Maximum aftershock prediction method based on online enhancement
By using online learning and data augmentation techniques, a personalized maximum aftershock prediction model is dynamically generated, overcoming the limitations of static models in existing technologies. This achieves high-precision and stable aftershock prediction, supporting rapid emergency response.
Patent Information
- Application Number
- CN202511635100.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-10
- Publication Date
- 2026-02-10
AI Technical Summary
Existing maximum aftershock prediction methods are static and cannot be adaptive or self-learning, resulting in low prediction accuracy and unreliability, especially when early data is incomplete and prediction errors are large.
Employing online learning and data/model enhancement mechanisms, personalized prediction models are dynamically generated through real-time data triggering, similar historical sequence retrieval, data enhancement, and model fine-tuning. These models utilize similar historical data and noise to simulate diverse scenarios and predict maximum aftershocks.
It achieves personalized and scenario-based prediction results, improves prediction accuracy and stability, reduces the risk of overfitting, enhances the model's generalization ability and real-time performance, and supports rapid emergency response within minutes.
Smart Images
Figure CN121500380A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of earthquake prediction technology, and in particular to a method for real-time prediction of the maximum aftershock following the main shock using machine learning and online learning techniques. Background Technology
[0002] Predicting the maximum aftershock is a crucial aspect of post-earthquake emergency response and disaster assessment. Following a strong mainshock, the magnitude of the maximum aftershock directly determines the severity of secondary disasters, significantly impacting the deployment of rescue forces, the inspection of damaged buildings, the resettlement of affected residents, and the stability of public morale. However, due to the extreme complexity of earthquake physics, accurately predicting the maximum aftershock remains a global challenge.
[0003] Currently, existing methods for predicting maximum aftershocks mainly rely on the following two technical approaches:
[0004] 1. Methods based on classical empirical statistical models
[0005] These methods originate from statistical analysis of historical earthquake sequences, establishing simple empirical formulas or statistical regularities.
[0006] Båth's Law: This is the most widely known rule of thumb, stating that the magnitude of the largest aftershock is usually about 1.0-1.2 magnitude units smaller than the mainshock. While this method is simple and intuitive, its shortcomings are quite obvious:
[0007] Poor accuracy: "1.2" is a global average, while actual earthquake sequences vary greatly under different tectonic environments. This "one-size-fits-all" prediction method ignores the specificity of earthquake sequences, resulting in a large range of prediction errors and low reliability.
[0008] Static invariance: This law is a static formula; once the mainshock occurs, its prediction results are fixed. It cannot utilize the continuous aftershock observation data generated after the mainshock to dynamically correct and optimize its prediction results, thus losing the possibility of improvement through data learning.
[0009] The probabilistic model based on the Gutenberg-Richter (GR) law: This method calculates the b-value (the ratio of large to small earthquakes) from the sequence of aftershocks that have occurred, and then estimates the probability of a future aftershock of a certain magnitude or higher.
[0010] Key drawbacks: Its predictions heavily rely on the completeness and quality of recorded aftershocks. In the early stages of the sequence, incomplete records of small-magnitude aftershocks (limitations of seismic network monitoring capabilities) lead to severe distortions in b-value estimations, making early probability predictions extremely unreliable. Similarly, it is typically used as a static model, failing to achieve real-time dynamic calibration.
[0011] Omori's law describes the decay of aftershock frequency over time. While it does not directly predict the maximum magnitude, it is often used as a reference for sequence evolution. It has indirectness and limitations; it cannot be used alone to predict the maximum aftershock magnitude, and its parameters (p-values) vary across different sequences, requiring combination with other methods. Furthermore, it also faces the problem of inaccurate parameter estimation due to the instability of early data.
[0012] 2. Methods based on static machine learning models
[0013] With the development of artificial intelligence technology, some studies have attempted to use machine learning models (such as support vector machines (SVM), random forests, etc.) to learn and predict the maximum aftershocks from historical earthquake data.
[0014] Basic process: Collect a large amount of historical earthquake sequence data, extract features (such as mainshock magnitude, location, etc.), and train a fixed prediction model.
[0015] Its inherent drawback is model rigidity; once the model is trained on historical data, its parameters are fixed. This means that the model cannot adaptively adjust to the unique behavior of the current sequence, and its predictions are "unchanging."
