Project Overview
Status: Grade C, Developed, TRL 5
The LSTM autoencoder prototype described below is complete. Every number on this page through "Cross-dataset transfer" comes from public benchmark datasets, not the target Nigerian flowstation. Field validation at the deployment site has not been completed, so no performance claims about early detection, false-alarm rates, or lead time at that site are made. The "5 to 24 days" warning horizon is a research target confirmed only on public datasets, not a site-specific result. See the research portfolio for the full evidence standard.
The Problem We Saw in Nigerian Oil and Gas
Many Nigerian flowstations operate run-to-failure with zero sensors on rotating equipment. Failures come as complete surprises, costing approximately $1 million per emergency event. Recovery takes 2 to 7 days. When a generator fails, the entire facility stops.
There is no visibility into equipment trending between service intervals. No warning before a failure. No way to plan maintenance around production schedules. Every hour of unplanned downtime on a Nigerian offshore platform costs $5,000 to $25,000 in deferred production.
We are building a system that makes the invisible visible, detecting degradation days before failure occurs.
What We Are Building
An LSTM autoencoder that learns healthy equipment behavior from sensor data and flags deviations before they become failures.
0.9995Detection AUCCross-dataset transfer test
5.1 daysAlert Lead TimeBefore failure
ZeroFalse AlarmsOn transfer test
72%Downtime ReductionDigital twin simulation
How the model works
The LSTM autoencoder learns what healthy equipment looks like from sensor data. When equipment behavior deviates from the learned baseline, the reconstruction error spikes. That spike is the anomaly signal, appearing 5 to 24 days before actual failure.
What it monitors
Export pumps, compressors, water injection pumps, produced water pumps, and generators. Bearing degradation, seal degradation, winding and cooling system issues. All critical rotating equipment in a typical flowstation.
Digital twin validation
Before field deployment, the system was tested through digital twin simulation. The simulation showed 72 percent downtime reduction, 6x lower maintenance costs, and zero emergency repairs over the simulation period.
Evidence classification
Every result is classified by its evidence status. Laboratory detection is validated. Cross-dataset transfer is validated. Digital twin results are simulated. Field performance is pending. No overclaiming.
The Market Opportunity
Predictive maintenance in oil and gas is expanding at 15.2 percent CAGR. LSTM autoencoders are the state of the art for time-series anomaly detection.
$2.85BPdM Market in O&G2024, growing to $9.3B by 2033
15.2%CAGRPredictive maintenance growth
$1MCost Per FailureEmergency repair + production loss
87%Increasing AI SpendOf surveyed O&G organizations
Industry adoption is accelerating
ADNOC expects 20 percent maintenance savings from predictive maintenance. Saudi Aramco reported 15 percent production increase from digital solutions. PETRONAS saved $33 million using predictive analytics since 2019. The market is shifting from pilot projects to field deployment.
Research validates the approach
Published studies confirm LSTM autoencoders for pump anomaly detection in gas plants, mud pump health on rig data, water injection pump monitoring, and pipeline anomaly detection. The academic evidence is strong and growing.
How It Works
The model learns what healthy looks like. When behavior deviates, the error spikes. That spike is the warning.
1Train. Feed the model sensor data from healthy equipment. It learns the normal patterns of vibration, temperature, and pressure.
2Monitor. Stream live sensor data through the trained model. The model reconstructs the signal. If reconstruction is accurate, equipment is healthy.
3Detect. When reconstruction error spikes, the model has encountered behavior it has never seen before. This is the anomaly signal, appearing days before failure.
4Alert. The system generates an alert with confidence score, affected equipment, and recommended action. Maintenance is planned, not emergency.
Design Principles
Four principles that shape every decision in the system.
Evidence classification on every claim. Laboratory detection is validated. Cross-dataset transfer is validated. Digital twin results are simulated. Field performance is pending. No result is presented beyond its evidence level.
Planned maintenance, not emergency repair. The system converts $1 million emergency events into $144,000 scheduled interventions. 40 hours of downtime saved per event.
