摘要
将伪标签算法引入地震类型识别领域,并设计伪标签神经网络法程序,对山东地区2019—2021年M L1.5以上的天然地震、爆破地震、塌陷地震三类事件开展试验。使用优选的有标签样本集预测无标签样本,将其标记为伪标签样本后加入联合训练,并对比传统BP神经网络法和支持向量机法,以初步验证伪标签算法在地震类型识别领域的可行性和在小样本条件下的适用性。试验结果表明:影响伪标签神经网络法分类效果的主要因素有已知样本数量和伪标签样本占比。当已知样本数量介于60~120个、伪标签样本占比20%~30%时,其识别效果最佳。在小样本条件下,伪标签神经网络法的识别率相较于传统BP神经网络法提高了2%~8%,与支持向量机法的识别率差值集中在±4%以内。因此,采用伪标签算法弥补部分地区样本库匮乏的不足,实现小样本地震类型识别,具备一定的应用价值。
This paper introduces the pseudo-label algorithm for earthquake type recognition and develops a pseudo-label neural network program to classify three types of earthquake events,namely,natural earthquakes,explosions,and collapses,occurring in the Shandong region from 2019 to 2021,with a magnitude above M L1.5.The algorithm uses a pseudo-labeling strategy to predict labels for unlabeled samples based on a selected set of labeled data.Once the unlabeled samples are assigned pseudo-labels,they are incorporated into the joint training process.The paper also compares the performance of the pseudo-label algorithm with traditional back propagation(BP)neural networks and support vector machines to preliminarily assess its feasibility and applicability,particularly under conditions of limited labeled data.Experimental results show that the classification performance of the pseudo-label neural network method is primarily influenced by the number of labeled samples and the proportion of pseudo-labeled samples.The optimal recognition performance is achieved when the number of labeled samples is between 60 and 120 and the proportion of pseudo-labeled samples is between 20%and 30%.Under small sample conditions,the recognition rate of the pseudo-label neural network method is increased by 2%-8%compared to traditional BP neural network methods,and the difference in recognition rate with the support vector machine method is generally within±4%.Therefore,the pseudo-label algorithm can help compensate for the shortage of sample data in certain areas,enabling earthquake type recognition under small sample conditions with practical application value.
作者
范晓易
王夫运
陈飞
陈传华
FAN Xiaoyi;WANG Fuyun;CHEN Fei;CHEN Chuanhua(Nanjing Earthquake Monitoring Center Station,Jiangsu Earthquake Agency,Nanjing 210000,Jiangsu,China;Geophysical Exploration Center,CEA,Zhengzhou 450000,Henan,China;Tai'an Earthquake Monitoring Center Station,Shandong Earthquake Agency,Tai'an 270000,Shandong,China)
出处
《地震工程学报》
北大核心
2025年第1期160-167,177,共9页
China Earthquake Engineering Journal
基金
山东省自然科学基金重点项目(ZR2020KF003)
中国地震局三结合项目(3JH-202301002)
国家重点研发计划-政府间国际合作重点专项(2018YFE0109700)。
关键词
伪标签算法
地震类型识别
神经网络法
小样本
pseudo-label algorithm
earthquake type recognition
neural network method
small samples
