A Multimodal IMU-RTLS Manufacturing Motion Dataset for Worker Activity Recognition

Worker Activity Recognition can improve operational visibility in human-centric manufacturing systems. However, existing approaches remain limited by scarce industrial datasets, isolated task recordings, and sensor-intensive configurations. This study presents two multimodal datasets and a cyber-physical architecture for complementary levels of worker motion. The Gross Worker Activity Recognition dataset targets operator proximity to workstation and the identification of logistic activities in reconfigurable manufacturing systems. It contains 814,065 positioning samples from eight production setups, each repeated twice, involving more than 40 workers. This dataset is benchmarked with different learning structure algorithms, and the Long Short-Term Memory is selected as a baseline model, achieving an F1-score of 86%. The Fine Worker Activity Recognition dataset targets the segmentation of five pump-assembly operations using two inertial measurement units and a radio-frequency-identification smart glove. It comprises approximately 7 h of recordings and more than one million samples from 50 complete assemblies performed by three workers. Transformer, ResNet, and SqueezeNet models are benchmarked, with SqueezeNet selected as the baseline after achieving the highest F1-score of 85.2%. Moreover, the results identify transition-rich activities as the principal recognition bottleneck and show how complementary motion levels support different manufacturing indicators.