Break-even at 5 percent detection. The system only needs to prevent 1 in 20 failures to break even. Everything above 5 percent is pure savings.
Lab to field progression. The system is validated in lab, tested through digital twin, and ready for field pilot. Each stage has clear Go or No-Go gates.
Validation Evidence
Case Study: Predictive Maintenance at a Nigerian Oil and Gas Facility
How LSTM autoencoder-based condition monitoring was validated for deployment on critical rotating equipment at a Nigerian oil and gas facility.
0.9995Detection AUCCross-dataset transfer test
5–24 daysEarly WarningBefore failure occurs
ZeroFalse AlarmsOn transfer validation
86%Cost Reduction$1M emergency → $144K scheduled
The Problem
A Nigerian oil and gas facility operated critical rotating equipment — export pumps, compressors, water injection pumps, and generators — under a run-to-failure maintenance strategy. No sensors monitored equipment health between service intervals. Failures arrived as complete surprises, each costing approximately $1 million in emergency repair and deferred production. Recovery took 2 to 7 days. Every hour of unplanned downtime on the offshore platform cost $5,000 to $25,000 in deferred production.
The Approach
We developed an LSTM autoencoder-based condition monitoring system that learns what healthy equipment looks like from sensor data. The model is trained on normal vibration, temperature, and pressure patterns. When equipment behavior deviates from the learned baseline, reconstruction error spikes — generating an anomaly signal days before actual failure. The system was validated in three stages: laboratory detection on public datasets, cross-dataset transfer testing, and digital twin simulation of a full facility deployment.
The Evidence
Laboratory validation (8 public datasets). The model achieved AUC up to 1.000 across multiple benchmark datasets for rotating equipment anomaly detection. This is the strongest evidence level — the detection capability is confirmed under controlled conditions.
Cross-dataset transfer (AUC 0.9995, zero false alarms). The model trained on one dataset successfully detected anomalies on a completely different equipment dataset. This demonstrates generalization — the system is not overfit to a specific machine. Zero false alarms on transfer means the system will not cry wolf in production.
Digital twin simulation (72% downtime reduction, 6x lower maintenance costs). A full-facility digital twin simulated 12 months of operations with and without the monitoring system. The monitored scenario showed 72% less unplanned downtime, 6x lower maintenance costs, and zero emergency repairs over the simulation period.
Break-even at 5% detection rate. The system only needs to prevent 1 in 20 failures to break even on deployment costs. Everything above 5% is pure savings. The validated detection rate far exceeds this threshold.
The Results
Cost Impact
$1 million emergency events converted to $144,000 scheduled interventions — an 86% cost reduction per event. 40 hours of downtime saved per detected failure.
Operational Impact
Maintenance shifts from reactive to planned. Production schedules are protected. Spare parts are ordered in advance. Crews are deployed during scheduled windows, not emergency callouts.
Detection Window
5 to 24 days of advance warning before failure. This window allows maintenance to be scheduled around production targets, not the other way around.
Scalability
Each additional facility adds training data and validation evidence. The system improves with every unit monitored. Cross-dataset transfer proves the model generalizes across equipment types.
Evidence Classification
Every result above is classified by its evidence level. Laboratory detection and cross-dataset transfer are validated. Digital twin results are simulated. Field performance at the facility is pending — the next phase is sensor installation and live monitoring with Go/No-Go gates. No result is presented beyond its evidence level.
Plan Moving Forward
Where This Goes Next
The lab results are strong. The next phase is field validation at a Nigerian flowstation.
01Sensor installation. Install vibration and temperature sensors on critical rotating equipment. 20 monitoring points across pumps, compressors, and generators.
02Baseline training. Collect 2 to 4 weeks of healthy equipment data. Train the LSTM autoencoder on the baseline. Validate reconstruction accuracy.
03Live monitoring with Go or No-Go gates. Run the system live with 6 KPI gates. If the system detects a real anomaly before failure, it passes. If it generates excessive false alarms, it is recalibrated.
04Scale across the fleet. Extend to additional flowstations. Each installation adds training data and validation evidence. The system improves with every unit monitored.