Evaluation of Minimal Data Size by Using Entropy, in a HMM Maintenance Manufacturing Use

Bernard Roblès (1) , Manuel Avila (1) , Florent Duculty (1) , Pascal Vrignat (1) , Stéphane Begot (1) , Frédéric Kratz (1)

1 PRISME - Laboratoire Pluridisciplinaire de Recherche en Ingénierie des Systèmes, Mécanique et Energétique

In this paper, we wish to find a minimal data size in order to better conceptualize industrial maintenance activities. We based our study on data given by a Synthetic Hidden Markov Model. This synthetic model is intended to produce real industrial maintenance observations (or "symbols"), with a corresponding degradation indicator. These time series events are shown as Markov chains, also called "signatures". The production of symbols is generated by using a uniform and a normal distribution. The evaluation is made by applying Shannon entropy on the HMM parameters. The results show a minimal number of data for each distribution studied. After a discussion about the use of a new "Sliding Window" of symbols usable in a Computerized Maintenance Management System, we developed two industrial applications and compare them with the best optimized "signature" previously found.

Mots clés: Hidden Markov Models; Shannon entropy; model optimization; learning algorithm; predictive maintenance

Domaines: Automatique

Bernard Roblès, Manuel Avila, Florent Duculty, Pascal Vrignat, Stéphane Begot, et al.. Evaluation of Minimal Data Size by Using Entropy, in a HMM Maintenance Manufacturing Use. MIM'2013 Manufacturing Modelling, Management and Control, Jun 2013, Saint Petersburg, Russia. pp.1. ⟨hal-00839857⟩