AI Development Company in Abu Dhabi

Oil and Gas

An Abu Dhabi upstream oil and gas operator (producing 850,000 barrels of oil equivalent per day across 8 onshore and offshore fields, managing $48 billion in production assets) was experiencing unplanned equipment downtime that cost $127 million annually. The most critical failures: electric submersible pumps (ESPs, which lift oil from wells — failure requires pulling the entire pump assembly, a 7-14 day operation costing $800,000-$2 million per failure), gas compressors (compression is essential for maintaining field pressure — a single compressor failure can reduce field production by 15,000 barrels per day at $80/barrel), heat exchangers (fouling reduces thermal efficiency — undetected fouling leads to energy waste and eventual shutdown for cleaning), and safety-critical equipment (pressure safety valves, gas detectors, fire and gas systems — where failure has HSE implications). The operator had a time-based maintenance programme: equipment was maintained on fixed schedules regardless of condition. This resulted in: over-maintenance of healthy equipment (estimated 35% of maintenance actions were unnecessary, consuming technician time and spare parts), under-maintenance of degrading equipment (the fixed schedule missed accelerated degradation caused by operating conditions — sand production from one field was causing ESPs to fail 40% faster than the scheduled maintenance interval), and reactive maintenance for failures that fell between scheduled interventions (62% of major failures occurred between maintenance windows). The operator had invested in a historian system (OSIsoft PI) that collected sensor data from 12,000 tags across the 8 fields — but this data was used only for real-time monitoring and post-failure root cause analysis, not for prediction.

We built a predictive maintenance system that transforms the operator's sensor data from a monitoring tool into a failure prediction engine. (1) Data integration: connected to the OSIsoft PI historian with real-time streaming — ingesting 12,000 sensor tags at 1-second resolution (pressure, temperature, vibration, flow rate, current, voltage, and derived measurements). Historical data: 7 years of sensor data (2.1 petabytes) plus maintenance records (28,000 work orders with failure descriptions, root causes, and repair actions). The maintenance records were semi-structured — we used NLP to extract structured failure modes, root causes, and affected components from free-text work order descriptions (in a mix of English and Arabic). (2) Equipment-specific failure models: rather than a single generic model, we built specialised models for each equipment class. ESP failure prediction: features include motor current, intake pressure, discharge pressure, vibration (radial and axial), motor temperature, and fluid properties (water cut, sand content, gas-liquid ratio). The model predicts remaining useful life (RUL) by identifying the degradation trajectory — comparing current vibration and current signatures against historical degradation patterns that preceded failure. Alert threshold: 14-day warning (sufficient time to order a replacement pump and schedule a pulling unit). Compressor health: features include discharge pressure, suction pressure, interstage pressures, vibration at each bearing, oil temperature, gas composition, and operating speed. The model detects: bearing degradation (vibration spectral analysis using FFT features), valve leakage (interstage pressure ratio deviations), and seal degradation (process gas in lube oil, detected through oil analysis integration). Heat exchanger fouling: monitoring the approach temperature (difference between process outlet and cooling medium inlet). As fouling progresses, approach temperature increases. The model predicts fouling rate and recommends optimal cleaning timing — balancing cleaning cost against energy penalty. Safety equipment: monitoring the response characteristics of safety-critical instruments. Pressure safety valve: lift pressure drift (detected through acoustic monitoring), gas detector: calibration drift (response time degradation), fire and gas system: voting logic integrity. (3) Risk-based maintenance optimiser: the failure predictions feed into an optimisation engine that generates maintenance schedules. The optimiser balances: failure probability (from the predictive models), failure consequence (production loss, HSE impact, environmental impact, repair cost — different for each equipment type and location), maintenance cost (mobilisation, spare parts, downtime during maintenance), and operational constraints (weather windows for offshore operations, field shutdown schedules, technician availability). The result: a dynamic maintenance schedule that shifts resources toward equipment with the highest risk and defers maintenance on healthy equipment. (4) Deployment on sovereign infrastructure: deployed on G42 Cloud with all data resident in UAE. The models run inference on real-time streaming data, with results displayed in a control room dashboard showing: equipment health scores (colour-coded fleet overview), failure predictions with confidence intervals and recommended actions, maintenance schedule recommendations, and HSE equipment status (highlighted separately for safety compliance). Arabic and English interface with role-based access: field operators see their assets, maintenance planners see scheduling recommendations, and management sees fleet-wide KPIs.

Unplanned downtime reduction: major equipment failures reduced from 47 per year to 14 — a 70% reduction. The remaining 14 failures were primarily in equipment types not yet covered by the model (specialised offshore equipment with insufficient historical failure data for reliable prediction). ESP mean time between failure: increased from 14 months (time-based maintenance) to 22 months (predictive maintenance) — the model identified operating conditions that accelerated ESP degradation and recommended operational adjustments (reduced pump speed in high-sand-production wells) in addition to maintenance timing. Production recovery: avoided production losses of $89 million annually (14-day advance warning enabled planned maintenance during low-production windows rather than emergency shutdowns). Over-maintenance reduction: unnecessary preventive maintenance decreased by 38%, freeing 12,000 technician-hours annually for value-added work. Spare parts inventory: optimised based on predicted failure rates — inventory carrying cost reduced by $4.2 million annually without increasing stockout risk. HSE improvement: zero safety equipment failures that would have resulted in an undetected hazardous condition — all safety equipment degradation was identified and addressed before functional failure. Compressor availability: increased from 91% to 97% across the fleet — the 6 percentage point improvement represents $28 million in additional annual production. Annual value: $125.2 million ($89M avoided production loss + $28M compressor availability + $4.2M inventory + $4M maintenance efficiency). Development investment: $780,000 over 28 weeks. ROI: 160x first-year return — reflecting the massive financial leverage of production uptime in upstream oil and gas.