ; The specific kernel function for the statistical features is as follows: K3(a,b) = a・b + c; In the formula, K1 is the kernel function specific to the frequency domain features, K2 is the kernel function specific to the time series features, K3 is the kernel function specific to the statistical features, a and b are different feature vectors of the same domain feature type, σ is the kernel width hyperparameter of K1, c is the offset hyperparameter of K2 and K3, and d is the order hyperparameter of K2.
4. The earthquake prediction method as described in claim 1, characterized in that, Different weights are assigned to different earthquake precursor data based on the correlation between different multidimensional heterogeneous features and earthquake prediction objectives. The specific steps include: Different multidimensional heterogeneous features are standardized to obtain the corresponding feature vectors; Unify the dimensions of multiple feature vectors; The unified feature vectors are input into the attention weight network to obtain the corresponding weights. The weights are then optimized using a grid search method. The attention weight network employs a multilayer perceptron.
5. The earthquake prediction method as described in claim 1, characterized in that, The initial fusion vector is shown below: F1=w1・φ1+w2・φ2+w3・φ3; In the formula, F1 is the initial fusion vector, w1, w2 and w3 are the weights corresponding to different multidimensional heterogeneous features, and φ1, φ2 and φ3 are the mapping vectors corresponding to different multidimensional heterogeneous features.
6. The earthquake prediction method as described in claim 1, characterized in that, The sub-model is either a single model or a composite model, and different single models are combined to obtain the composite model. Weighting the probability values of different forecast results involves the following steps: Determine the AUC value for each sub-model; The weights are normalized by weighting multiple AUC values to obtain the corresponding weights; Multiply the weights by their corresponding probability values and sum them to obtain the overall confidence level of the final forecast result.
7. The earthquake prediction method as described in claim 1, characterized in that, The real-time acquisition of multi-source earthquake precursor data in the target area includes the following steps: Real-time acquisition of multi-source earthquake precursor data from different regions through APIs on earthquake monitoring websites and institutions; Multi-source earthquake precursor data from different regions are partitioned using a locality-sensitive hash function to obtain multiple subsets. A similarity metric is used to evaluate the data in the subsets, and based on the evaluation results, multi-source earthquake precursor data from the same region are stored in the same Omen. Adaptive grouping of multi-source earthquake precursor data from different sources yields a multidimensional heterogeneous database; Real-time acquisition of multi-source earthquake precursor data for the target area from a multi-dimensional heterogeneous database.
8. An earthquake prediction system, characterized in that, include: The acquisition module is used to acquire multi-source earthquake precursor data of the target area in real time; The extraction module is used to extract features from multi-source earthquake precursor data to obtain corresponding multi-dimensional heterogeneous features. And specific kernel functions are set for different heterogeneous feature types; The fusion module is used to determine the corresponding dedicated kernel function based on the primary and secondary status of different heterogeneous feature types in the current multidimensional heterogeneous features. The dedicated kernel function is used to perform high-dimensional mapping of the current multidimensional heterogeneous features to obtain a mapping vector. Based on the correlation between different multidimensional heterogeneous features and earthquake prediction targets, different weights are assigned to different earthquake precursor data. Multiple mapping vectors are linearly fused with multiple weights to obtain a preliminary fusion vector. The preliminary fusion vector is then dimensionality-reduced to obtain a fusion feature vector. The prediction module is used to input the fused feature vector into the sub-models corresponding to different earthquake prediction targets to obtain the probability values of the corresponding prediction results; the probability values of different prediction results are weighted to obtain the final prediction result and the corresponding confidence level; Different earthquake prediction targets include the time, magnitude, and location of the earthquake.
9. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the earthquake prediction method according to any one of claims 1-7.
10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the earthquake prediction method according to any one of claims 1-7.