Human reliability data, human error and accident models - illustration through the Three Mile Island accident analysis

Our first objective is to provide a panorama of Human Reliability data used in EDF's Safety Probabilistic Studies, and then, since these concepts are at the heart of Human Reliability and its methods, to go over the notion of human error and the understanding of accidents. We are not sure today that it is actually possible to provide in this field a foolproof and productive theoretical framework. Consequently, the aim of this article is to suggest potential paths of action and to provide information on EDF's progress along those paths which enables us to produce the most potentially useful Human Reliability analyses while taking into account current knowledge in Human Sciences. The second part of this article illustrates our point of view as EDF researchers through the analysis of the most famous civil nuclear accident, the Three Mile Island unit accident in 1979. Analysis of this accident allowed us to validate our positions regarding the need to move, in the case of an accident, from the concept of human error to that of systemic failure in the operation of systems such as a nuclear power plant. These concepts rely heavily on the notion of distributed cognition and we will explain how we applied it. These concepts were implemented in the MERMOS Human Reliability Probabilistic Assessment methods used in the latest EDF Probabilistic Human Reliability Assessment. Besides the fact that it is not very productive to focus exclusively on individual psychological error, the design of the MERMOS method and its implementation have confirmed two things: the significance of qualitative data collection for Human Reliability, and the central role held by Human Reliability experts in building knowledge about emergency operation, which in effect consists of Human Reliability data collection. The latest conclusion derived from the implementation of MERMOS is that, considering the difficulty in building 'generic' Human Reliability data in the field we are involved in, the best data for the analyst consist of the knowledge built up through already existing probabilistic analyses. (C) 2003 Elsevier Ltd. All rights reserved.