[0016] Summary of technical issues:
[0017] In summary, the core problem with existing technologies, whether classical statistical models or static machine learning models, lies in their "staticity." They cannot build a predictive system that can adapt, learn, and optimize itself, resulting in technical bottlenecks such as limited prediction accuracy and low reliability in the task of predicting maximum aftershocks. Summary of the Invention
[0018] The purpose of this invention is to overcome the shortcomings of the prior art and provide a prediction model that can be dynamically optimized, thereby improving the accuracy and timeliness of maximum aftershock prediction.
[0019] To achieve the above objectives, this invention provides the following technical solution: a maximum aftershock prediction method based on online enhancement, the core of which lies in the integration of two major mechanisms: "online learning" and "data / model enhancement." The specific steps are as follows:
[0020] Step S1: Data Preparation
[0021] Data sources include global earthquake catalogs, such as the USGS / ANSS Integrated Catalog of the U.S. Geological Survey, the ISC Catalog of the International Seismological Centre, and the CENC Catalog of the China Earthquake Networks Center. Preferred catalogs are those with a unified magnitude scale (e.g., moment magnitude, Mw) and have undergone recalibration. For specific high-risk areas, higher-precision regional seismic network data can be supplemented to obtain a more complete record of minor earthquakes and more accurate location results.
[0022] Definition of mainshock: Within a specific spatiotemporal window, the earthquake event with the largest magnitude is defined as the mainshock.
[0023] Aftershock association rules: Using spatiotemporal window methods (such as the Gardener & Knopoff algorithm) or more advanced structure-based background earthquake swarm association methods, earthquake events that occur within the source region of the mainshock after the mainshock are associated as aftershocks.
[0024] Extracting mainshock features, including but not limited to:
[0025] Mainshock Magnitude The moment magnitude of the main shock;
[0026] Mainshock depth The focal depth of the main shock;
[0027] Solution of the focal mechanism of the mainshock: fault plane solution (strike, dip, slip angle), seismic moment tensor, etc.;
[0028] Tectonic environment characteristics: distance between the mainshock and known active fault zones, long-term average b-value of the region, crustal type (continental / oceanic), etc.
[0029] Step S2: Construct the initial model
[0030] Using historical earthquake data and their corresponding, ultimately observed maximum aftershock magnitudes as labels, an initial maximum aftershock prediction model is trained. Considering the limited number of similar historical earthquake sequences available, this model is preferably a machine learning model suitable for small-sample learning, such as Support Vector Machine (SVM), Random Forest, or the advanced TabPFN (Tabular Prior-Data Fitted Network). This initial model is pre-trained and stored as a general, basic predictor.
[0031] Step S3: Online Enhanced Prediction
[0032] This is the core of the invention. When a new mainshock occurs, the system no longer directly uses the static initial model, but instead initiates the following online enhancement process to dynamically generate a "personalized" prediction model for that specific mainshock:
[0033] S3.1: Real-time data triggering: Whenever a new mainshock event is recorded, the system immediately and automatically triggers the subsequent model update process.
[0034] S3.2: Online Data Selection: For a newly occurring mainshock, the system starts from its feature vector (such as magnitude, depth, focal mechanism, tectonic environment, etc.) and retrieves the K most similar historical mainshock sequences from the historical database. Similarity metrics can be Euclidean distance, cosine similarity, or a domain-knowledge-based weighted distance function. The purpose of this step is to find an "analogous sample set" for the new mainshock, rather than using all historical data.
[0035] S3.3: Online Data Augmentation: Data augmentation techniques are applied to the K similar historical data sequences selected in S3.2 to construct a larger and more diverse "virtual training set," thereby simulating various scenarios that may occur under this type of earthquake environment. Augmentation methods include:
[0036] Add random noise: Add Gaussian noise or uniform noise that conforms to geophysical laws to the mainshock characteristics (such as depth and magnitude) and the maximum aftershock label.
[0037] Feature space interpolation / extrapolation: Performs linear or nonlinear interpolation between selected similar sequences to generate synthetic sequence data with transitional features.
[0038] SMOTE oversampling: For small sample classes, a synthetic minority class oversampling technique is used to generate new synthetic samples.
[0039] S3.4: Model Fine-tuning Training: Using the enhanced training dataset obtained in S3.3, the initial model pre-trained in step S2 is rapidly fine-tuned. This process employs incremental learning or transfer learning strategies to update the model weights with a low learning rate, enabling it to quickly adapt to the specific type of the new main shock while retaining general knowledge.
[0040] S3.5: Dynamic Prediction: Input the feature vector of the new mainshock into the finely tuned "personalized" model to predict the magnitude of the maximum possible aftershock in the future, and output the prediction results and uncertainty range.
