Assessment of a safety-critical system including software: a Bayesian belief network for evidence sources

Assessment of safety critical systems including software cannot rely only on conventional techniques, based on statistics and dependability models. In such systems, the predominant faults usually are design faults, which are very hard to predict. Therefore, the assessment can only be qualitative, and is performed by experts, who take into account various evidence sources. The aim of the SERENE European project is to improve the understandability, and repeatability of such assessments, thanks to a representation of the expert's reasoning by a mathematical model (a Bayesian belief network). The subject of this paper is the presentation of the BBN built by EDF to model one of its assessment approaches, valid for the products for which EDF writes the requirements specification, and then monitors the development made by an external supplier. No doubt that, before it yields reliable forecasts, this kind of model will require many years of calibration, by comparison between the predictions it gives, and the real, observed safety level of the evaluated systems. However, the authors think that in the short term, they can bring a rationale in the discussions between experts. They will also help in determining which are the most influential variables in the design process of a system, which is a necessary prerequisite for setting up any kind of field experience collection.