Federated Smart Work Package Framework with Triplet Loss for Mental Fatigue Monitoring

Abstract

Monitoring crane operators’ mental fatigue is critical for construction occupational health and safety, as their work demands constant alertness. Fatigue manifests as a highly personalized phenomenon, complicating detection efforts. While deep learning models excel at identifying fatigue patterns, strict privacy regulations such as the European Union’s General Data Protection Regulation hinder the use of centralized data repositories for training. These challenges underscore the urgent need for personalized and privacy-preserving fatigue detection frameworks, particularly for high-risk roles such as crane operation. Previous studies have explored approaches such as the federated transfer learning-enabled smart work packaging (FedSWP). Still, these methods often rely on large volumes of labeled user-specific data for fatigue status, which is impractical in the real world. This paper introduces an adaptive and lightweight federated smart work package framework with triplet loss (FedSWP-TL), significantly reducing the required amount of manually labeled mental fatigue data from individual crane operators. By leveraging triplet networks and efficient methods such as compressive aggregation and a lighter MobileVit network architecture, the FedSWP-TL demonstrates enhanced generalization and mobility capabilities even with minimal labeled samples. Results on the YAWDD, DROZY, and ConPPMF data sets demonstrate FedSWP-TL’s adaptability to diverse groups, achieving recall improvements of 0.07, 0.11, and 0.09, respectively, over baseline methods while maintaining less overhead for 2.32M parameters. This underscores its potential for real-world deployment in scenarios plagued by target-labeled data scarcity and constrained resources.