Dynamic Causal Hidden Markov Model Risk Assessment

Abstract

Understanding system vulnerabilities to risk factors during operation is essential for developing dependable systems. By implication, assessing in-use risk factors requires monitoring system parameters that contribute to making probabilistic inferences. We argue, however, that naïve use of statistical data without regard to causality can yield surprising and often erroneous risk predictions. Making reliable risk predictions is further complicated by lack of full awareness of system states and the existence of unobservable parameters in complex systems. Overly conservative risk assessment leads to increased life-cycle cost and reduced system availability resulting from overly aggressive preventive maintenance or replenishment strategies, while overly optimistic risk assessment can lead to even higher life-cycle cost and potential harm when otherwise preventable failures occur. This paper discusses a causality-aware, dynamic risk assessment model based on hidden Markov model construct. This model employs the concept of hidden system states that account for otherwise unexplainable observations. The model is continuously evaluated during system operation and updated when new observations warrant reevaluation.