[论文解读] A Review of Physics-Informed Machine Learning Methods with Applications to Condition Monitoring and Anomaly Detection

Yuandi Wu, Brett Sicard+1 more|arXiv (Cornell University)|Jan 22, 2024

Non-Destructive Testing Techniques被引用 6

一句话总结

本综述汇总了面向条件监测和异常检测的物理信息机器学习（PIML）方法，阐述了物理定律如何嵌入到机器学习模型中，并概述了常见架构、优点及案例研究。

ABSTRACT

This study presents a comprehensive overview of PIML techniques in the context of condition monitoring. The central concept driving PIML is the incorporation of known physical laws and constraints into machine learning algorithms, enabling them to learn from available data while remaining consistent with physical principles. Through fusing domain knowledge with data-driven learning, PIML methods offer enhanced accuracy and interpretability in comparison to purely data-driven approaches. In this comprehensive survey, detailed examinations are performed with regard to the methodology by which known physical principles are integrated within machine learning frameworks, as well as their suitability for specific tasks within condition monitoring. Incorporation of physical knowledge into the ML model may be realized in a variety of methods, with each having its unique advantages and drawbacks. The distinct advantages and limitations of each methodology for the integration of physics within data-driven models are detailed, considering factors such as computational efficiency, model interpretability, and generalizability to different systems in condition monitoring and fault detection. Several case studies and works of literature utilizing this emerging concept are presented to demonstrate the efficacy of PIML in condition monitoring applications. From the literature reviewed, the versatility and potential of PIML in condition monitoring may be demonstrated. Novel PIML methods offer an innovative solution for addressing the complexities of condition monitoring and associated challenges. This comprehensive survey helps form the foundation for future work in the field. As the technology continues to advance, PIML is expected to play a crucial role in enhancing maintenance strategies, system reliability, and overall operational efficiency in engineering systems.

研究动机与目标

• 激发在工程系统中使用PIML以应对数据稀缺性和对物理一致预测的需求。
• 整理将物理知识嵌入到ML模型中的方法（特征、正则化、架构）。
• 从准确性、可解释性、计算效率和泛化能力等方面评估PIML方法的优点与局限性。
• 突出代表性案例研究，展示PIML在条件监测和异常检测中的有效性。

提出的方法

• 对PIML整合框架进行分类：物理嵌入特征空间、基于数据对物理模型的改进、物理信息正则化，以及物理引导的架构设计。
• 描述物理定律如何被纳入模型输入、损失函数和网络架构（例如PINNs、基于物理的特征扩增）。
• 讨论通过基于物理的仿真进行合成数据生成以及半监督或迁移学习以扩充训练数据。
• 通过表格和图形总结文献，勾勒出架构和应用领域。
• 给出有限元模型及其他物理工具生成训练数据或引导特征形成的示例。
• 在可解释性、计算和泛化能力方面解释每种方法的权衡。

实验结果

研究问题

• RQ1 在条件监测和异常检测中，将物理学与数据驱动模型整合的主要方法是什么？
• RQ2 每种PIML框架的优点、局限性及实际考虑因素（计算成本、可解释性、数据需求）有哪些？
• RQ3 PIML方法在不同工程应用和故障情景中的表现如何？
• RQ4 关于PIML的架构设计和数据合成，最近文献中出现了哪些趋势？

主要发现

• 通过在学习中强制执行物理原理，PIML 提供了更好的可解释性和鲁棒性。
• 特征空间物理嵌入和物理引导的架构可以降低对数据的需求并提高泛化性。
• 在真实标注数据稀缺时，使用基于物理模型的合成数据和迁移学习可提升训练效果。
• 在条件监测任务中使用了多种架构（PINNs、物理信息CNN/RNN/GNN、以及物理引导的正则化）。
• 有限元模型和其他物理仿真器常作为数据源或特征形成与模型标定的引导。