Scenario-based generalization bound for anomaly detection support vector machine ensembles

This work proposes an ensemble of robust Support-Vector-Machine (SVM) classifiers to monitor systems health-states given uncertain measurements from multiple sensors. Scenario optimization is a well-established theory to solve optimization problems in the presence of uncertainty and it is used to render a formal bound on the SVM misclassification probability. This probabilistic certificate of generalization (robustness) holds non-asymptotically and for any stationary random mechanism generating the data. A novel selection strategy is introduced seeking an ensemble design which maximize accuracy in the prediction and robustness given by Scenario theory. The framework is tested on a Prognostics and Health Management (PHM) challenge problem launched by the ARAMIS group in 2020 where a set of sensor measurements and labels are provided to abnormal operational states. The points of strength and the weaknesses of the proposed framework are presented and discussed.