Online Intrusion Detection for Industrial Cyber-Physical Systems Based on Self-Supervised Predictive Model

Industrial cyber-physical systems (CPS) demand a reliable intrusion detection despite evolving operational conditions and a lack of labeled attack data. This article proposes a novel context–target predictive self-supervised framework for online intrusion detection in industrial CPS. The models normal behavior by predicting future sensor states from past context, allowing detection of anomalies as large prediction errors. To address concept drift, we incorporate a trust-aware online adaptation mechanism and a robust, statistics-based thresholding strategy that adjusts in real time. We validate the method on a fluid catalytic cracking unit simulation, where it achieves high precision and recall under fully online, unlabeled conditions, outperforming baseline models. The proposed framework provides an adaptive and lightweight intrusion detection systems suitable for autonomous industrial CPS monitoring without human intervention or costly labeling.