Scholarlink

<p>Design Failure Mode and Effects Analysis (DFMEA) traditionally relies on static tables, limiting its ability to capture failure propagation across design hierarchies or support reasoning under out-of-distribution conditions. These constraints hinder knowledge reuse and evidence-based decision-making in early design phases. To address this, we formalize failure knowledge as a directed Knowledge Graph grounded in fault-tree logic and introduce AfGNN (Adaptive Failure Graph Neural Network). AfGNN achieves robust prediction on rare, long-tail failure patterns through three integrated innovations: (1) Causal-Enhanced Soft-Label Embedding (CESLE), which integrates semantic similarity with causal weights to distinguish genuine relationships from statistical correlations; (2) a Depth-Adaptive Causal Propagation Framework, synergizing dynamic subgraph sampling with depth-decay attention to balance efficiency and fidelity while suppressing noise in deep layers; and (3) a formalized computational workflow that transforms DFMEA into a reusable causal graph, enabling systematic reasoning over incomplete failure records. Evaluated on five public datasets and a self-constructed automotive failure KG, AfGNN surpasses all GNN-based baselines and competes with LLM-based methods on general benchmarks, while substantially outperforming all baselines on the domain-specific FMEA dataset (MRR, Hits@1, Hits@10). This framework enables engineers to reason about rare multi-failure cascading effects without relying on complete historical data, advancing failure knowledge management and reliability-oriented design decision-making.</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

AfGNN: adaptive graph neural networks for causal failure reasoning in DFMEA

  • 期刊论文

  • 发表日期：

    2026年6月29日

查看全文
  • Junchen Liu,
  • Fei Li,
  • Jingsong Bao

摘要

Design Failure Mode and Effects Analysis (DFMEA) traditionally relies on static tables, limiting its ability to capture failure propagation across design hierarchies or support reasoning under out-of-distribution conditions. These constraints hinder knowledge reuse and evidence-based decision-making in early design phases. To address this, we formalize failure knowledge as a directed Knowledge Graph grounded in fault-tree logic and introduce AfGNN (Adaptive Failure Graph Neural Network). AfGNN achieves robust prediction on rare, long-tail failure patterns through three integrated innovations: (1) Causal-Enhanced Soft-Label Embedding (CESLE), which integrates semantic similarity with causal weights to distinguish genuine relationships from statistical correlations; (2) a Depth-Adaptive Causal Propagation Framework, synergizing dynamic subgraph sampling with depth-decay attention to balance efficiency and fidelity while suppressing noise in deep layers; and (3) a formalized computational workflow that transforms DFMEA into a reusable causal graph, enabling systematic reasoning over incomplete failure records. Evaluated on five public datasets and a self-constructed automotive failure KG, AfGNN surpasses all GNN-based baselines and competes with LLM-based methods on general benchmarks, while substantially outperforming all baselines on the domain-specific FMEA dataset (MRR, Hits@1, Hits@10). This framework enables engineers to reason about rare multi-failure cascading effects without relying on complete historical data, advancing failure knowledge management and reliability-oriented design decision-making.