Ph. D. Project
Dynamic Probabilistic Graphic Model applied to the prognosis and the remain useful life estimation of flow distribution system.
2017/10/17 - 2020/11/30
This thesis aims to contribute to the development of research in Health Management: management of the state of health of complex systems. In a context of operational management and system reliability, we propose to study how Dynamic Probabilistic Graphic Model (MGPD) allows the diagnosis of the state of health of a system, the prognosis of its performance and the evolution of degradations and the estimation of its remain usefull life. This thesis aims to contribute to the following scientific topics:
- modeling of complex systems for decision support in an uncertain universe: by proposing a method adapted to the new challenges of the dependability to evaluate systems of distribution of flows;
- propagation of uncertainties in the modeling of complex systems: in particular the uncertainty about the evolution of the operating environment, the uncertainty about the propagation of events in the probabilistic safety assessment, and in particular uncertainties epistemic linked to lack of knowledge;
- diagnosis of a system and the prognosis of performance, degradation and remain usefull life of a system.

For application support, we are interested in flow distribution systems: for information, energy, or materials. This support presents the interest of complexity by the number of variables but also by their interdependencies. Moreover, it is of societal interest because these systems correspond to the critical infrastructures supporting the functioning of our societies: drinking water distribution network, electricity distribution networks, heat, oil or gas distribution network.

The objective of this thesis is to study how an MGPD allows to model: the degradation of the components and the impact of these degradations on the performance of the system. The degradation of the components is generally unknown and requires a shutdown of the system to be observed which is impossible during operation of the system. However, a set of observable quantities on the system can characterize the degradation and consequently the residual life of the system.

The MGPDs allow a multi-state approach adapted to the modeling of this type of systems (Weber and Simon 2016). In addition (Thanh 2014, Tobon-Mejia 2012) propose the use of HMM (Hidden Markov Model) or HSMM (Hidden Semi-Markov Model) to model the unobservable process of degradation and link it to observations of their consequences.

We proposed contributions for the modeling of component reliability by integrating the impact of the environment and operational constraints by IO-HMM (Ben Salem 2006) as well as for modeling the impact of component reliability on the functional levels of systems by Bayesian networks (Medina-Oliva 2015). The learning and inference algorithms of the MGPDs available today make these complex models exploitable for prognosis.
We also worked on the modeling of epistemic uncertainty and incomplete knowledge in reliability and risk control (Simon 2008) and more specifically for the modeling of the performance of multi-state systems (Simon and Weber 2009), which corresponds Well to the problem of knowledge of the process of degradation of the components. Finally, we have integrated the dynamic aspect in this modeling of uncertain processes (Weber and Simon 2008).

The aim of this thesis is to translate and capitalize the experience of this previous work in a context of prognosis on the basis of a more effective MGPD given the knowledge available on the system. We propose to extend the classical modeling of models from the HMM family to the MGPD to allow a temporal propagation of the uncertainty in the form of probability intervals in order to solve the prognostic problem. This research includes the extension of learning and inference algorithms. The variants of the HMM model will be considered to integrate the operational context into the prognosis.

The issue is major because the prognosis of the degradation of a system makes it possible to define strategies of control or maintenance in relation to the remain usefull life of the system. Thus it reduce the probability of occurrence of a shutdown due to malfunction of the system and by reducing the rate of degradation with a preventive maintenance plan or by proactively maintenance interventions.

Weber P., Simon C., Systems Dependability Assessment: Benefits of Bayesian Networks Models. Wiley-ISTE, 2016, ISBN: 978-1-84821-992-2
Medina-Oliva G., Weber P., Iung B. Industrial system knowledge formalization to aid decision making in maintenance strategie sassessment. EAAI, 37, 2015.
Thanh Trung Le, Florent Chatelain, Christophe Berenguer. Hidden Markov Models for diagnostics and prognostics of systems under multiple deterioration modes. ESREL 2014, Poland
Diego Tobon-Mejia, Kamal Medjaher, Noureddine Zerhouni, Gérard Tripot. A data-driven failure prognostics method based on mixture of gaussians hidden markov models. IEEE Trans. on Reliability, 2012, 61 (2) <10.1109/TR.2012.2194177>
Simon C., Weber P. Evidential networks for reliability analysis and performance evaluation of systems with imprecise knowledge. IEEE Trans. on Reliability, 58(1), 2009, DOI : 10.1109/TR.2008.2011868
Simon C., Weber P., Evsukoff A.G. Bayesian networks inference algorithm to implement Dempster Shafer theory in reliability analysis. Special Issue "Bayesian networks in dependability", in RESS, 93(7) 2008, DOI : 10.1016/j.ress.2007.03.012
Weber P., Simon C. Dynamic evidential networks in system reliability analysis: A Dempster Shafer approach. 16th Mediterranean Conference on Control and Automation, Ajaccio, France, 2008.
Ben Salem A., Muller A., Weber P. Dynamic Bayesian Networks in system reliability analysis. 6th IFAC SAFEPROCESS, Beijing, P.R. China, 2006.
Prognosis, remain usefull life, reliability, dynamic probabilistic model.
Eco-Technic systems engineering