Logistics Intelligence Series | Module 5 of 6

Diesel consumption anomalies:
What your tank reveals about your fleet

Every truck has a statistically expected consumption - based on route, load and driving behaviour. An Isolation Forest model identifies which truck consumes unexpectedly high fuel on which route - and points to the cause: technical, route or driver.

01 The invisible problem - why averages lie

Your fleet average of 31.2 l/100km looks fine - but it hides the outliers

Most fleets monitor diesel consumption as a fleet average. As long as it stays stable, everything seems in order. But: an average across 150 trucks smooths everything away. LKW-089 has been consuming 4.2 litres more than expected for 6 weeks - that is €340/month in excess cost, hidden in the noise of the fleet figure.

The problem is not consumption itself - it is that nobody asks: "How much should exactly this truck consume on exactly this route with exactly this load?" If you can answer that question, every deviation becomes a signal.

2.1m litres of diesel × €1.42 = €2.98m/year

Even a 3% reduction through anomaly detection saves €89,400 - and prevents breakdowns.

Fuel data + telematics
Expected-value model
Isolation Forest
Anomaly classification
Alert + Ursache

02 Data foundation - fuel cards, telematics, topography

78,000 trip segments with individual consumption profiles

We combine fuel card billing (litres per trip) with telematics data (speed profile, elevation, braking frequency) and trip master data (load, route type). This creates an individual expected value for each truck - and every deviation becomes measurable.

▸ Output
Dataset: 7.800 Wochen-Records
Fahrzeuge: 150
Ø Verbrauch: 31.2 l/100km
Anomalie-Wochen: 487 (von 18 Fahrzeugen)
7.800
Weekly data points
31.2 l
Avg. consumption /100km
18
Trucks with anomaly
€2.98M
Diesel costs / year

03 Expected-value model - what should this truck consume?

A gradient boosting regression model as the reference line

Before we can detect anomalies, we need an answer to the question: "What is the normal consumption for exactly this truck, on exactly this route, with exactly this load?" For this we train an XGBoost regression model on normal weeks (without known defects).

0.68 l
MAE expected model
0.912
R² score
±2.2%
Typical variation
>5%
Anomaly threshold
↳ Why This Matters

The model predicts what every truck should consume with ±0.68 l/100km accuracy. Anything exceeding 5% deviation over 2+ weeks is a signal. For comparison: an injector defect causes +10–15% - the model sees it immediately, while it disappears in the fleet average.

Consumption: expected vs. actual - distribution of deviations

04 Isolation Forest - detecting anomalies automatically

The algorithm that isolates outliers rather than learning normality

Isolation Forest works differently from classical methods: instead of learning what is "normal", it learns how easily a data point can be separated from the others. Anomalies are by definition "more easily isolable" - they require fewer decisions in the tree.

0.874
Precision
0.891
Recall
624
Anomaly weeks detected
18
Vehicles affected
Anomaly detection: residuals over time - 3 example trucks
↳ What the Model Finds

LKW-089 (orange): from week 22 consumption rises steadily - typical of a creeping injector defect. LKW-034 (red): sudden spike in week 31 - turbo leak. LKW-127 (blue): sporadic outliers - cruise control malfunction occurring only on certain routes. Three different patterns, one model.

05 Anomaly classification - what lies behind it?

The model not only detects that something is wrong - it also indicates why

Based on the anomaly pattern (gradual vs. sudden, constant vs. sporadic, load-dependent vs. route-dependent) a second model classifies the probable cause:

Anomaly distribution by cause type
LKW-089 · Baujahr 2020 · Euro6d
Mechanical – Injector defect
+4.8 l/100km vs. expected
Pattern: Creeping increase since week 22 (+0.15 l/week). Load-independent. Route-independent.
Cost to date: €2,840 excess consumption + impending engine damage.
Recommendation: Immediate workshop – check injectors, measure common-rail pressure.
LKW-034 · Baujahr 2019 · Euro6d
Mechanical – Turbocharger leak
+3.2 l/100km vs. expected
Pattern: Sudden increase in week 31. Stronger under high load (>20t).
Cost to date: €1,960 excess consumption.
Recommendation: Check boost pressure + visual turbo inspection. Do not continue under full load.
LKW-156 · Baujahr 2022 · Euro6e
Driving style – New driver
+2.1 l/100km vs. expected
Pattern: Increase correlates with driver change (week 18). Harsh-Braking +60%, more erratic speed profile.
Cost to date: €3,150 excess consumption (longer period).
Recommendation: Eco-driving training. Expected saving: €2,400/year.
LKW-042 · Baujahr 2021 · Euro6e
Route – Diversion due to roadworks
+1.7 l/100km vs. expected
Pattern: Since week 28 only on a specific route (FRA→STR). Altitude +40%, motorway share –25%.
Cost to date: €890 excess consumption.
Recommendation: Check alternative route via A6. A5 roadworks planned until Q2/2025.
↳ The Hidden Risk

The injector defect on LKW-089 does not just cause €2,840 in excess fuel - it is a precursor to engine damage (€8,000–€15,000). The model detected it in week 24. Without the model it would have been noticed at the next inspection in week 38 - 14 weeks and €1,200 in excess consumption later. Or as a breakdown on the motorway.

06 Feature importance - what really drives consumption

Surprising insights from the expected-value model

Feature importance - expected-value model (XGBoost)
↳ What the Fleet Manager Needs to Know

Ambient temperature explains 14% of consumption - and cannot be influenced. But: motorway share (21%) and load (18%) can be. The model shows which route configurations systematically cost more diesel than necessary - and where replanning has the greatest leverage.

07 Business impact - what early detection saves

Diesel costs + avoided repairs + fewer breakdowns

Annual savings potential by category
€96.260
Total savings / year
3.2%
Diesel cost reduction
CategoryAmount/yearMechanismConfidence
Consequential damage avoided€40.9506 consequential damages à €6,500 preventedHigh
Excess diesel consumption€22.3108 weeks earlier detection × 18 trucksHigh
Eco-driving€12.0005 drivers after training: –€2,400/yearMedium
Breakdown prevention€11.2004 breakdowns à €2,800 preventedHigh
Other (routes, parking heater)€9.800Route and parking heater optimizationMedium
↳ The Surprising Main Lever

The diesel itself is not the biggest item - it is the avoided consequential damage (€40,950). An injector defect that goes undetected for 10 weeks damages the engine. Anomaly detection is therefore not just a fuel-saving tool but a predictive maintenance system that pays for itself through avoided workshop costs.

08 Next steps

From weekly report to real-time alert

① Fuel card integration

Automated import from DKV/UTA/Shell Card. Combined with telematics data (Fleetboard, TomTom). Weekly batch analysis.

② Alert system

Automatic email to workshop manager when anomaly score exceeds threshold. Including cause classification and recommended action.

③ Driver cockpit

Individual consumption feedback for each driver: "This week 1.4 l/100km below your average - well done." Gamification rather than surveillance.