Deep Natural Language Processing for Automatic Root Cause Analysis of Non Productive Time Events in Drilling Reports

Abstract In drilling operations, the operational inefficiencies, such as stuck pipe, lost circulation, tool failure, and wellbore instability, represents critical challenges in the oil and gas industry. In this study, we present a robust AI-driven framework for sequence mining and information extraction from unstructured drilling reports in oil wells by leveraging deep learning in natural language processing (NLP) techniques. The primary objective is to develop a workflow to identify and mitigate the operational inefficiencies caused by non- productive time (NPT) through automated root cause analysis, facilitating proactive decision-making during drilling operations. The unstructured daily drilling reports (DDRs) textual data undergoes a systematic preprocessing procedure of data cleaning, tokenization, and vectorization using embedding models. These steps transform words and sentences into numerical vectors, enabling large language models (LLMs) to extract deeper contextual insights. This transformation is essential for detecting patterns, trends, and relationships within the unstructured reports, forming the foundation for deeper analysis. The methodology integrates change point detection (CPD) to identify anomalies in multivariate time-series drilling parameters, including mud volume, torque, and weight on bit. Concurrently, a semi-supervised NLP system categorizes sentences in DDRs into key categories: Symptom, Event, and Action. This combined structure allows for the creation of event sequences that map operational behavior from symptom onset through mitigation, making it possible to uncover recurring failure patterns and support early intervention. By aligning symptoms, events, and corrective actions, the framework enables predictive insight and knowledge transfer across wells. Applied to 701 DDRs from 10 wells, the system identified 803 NPT events and demonstrated the ability to detect issues up to 12 hours in advance. In field applications, the methodology contributed to an average NPT reduction of 18.5%, equivalent to approximately 3.2 days saved per well and cost savings exceeding $300,000 per well. A visual dashboard provides real-time alerts and historical insights for field engineers, enhancing decision-making and future well planning. The strength of this framework lies in its hybrid integration of structured and unstructured data, facilitating both real-time and retrospective analysis. Furthermore, its scalability allows deployment across a range of drilling environments with varying geological challenges. By transitioning from reactive troubleshooting to proactive monitoring and planning, this study offers a transformative tool for improving drilling performance, reducing costs, and supporting informed decision-making in complex subsurface conditions.