Development of natural language processing-based solutions for risk analysis : application to a hydropower company and an O&G industry

Abstract

Risk Analysis (RA) is crucial to prevent and mitigate potential risk events; however, there are several challenges related to RA. For instance, accident investigation reports are useful sources of information to support safety professionals to propose measures to prevent or mitigate identified occupational accident root causes. Nevertheless, reports’ low quality and lack of detail may limit their usefulness. Moreover, the quality of Quantitative Risk Analysis (QRA) strongly relies on the identification of all potential hazards with major consequences related to the operation of an industrial system, which is usually performed by multiple experts and consumes a considerable amount of time and effort. Since valuable knowledge about an industrial system is stored in the form of textual data, Natural Language Processing (NLP) techniques can be helpful since it can be applied to extract, organize, and classify information from text. Although several studies contributed to the advance of RA, most studies applying NLP focus primarily on automatically identifying patterns from reactive data, such as accident reports, and do not consider the quality of information contained in these documents. In addition, different forms of text data store relevant knowledge about industrial systems and their respective risks, especially proactive data such as documents resulting from preliminary risk studies, and adoption of these data could support preventive risk studies. The main purpose of this study is to develop NLP-based solutions to different issues faced in RA. Thus, this thesis presents two methodologies to (i) identify issues in a hydropower company’s accident investigation reports that may compromise their usefulness as a decision support tool (ii) automatically identify risk features from documents to support the initial stage of QRA in Oil and Gas (O&G) industries. Occupational safety technicians can benefit from the methodology that helps to identify issues and propose improvements to the accident reports. In addition, the second methodology can help experts to identify and assess hypothetical accidental scenarios related to the operation of an industrial facility. Thus, this thesis may contribute to the prevention and mitigation of occupational and/or major accidents and consequently avoid/reduce property damage, economic and social disruption, environmental degradation, and human losses.