Scholarlink

<p>The underground mining environment generates highly heterogeneous microseismic (MS) signals, whose accurate identification is crucial for source localization and failure mechanism analysis. Limited data in the early monitoring stages restrict recognition performance. Existing methods often perform poorly under small-sample conditions due to low identification rates and limited generalizability. To address these challenges, this study proposes a transfer-learning-enhanced Bidirectional LSTM and Deep Convolutional Network (Tr-BiLSTM-DCNN) model. Using Mel-spectrograms to characterize signals, the model integrates BiLSTM-based bidirectional temporal modeling with DCNN-based multi-scale spatial feature extraction, constructing a spatiotemporal feature representation. Training follows a two-phase strategy: pretraining on large cross-mine datasets with hyperparameter optimization, followed by domain-adaptive fine-tuning via transfer learning on small-sample target mine data. The model achieves 94.44% test accuracy under limited data conditions, representing an 80.85% relative improvement over non-transfer baselines and outperforming conventional CNN and LSTM approaches. This framework provides an effective few-shot learning solution for mine MS monitoring and demonstrates strong potential for engineering applications in dynamic disaster early warning.</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

A novel transfer learning-enhanced BiLSTM-DCNN architecture for mine microseismic signal identification with small data set

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月20日

查看全文
  • Yong Zhao,
  • Shihui Jiao,
  • Tianhong Yang,
  • Yao Wang,
  • Haiyan Xu,
  • Zhengdong Liu

摘要

The underground mining environment generates highly heterogeneous microseismic (MS) signals, whose accurate identification is crucial for source localization and failure mechanism analysis. Limited data in the early monitoring stages restrict recognition performance. Existing methods often perform poorly under small-sample conditions due to low identification rates and limited generalizability. To address these challenges, this study proposes a transfer-learning-enhanced Bidirectional LSTM and Deep Convolutional Network (Tr-BiLSTM-DCNN) model. Using Mel-spectrograms to characterize signals, the model integrates BiLSTM-based bidirectional temporal modeling with DCNN-based multi-scale spatial feature extraction, constructing a spatiotemporal feature representation. Training follows a two-phase strategy: pretraining on large cross-mine datasets with hyperparameter optimization, followed by domain-adaptive fine-tuning via transfer learning on small-sample target mine data. The model achieves 94.44% test accuracy under limited data conditions, representing an 80.85% relative improvement over non-transfer baselines and outperforming conventional CNN and LSTM approaches. This framework provides an effective few-shot learning solution for mine MS monitoring and demonstrates strong potential for engineering applications in dynamic disaster early warning.