This paper is concerned with the discovering of temporal knowledge from a sequence of timed observations provided by a system monitoring of dynamic process. The discovering process is based on the Stochastic Approach framework where a series of timed observations is represented with a Markov chain. From this representation, a set of timed sequential binary relations between discrete event classes is discovered with an abductive reasoning and represented as abstract chronicle models. To reduce the search space as close as possible to the potential relations between the process variables, we propose to characterize a set of series of timed observations with a unique measure of the homogeneity of the crisscross of class occurrences and to use this measure to prune abstract chronicle models.

This paper aims at showing a method to discover signatures (or models of chronicles) from a discrete event sequence (alarms) generated by a monitoring cognitive agent (MCA). When the counting process of the events generated by a couple (process, MCA) behaves like a Poisson process, this couple can be considered as stochastic discrete event generator SDEG/sub (pr, MCA)/ and modeled as a superposition of Poisson and an homogeneous discrete time Markov chain. The 'BJT' algorithm uses these two representations in order to help in the discovering of signatures. The results obtained on an industrial process monitored with a Sachem system have been validated by experts, confirming so the relevance of the approach within an industrial frame.