Multiple Error Filtering in Cyclic Systems

Voting algorithms are used to arbitrate between the variant results in fault-tolerant systems. Traditional voters produce incorrect outputs in multiple error conditions. This paper introduces a class of voters, called predictor voters, which can resolve some multiple error conditions. These voters use analysis of a sequence of results in cyclic systems to select the most likely correct variant result as the voter output. Large discontinuities between successive results in cyclic systems are indicative of faults. The voting algorithms have the effect of filtering discontinuities to improve availability. Three different versions of predictor voters are described. Fault-injection simulation tests are used to investigate their safety and availability performance in triple error scenarios. Experimental results show that predictor voters give safety behaviour between majority and median voters. Predictor voters with order three and above give higher availability than the median voter. Predictor voters are suitable for use in systems in which some incorrect outputs can be tolerated in order to maintain functionality over a long period of time.