Secure Artificial Intelligence for Asset Performance in Critical Infrastructure: A Critical Narrative Review of Safety, Operational Excellence and Continuous Quality Improvement

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Critical infrastructure operators increasingly deploy artificial intelligence (AI) to detect degradation, predict failures, optimise maintenance, identify cyber anomalies and support operational decisions. These capabilities create a plausible route to higher availability, reliability and quality, yet they also couple asset-performance decisions to data integrity, model uncertainty, software supply chains, human oversight and operational technology security. This critical narrative review examines how secure AI can contribute to asset performance without weakening safety or resilience. Literature spanning predictive maintenance and prognostics, industrial AI, digital twins, industrial cyber-physical security, adversarial machine learning, explainability, uncertainty, machine-learning operations and human factors was critically synthesised. The evidence is strongest for AI as a decision-support layer for condition monitoring, fault diagnosis and targeted maintenance where data provenance is controlled, failure modes are sufficiently represented and recommendations remain bounded by engineering constraints. Evidence for fully autonomous optimisation in high-consequence infrastructure is less mature because benchmark accuracy does not directly establish operational utility, transferability or safe behaviour under distribution shift and hostile manipulation. Digital twins can improve contextual diagnosis and testing, but their value depends on fidelity, synchronisation, governance and protection of the data-model-actuation pathway. Cybersecurity and safety therefore cannot be treated as independent assurance domains: poisoned training data, adversarial inputs, compromised updates or unavailable models can become physical reliability and quality risks. Explainability alone is also insufficient for assurance; uncertainty calibration, abstention, independent validation, auditability and meaningful human authority are required. The synthesis proposes a secure asset-performance loop in which AI recommendations are evaluated against safety, security, service and quality constraints, with outcomes fed back into monitored model and process improvement. The central implication is that asset performance should be maximised as a constrained socio-technical objective rather than as an unconstrained prediction or utilisation metric.