Multimodal Measurement and Diagnosis for Traceable AI for Water and Energy Infrastructure: Lifecycle Evidence and Accountable Decisions

Abstract

Water and energy infrastructure produce heterogeneous evidence across sensors, images, laboratory tests, maintenance records, operator notes, and network topology. This review develops a traceable multimodal AI architecture for converting that evidence into lifecycle diagnosis. Temporal and graph representations combine local equipment behavior with system dependencies, while uncertainty-aware fusion preserves conflicts that may indicate degradation, attack, sensor failure, or an invalid data pipeline. A cloud-edge design keeps time-critical inference near assets and supports federated learning when operators cannot centralize sensitive data. Provenance links every feature and model update to its source; policy controls, tamper-evident logs, and human override connect diagnosis to accountable action. The synthesis evaluates missing modalities, domain shift, calibration, false-alarm burden, and revalidation after equipment, software, or organizational change. No new infrastructure trial is claimed. The article contributes a lifecycle AI-governance blueprint in which multimodal performance, privacy, traceability, and operational resilience are designed together rather than added as separate compliance layers.