[0041] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0042] It achieves truly personalized and scenario-based predictions: overcoming the limitations of the traditional single model's "one-size-fits-all" approach, by dynamically matching similar historical cases with each new mainshock and fine-tuning the model, the prediction results are more in line with the specific geological and focal background of the sequence, significantly improving the physical rationality and accuracy of the predictions.
[0043] It solves the prediction challenge under small sample conditions: through the "retrieval-enhancement" mechanism, it effectively amplifies the limited but high-value similar sequence information, enabling the model to obtain robust training even when data is scarce, and greatly improving the performance and stability of the prediction model during the critical early warning window period.
[0044] Enhanced model generalization and robustness: Data augmentation techniques introduce reasonable noise and diversity, simulate unobserved scenarios, make the fine-tuned model insensitive to data uncertainty, reduce the risk of overfitting, and improve its adaptability to complex real-world situations.
[0045] The system has improved the automation and real-time performance of the prediction system: from the triggering of the main shock to the generation of personalized predictions, the entire process can be completed within minutes without human intervention, providing unprecedented rapid and scientific support for emergency response decision-making. Attached Figure Description
[0046] Figure 1 This is a flowchart illustrating the overall process structure of the method of the present invention. Detailed Implementation
[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0048] General Overview
[0049] This invention aims to provide a method for rapidly generating "personalized" maximum aftershock prediction models for newly occurring mainshock events. Its core idea is to abandon single, static prediction models and instead adopt a dynamic prediction paradigm based on case analogy and online enhancement. The specific implementation steps of this invention are illustrated below using a hypothetical newly occurring mainshock (denoted as M_new) as an example.
[0050] Specific implementation steps
[0051] Step 1: Data Preparation and Initial Model Building (Offline Phase)
[0052] Data collection and processing:
[0053] Download historical earthquake data from 1980 to the present, with a magnitude (Mw) greater than or equal to 5.0, from the ANSS Comprehensive Catalog of the U.S. Geological Survey (USGS).
[0054] The classic spatiotemporal window method proposed by Gardner and Knopoff is used to decluster the earthquake sequences, identifying each mainshock and its corresponding aftershock sequence. For example, for a magnitude 6.0 mainshock, the spatial window radius can be set to 20 kilometers, and the time window length can be set to 365 days.
[0055] For each successfully identified mainse sequence, the following features are extracted to form a feature vector:
[0056] M_main: Main shock moment and magnitude.
[0057] Depth_main: Mainshock depth (km).
[0058] Strike, Dip, Rake: Fault plane solutions (degrees) of the mainshock.
[0059] Dist_to_Fault: Distance (in kilometers) between the mainshock and the nearest known active fault zone.
[0060] The label M_max_aftershock for each sequence is the moment magnitude of the largest aftershock actually observed in that sequence.
[0061] Initial model training:
[0062] The processed historical dataset is randomly divided into training and test sets in a 7:3 ratio.
[0063] TabPFN (Tabular Prior-Data Fitted Network) was chosen as the initial prediction model. TabPFN is a neural network designed specifically for small-sample tabular data. By pre-training on a large amount of synthetic data, it can achieve fast and accurate inference on a very small number of samples.
[0064] The TabPFN model is trained using the training set data (essentially adapting it to the data distribution of the current task) and its baseline performance is validated on the test set. The trained model is saved as base_model.pkl as the starting point for subsequent online augmentations.
[0065] Step 2: Online Enhancement Prediction (Online Phase)
[0066] Suppose the system detects a new mainshock event M_new (magnitude). =6.5, depth 10km, focal mechanism solution known).
[0067] S201: Real-time data trigger:
[0068] The system obtains preliminary information about M_new in real time through the API interface. Once it is determined that the magnitude exceeds a preset threshold (e.g., If the value is ≥5.5, the subsequent process will be triggered immediately.
[0069] S202: Online Data Selection (Similar Sequence Retrieval):
[0070] The system extracts the feature vector F_new = [M_main=6.5, Depth_main=10, Strike=120, Dip=45, Rake=80, Dist_to_Fault=5] of M_new.
[0071] The system calculates the weighted Euclidean distance between F_new and the eigenvectors of all mainshocks in the historical database. The weights can be set based on domain knowledge; for example, higher weights can be assigned to the mainshock magnitude and focal mechanism.
[0072] The system retrieves the K=50 historical mainshock sequences that are closest to F_new, forming a similar sequence set S_similar.
