Cyber-Physical Systems (CPS) are integral components of modern, interconnected environments, playing a crucial role in various applications such as smart infrastructure and autonomous systems. These systems operate in complex and ever-evolving settings, where resilience is an essential characteristic. Recently, several Machine Learning (ML) techniques have been proposed to design and implement resilience in CPS. However, this approach has often prioritized performance gains and adaptability enhancements, leading to a disregard for the potential dangers linked to centralized data processing, including data breaches and privacy infringements. To harness the full potential of ML in CPS, this paper introduces a distributed intelligence framework that equally prioritizes security, data privacy, and adaptability. The proposed framework is implemented through the integration of Federated Machine Learning techniques, where the CPS architecture is decentralized, allowing data processing to occur locally on individual nodes or devices. This decentralized approach facilitates the aggregation of insights from multiple sources without the need for centralized data processing, thereby minimizing the risks associated with data breaches and privacy violations. We further validate the viability of the proposed framework through its successful implementation in a real-world industrial CPS application, specifically focused on fault prediction within industrial CPS environments. In addition to its privacy and security benefits, our approach also achieved promising results in terms of accuracy and precision.

网络物理系统（CPS）是现代互联环境的组成部分，在智能基础设施和自治系统等各种应用中发挥着至关重要的作用。这些系统在复杂且不断变化的环境中运行，其中弹性是一个基本特征。最近，已经提出了几种机器学习（ML）技术来设计和实现 CPS 中的弹性。然而，这种方法通常优先考虑性能提升和适应性增强，导致忽视与集中式数据处理相关的潜在危险，包括数据泄露和隐私侵犯。为了充分发挥 ML 在 CPS 中的潜力，本文引入了一种分布式智能框架，该框架同样优先考虑安全性、数据隐私和适应性。所提出的框架是通过集成联合机器学习技术来实现的，其中 CPS 架构是分散的，允许数据处理在各个节点或设备上本地进行。这种去中心化的方法有助于聚合来自多个来源的见解，而无需集中数据处理，从而最大限度地降低与数据泄露和隐私侵犯相关的风险。我们通过在现实世界的工业 CPS 应用中的成功实施，进一步验证了所提出的框架的可行性，特别关注工业 CPS 环境中的故障​​预测。除了隐私和安全方面的优势外，我们的方法在准确性和精确度方面也取得了有希望的结果。