A novel hybrid strategy based on Swarm and Heterogeneous Federated Learning using model credibility awareness for activity recognition in cross-silo multistorey building

The novel HAR-SHFDL system leverages a Swarm Heterogeneous Federated Deep Learning framework for smartphone-based Human Activity Recognition (HAR). Unlike traditional Federated Learning (FL) with high communication overhead, HAR-SHFDL utilizes a swarm group architecture where smartphone users on each floor elect a swarm representative to communicate with the heterogeneous federated learning. Within these swarm groups, individuals train models and extract logits, which are weighted and averaged based on performance. The swarm representative transmits these logits to the Building Side Unit (BSU) on each floor, which refines and checks logits credibility before sending them to a heterogeneous federated server. This heterogeneous server aggregates information from all BSUs to update the global HAR model. Experiments show HAR-SHFDL outperforms existing FL frameworks in accuracy, precision, recall, and F1-score, with accuracy improvements up to 6.23%. The importance of HAR-SHFDL extends beyond human activity recognition. Its core strengths communication efficiency, robust model aggregation, and federated learning can be applied to various practical applications, such as optimizing building management systems, enhancing security, improving healthcare and fitness monitoring, and developing federated recommendation systems and anomaly detection in sensor networks. HAR-SHFDL offers a versatile framework for building intelligent systems that enhance daily life and promote efficiency and security.