Application des Modèles de Markov Cachés Dans la Classification des Défauts de Machines Tournantes

Abstract

In this paper, we test the application of hidden Markov models (HMM) in the classification of defects in rotating machines. The HMMs are a modeling tool that has proven itself particularly in the field of speech processing, image processing and analysis of biological sequences. They are probabilistic finite state automata. In this work, we have considered only the discrete HMMs. With temporal indicators extracted from vibration signals, we constructed matrices characterizing each class representing a state of health of the machine in question (supervised classification) as an attribute which is at the same time observing or observable state of HMM, which corresponds to a hidden state has determined by a probabilistic approach, this approach is called Markovian modeling . Learning of the HMM was performed by the Baum-Welch algorithm based on maximum likelihood (LL). The resulting system has been designed purpose in the form of a toolbox HMM-MAG 10. The results obtained demonstrate the reliability and effectiveness of this model nevertheless reserves conditioning this application remain posted on the choice of indicators and their sensitivity to the signal, the construction of the dataset, the number of states etc.. ...