Abstracting 'continuous' system behaviours into discrete-event representations (i.e., timed automata) for diagnosis purposes is demonstrated in this paper. As complex system dynamics are often partially known, the resulting imprecision on continuous variables is represented by means of intervals partitioning the state space according to landmarks defined by expert knowledge. Based on a continuous model simulation, an algorithm assigns discrete labels to landmark crossing by continuous variables, then, generates a timed automaton that can be further analysed by a model-checker. This procedure allows one to summarize a continuous system simulation output as a set of transitions among discrete states with qualitative interpretation (e.g., high, medium, low). In order to reduce explosion in the number of states, the generated timed automaton is specifically determined according to the property of interest for the user (e.g., reachability of some unwanted states). This approach has been applied to predict possible dysfunctions of a wastewater treatment process and validated using real-life data.