Scholarlink

<p>Accurate prediction of rock failure under unknown loading conditions is a major challenge in geotechnical engineering. This study proposes an AI-integrated framework combining a real-world Rock dataset, a deep learning-based displacement estimation model (DICNet-Rock), and a novel morphological warning indicator (WID) to detect early precursors of rock failure. The Rock dataset, derived from 421 uniaxial compression tests, captures natural rock surface textures and real-world imaging noise. DICNet-Rock was trained and validated on this dataset, achieving superior displacement prediction accuracy (MSE &lt; 0.002) and real-time processing speed (~ 10 frames per second), outperforming both traditional and existing AI-based digital image correlation methods. The WID indicator quantifies displacement concentration trends and distinguishes meaningful instability signals from random fluctuations. Single-sample and statistical analyses demonstrated that WID signal frequency and clustering reliably correlate with time to failure, providing consistent early warnings. Approximately 95% of samples released initial WID signals 40–60&#xa0;s before failure, while 83% accumulated 20 signals within 20&#xa0;s prior to failure. The method’s robustness, real-time capability, and independence from load measurements highlight its potential for integration into automated early warning systems for mining, tunneling, and slope stability monitoring.</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

AI-Enhanced Image Processing for Real-Time Brittle Rock Failure Forecasting: Development and Validation Using a Large Experimental Dataset

  • 期刊论文

  • 发表日期：

    2026年1月6日

查看全文
  • Fuqiang Gao,
  • Liqun Zhong,
  • Lei Yang,
  • Zhiguo Lu,
  • Wenju Liu

摘要

Accurate prediction of rock failure under unknown loading conditions is a major challenge in geotechnical engineering. This study proposes an AI-integrated framework combining a real-world Rock dataset, a deep learning-based displacement estimation model (DICNet-Rock), and a novel morphological warning indicator (WID) to detect early precursors of rock failure. The Rock dataset, derived from 421 uniaxial compression tests, captures natural rock surface textures and real-world imaging noise. DICNet-Rock was trained and validated on this dataset, achieving superior displacement prediction accuracy (MSE < 0.002) and real-time processing speed (~ 10 frames per second), outperforming both traditional and existing AI-based digital image correlation methods. The WID indicator quantifies displacement concentration trends and distinguishes meaningful instability signals from random fluctuations. Single-sample and statistical analyses demonstrated that WID signal frequency and clustering reliably correlate with time to failure, providing consistent early warnings. Approximately 95% of samples released initial WID signals 40–60 s before failure, while 83% accumulated 20 signals within 20 s prior to failure. The method’s robustness, real-time capability, and independence from load measurements highlight its potential for integration into automated early warning systems for mining, tunneling, and slope stability monitoring.