Scholarlink

<p>Accurate real-time prediction of thrust and torque is crucial for tunnel boring machines (TBM) tunneling safety and efficiency. This paper proposes Att-Wave-ConvNet (AWCNet), a TBM thrust and torque prediction model based on TimesNet, designed for longer real-time forecasting. It replaces fast Fourier transform and complex convolution with discrete wavelet packet transform and two-dimensional elevated convolution, while integrating a multi-head self-attention mechanism to enhance long-sequence feature extraction. Experiments on field data from a challenging hard-rock tunneling project show that AWCNet achieves reliable prediction even at a 60-s length, outperforming TimesNet, Dlinear, Informer and Transformer. Compared to existing methods, it reduces lowest root mean square error (RMSE) and mean absolute error (MAE) by 0.169 and 0.047, improves <i>R</i><sup>2</sup> by 4.8%, and shortens prediction runtime by 26.07%. The results confirm that AWCNet ensures both accuracy and efficiency for real-time TBM load prediction, demonstrating its practical applicability over extended durations.</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

AWCNet: A TimesNet-based model for real-time TBM thrust and torque prediction over extended durations

  • 期刊论文

  • 发表日期：

    2025年9月22日

查看全文
  • Zhaoyang Li,
  • Wei Tang,
  • Huxiu Xu,
  • Yiding Zhong,
  • Kecheng Qin,
  • Huayong Yang,
  • Jun Zou

摘要

Accurate real-time prediction of thrust and torque is crucial for tunnel boring machines (TBM) tunneling safety and efficiency. This paper proposes Att-Wave-ConvNet (AWCNet), a TBM thrust and torque prediction model based on TimesNet, designed for longer real-time forecasting. It replaces fast Fourier transform and complex convolution with discrete wavelet packet transform and two-dimensional elevated convolution, while integrating a multi-head self-attention mechanism to enhance long-sequence feature extraction. Experiments on field data from a challenging hard-rock tunneling project show that AWCNet achieves reliable prediction even at a 60-s length, outperforming TimesNet, Dlinear, Informer and Transformer. Compared to existing methods, it reduces lowest root mean square error (RMSE) and mean absolute error (MAE) by 0.169 and 0.047, improves R2 by 4.8%, and shortens prediction runtime by 26.07%. The results confirm that AWCNet ensures both accuracy and efficiency for real-time TBM load prediction, demonstrating its practical applicability over extended durations.