Dynamic Diagnosis and Predictive Maintenance of Train Onboard Systems – DIADEM

The first component of the work achieved within the framework of DIADEM is the dynamic diagnosis via pattern recognition approaches. This kind of methods consists in predicting the current operating state of the system and analyzing its evolution in the course of time. The innovative character of the current project lies in the (potentially) dynamic nature of the diagnosis, which is taken into account through the classes’ parameters which are allowed to evolve in the course of time. The dynamic classification problem is addressed by using the generic formalism of mixture models where, in contrast to the classical situation of stationary data, the classes’ parameters are stochastically modeled (for instance, autoregressive processes are used to model the classes means).
The second one is dedicated to predictive maintenance (also called prognosis). Due to the increasing number of onboard diagnosis devices, providing more and more precise monitoring for the main sensitive sub-systems of rolling stocks , manufacturers (and operators) are assigned to rethink and strongly improve the way their maintenance strategies are settled. In this way, the developpement of more precise degradation models allows, in collaboration with diagnosis processes, to estimate and regularly update the future behavior of the system and thus to foresee when its state would reach an unwanted degradation level. With such an information (called Remaining Useful Life or Remaining Operating Time) it is possible to optimise the maintenance actions schedule, determining the best time to maintain a system in respect of both its life expectancy and operating constraints. If Dynamic Bayesian Networks are commonly used for «standard« industrial maintenance applications, the integration of prognosis approaches to this formalism represent one of the main innovations of this task.

Considering an approache based on the use of sojourn time distributions, a first prognosis algorithm was proposed for the estimation of the remaining usefull life (RUL) of perdiodically observed systems with discrete and finite states space. In this work, sojourn time distributions in each state, caracterizing the system degradation dynamic, can be learnt from return of experience databases. A specific dyanmic bayesian network structure was proposed to integrate this prognosis algorithm in tha bayesian network frame. The obtained model provide an intial estimation of the RUL, updated after each new diagnosis of the system operating state. Finally, an extension of the Graphical Duration Models, used for the degradation process modelling, was proposed, named GDM-CSTD. This new approach integrates the notion of conditional sojourn time distributions, allowing to take into account some multi-dynamics degradation processes. Both the prognosis algorithm and the bayesian network structure were adapted to the GDM-CSTD extension.
A bibliographical study on the existing dynamic diagnosis approaches leads us to the implementation of two classification approaches: the dynamic logistic regression where the logistic coefficients are stochastically modeled and a Gaussian mixture model whose classes centers evolves according to a polynomial regression model. Some limitations of the latter approaches lead us to consider a more general formalism based on state space models which resulted in a specific algorithm for classifying non stationary sequential data. In parallel with this work, a generic methodology for multivariate time series segmentation has been developed to extract relevant features from raw signals.

Dynamic Bayesian Network integrating RUL estimation algorithms for discrete and finite states space systems.
Degradation processes modeling taking into account multi-dynamics by use of a conditional sojourn time distributions approach.
Proposal of a generic formalism based on state space models for dynamic unsupervised classification. Development of a sequential classification algorithm to deal with potentially non stationary data. Feature extraction via multivariate time series segmentation approach based on a hidden logistic process.

H. El Assaad, A. Samé, G. Govaert, P. Aknin (2015). A variational Expectation Maximisation algorithm for temporal data clustering, Computational Statistics & Data Analysis (CSDA). En révision.
J. Foulliaron, L. Bouillaut, P. Aknin, A. Barros. A dynamic Bayesian network approach for modelling and optimizing predictive maintenance policies. Article en phase de relecture pour une publication dans Proceedings of the Institution of Mechanical Engineers, Part O, Journal of Risk and Reliability.
J. Foulliaron, L. Bouillaut, P. Aknin, A. Barros. An extension of Graphical Duration Models integrating conditional sojourn time distributions. Article à soumettre dans Proceedings of the Institution of Mechanical Engineers, Part O, Journal of Risk and Reliability.
J. Foulliaron, L. Bouillaut, A. Barros, P. Aknin (2015). ‘Dynamic bayesian networks for reliability analysis: from a Markovian point of view to semi-markovian approaches’, 9th IFAC Symposium on Fault Detection, Supervision and Safety of Technical Processes SafeProcess’15, Paris, France.
J. Foulliaron, L. Bouillaut, P. Aknin, A. Barros (2015). ‘A specific semi-markovian dynamic bayesian network estimating residual useful life’, 16th Conference of the Applied Stochastic Models and Data Analysis ASMDA’15, Piraeus, Greece.
J.Foulliaron, L.Bouillaut, P.Aknin, A.Barros (2014). ‘A specific Dynamic Bayesian Network for a prognosis based maintenance strategy’, 2nd International Conference on Railway Technology: Research, Development and Maintenance Railways’14, Ajaccio, France.
J.Foulliaron, L.Bouillaut, A.Barros, P.Aknin, R.Rozas (2013). ‘A prognostic algorithm based on probabilistic graphical models for a periodically observed system’, 11th ESRA European Safety & Reliability International Conference ESREL’13, Amsterdam, The Netherlands.