Predictive Statistical Monitoring(Improvements on online learning and background knowledge)

Monitoring is a highly used technique to check cyber-physical systems(CPS), where the corresponding model is either unavailable or not worthwhile to obtain.q Based on observing executions of a system, a monitor analyses if observations comply with expected behavior. Predictive monitoring of CPSs is an attempt to add prediction capability to this verification platform, where the history of a system is learned by training over previous traces. We seek to enhance the flexibility of the statistical monitor to make it feasible to be used in the CSPs. These models can be used in safety-critical systems to inform the imminent fault in advance. In this thesis, we proposed a statistical predictive run-time verification that is equipped with dynamic horizon allocation. Therefore, based on the received observations, the monitor can increase the prediction window when it gets a trace similar to the one already seen in the historical data log. On the other hand, the monitor switches to a more conservative short window prediction zone when receiving an unseen data stream. This feature adds flexibility to the solution allowing a solution to be used in a different environment with different degrees of predictability. Another contribution is to use background knowledge(BK) to enhance the prediction capability of a system while running the monitor online. It was demonstrated that BK contributes positively to increasing the prediction zone while decreasing the error. We also suggested that the hidden Markovian model(HMM), used as the system's history, can be updated while receiving the new trace. We considered both adaptations of the parameters and modification of the structure on the fly.