Scholarlink

<p>Bolt looseness in pipeline structures poses a significant risk to urban infrastructure safety. To address this, a novel method is proposed that combines Anti-noise Differential Mel-scale Frequency Cepstral Coefficients (AntiD-MFCC) with a lightweight deep learning model for automatic looseness detection. Two key innovations are introduced: (1) A noise-robust acoustic feature extraction method that integrates Regularized Contig Shrinkage denoising with MFCC, first-order, and second-order difference features, resulting in significantly enhanced robustness under varying noise levels; (2) A deep classification architecture named EfficMonitNet, was designed to improve detection accuracy and noise resistance compared to conventional deep learning models. Laboratory experiments were conducted on flange-connected steel pipelines under five torque-defined looseness levels. Results show that the proposed method achieves high classification accuracy and outperforms traditional convolutional neural network (CNN) and backpropagation (BP) neural networks, particularly in noisy environments. This study demonstrates a feasible and intelligent approach for real-time bolt monitoring in urban pipeline systems, offering a new paradigm for non-contact structural health assessment.</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

Automatic detection of loose bolts in pipeline structures based on anti-noise mel cepstrum differential features and deep learning

  • 期刊论文

  • 发表日期：

    2026年1月28日

查看全文
  • Yixuan Chen,
  • Hongyi Chen,
  • Jingyi Wei,
  • Jianchao Wu

摘要

Bolt looseness in pipeline structures poses a significant risk to urban infrastructure safety. To address this, a novel method is proposed that combines Anti-noise Differential Mel-scale Frequency Cepstral Coefficients (AntiD-MFCC) with a lightweight deep learning model for automatic looseness detection. Two key innovations are introduced: (1) A noise-robust acoustic feature extraction method that integrates Regularized Contig Shrinkage denoising with MFCC, first-order, and second-order difference features, resulting in significantly enhanced robustness under varying noise levels; (2) A deep classification architecture named EfficMonitNet, was designed to improve detection accuracy and noise resistance compared to conventional deep learning models. Laboratory experiments were conducted on flange-connected steel pipelines under five torque-defined looseness levels. Results show that the proposed method achieves high classification accuracy and outperforms traditional convolutional neural network (CNN) and backpropagation (BP) neural networks, particularly in noisy environments. This study demonstrates a feasible and intelligent approach for real-time bolt monitoring in urban pipeline systems, offering a new paradigm for non-contact structural health assessment.