Using Accident Anatomy Analysis and Small Language Models to Support Systematic Lessons Learned Analysis and Improve the Completeness of Hazard Analysis

Over a period of 36 years, risk analysis was made of 105 process plants using HAZOP, action error analysis, mechanical integrity audit or pre-commissioning audit and QRA. Follow up studies were made at the end of the period. Altogether 82 major hazards accidents occurred, with a total of 168 fatalities as consequence. Of these 50% were predicted but not prevented due to inadequate safety management. The other 50% were due to lack of knowledge of some rare hazard types at the time of risk assessment. To improve completeness of hazard identification the Systematic Lessons Learned method was used which consists of cross referencing a large database of accident case histories. To ease the preparation of the database, a program, Cassandra, was developed for semantic analysis of accident case histories and to allow natural language retrieval of supplementary information for hazard identification The risk analyses were made using HAZOP with the workshop teams including the plant or design engineers and senior operators, with the author as facilitator. Additionally, generic action error analyses for operation of process unit types to support human error analysis (Taylor 2016). Completeness of analyses was checked using automated HAZOP analysis (Taylor 2017) and automated cross checking between manually and automatically completed results.