Modeling accident scenarios from databases with missing data: A probabilistic approach for safety-related systems design

This paper proposes a new probabilistic approach for safety-related systems design based on modeling accident scenarios with databases where missing data is a major concern. The method is based on (1) multivariate imputation by chained equations that addresses the problem of missing data in industrial databases, (2) Bayesian network learning approach that addresses the synthesis of the Fault tree (FT) and the Event tree (ET) diagrams and (3) the bowtie diagram that addresses the synthesis of the complete accident scenarios. An experimental application example of a database with missing data related to an accident that occurred in the BP Texas city refinery illustrates the effectiveness of the method. At first, a complete database is submitted to random data extractions and then the proposed method is applied. Experimental results confirm that it is possible to identify the relationship among observed and partially observed critical/undesirable events related to the critical faults even in conditions of missing data. The method might be used for the design of safety-related systems as it is able to support: (1) IEC 61511 and IEC 61508 standards, (2) uncertainty of databases with missing data; and (3) ensuring safe-diagnosability property regarding dynamical aspects of actual systems.