一种基于物理约束的多模态大型语言模型(LLM),用于处理数据不平衡且样本不足的压水堆冷却剂系统的故障诊断
A physics-constrained multimodal LLM for fault diagnosis of pressurized water reactor coolant systems with imbalanced and under-sampled data
ES评分 0
| DOI |
10.1016/j.jii.2026.101143 |
| 刊名 |
|
| 年,卷(期) |
2026, 52() |
| 作者 |
|
| 作者单位 |
南华大学
|
| 摘要 |
压力级水反应堆冷却液系统(PWRC)的故障诊断仍极具挑战性,原因包括故障事件严重采样不足、多模态异质性强,以及在安全关键环境中需要严格的物理可解释性。为应对这些局限性,本文提出了一种由物理约束的多模态LLM(PCM-LLM),将第一原理约束与深度多模态表征学习相结合。首次制定出基于物理引导的数据增强策略,在严格满足质量能量保护和热-水分边界条件的同时,扩大了有限的冷却剂损耗事故(LOCA)样品。异质传感器流(包括压力、温度和流量)随后被编码为统一的多模态张力,并通过分层的跨模态注意力机制融合而成。为了确保物理上有意义的推理,质量平衡、能量保持、动量方程和中子动力学被嵌入为可分的软约束,引导优化实现物理上合理的决策边界。此外,基于LLM的神经符号模块可生成物理上一致的诊断解释以及具有不确定性感知的推理链。在不平衡的PWRCS数据集上的实验结果表明,PCM-LLM的准确率达到92.5%,实现0.942 AUC-ROC,以及超过0.90的物理一致性,优于最先进的深度学习和物理知识基准。该方法在少量条件和高噪声环境下表现出强大的稳定性,同时提供符合工程原理的可解释的缺陷因果关系。本研究通过将数据驱动的智能与严格的物理约束相结合,建立了核安全诊断的新范式。
|
| Abstract |
Fault diagnosis of pressurized water reactor coolant systems (PWRCS) remains extremely challenging due to severe under-sampling of fault events, strong multimodal heterogeneity, and the need for strict physical interpretability in safety-critical environments. To address these limitations, this paper proposes a Physics-Constrained Multimodal LLM (PCM-LLM) that integrates first-principles constraints with deep multimodal representation learning. A physics-guided data enhancement strategy is first developed, expanding limited loss-of-coolant accidents (LOCA) samples while strictly satisfying mass-energy conservation and thermal-hydraulic boundary conditions. Heterogeneous sensor streams, including pressure, temperature, and flow, are then encoded into unified multimodal tensors and fused through a hierarchical cross-modal attention mechanism. To ensure physically meaningful reasoning, mass balance, energy conservation, momentum equations, and neutron kinetics are embedded as differentiable soft-constraints, guiding the optimization toward physically plausible decision boundaries. Furthermore, an LLM-based neuro-symbolic module generates physically consistent diagnostic explanations and uncertainty-aware reasoning chains. Experimental results on an imbalanced PWRCS dataset demonstrate that PCM-LLM achieves 92.5% accuracy, 0.942 AUC-ROC, and >0.90 physical consistency, outperforming state-of-the-art deep learning and physics-informed baselines. The method shows strong robustness under few-shot conditions and high-noise environments, while providing interpretable fault causality aligned with engineering principles. This study establishes a new paradigm for nuclear safety diagnostics by unifying data-driven intelligence with rigorous physical constraints.
|
| 关键词 |
加压水反应堆冷却剂系统(PWRCS) ; 物理约束型多模态LLM(PCM-LLM) ; 冷却液事故(LOCA) ; 可分量的软约束 ; 条件不多
|
| KeyWord |
Pressurized water reactor coolant system (PWRCS)Physics-constrained multimodal LLM (PCM-LLM)Loss-of-coolant accidents (LOCA)Differentiable soft-constraintsFew-shot conditionsDifferentiable soft-constraintsFew-shot conditions
|
| 基金项目 |
|
| 页码 |
101143-101143 |
Keshun Y, Haidon S, Zhichao C, et al..
一种基于物理约束的多模态大型语言模型(LLM),用于处理数据不平衡且样本不足的压水堆冷却剂系统的故障诊断 [J].
Journal of Industrial Information Integration.
2026; 52; ().
101143 - 101143.