Abstract
Early detection of leaks, bursts, process faults, and cyber-physical attacks is a direct lever for the sustainable management of water infrastructures: it reduces non-revenue water, the energy spent on treating and pumping lost water, and service interruptions (Sustainable Development Goal 6). Anomaly detectors for water distribution networks (WDNs) and wastewater treatment plants (WWTPs) are increasingly deployed through Artificial Intelligence of Things (AIoT) architectures, yet a detector that performs well on one system cannot be assumed to transfer to another with a different topology, sensor layout, demand pattern, and label availability. This systematic review examines the literature through this transfer problem, with graph neural networks (GNNs) as topology-aware models, tabular foundation models (TFMs; in the retrieved evidence, the TabPFN family) as candidate local screeners, federated learning as privacy-preserving adaptation, and digital twins and explainable artificial intelligence as validation and trust layers. It follows PRISMA 2020 and is registered retrospectively in the Open Science Framework (DOI 10.17605/OSF.IO/D7RXU). Peer-reviewed articles, theses, and preprints in English or Spanish were identified through modular block searches (Google Scholar, publisher sites, three Spanish repositories), author-supplied sources, and citation chasing, verified against primary records on 6–7 September 2026, screened in two stages, coded with a predefined framework, appraised with a six-item checklist, and synthesized descriptively because tasks, datasets, and metrics are not commensurable. The corpus comprises 150 unique works: 76 water-specific experimental studies, 21 benchmark or testbed studies, and 53 method, review, or contextual works. Graph-based approaches have the stronger water-specific evidence for topology-dependent tasks in WDNs, whereas TabPFN-family TFMs rest on general tabular-learning evidence and one water-adjacent prediction study, with no validated water anomaly-detection use. Only 6 of the 76 experimental studies evaluate an explicit transfer mechanism, federated evidence is limited to one water study, and no study reports an operator evaluation of explanations. The review contributes a transfer taxonomy with study counts, an evidence-class scheme, a gap map of demonstrated and proposed components on a transfer-aware AIoT layout, and an evaluation agenda based on source–target protocols, and identifies the evidence utilities need before transferable detectors can support sustainable operation at scale.