Causal Modelling of Lower Consequence Rail Safety Incidents
MetadataShow full item record
The Safety Risk Model (SRM) is a key source of information for the GB rail industry. It is a structured representation of the 120 hazardous events that can lead to injury or death during the operation of the railway and is used to estimate the risk to passengers, workers and third parties. The SRM includes both rare but high consequence events such as train collisions and more frequent but lower consequence events such as passenger accidents at stations. In aggregate, these lower consequence events make an important contribution to the overall risk, which is measured by a weighted sum of injuries of different severity. Where possible, the SRM is derived from historical incident data, but the derivation of the model parameters still present challenges, which differ for different subsets of events. High consequence events occur rarely so it is necessary to use expert judgement in detailed models of these incidents. In comparison, the low consequence events occur more frequently, but both records of incidents and the models in the SRM are less detailed. The frequency of these low consequence events is sufficient to allow both the absolute risk and trends in the overall risk to be monitored directly. However, without explicit causal factors in the data or the model, the models are less able to support risk management directly, since this requires estimates of the risk reduction possible from particular interventions and control measures. Moreover, such estimates must be made locally, taking account of the local conditions, and at each location even the low consequence events are infrequent. In this paper we describe an approach to modelling the causes of low consequence events in a way that supports the management of risk. We show both how to extract more information from the available data and how to make use of expert judgement about contributory factors. Our approach uses Bayesian networks: we argue their advantages over fault and event trees for modelling incidents that have many contributory causes. Finally, we show how the new approach improves safety management, both by estimating the contribution of the underlying causes to this risk and by predicting how possible management interventions and control measures would reduce this risk.