(55ah) Application of Large Language Models for Optimising Collection, Insight and Actions from Process Safety Observations

Process safety observations should be a primary source of ultra weak signals for operators of high hazard facilities. They provide early warnings of deviations from safe acts and conditions, which can provide important insight into the propensity for near misses and actual events.

However, in many companies, these observations are often a low value activity, undertaken by an unengaged workforce, who see this activity to be driven by management targets and corporate HSE programmes. This is compounded by the poor level of feedback generally provided to the observers. Equally, HSE departments struggle to create real insight from HSE observations due to their volume and the complexity of systematically analysing large volumes of text.

This paper describes a tool that we have developed that uses prompt engineered large language models to provide recommended consequences and actions for each observation. Furthermore, additional models predict the observation type and the most appropriate category on the safety critical barrier model. This surfaces the information required to visualise HSE observation data on a pseudo-barrier diagram visualisation: providing an overview of the observation type (unsafe/safe act or unsafe/safe condition) by barrier type (e.g. Process Containment). The models have been trained on an industrial dataset of existing observations, with future plans to augment this dataset with the outputs of the tool to iteratively improve the quality of model prediction.

The tool is designed to improve the entire submission and reporting cycle of HSE observations. High quality observations are encouraged by providing predictions of action, consequence, observation type and observation category that can be edited by the user. Users also have a holistic view of the analysed observations in the barrier visualisation, providing a valuable feedback channel to the workforce. The tool is developed with the ethos that AI/ML models should improve and expediate workflows, whilst recognising that the expertise, and engagement, of the workforce is of paramount importance and irreplaceable.

The benefits of this approach are: to transform unengaged workers in “human sensors” of ultra weak signals; to surface systemic weaknesses in safety critical barrier integrity at an early stage of degradation; and to provide highly meaningful feedback to the workforce to encourage them to take the time and effort to submit high quality observations.