Scholarlink

<p>Accurate identification of rock failure modes plays a crucial role in safety assessment and disaster early warning for rock engineering. However, traditional acoustic emission (AE) analysis methods exhibit limitations in complex signal processing and model interpretability, which hinders their ability to meet practical engineering demands for dynamic interpretation of fracture mechanisms. Here we proposed a framework that combined the high-precision advantage of the XGBoost (eXtreme Gradient Boosting) algorithm with the interpretability analysis of SHAP (Shapley Additive exPlanations) for rock failure mode identification. Based on the commonly used Brazilian splitting test and Z-shaped shear test methods, the polarity characteristics of the P-wave initial motion amplitude of AE signals were used to classify and label the tensile and shear failure events during the rock failure process, and a dataset of AE characteristics of typical failure modes was then constructed. The data quality was optimized through Kernel Principal Component Analysis (KPCA) combined with an adaptive Interquartile Range (IQR) threshold method. Furthermore, the comparative analysis with algorithms such as decision tree and random forest demonstrated that XGBoost performed well in learning nonlinear acoustic emission characteristics, achieving an accuracy of 89% on the test dataset. SHAP value quantification revealed classification contributions of AE parameters. An innovative interpretable model was built to correlate AE characteristics with rock failure mechanisms. Finally, the proposed XGBoost–SHAP interpretable model was applied to the uniaxial compression test dataset to verify its generalization ability with an accuracy of 74%. The results show that this method can effectively identify the failure modes under different loading conditions. By learning AE characteristics, the framework achieved high-accuracy identification of rock failure modes and transparent analysis of decision-making processes. This study established a quantitative correlation between AE characteristics and failure mechanisms through an interpretable machine learning framework, providing a data-driven solution for failure mode recognition in engineering 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

Establishment and Verification of an XGBoost–SHAP Interpretable Model for Rock Failure Mode Identification Based on Acoustic Emission Signals

  • 期刊论文

  • 发表日期：

    2026年1月12日

查看全文
  • Zhenghu Zhang,
  • Meihui Guo,
  • Ke Ma,
  • Jianhui Deng

摘要

Accurate identification of rock failure modes plays a crucial role in safety assessment and disaster early warning for rock engineering. However, traditional acoustic emission (AE) analysis methods exhibit limitations in complex signal processing and model interpretability, which hinders their ability to meet practical engineering demands for dynamic interpretation of fracture mechanisms. Here we proposed a framework that combined the high-precision advantage of the XGBoost (eXtreme Gradient Boosting) algorithm with the interpretability analysis of SHAP (Shapley Additive exPlanations) for rock failure mode identification. Based on the commonly used Brazilian splitting test and Z-shaped shear test methods, the polarity characteristics of the P-wave initial motion amplitude of AE signals were used to classify and label the tensile and shear failure events during the rock failure process, and a dataset of AE characteristics of typical failure modes was then constructed. The data quality was optimized through Kernel Principal Component Analysis (KPCA) combined with an adaptive Interquartile Range (IQR) threshold method. Furthermore, the comparative analysis with algorithms such as decision tree and random forest demonstrated that XGBoost performed well in learning nonlinear acoustic emission characteristics, achieving an accuracy of 89% on the test dataset. SHAP value quantification revealed classification contributions of AE parameters. An innovative interpretable model was built to correlate AE characteristics with rock failure mechanisms. Finally, the proposed XGBoost–SHAP interpretable model was applied to the uniaxial compression test dataset to verify its generalization ability with an accuracy of 74%. The results show that this method can effectively identify the failure modes under different loading conditions. By learning AE characteristics, the framework achieved high-accuracy identification of rock failure modes and transparent analysis of decision-making processes. This study established a quantitative correlation between AE characteristics and failure mechanisms through an interpretable machine learning framework, providing a data-driven solution for failure mode recognition in engineering monitoring.