机器学习模型可解释性研究及其在PHM中应用现状综述  

A Review on Interpretable Machine Learning and Its Application on PHM

作　　者：周登极[1] 郝佳瑞 黄大文 ZHOU Deng-ji;HAO Jia-rui;HUANG Da-wen(School of Mechanical Engineering,Shanghai Jiao Tong University,Shanghai 200240,China)

机构地区：[1]上海交通大学动力机械及工程教育部重点实验室,上海200240

出　　处：《系统工程》2022年第6期1-10,共10页Systems Engineering

基　　金：国家青年人才项目(2019QNRC001);国家自然科学基金资助项目(51706132)。

摘　　要：得益于机器学习理论及技术的高速发展,数据驱动的预测及健康管理(Prognostic and Health Management, PHM)系统能从高维的多源异构数据中学习设备不同状态的特征和趋势,进而完成诊断和预测任务。然而,复杂机器学习模型的不透明性也引发了使用者对模型的信任问题,且设计者难以根据自身知识针对性改进模型。为此,本文总结了机器学习可解释性研究进展,归纳了其研究范式,概述了模型可解释性在PHM领域中的应用,并对比了当前可解释性研究在机器学习和PHM中的研究模式,提出未来的一些研究设想。Nowadays, machine learning theory and techniques have been extensively developed. Therefore, data-driven Prognostic and Health Management system applying the machine learning algorithm is able to learn the operation pattern from multi-sourced heterogeneous data to facilitate the Fault Diagnostic and Condition Forecasting procedures for industrial equipment. However, transparency of complex machine learning algorithm has lead to trust problem and difficulty in model refinement using experts’ knowledge. To tackle the above issues, research status and paradigm of interpretable machine learning were summarized, then its applications on PHM were introduced. At the end of the paper, different methods of model interpretability research on machine learning and PHM were compared and some future trends on PHM research were proposed.

关 键 词：机器学习 可解释性 PHM 深度学习 故障诊断 特征选择 

分 类 号：N945[自然科学总论—系统科学]