基于PRIME-DP预训练模型的滑坡微地震等非天然地震分类研究

1.中国长江三峡集团科学技术研究院,武汉 430010;2.中国地震局地球物理研究所,北京 100081

地震信号分类;地震预训练模型;地震大模型

Classification of Non-Natural Seismic Events with Landslide Microseismicity as Representative Based on PRIME-DP Pre-Training Model
CAI Yuqi1,2,SHAO Bo1,YU Ziye2,YAO Xianglong1,LIU Lu1,OUYANG Jinhui1

1.Institute of Science and Technology , China Three Gorges Corporation , Wuhan 100089, China;2.Institute of Geophysics , China Earthquake Administration , Beijing 100081, China

Classification of seismic signal;Pre-training seismic model;Seismic large model

DOI: 10.13512/j.hndz.2025.01.02

备注

天然地震和非天然事件分类对于地震编目、地震预警、滑坡灾害分析等具有重要意义。目前的研究方法中传统方法需要构建P/S振幅比、谱比等参数,这限制了算法对于复杂类别分类的可能。而基于深度神经网络的算法虽然可以对复杂的事件进行分类,但是目前研究中多集中于爆破、塌陷事件的识别,对于滑坡微地震事件等更多类型分类则研究较少。更重要的是深度神经网络训练需要大量人工标注数据,而非天然地震事件标注数据较为缺乏,因此难以训练高精度、高泛化能力的模型。因此我们基于开源的PRIME-DP地震数据处理预训练模型训练了用于天然地震、爆破、塌陷、滑坡微地震事件的多分类模型。并使用预训练模型数据和262条微地震事件进行了迁移学习训练。训练结果表明,我们的模型相比于基于STFT特征的深度学习分类模型精度从79.4%提升到了94.8%。这说明基于预训练模型来进行地震分类可以有效的提升最终精度。
Classification of natural earthquakes and non-natural seismic events is of great significance for earthquake cataloguing, earthquake early warning, and landslide disaster analysis. The existing traditional research methods need to construct parameters such as the P/S amplitude ratio and spectrum ratio,which limits the possibility of the algorithm to classify complex categories. Although the algorithms based on deep neural networks can classify complex events, most of the current research focuses on the identification of blasting and collapse events, and there is less research on the classification of more types such as landslide microseismicity. More importantly, deep neural network training requires a large amount of manually labeled data, which is relatively scarce for non-natural seismic event, so it is difficult to train models with high precision and high generalization capabilities. Therefore,this paper trained multi-classification models for natural earthquakes,blasting,collapse, and landslide microseismicity events based on the open-source PRIME-DP seismic data processing pre-training model. The transfer learning and training were performed by using the pre-training model data and 262 microseismicity events. The training results show that compared with the deep learning classification model based on STFT features, the accuracy of the proposed model is improved from 79.4% to 94.8%. This indicates that seismic event classification based on the pre-training model can effectively improve the final accuracy.
·