Industry 5.0 is a new era for intelligent manufacturing characterized by human-machine collaboration, adaptability and decentralized intelligence. Integration of heterogeneous and multi-modal sensor data becomes vital in such environments for real-time reliable decision-making on various scales. However, conventional data fusion techniques, founded on static rules and a centralized architecture, fail to accommodate dynamic sensor behavior and contextual variability and the necessity for human oversight. To address these limitations, this work proposes federated, adaptive and human-in-the-loop Data Fusion (FADF), a novel multi-layered framework that integrates multi-criteria decision analysis, Markov decision processes, reinforcement learning and federated learning to enable real-time, context-aware sensor data fusion in edge-based architectures. FADF dynamically adjusts fusion rules by considering sensor reliability, data quality and human feedback, resulting in enhanced accuracy and adaptability. Experimental validation using a synthetic IIoT dataset shows that FADF outperforms conventional methods (Kalman filter, CNN fusion and federated averaging) achieving a 9.2% improvement in F1-score, demonstrating its robustness, scalability and relevance for real-world industry 5.0 applications.