[0073] S203: Online Data Enhancement
[0074] For each sequence in S_similar (whose data is (F_i, M_max_i)), perform the following enhancement operation:
[0075] Add random noise: Add Gaussian noise with a mean of 0 and a standard deviation of 1% of the feature's own standard deviation to the continuous features (M_main, Depth_main). Also add small noise to the label M_max_i.
[0076] Linear interpolation: Randomly select two different sequences (F_a, M_max_a) and (F_b, M_max_b) from S_similar, and generate a new synthetic sample through linear interpolation: F_synthetic = α * F_a + (1-α) * F_b, M_max_synthetic = α * M_max_a + (1-α) * M_max_b, where α is a random number in the interval [0,1].
[0077] Using the above methods, the original 50 samples are expanded to 500 augmented samples, forming the augmented training set D_augmented.
[0078] S204: Model Fine-tuning (Personalized Model Generation):
[0079] Load the pre-trained initial model base_model.pkl.
[0080] The model is fine-tuned using an augmented training set D_augmented. The fine-tuning process uses a small learning rate (e.g., 1 / 10 of the initial learning rate) and the training epochs are controlled to 5-10 rounds to prevent overfitting to small sample augmented data.
[0081] After fine-tuning, we obtain the personalized prediction model Model_personalized, which is specifically customized for M_new.
[0082] S205: Dynamic Prediction
[0083] The feature vector F_new of M_new is input into the fine-tuned Model_personalized.
[0084] The model outputs a continuous magnitude value, which is the predicted maximum aftershock magnitude M_pred of the M_new sequence. Simultaneously, the model can output the confidence interval of the predicted value.
[0085] Beneficial effects verification
[0086] To verify the effectiveness of this invention, we conducted a retrospective test. One hundred mainshock sequences from the USGS catalog that were not used in the training were selected as the test set. Predictions were made using: (1) the traditional Båth law; (2) the static TabPFN model without online enhancement; and (3) the online enhancement method proposed in this invention.
[0087] The results show that the method proposed in this invention has significantly lower root mean square error (RMSE) and mean absolute error (MAE) than the previous two methods. Especially in the early stages of the sequence, this invention effectively reduces the uncertainty of prediction by utilizing prior knowledge of similar sequences, demonstrating superior stability and accuracy.
[0088] In summary, this invention, through an online process of "retrieval-enhancement-fine-tuning," has achieved a leap from "general models" to "case-specific customization" in maximum aftershock prediction, providing an effective technical means to solve the problem of rapid and accurate prediction during the critical post-earthquake period.
Claims
1. A maximum aftershock prediction method based on online enhancement, characterized in that, Includes the following steps: Data preparation steps: Obtain historical earthquake data from global and regional earthquake catalogs, define the mainshock and aftershock sequences, extract mainshock characteristics, and construct a historical earthquake database; Initial model building steps: Using the historical earthquake database, train an initial maximum aftershock prediction model; Online augmentation prediction steps, the steps including: Real-time trigger sub-step: In response to the occurrence of a new mainshock event, the model update process is triggered; Data selection sub-step: Based on the characteristics of the newly occurred mainshock, retrieve K similar historical mainshock sequences from the historical earthquake database; Data augmentation sub-step: Perform data augmentation on the data corresponding to the K similar historical mainshock sequences retrieved, and generate an augmented training set; Model fine-tuning sub-step: Using the enhanced training set, fine-tune the initial maximum aftershock prediction model to obtain a personalized prediction model for the new main shock; Prediction sub-step: Using the personalized prediction model, predict the magnitude of its maximum aftershock based on the characteristics of the new mainshock.
2. The method according to claim 1, characterized in that, In the data selection sub-step, similarity retrieval is performed based on one or more of the following characteristics of the newly occurring mainshock: mainshock magnitude, mainshock depth, mainshock focal mechanism solution, and distance from known active fault zones.
3. The method according to claim 1, characterized in that, The data augmentation sub-step employs one or more of the following data augmentation techniques: adding random noise to continuous features; interpolating in the feature space of similar sequences to generate synthetic samples; and using synthetic minority class oversampling (SMOTE) to balance the data distribution.
4. The method according to claim 1, characterized in that, The model fine-tuning sub-step employs transfer learning or incremental learning techniques to update the model parameters of the initial model at a learning rate lower than that used when training the initial model.
5. The method according to claim 1, characterized in that, The initial maximum aftershock prediction model is a machine learning model suitable for few-shot learning, preferably a support vector machine (SVM) or a TabPFN (Tabular Prior-Data Fitted Network).