Process Model-guided Post-filtering for Wearable Activity Recognition in Ergonomic Work Processes

Ergonomic assessment of manual work processes requires continuous recognition of workers' physical activities – a task complicated by privacy constraints, sensor complexity, and noisy movement data. This paper proposes a process model-guided post-filtering approach that leverages existing process knowledge to refine the output of wearable activity classifiers, regardless of the underlying recognition method. Evaluated across two industry-derived use cases using body area networks with 3, 5, and 7 sensors, the approach yields dramatic accuracy improvements – from below 50 % unfiltered to over 82 % with as few as three sensors. Results demonstrate that explicit process model knowledge can substantially compensate for reduced sensor setups, lowering hardware costs and privacy risks without sacrificing recognition quality. This proof-of-concept establishes process model filtering as a promising component in ergonomic monitoring pipelines for structured manual work.