Natural Language Processing and Text Mining Approaches in Production Shortfalls Analytics: Methodology, Case-Study and Value in the North Sea

Maximizing operational efficiency is a critical challenge in oil and gas production, particularly important for mature assets in the North Sea. The causes of production shortfalls are numerous, distributed across a wide range of disciplines, technical and non-technical causes. The primary reason to apply Natural Language Processing (NLP) and text mining on several years of shortfall history was the need to support efficiently the evaluation of digital transformation use-case screenings and value mapping exercises, through a proper mapping of the issues faced. Obviously, this mapping contributed as well to reflect on operational surveillance and maintenance strategies to reduce the production shortfalls. This paper presents a methodology where the historical records of descriptions, comments and results of investigation regarding production shortfalls are revisited, adding to existing shortfall classifications and statistics, in particular in two domains: richer first root-cause mapping, and a series of advanced visualizations and analytics. The methodology put in place uses natural-language pre-processing techniques, combined with keyword-based text-mining and classification techniques. The limitations associated to the size and quality of these language datasets will be described, and the results discussed, highlighting the value of reaching high level of data granularity while defeating the ‘more information, less attention’ bias. At the same time, visual designs are introduced to display efficiently the different dimensions of this data (impact, frequency evolution through time, location in term of field and affected systems, root causes and other cause-related categories). The ambition in the domain of visualization is to create User Experience-friendly shortfall analytics, that can be displayed in smart rooms and collaborative rooms, where display's efficiency is higher when user-interactions are kept minimal, number of charts is limited and multiple dimensions do not collide. The paper is based on several applications across the North Sea. This case study and the associated lessons learned regarding natural language processing and text mining applied to similar technical concise data are answering several frequently asked questions on the value of the textual data records gathered over years.