Scholarlink

<p>An on-site earthquake early warning model utilizing a long short-term memory (LSTM) neural network is proposed, diverging from traditional methods by focusing on acceleration response spectrum Sa, the ground motion intensity measure correlated with structural responses. A three-channel acceleration waveform is taken as the model input, and an acceleration response spectrum serves as output. The model is trained using strong motion acceleration data acquired from Japan’s K-NET network. On the test set, the mean squared error (MSE) of the predictions yielded by the proposed model decreases as the input time window increases. In the temporal window spanning from 1–10 s, an MSE reduction of 72.35% is observed. The MSE is 1.92×10<sup>−4</sup> g 10 s after the P-wave is triggered. When subjected to generalization testing with cross-regional and cross-instrument-type Chinese intensity meter data, the model still exhibits the same trend as that observed on the test set. The MSE decreases by 74.16% 10 s after the P-wave is triggered (compared to the value obtained 1 s after the P-wave is triggered). The MSE is 1.93×10<sup>−4</sup> g 10 s after the P-wave is triggered in the cross-domain dataset. The results demonstrate that the proposed model exhibits good generalization performance.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Predicting acceleration response spectrum from P-wave arrivals via a long short-term memory neural network

  • 期刊论文

  • 发表日期：

    2026年4月25日

查看全文
  • Haozhen Dai,
  • Yueyong Zhou,
  • Shanyou Li,
  • Yongxiang Wei,
  • Heyi Liu,
  • Jindong Song

摘要

An on-site earthquake early warning model utilizing a long short-term memory (LSTM) neural network is proposed, diverging from traditional methods by focusing on acceleration response spectrum Sa, the ground motion intensity measure correlated with structural responses. A three-channel acceleration waveform is taken as the model input, and an acceleration response spectrum serves as output. The model is trained using strong motion acceleration data acquired from Japan’s K-NET network. On the test set, the mean squared error (MSE) of the predictions yielded by the proposed model decreases as the input time window increases. In the temporal window spanning from 1–10 s, an MSE reduction of 72.35% is observed. The MSE is 1.92×10−4 g 10 s after the P-wave is triggered. When subjected to generalization testing with cross-regional and cross-instrument-type Chinese intensity meter data, the model still exhibits the same trend as that observed on the test set. The MSE decreases by 74.16% 10 s after the P-wave is triggered (compared to the value obtained 1 s after the P-wave is triggered). The MSE is 1.93×10−4 g 10 s after the P-wave is triggered in the cross-domain dataset. The results demonstrate that the proposed model exhibits good generalization performance.