Deployment-oriented defence-in-depth for legacy industrial control systems under operational constraints: A systematic review of detection and compensating controls

Industrial Control Systems (ICS) in critical infrastructure remain highly exposed to cyber threats while operating under technical, operational, and regulatory constraints that often limit direct modification, update, or replacement. Existing surveys either provide broad overviews of ICS security or focus narrowly on Artificial Intelligence (AI)/Machine Learning (ML) methods, but rarely translate the evidence into deployment-oriented guidance for protecting legacy systems in brownfield environments. This paper presents a systematic review of 84 peer-reviewed studies published between 2021 and 2026, examining how AI/ML-based anomaly detection and compensating security controls can be combined within defence-in-depth architectures for legacy ICS. The review organises detection approaches by deployment zone within the Purdue architecture and learning paradigm, and relates major categories of compensating controls, such as network segmentation, unidirectional gateways, virtual patching, and Operational Technology (OT)-adapted Zero Trust architectures, to the IEC 62443 Security Levels. The evidence shows that no single detection architecture dominates across deployment scenarios, that benchmark performance alone does not establish operational readiness, and that AI/ML detection works best as a layer augmenting foundational controls rather than replacing them. It also identifies persistent weaknesses in cross-dataset validation and the continuing gap between reported model performance and real-world deployability. This review translates the literature into deployment-oriented guidance for asset owners, plant operators, and OT security architects, showing how detection and compensating controls can be combined across Purdue zones and IEC 62443 Security Levels to improve visibility, containment, and resilience without compromising operational continuity in legacy industrial environments.