Supporting Predictive Maintenance through a semi-supervised labelling of robot cycles

Predictive maintenance has always been a difficult problem in real word modern industries. With the new Industry 4.0 the manufacturing environments are becoming digital factories and this context produce vast volumes of raw data. Enhancing manufacturing intelligence brings a wide range of benefits, predictive diagnostics is one of the most important. To sustain this well known issue, this work present the design and development of a semi-supervised data-driven methodology, to characterize multi-cycle processes and support robot cycle labelling. The latter exploits the best developed clustering algorithms, discovering automatically clusters of production cycles through time-independent common properties. Later, a self-tuning strategy has been integrated to help the selection of the best approach, input data and parameter. Finally, each cluster is locally characterized from a set of most relevant features.