Fuel performance is often managed as a monthly financial variance. By then, the useful operational context has disappeared. The business sees how many liters were purchased, but not precisely which truck, road segment, queue, payload, grade, idle event, or fueling exception created the difference.
A mining fuel-intelligence system should explain consumption at the level where a supervisor can act. That requires more than a dashboard. It requires a reconciled fuel ledger, a spatial operating model, reliable telemetry, anomaly detection, and optimization that respects production constraints.
The value is measurable
Vendor case studies are not independent trials, but they show the size and shape of operational opportunities:
- Veridapt reports that removing one minute of fueling wait represented about 990 fleet hours per year and approximately US$1.7 million in production profit for one site.
- A separate Veridapt case attributes 1.7 million liters and US$2.55 million in annual fuel savings to a mining deployment.
- Cascadia Scientific reports 113,257 gallons of diesel and US$380,544 saved at a South American open-pit mine.
- COSMOS reports 97% consumption-prediction accuracy, 30% less fueling time, and 15.6 additional effective truck hours per month in customer cases. Its modeled example for a 50-truck fleet estimates 2.6 million liters, US$2 million, and 9,600 tonnes of carbon dioxide at a 5% improvement.
These figures use different baselines, fleets, and accounting methods. They should inform a site hypothesis, not a guaranteed proposal.
Build one auditable fuel ledger
Reconcile the complete chain:
- Supplier invoice and delivery note.
- Receipt at bulk storage.
- Tank level, density, temperature, and meter readings.
- Transfer to bowsers or satellite tanks.
- Dispense event with time, location, operator, and asset.
- Vehicle tank and engine consumption estimate.
- Production context for the same operating period.
Every movement receives a source, unit, timestamp, confidence, and reconciliation status. Normalize liters, gallons, mass, temperature compensation, and time zones. Preserve meter resets and calibration history.
Without this foundation, an anomaly model learns accounting noise.
Model expected consumption at road-segment level
Truck consumption changes with payload, gradient, rolling resistance, road condition, speed, acceleration, queueing, idle time, weather, engine state, and operator behavior. A route-level average hides the reason.
The 2026 Pingshuo study illustrates a more granular approach. Researchers analyzed 86,409 records for three truck types over 9,601 road segments. The reported XGBoost model achieved an R-squared of 0.94 and mean absolute error of 0.37 on its study data.
A separate peer-reviewed haul-monitoring study analyzed 1,780 cycles across 150 shifts over 90 days and identified gradients as a dominant factor. Together, the studies reinforce that route geometry and operating context belong in the model.
Use the model to calculate an expected-consumption interval, not a single perfect number. Compare actual consumption with that interval by truck, shift, segment, and operator. Then classify the likely cause: road, payload, idle, maintenance, measurement, leakage, or dispatch.
Detect loss without accusing people
Fuel anomalies may reflect theft, but they may also reflect sensor drift, incorrect asset identity, meter lag, return-to-tank activity, maintenance drain, temperature, or incomplete data. The system should rank evidence and request verification.
A useful anomaly packet includes:
- The exact transfer and telemetry records.
- Expected and observed quantity with confidence.
- Comparable events for the same asset and location.
- Sensor-health and calibration status.
- The missing data that could change the conclusion.
- A recommended investigation step.
Avoid labels such as fraud at the model layer. Human investigators need neutral evidence and documented due process.
Optimize fueling and dispatch together
Cheaper fuel performance can still reduce production if it creates queues or unnecessary travel. The optimizer should consider:
- Predicted remaining fuel and uncertainty.
- Queue and fueling duration.
- Current truck location and assignment.
- Shift change, maintenance, and road closures.
- Production target and shovel availability.
- Safe minimum fuel and emergency reserve.
The Veridapt one-minute example makes the interaction clear: a small fueling delay can represent hundreds of annual truck hours. The objective should therefore be cost and production value together, not liters alone.
KPI hierarchy
Measurement Layers
Data
matched transaction rate, sensor uptime, timestamp error.
Model
consumption MAE, prediction-interval coverage, anomaly precision.
Fuel
liters per tonne, unexplained variance, loss per transfer.
Haulage
queue minutes, cycle time, truck hours, tonnes per hour.
Financial
fuel cost per tonne, avoided loss, incremental production margin.
Environmental
fuel and emissions per tonne.
| Level | Metrics |
|---|---|
| Data | matched transaction rate, sensor uptime, timestamp error |
| Model | consumption MAE, prediction-interval coverage, anomaly precision |
| Fuel | liters per tonne, unexplained variance, loss per transfer |
| Haulage | queue minutes, cycle time, truck hours, tonnes per hour |
| Financial | fuel cost per tonne, avoided loss, incremental production margin |
| Environmental | fuel and emissions per tonne |
Report absolute savings and normalized performance. A 5% fuel improvement during a lighter production month may not be an operational gain.
A 12-week implementation
Weeks 1 to 3: map the fuel chain and reconcile one location. Weeks 4 to 6: build asset, route, and segment features and establish an expected-consumption baseline. Weeks 7 to 9: run anomalies in shadow mode and have operations label causes. Weeks 10 to 12: issue supervised recommendations for a defined fleet.
Predefine the financial method. State fuel price, production margin, excluded events, comparison period, and confidence range. Separate measured savings from modeled opportunity.
The conclusion
Mining fuel is not just a procurement category. It is a high-frequency operational signal connected to road condition, equipment health, dispatch, and production.
The best AI system creates a traceable line from every liter to every cycle, explains variance at the point of action, and optimizes without compromising safety or tonnes moved.
Research note
Research is current through September 5, 2026. Academic results are specific to their data and mine context. Commercial case metrics are vendor reported. Currency, fleet, fuel price, production margin, and emissions factors must be recalculated for each site.
Continue the research
- The complete mining AI model stack
- Mining voice AI for frontline operations
- RAG vs fine-tuning vs bespoke AI models
- Why an enterprise evaluation set becomes an AI moat
- AI fuel consumption optimization for manufacturing
Building a Production-Ready System
Conscious Engines builds mining fuel optimization systems that reconcile supplier, storage, bowser, dispenser, truck, route, payload, and production data. The model explains expected consumption and loss at the level where operations can act, then measures savings without sacrificing tonnes moved.