A transfer-federated learning framework integrating toxicological risk modeling for heterogeneous UAV, UGV, and IoT air-quality monitoring

Existing air quality monitoring systems rely on static sensors, failing to capture personal exposure dynamics and hindering secure cross-platform data sharing. To address these limitations, this study introduces a novel Transfer-Federated Learning (TFL) framework for collaborative learning among heterogeneous devices—an unmanned ground vehicle (UGV), an unmanned aerial vehicle (UAV), and fixed IoT stations—while preserving data privacy. The TFL approach uniquely adapts a pretrained model to clients with different sensor configurations through layer-wise transfer and FedAvg aggregation. Lightweight models (Tiny-TCN, TFT-lite) predict both pollutant concentrations (PM₂.₅, PM₁₀, CO₂, CH₂O, VOCs) and mechanistic toxicological risks (inhalation dose and hazard quotient). Field tests across three industrial sites demonstrated high cross-platform accuracy (global R² = 0.984) and real-time inference (<2 s), reducing personal exposure estimation error by over 40% compared to static fixed-sensor baselines. The core contribution lies in the system-level integration of transfer learning and federated learning across heterogeneous platforms, with toxicological metrics used to translate predictions into actionable occupational risk indicators. This work establishes a new paradigm for privacy-preserving, health-aware monitoring in industrial environments.