Scholarlink

<p>Seismic ambient noise cross-correlation has become an essential tool for monitoring subsurface dynamics, offering continuous data acquisition and broad spatial coverage. To address the dense and repetitive of geological anomaly events, we propose a few-shot anomaly detection method integrating deep temporal modeling with statistical distance metrics. By integrating Transformer-based temporal modeling with Mahalanobis distance metrics for anomaly detection, a time–frequency few-shot learning framework for multi-type anomaly recognition under limited samples is developed. Additionally, an adaptive learning rate adjustment strategy is proposed, which detects data drift by measuring the Euclidean distance between new data samples and historical class prototypes. The experimental results demonstrate that the baseline model achieves accuracies of 96.5% and 98.5% on 4-way 1-shot and 4-way 5-shot tasks, respectively, with 200 training samples. Even under extreme conditions (40 samples, SNR = 6), the model maintains strong performance (≈ 85% accuracy). Upon introducing the dynamic update mechanism, the model achieves an average accuracy of 94.5% during continual learning. Each iteration takes 1.1&#xa0;s, with parameter optimization requiring only 0.6&#xa0;s, thus satisfying near-real-time processing requirements for the model update process. The proposed method is validated using continuous 320-day monitoring data from an operational seismic network, achieving an average classification accuracy of 91.8% under complex environmental interference. Visualization of the model’s attention distribution highlights high-energy regions in the time–frequency domain, offering physically grounded interpretability for anomaly analysis.</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

Few-shot learning for dynamic anomaly detection in seismic ambient noise monitoring

  • 期刊论文

  • 发表日期：

    2026年5月26日

查看全文
  • Xiaomin Wu,
  • Fang Ye

摘要

Seismic ambient noise cross-correlation has become an essential tool for monitoring subsurface dynamics, offering continuous data acquisition and broad spatial coverage. To address the dense and repetitive of geological anomaly events, we propose a few-shot anomaly detection method integrating deep temporal modeling with statistical distance metrics. By integrating Transformer-based temporal modeling with Mahalanobis distance metrics for anomaly detection, a time–frequency few-shot learning framework for multi-type anomaly recognition under limited samples is developed. Additionally, an adaptive learning rate adjustment strategy is proposed, which detects data drift by measuring the Euclidean distance between new data samples and historical class prototypes. The experimental results demonstrate that the baseline model achieves accuracies of 96.5% and 98.5% on 4-way 1-shot and 4-way 5-shot tasks, respectively, with 200 training samples. Even under extreme conditions (40 samples, SNR = 6), the model maintains strong performance (≈ 85% accuracy). Upon introducing the dynamic update mechanism, the model achieves an average accuracy of 94.5% during continual learning. Each iteration takes 1.1 s, with parameter optimization requiring only 0.6 s, thus satisfying near-real-time processing requirements for the model update process. The proposed method is validated using continuous 320-day monitoring data from an operational seismic network, achieving an average classification accuracy of 91.8% under complex environmental interference. Visualization of the model’s attention distribution highlights high-energy regions in the time–frequency domain, offering physically grounded interpretability for anomaly analysis.