Predictive monitoring for signal temporal logic with probabilistic guarantees

Abstract

Monitoring is an effective approach for identifying safety violations for complex cyber-physical systems. In this poster, we consider safety specifications expressed in Signal Temporal Logic (STL). STL is a logic for specifying timed properties of real-valued signals, and there has been significant work on offline and online monitoring of STL formulas on signals. Boolean monitoring techniques solve the problem of determining if a given STL formula is satisfied by a signal, while robust monitoring techniques seek to compute a quantitative degree of satisfaction of the formula. Online techniques can compute satisfaction or violation of the formula when the entire signal is not available, but existing online techniques can only provide worst-case estimates of satisfaction (or violation). In this poster, we propose algorithms to predict the satisfaction or violation of an STL formula when only partial information of a trace (i.e. its prefix) is available. The output of our algorithm is a predicted interval for the robust satisfaction value, along with a probabilistic guarantee on the correctness of the prediction. We demonstrate the utility of our approach on monitoring a safety-critical signal in the context of unmanned aerial vehicle.