Learning how to monitor: Pairing monitoring and learning for online system verification

In several domains, the execution of a system is associated with the generation of continuous streams of data. Such streams may contain important telemetry information, which can be used to perform tasks like predictive maintenance and preemptive failure detection, in order to issue early warnings. In critical contexts, formal methods have been recognized as an effective approach to ensure the correct behaviour of a system. However, they have at least two significant weaknesses: (i) a complete, hand-made specification of all the properties that have to be guaranteed during the execution of the system turns out to be often out of reach when complex systems have to be handled and, for the same complexity reasons, (ii) it may be difficult to derive a complete model of the system against which to check the properties of interest. In this paper, to overcome these limitations, we extend a recently presented framework that pairs monitoring with machine learning, in order to allow for the preemptive detection of critical system behaviours in an on-line setting. The framework is tested on a practical use-case based on the public NASA C-MAPSS dataset, and is shown to obtain promising performance in terms of its ability to forecast the approach of failures, and to provide interpretable results.