Time-Sliced Temporal Evidential Networks: The case of Evidential HMM with application to dynamical system analysis

Diagnostics and prognostics of health states are important activities in the maintenance process strategy of dynamical systems. Many approaches have been developed for this purpose and we particularly focus on data-driven methods which are increasingly applied due to the availability of various cheap sensors. Most data-driven methods proposed in the literature rely on probability density estimation. However, when the training data are limited, the estimated parameters are no longer reliable. This is particularly true for data in faulty states which are generally expensive and difficult to obtain. In order to solve this problem, we propose to use the theory of belief functions as described by Dempster, Shafer (Theory of Evidence) and Smets (Transferable Belief Model). A few methods based on belief functions have been proposed for diagnostics and prognostics of dynamical systems. Among these methods, Evidential Hidden Markov Models (EvHMM) seems promising and extends usual HMM to belief functions. Inference tools in EvHMM have already been developed, but parameter training has not fully been considered until now or only with strong assumptions. In this paper, we propose to complete the generalization of HMM to belief functions with a method for automatic parameter training. The generalization of this training procedure to more general Time-Sliced Temporal Evidential Network (TSTEN) is discussed paving the way for a further generalization of Dynamic Bayesian Network to belief functions with potential applications to diagnostics and prognostics. An application to time series classification is proposed.