Scholarlink

<p>Traditional compaction evaluation methods often suffer from low reliability and limited robustness due to insufficient consideration of material-specific properties. To address this challenge, this study proposes a hybrid intelligent evaluation method that integrates advanced signal processing techniques and deep learning models. Specifically, vibration signals collected during compaction are decomposed using Variational Mode Decomposition (VMD) into low- and high-frequency components, each reflecting distinct physical phenomena. The low-frequency component is analyzed via Short-Time Fourier Transform (STFT) to assess compaction quality, while the high-frequency component is characterized using Symmetric Dot Pattern (SDP) to identify material gradation features. Two ResNeSt deep learning networks with split-attention mechanisms are developed to respectively handle gradation classification and compaction prediction, achieving accuracies of 85% and 98%. Extensive laboratory and field experiments validate the effectiveness of the proposed method. By incorporating material gradation information, this approach effectively improves the reliability and interpretability of compaction quality assessment, overcoming the limitations of traditional methods that often neglect material characteristics.</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

Integrating material characteristics into compaction assessment: a VMD-SDP/STFT and deep learning approach for rockfill quality monitoring

  • 期刊论文

  • 发表日期：

    2026年2月19日

查看全文
  • Zhenfeng Qiu,
  • Dingjun Ke,
  • Yuping Ma,
  • Ting Cao,
  • Yi Feng,
  • Jun Fang

摘要

Traditional compaction evaluation methods often suffer from low reliability and limited robustness due to insufficient consideration of material-specific properties. To address this challenge, this study proposes a hybrid intelligent evaluation method that integrates advanced signal processing techniques and deep learning models. Specifically, vibration signals collected during compaction are decomposed using Variational Mode Decomposition (VMD) into low- and high-frequency components, each reflecting distinct physical phenomena. The low-frequency component is analyzed via Short-Time Fourier Transform (STFT) to assess compaction quality, while the high-frequency component is characterized using Symmetric Dot Pattern (SDP) to identify material gradation features. Two ResNeSt deep learning networks with split-attention mechanisms are developed to respectively handle gradation classification and compaction prediction, achieving accuracies of 85% and 98%. Extensive laboratory and field experiments validate the effectiveness of the proposed method. By incorporating material gradation information, this approach effectively improves the reliability and interpretability of compaction quality assessment, overcoming the limitations of traditional methods that often neglect material characteristics.