可解释人工智能在工业智能诊断中的挑战和机遇:归因解释
Challenges and Opportunities of XAI in Industrial Intelligent Diagnosis:Attribution Interpretation
严如强 1周峥 1杨远贵 1李亚松 1胡晨烨 1陶治宇 2赵志斌 1王诗彬 1陈雪峰1
作者信息
- 1. 西安交通大学机械工程学院 西安 710049
- 2. 广州航新航空科技股份有限公司 广州 510663
- 折叠
摘要
针对目前迅猛发展的工业智能诊断方法缺乏可解释性开展综述,指出模型无关的归因解释技术在工业智能诊断中的研究现状和潜在研究方向.分析可解释性技术的主要观点和作用,针对工业智能诊断的两个特性问题—非线性高维观测、知识表征精度低,归因解释技术可以提供有效的前向理解智能模型逻辑结构、反向优化模型设计的工具.从注意力机制、显著性分析、规则提取、代理模型四个方面,概述其主要观点与作用,介绍现有方法的研究现状,并总结分析不同归因解释技术的优势与不足.通过四个案例分析,阐述不同归因解释技术在智能诊断中的效果.最后展望归因解释技术在工业智能诊断中的研究方向,包括可解释性量化、反馈模型设计、模型复杂性与可解释性平衡、高维特征的归因解释,期望为可解释人工智能技术在工业智能诊断中的发展提供方向建议.
Abstract
The purpose is to figure the lack of interpretability for current industrial intelligence diagnosis methods,review the development situation of model-agnostics attribution analysis in industrial intelligence diagnosis and point out the potential development direction.The main viewpoints and functions of interpretable techniques are analyzed.Aiming at two characteristic problems of industrial intelligence diagnosis,i.e.,nonlinear high-dimensional observation and inaccurate knowledge representation,attribution interpretation provides effective methods for understanding forward logical structure and reverse optimizing design of intelligent models.The core concepts,existing works and pros and cons of attention mechanism,saliency analysis,rule extraction,and proxy model are systematically summarized.Four case studies are used to illustrate the result of attribution interpretation techniques.Finally,potential research directions of attribution interpretation technology in industrial intelligent diagnosis are discussed,including quantification of interpretability,feedback to model design,balance between model complexity and interpretability,and attribution analysis in high dimension.Through this review,we hope to provide a suggestion to conduct further development of interpretable intelligence in industrial fault diagnosis.
关键词
工业智能诊断/可解释性/模型无关/归因分析Key words
industrial intelligence diagnosis/interpretability/model-agnostic/attribution analysis引用本文复制引用
基金项目
国家自然科学基金重点资助项目(51835009)
出版年
2024
NETL
NSTL
万方数据