Magnitude estimation for earthquake early warning based on machine learning and transfer learning
Magnitude estimation is a crucial component of earthquake early warning(EEW).Rapid and reliable magnitude estimation can provide effective EEW information.Traditional methods for estimating magnitude primarily rely on establishing empirical prediction equations using single feature extracted from P-wave signal,leading to significant magnitude estimation errors.To improve the reliability of EEW magnitude estimation and explore the feasibility of machine learning approach in the Sichuan-Yunnan region of China,this study establishes a machine learning pre-trained model(CRnet-M)for magnitude estimation based on strong motion data recorded by the Japanese K-NET network.The CRnet-M model combines convolutional neural network and recurrent neural network.Subsequently,leveraging transfer learning and strong motion data from the Sichuan-Yunnan region,this study fine-tunes and trains the pre-trained CRnet-M model,thus establishing a magnitude estimation model(TLCRnet-M)specifically tailored for the Sichuan-Yunnan region of China.The research findings indicate that for the Japanese test dataset,following the triggering of P-wave by 3 s,the pre-trained CRnet-M model exhibits a smaller magnitude estimation error compared to traditional magnitude estimation methods in EEW.The percentage of absolute error within the range of 0 to 0.5 magnitude units reaches 86.89%.For the Sichuan-Yunnan region test dataset,following the triggering of P-wave by 3 s,the TLCRnet-M model with transfer learning enhances the reliability of magnitude estimation compared to the traditional EEW magnitude estimation methods and the CRnet-M model without transfer learning.The percentage of absolute error within the range of 0 to 0.5 magnitude units is 76.25%.The method proposed in this study enhances the reliability of EEW magnitude estimation to a certain extent and holds significance for EEW systems.
earthquake early warningmachine learningneural networktransfer learningmagnitude estimationP wave
NETL
NSTL
万方数据
