于紫凝,李海峰,景锡龙,池成全,郑海永. 2024. 基于四分量钻孔应变数据的神经网络地震活动性预测分析. 地震学报,46(2):327−339. doi: 10.11939/jass.20230122
引用本文: 于紫凝,李海峰,景锡龙,池成全,郑海永. 2024. 基于四分量钻孔应变数据的神经网络地震活动性预测分析. 地震学报,46(2):327−339. doi: 10.11939/jass.20230122
Yu Z N,Li H F,Jing X L,Chi C Q,Zheng H Y. 2024. Borehole strain data based seismicity prediction analysis using a neural network. Acta Seismologica Sinica46(2):327−339. doi: 10.11939/jass.20230122
Citation: Yu Z N,Li H F,Jing X L,Chi C Q,Zheng H Y. 2024. Borehole strain data based seismicity prediction analysis using a neural network. Acta Seismologica Sinica46(2):327−339. doi: 10.11939/jass.20230122

基于四分量钻孔应变数据的神经网络地震活动性预测分析

Borehole strain data based seismicity prediction analysis using a neural network

  • 摘要: 首先利用四分量钻孔应变数据独有的自洽特性,构建震前应变特征数据集;之后基于一维卷积神经网络框架,设计地震震级与方位的预测模型;然后通过混淆矩阵计算准确率、召回率以及F1分数,对模型预测结果进行评价与修正;最后对我国西南地区的永胜、昭通、姑咱及腾冲四个台站的钻孔应变特征分别进行训练与验证,并讨论了不同特征窗长对预测效果的影响。训练完成后的模型效果在测试集上均表现优异,四个台站对震级和方位预测的平均准确率分别可达85%和80%左右,说明四分量钻孔应变数据特征与地震的发生有着很强的相关性,通过卷积神经网络对地震前兆特征进行挖掘具有很大研究潜力,本文提出的预测策略也为未来短临地震的精确预测研究打下基础。

     

    Abstract:
    With the development of seismological observational techniques, a number of case studies indicate that the seismogenic process of major earthquakes is often accompanied by deformation anomalies. Strain data serve as an indicator of crustal deformation, which reflects changes in subsurface stress and holds significant importance for seismology research. However, the research on the extraction of pre-earthquake anomalies from borehole strain data is currently limited to the stages of case analysis and small-sample statistical analysis. Thus, it is very meaningful to use a new technique of data mining to analyze the association between strain anomalies and earthquakes.
    In order to dig out more strain information and correlation information between multiple strains, according to the scholars’ many analyses of the correlation between areal strains, shear strains and the self-consistent coefficients of the four-component strains, it is found that the borehole strain may reflect the preparatory process of earthquake nucleation. Therefore, this study calculates the Pearson correlation coefficients and self-consistent coefficients between these strains to finally constitute a 24-dimensional feature dataset. Subsequently, we divide earthquake events based on magnitude into three classes: no earthquake, earthquakes with 3.0≤MS<5.0, and earthquakes with MS≥5.0. Simultaneously, these earthquakes are also classified into five groups based on their orientation relative to the borehole strainmeter, forming an earthquake orientation dataset. Thereby, the labels for earthquake samples are generated according to these two classification criteria.
    Next, the study employs a one-dimensional convolutional neural network (1D-CNN) framework to develop a short-term prediction model for the magnitude and location of earthquakes. The CNN can leverage the advantages of its convolutional layers’ parameter-sharing mechanism to effectively capture local features in the data. This 1D-CNN model consists of three parts: the input layer, hidden layers, and output layer. Strain feature samples from the dataset are used as inputs to the model, which outputs predicted values for earthquake magnitude and location labels. The hidden layer of the model is divided into two parts: the convolutional region and the fully connected region. We apply the cross-entropy loss function and the Adam optimizer for model compilation, and a learning rate decay strategy is used to dynamically adjust the learning rate. Parameter optimization is conducted using a random search algorithm. To evaluate this prediction model, we use a confusion matrix to calculate accuracy, recall, and F1 scores for examining the efficiencies for each class. Furthermore, the study adjusts the model’s prediction window size for earthquakes above magnitude 5.0 to balance the distribution of samples across different magnitude classes.
    Finally, the borehole strain data from stations Guzan, Yongsheng, Zhaotong, and Tengchong in southwest China are respectively utilized for training and testing in this study. The results indicate that the pre-earthquake strain features from stations Yongsheng, Guzan, Zhaotong, and Tengchong exhibited a high level of accuracy up to 80% in predicting both magnitude and direction. Furthermore, the predictive results are independently validated for five typical major earthquakes in China. The findings demonstrate that the predicted magnitude and direction labels for these five earthquakes corresponded to the actual events, suggesting that the model successfully captures valuable pre-earthquake strain information and possesses predictive capabilities. Finally, we analyze the impact of the time window and prediction window sizes of input data on the model’s prediction accuracy across different magnitude classes. The results reveal that for all three magnitude classes, a longer time window leads to higher predictive accuracy of the model. Moreover, the results of major earthquakes are overall higher and more random than that of moderate earthquakes. It may reflect that the borehole strain has the short-term predictive capability for major earthquakes, and there are differences in pre-earthquake features among different major earthquakes.
    The 1D-CNN models built in this study could effectively predict earthquake magnitude and approximate location using data from the respective stations. This demonstrates that the convolutional neural network architecture has the capability to extract meaningful pre-earthquake anomaly features from borehole strain data. This provides new insights and methods for research in earthquake precursory observations, laying a foundation for accurate earthquake predictions in the future. However, considering practical applicability, there are limitations to the methodology of this study. Future research will devise more robust networks to address sample imbalances by combining labels and achieve simultaneous predictions for the three seismic elements. Alternatively, a combination of multiple strain stations could be utilized, incorporating additional earthquake location information (such as epicentral distance and angles) to enhance directional prediction accuracy through partitioning high-resolution spatial grids. Moreover, experimental validations will be conducted on strain observation stations in seismically weak areas or other multi-seismic regions to establish a robust short-term earthquake prediction model suitable for strong seismic events.

     

/

返回文章
返回