The developments of reduced manning on offshore facilities and increased information transfer from offshore to land continue and may also be a prerequisite for the future survival of the oil and gas industry. A general requirement from the operators has emerged in that all relevant information from offshore-located systems should be made available so that it can be analysed on land. This represents a challenge to safety in avoiding negative impacts and potential accidents for these facilities. The layered Purdue model, which helps protect OT systems from unwanted influences through network segregation, is undermined by the many new connections arising between the OT systems and the surroundings. Each individual connection is not necessarily a problem; however, in aggregate, they add to the overall complexity and attack surface thereby exposing the OT systems to increased cyber risk. Since the OT systems are critical to controlling physical processes, the added connections represent a challenge not only to security but also to safety.

Previous research has investigated how automated annotation and technical language processing of maintenance work orders can support maintenance decisions for different types of equipment, e.g., Sexton et al. (2018) and Brundage et al. (2021). The objective of this paper is to investigate how technical language processing can be utilized for efficient and automated classification of failure events for safety critical equipment on petroleum facilities (e.g., gas detectors and shutdown valves). For such equipment, maintenance decisions are heavily influenced by the consequence of a failure, e.g., if the failure is classified as critical or degraded, or dangerous or safe. To explore the possibilities for automated classification of failure events, the paper will first present the fundamental concept hierarchy needed during annotation of safety critical equipment on petroleum facilities. This hierarchy is based on previous work where taxonomies from the ISO 14224:2016 standard (ISO 14224, (2016)) have been operationalised and detailed for safety critical equipment. The concept hierarchy is tested by a group of experts on a set of notifications to ensure that it is fit for purpose. The results from the test group are then used to refine the concept hierarchy before a larger group of persons, having different roles in the petroleum industry, is engaged in the entity typing of notifications. This entity typing will then provide the basis for mapping informal technical language to a computable format. By combining this knowledge with available structured information, such as failure mode and detection method, it is expected that an automated, standardised, and more efficient classification of failure events will be possible. Our database of several thousand manually classified notifications collected during the last decade, will finally enable us to evaluate the performance of the automated classification.