A Bradley Curve-AI Framework for Systematic Safety Culture Improvement in Nigeria's Upstream Oil and Gas Industry

Abstract Persistent Health, Safety and Environment (HSE) underperformance in Nigeria's upstream oil and gas sector reflects deep-rooted weaknesses in safety culture, not just technical or regulatory failures. This paper presents a four-phase integrated framework that combines the DuPont Bradley Curve safety culture maturity model with three machine learning algorithms to create a practical, data-driven pathway for systematic HSE improvement in Nigerian upstream operations. The framework is built on a master dataset of 264 engineered records spanning 14 years (2010 to 2023), compiled from verified primary sources including the Nigerian Upstream Petroleum Regulatory Commission (NUPRC), the National Oil Spill Detection and Response Agency (NOSDRA), and operator sustainability disclosures across all four terrain types. Three machine learning models are developed and validated: a Random Forest incident cause classifier, an XGBoost Lost Time Injury Frequency (LTIF) predictor, and a K-Means unsupervised stage classifier. SHapley Additive exPlanations (SHAP) analysis of the XGBoost model yields the study's most significant finding: Bradley Curve stage accounts for 52.9 per cent of total LTIF predictive power, more than all other operational and technical features combined. This is the first quantitative demonstration, using Nigerian upstream data, that safety culture maturity is the dominant determinant of injury frequency. K-Means clustering shows that 51 per cent of Nigerian upstream operational records are at the Reactive or Dependent stage, with onshore Joint Venture operations concentrated at Stages 1 and 2. No Nigerian operator in the dataset has achieved sustained Interdependent-stage performance across the study period. An original Environmental Incident Intensity Index (EIII) is introduced as a production-normalised environmental performance metric. The EIII is the third-ranked feature in the Random Forest classifier, confirming that environmental and injury performance share common cultural and operational determinants. A targeted 24.7 per cent LTIF reduction is projected within a single operational cycle through improvements to three leading indicators that require no capital investment: Permit to Work (PTW) compliance, leadership safety observation frequency, and near-miss reporting rate. The framework aligns with Section 102 of the Petroleum Industry Act (PIA) 2021 and can be implemented using open-source Python tools and existing regulatory reporting data.