基于物理软约束的可解释多模态深度学习模型,用于核电厂中压水反应堆冷却液系统的小样本故障诊断

A physical soft-constraint interpretability-based multimodal deep learning model for small-sample fault diagnosis of pressurized water reactor coolant systems in nuclear power plants

ES评分 0

DOI 10.1016/j.eswa.2026.133978
刊名
年,卷(期) 2027, 333()
作者
作者单位 南华大学

摘要
冷却液事故损失(LOCA)是加压水反应堆(PWR)中一次关键的设计基体事故,目前智能诊断面临故障样本稀缺和模型可解释性差的挑战。为此,本文提出了一种物理软约束的可解释多模态深度学习(PSCI-MDL)模型。首先,采用压力流比、非负数和温度边界约束的多尺度数据增强方法,将30个原始样本扩展为330个物理一致性样本。随后构建了多模态功能融合网络,将Mamba集成用于时间序列建模,CNN用于参数化特征,并实现物理特征处理。通过离散微分方程,将物理软约束作为正则术语嵌入其中,将保护定律转化为可微分损耗函数,使模型与工程原理保持一致。故障解释器通过梯度分析和物理规则引擎生成自然语言诊断推理。实验表明,该方法将LOCA的诊断准确率从45%提高到99.5%以上,并减少了超过32%的物理约束损失,表明该模型的预测更符合核安全分析标准所规定的基本保护定律(质量、能量和动量)。可解释性验证表示冷腿LOCA的压力特征权重为0.48(±0.03),而“严重原压降(低于30%)”等解释表明,LOCA在专家判断下实现96.7%的一致性。本工作为核电站的小样本故障诊断提供了高精度且可解释的解决方案。
Abstract
Loss of Coolant Accident (LOCA) is a critical design-basis accident in Pressurized Water Reactors (PWRs), with current intelligent diagnosis facing challenges of scarce fault samples and poor model interpretability. To address this, this paper proposes a Physical Soft-constraint Interpretable Multimodal Deep Learning (PSCI-MDL) model. First, a multi-scale data enhancement method incorporating pressure-flow ratio, non-negativity, and temperature boundary constraints expands 30 original samples to 330 physically consistent ones. A multimodal feature fusion network is then built, integrating Mamba for time-series modelling, CNN for parametric features, and physical feature processing. Innovatively, physical soft constraints are embedded as regularization terms via discretized differential equations, translating conservation laws into differentiable loss functions to align the model with engineering principles. A fault interpreter is developed to generate natural-language diagnostic reasoning through gradient analysis and a physics rule engine. Experiments show the method improves LOCA diagnosis accuracy from 45% to over 99.5% and reduces physical constraint loss by more than 32%, indicating that the model's predictions more closely adhere to the fundamental conservation laws (mass, energy, and momentum) that are mandated by nuclear safety analysis standards. Interpretability validation indicates a pressure feature weight of 0.48 (+/- 0.03) for cold-leg LOCA, and explanations such as "severe primary pressure drop (below 30%) suggests LOCA" achieve 96.7% consistency with expert judgment. This work offers a high-accuracy, interpretable solution for Small-sample Fault Diagnosis in nuclear power plants.
关键词
冷却剂事故损失(LOCA) ; 物理软约束 ; 多模态深度学习 ; 可解释性 ; 小样本故障诊断
KeyWord
Loss of Coolant Accident (LOCA)Physical Soft-constraintMultimodal Deep LearningInterpretabilitySmall-sample Fault Diagnosis
基金项目
页码 133978-133978
  • 参考文献
  • 相关文献
  • 引用本文

Keshun Y, Haidon S, Yajun W, et al. 基于物理软约束的可解释多模态深度学习模型,用于核电厂中压水反应堆冷却液系统的小样本故障诊断 [J]. Expert Systems with Applications. 2027; 333; (). 133978 - 133978.

  • 文献评论

相关学者

相关机构