Feasibility analysis of unsupervised industrial activity recognition based on a frequent micro action
Behrooz Azadi, Michael Haslgrübler, Georgios Sopidis, Michaela Murauer, Bernhard Anzengruber, Alois Ferscha · 2019
The ubiquity of wearable sensors has contributed a lot in human activity recognition. Although activities of daily living have been studied over the past decades, there is a lack of such efforts on workers' activities in manufacturing industry. In this paper, we look at a simple yet frequent task for manufacturing, screwing, and investigate the ability to recognize this repetitive task through wearable sensors and unsupervised learning, by describing an experiment for industrial activity recognition and showing how effective clustering analysis is in detecting such a frequent micro action, by comparing different techniques throughout a machine learning pipeline. The achieved results demonstrated that unsupervised learning is a good solution to deal with large amount of unlabelled sensory data. We also show the different stages of pre-preprocessing, data normalization, segmentation, dimensional reduction and algorithm selection and how this affects the overall outcome by provide an in depth view, on state of the art techniques, by applying it on a relevant industrial problem.