Scholarlink

<p>Accurate fault diagnosis of tribo-mechanical machinery is essential for preventing unexpected failures and reducing maintenance costs. However, most existing deep learning approaches, including CNN, GNN, and transformer-based models, rely on sample-level feature learning and often fail to generalize under non-stationary operating conditions such as varying load, speed, and progressive wear evolution. Moreover, the absence of explicit physical constraints limits interpretability and robustness in real-world industrial environments. To address these limitations, this study proposes a physics-informed graph neural operator (PI-GNO) integrated with a hierarchical ensemble fault reasoning and decision fusion framework. The proposed method reformulates fault diagnosis as a function-to-function learning problem, where multi-sensor signals are modeled as continuous degradation fields over a physically constructed machinery graph. Unlike conventional discrete classifiers, PI-GNO learns continuous spatiotemporal degradation dynamics while embedding tribology-driven physical laws into the learning process. Physics-informed constraints enforce monotonic wear evolution, friction–temperature coupling, and energy dissipation consistency, ensuring physically realistic modeling of degradation behavior. The continuous operator output is further mapped into structured diagnostic decisions through a hierarchical ensemble strategy consisting of global health assessment, fault mechanism identification, and localized severity estimation. This multi-level design enables coarse-to-fine reasoning and improves interpretability across different abstraction levels. The framework is implemented using Python, PyTorch, and PyTorch Geometric and evaluated on the PRONOSTIA run-to-failure dataset. A 70/15/15 train–validation–test split with bearing-wise separation is used, along with cross-condition testing on unseen load and speed settings. Additional validation is conducted on XJTU-SY and CWRU datasets to assess generalization capability. Experimental results show that PI-GNO achieves 98.6% accuracy, outperforming state-of-the-art baselines, while robustness tests under 2&#xa0;dB noise retain 83.4% accuracy. These results confirm strong generalization, physical consistency, and interpretability.</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

Graph neural operator-based dynamic degradation modeling and hierarchical decision fusion for robust multi-sensor fault diagnosis in tribo-mechanical systems

  • 期刊论文

  • 发表日期：

    2026年7月11日

查看全文
  • Lakshmaiya Natrayan

摘要

Accurate fault diagnosis of tribo-mechanical machinery is essential for preventing unexpected failures and reducing maintenance costs. However, most existing deep learning approaches, including CNN, GNN, and transformer-based models, rely on sample-level feature learning and often fail to generalize under non-stationary operating conditions such as varying load, speed, and progressive wear evolution. Moreover, the absence of explicit physical constraints limits interpretability and robustness in real-world industrial environments. To address these limitations, this study proposes a physics-informed graph neural operator (PI-GNO) integrated with a hierarchical ensemble fault reasoning and decision fusion framework. The proposed method reformulates fault diagnosis as a function-to-function learning problem, where multi-sensor signals are modeled as continuous degradation fields over a physically constructed machinery graph. Unlike conventional discrete classifiers, PI-GNO learns continuous spatiotemporal degradation dynamics while embedding tribology-driven physical laws into the learning process. Physics-informed constraints enforce monotonic wear evolution, friction–temperature coupling, and energy dissipation consistency, ensuring physically realistic modeling of degradation behavior. The continuous operator output is further mapped into structured diagnostic decisions through a hierarchical ensemble strategy consisting of global health assessment, fault mechanism identification, and localized severity estimation. This multi-level design enables coarse-to-fine reasoning and improves interpretability across different abstraction levels. The framework is implemented using Python, PyTorch, and PyTorch Geometric and evaluated on the PRONOSTIA run-to-failure dataset. A 70/15/15 train–validation–test split with bearing-wise separation is used, along with cross-condition testing on unseen load and speed settings. Additional validation is conducted on XJTU-SY and CWRU datasets to assess generalization capability. Experimental results show that PI-GNO achieves 98.6% accuracy, outperforming state-of-the-art baselines, while robustness tests under 2 dB noise retain 83.4% accuracy. These results confirm strong generalization, physical consistency, and interpretability.