Scholarlink

<p>Predicting penetration rates is crucial for optimizing tunnel boring machine (TBM) operations and minimizing jamming risks. This paper introduces TBMTSMixer, a novel hybrid model based on time-series mixer (TSMixer). TBMTSMixer enhances performance through a cross-dimensional module and a fusion external attention mechanism. The cross-dimensional linear transformation in the residual block flattens the input vector, applies linear transformations and activation functions for upscaling and downscaling operations, and restores original dimensions. The fusion external attention mechanism generates keys and values in the temporal and feature dimensions through linear transformation, applying the softmax function to calculate importance weights. The weighted results are concatenated and integrated through a fusion layer. TBMTSMixer excels in 30-s, 60-s, and 90-s prediction tasks, particularly in the 30-s task, achieving an MSE of 0.232, MAE of 0.272, and an R<sup>2</sup> of 0.738, outperforming eight comparative models. It demonstrates superior stability, convergence, and generalization, making it highly advantageous for practical engineering applications.</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

Penetration Rate Forecasting for Hard Rock TBMs Using the TBMTSMixer Model

  • 期刊论文

  • 发表日期：

    2025年6月28日

查看全文
  • Long Li,
  • Binbin Zheng

摘要

Predicting penetration rates is crucial for optimizing tunnel boring machine (TBM) operations and minimizing jamming risks. This paper introduces TBMTSMixer, a novel hybrid model based on time-series mixer (TSMixer). TBMTSMixer enhances performance through a cross-dimensional module and a fusion external attention mechanism. The cross-dimensional linear transformation in the residual block flattens the input vector, applies linear transformations and activation functions for upscaling and downscaling operations, and restores original dimensions. The fusion external attention mechanism generates keys and values in the temporal and feature dimensions through linear transformation, applying the softmax function to calculate importance weights. The weighted results are concatenated and integrated through a fusion layer. TBMTSMixer excels in 30-s, 60-s, and 90-s prediction tasks, particularly in the 30-s task, achieving an MSE of 0.232, MAE of 0.272, and an R2 of 0.738, outperforming eight comparative models. It demonstrates superior stability, convergence, and generalization, making it highly advantageous for practical engineering applications.