Scholarlink

<p>Accurate prediction and interpretation of the Factor of Safety (Fs) are critical for making risk-based decisions in open-pit mines. Historical surrogate models have typically treated geotechnical variables in a geospatial sense (2D) and used separate historical monitoring data in a temporal sense (3D). This limits the ability to determine how the instability mechanism changes across different mining phases. To mitigate this issue, we propose ST-GeoNet, the first dual-branch architecture that integrates phase-aware Transformer temporal attention with the Convolutional Block Attention Module (CBAM) spatial refinement for interpretable Factor of Safety prediction in multi-phase open-pit slopes. Unlike prior spatial-only surrogates, ST-GeoNet jointly processes high-resolution 2D geotechnical fields and multi-phase monitoring sequences to capture both spatial heterogeneity and time-dependent kinematic signatures. The proposed architecture employs a Convolutional Backbone (with Convolutional Block Attention Module (CBAM) spatial feature refinement) and a lightweight Transformer Encoder (with temporal attention) for phase-aware sequence modeling. The gated fusion module takes two context vectors and combines them into a single scalar prediction of Fs, while also exporting interpretable attention maps for engineering diagnostics. The model was developed and tested on a physics-based data set of 8,000 scenarios, including geomechanical properties calibrated to published experimental results for oil sands. Results from multiple randomly seeded evaluations demonstrate that ST-GeoNet has a mean test RMSE of 0.119 and <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\({R}^{2}\)</EquationSource> </InlineEquation> of 0.905, representing 16.1%, 11.7%, and 30.6% improvement in RMSE over matched convolutional neural networks (CNN), long short-term memory (LSTM), and extreme gradient boosting (XGBoost) baseline models, respectively. Ablation studies show that both spatial and temporal attention independently contribute to prediction accuracy, with temporal attention particularly important in near-failure regimes. Robustness tests indicate that ST-GeoNet maintains stable predictive accuracy across varying groundwater scales, when input parameters are perturbed by noise, and, in extreme cases, when monitoring is very sparse. Most importantly, the spatial attention heatmaps align closely with numerically computed slip surfaces, and the temporal weights identify phases of high-risk excavation, transforming the surrogate model from a black-box predictor to a transparent decision-support model. With ST-GeoNet, a new benchmark for scenario-aware, interpretable slope stability assessment in dynamic mining environments has been established.</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

ST-GeoNet: a spatio-temporal attention network for interpretable factor of safety prediction in open-pit oil sands slopes

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月30日

查看全文
  • Khayitov Odiljon,
  • Baymirzaev Bakhtiyor,
  • Saidova Lola,
  • Karamov Alisher,
  • Khasanov Adkham,
  • Khasanov Obid,
  • Sayyidkosimov Sayyidjabbor,
  • Ochilov Golibjon,
  • Yusupkhodjaeva Eleonora,
  • Turdiev Shakhboz

摘要

Accurate prediction and interpretation of the Factor of Safety (Fs) are critical for making risk-based decisions in open-pit mines. Historical surrogate models have typically treated geotechnical variables in a geospatial sense (2D) and used separate historical monitoring data in a temporal sense (3D). This limits the ability to determine how the instability mechanism changes across different mining phases. To mitigate this issue, we propose ST-GeoNet, the first dual-branch architecture that integrates phase-aware Transformer temporal attention with the Convolutional Block Attention Module (CBAM) spatial refinement for interpretable Factor of Safety prediction in multi-phase open-pit slopes. Unlike prior spatial-only surrogates, ST-GeoNet jointly processes high-resolution 2D geotechnical fields and multi-phase monitoring sequences to capture both spatial heterogeneity and time-dependent kinematic signatures. The proposed architecture employs a Convolutional Backbone (with Convolutional Block Attention Module (CBAM) spatial feature refinement) and a lightweight Transformer Encoder (with temporal attention) for phase-aware sequence modeling. The gated fusion module takes two context vectors and combines them into a single scalar prediction of Fs, while also exporting interpretable attention maps for engineering diagnostics. The model was developed and tested on a physics-based data set of 8,000 scenarios, including geomechanical properties calibrated to published experimental results for oil sands. Results from multiple randomly seeded evaluations demonstrate that ST-GeoNet has a mean test RMSE of 0.119 and \({R}^{2}\) of 0.905, representing 16.1%, 11.7%, and 30.6% improvement in RMSE over matched convolutional neural networks (CNN), long short-term memory (LSTM), and extreme gradient boosting (XGBoost) baseline models, respectively. Ablation studies show that both spatial and temporal attention independently contribute to prediction accuracy, with temporal attention particularly important in near-failure regimes. Robustness tests indicate that ST-GeoNet maintains stable predictive accuracy across varying groundwater scales, when input parameters are perturbed by noise, and, in extreme cases, when monitoring is very sparse. Most importantly, the spatial attention heatmaps align closely with numerically computed slip surfaces, and the temporal weights identify phases of high-risk excavation, transforming the surrogate model from a black-box predictor to a transparent decision-support model. With ST-GeoNet, a new benchmark for scenario-aware, interpretable slope stability assessment in dynamic mining environments has been established.