基于可解释无监督机器学习的信息物理系统维护

摘要

信息物理系统（CPS）在当代技术格局中扮演着关键角色，对稳健、透明的机器学习（ML）模型的需求也因此变得迫切。本文探讨如何把可解释人工智能（XAI）的原则融入无监督机器学习（UML）方法，以增强对信息物理系统内部复杂关系的可解释性与理解。研究的重点包括：把自组织映射（SOM）作为具有代表性的无监督学习算法加以应用，以及引入可解释的机器学习方法。信息物理系统的数据由物理过程与数字部件融合而成，本身就极为复杂，本文对由此带来的挑战作了深入讨论。无监督学习中传统的黑箱做法往往妨碍人们理解模型给出的结论，因而并不适合关键的信息物理系统应用。为此，本文提出一个新的框架：在利用自组织映射这一强大的无监督方法的同时，借助可解释人工智能技术保证其可解释性。本文全面梳理了现有的可解释人工智能方法及其向无监督学习范式的迁移，并着重讨论如何为信息物理系统领域中习得的模式建立透明的表示。所提方法力求通过生成人类可理解的可视化结果与解释来提升模型的可解释性，从而弥合先进机器学习模型与领域专家之间的鸿沟。

Abstract

As cyber-physical systems (CPS) continue to play a pivotal role in modern technological landscapes, the need for robust and transparent machine learning (ML) models becomes imperative. This research paper explores the integration of explainable artificial intelligence (XAI) principles into unsupervised machine learning (UML) techniques for enhancing the interpretability and understanding of complex relationships within CPS. The key focus areas include the application of self-organizing maps (SOMs) as a representative unsupervised learning algorithm and the incorporation of interpretable ML methodologies. The study delves into the challenges posed by the inherently intricate nature of CPS data, characterized by the fusion of physical processes and digital components. Traditional black-box approaches in unsupervised learning often hinder the comprehension of model-generated insights, making them less suitable for critical CPS applications. In response, this research introduces a novel framework that leverages SOMs, a powerful unsupervised technique, while concurrently ensuring interpretability through XAI techniques. The paper provides a comprehensive overview of existing XAI methods and their adaptation to unsupervised learning paradigms. Special emphasis is placed on developing transparent representations of learned patterns within the CPS domain. The proposed approach aims to enhance model interpretability through the generation of human-understandable visualizations and explanations, bridging the gap between advanced ML models and domain experts.