Bernard Roblès

Statistical evaluation of Hidden Markov Models topologies, based on industrial synthetic model

Abstract Prediction of physical particular phenomenon is based on knowledges of the phenomenon. Theses knowledges help us to conceptualize this phenomenon throw different models. Hidden Markov Models (HMM) can be used for modeling complex processes. We use this kind of models as tool for fault diagnosis systems. Nowadays, industrial robots living in stochastic environment need faults detection to prevent any breakdown. In this paper, we wish to evaluate three Hidden Markov Models topologies of Vrignat et al. (2010), based on upstream industrial synthetic Hidden Markov Model. Our synthetic model gives us simulation such as real industrial Computerized Maintenance Management System. Evaluation is made by two statistical tests. Therefore, we evaluate two learning algorithms: Baum-Welch Baum et al. (1970) and segmental K-means Viterbi (1967). We also evaluate two different distributions for stochastic generation of synthetic HMM labels. After a brief introduction on Hidden Markov Model, we present some statistical tests used in current literature for model selection. Afterwards, we support our study by an example of simulated industrial process by using synthetic HMM. This paper examines stochastic parameters of the various tested models on this process, for finally come up with the most relevant model and the best learning algorithm for our predictive maintenance system.

Hidden Markov model framework for industrial maintenance activities

This article deals with modelization of industrial process by using hidden Markov model. The process is seen as a discrete event system. We propose different structures based on Markov automata, called topologies. A synthetic hidden Markov model is designed in order to match to a real industrial process. The models are intended to decode industrial maintenance observations (also called “symbol”). Symbols are produced with a corresponding degradation level (also called “state”). These 2-tuple (symbol, state) are known as Markov chains, also called “a signature.” Hence, these various 2-tuple are implemented in the proposed topologies by using the Baum–Welch learning algorithm (decoding by forward variable) and the segmental K-means learning (decoding by Viterbi). We assess different frameworks (topology, learning and decoding algorithm, distribution) by relevancy measurements on model outputs. Then, we determine the most relevant framework for use in maintenance activities. Afterward, we try to minimize the size of the learning data. Thus, we could evaluate the model by using “sliding windows” of data. Finally, an industrial application is developed and compared with this framework. Our goal is to improve worker safety, maintenance policy, process reliability and reduce CO2 emissions in the industrial sector.