Autonomous AI Safety Layer Detection of Hazardous Mud-Weight Events

Abstract Real-time fluid sensors stream thousands of data points per day, yet mud-weight decisions still rely heavily on manual checks. We built a lightweight, two-layer warning system that monitors density and key mud parameters (temperature, water, oil, solids, plastic viscosity) directly from an Offshore Drilling-Fluid Monitoring System. Performance is demonstrated on a continuous 48-h dataset containing several operationally relevant events. Layer 1 applies rolling-window statistics to density: |Δρ| > 0.003 g cm-3 and >2σ from the 30-sample mean triggers a rule alert. Layer 2 trains an online Isolation Forest (200 trees, 2% contamination) on the six-feature vector [ρ, Temperature, Water%, Oil%, Solids%, PV]; records (