Assessing the reliability of diverse fault-tolerant software-based systems

Abstract

We discuss a problem in the safety assessment of automatic control and protection systems. There is an increasing dependence on software for performing safety-critical functions, like the safety shut-down of dangerous plants. Software brings increased risk of design defects and thus systematic failures; redundancy with diversity between redundant channels is a possible defence. While diversity techniques can improve the dependability of software-based systems, they do not alleviate the difficulties of assessing whether such a system is safe enough for operation. We study this problem for a simple safety protection system consisting of two diverse channels performing the same function. The problem is evaluating its probability of failure in demand. Assuming failure independence between dangerous failures of the channels is unrealistic. One can instead use evidence from the observation of the whole system's behaviour under realistic test conditions. Standard inference procedures can then estimate system reliability, but they take no advantage of a system's fault-tolerant structure. We show how to extend these techniques to take account of fault tolerance by a conceptually straightforward application of Bayesian inference. Unfortunately, the method is computationally complex and requires the conceptually difficult step of specifying ‘prior’ distributions for the parameters of interest. This paper presents the correct inference procedure, exemplifies possible pitfalls in its application and clarifies some non-intuitive issues about reliability assessment for fault-tolerant software.