Fleet Fuel Management: Combating Diesel Theft in South Africa's Hub

Diesel is the largest single line item on a Gauteng fleet's operating budget, and with diesel prices at historic highs through 2026, every stolen litre is a measurable loss. Fuel theft is not new on the Reef. What has changed is the mathematics. A driver siphoning ten litres a shift used to cost the operator the price of breakfast. The same ten litres today is the price of a school uniform per shift, and the annual loss across a 40-truck fleet runs to seven figures. The good news is that catching diesel theft has become a structured engineering problem rather than a guessing game. This guide walks through the stack.

Direct answer

Modern fleet fuel management combines a tank-level sensor, a CAN bus engine feed, GPS event data, and a fuel-control algorithm that compares actual consumption against expected consumption. When the three data streams disagree, the platform raises a fuel event. The dispatcher then reviews the event against the truck's location and operational history to confirm or dismiss the loss.

The three diesel-theft patterns Gauteng fleets actually see

Most fuel loss falls into one of three operational patterns. Each leaves a different fingerprint in the data. Mixing them up causes wasted investigations and missed losses.

The first pattern is driver siphoning during normal operations. The driver pulls over at a quiet spot, runs a hose from the tank to a jerry can, and sells the diesel privately. The fingerprint is a sudden tank-level drop with no corresponding kilometres travelled and the engine at idle or off. The loss is small per event (typically 5 to 30 litres) but cumulative.

The second pattern is depot collusion. A pump attendant at a third-party fuel station, or an internal stores clerk at the operator's own depot, dispenses fewer litres than the receipt records. The fingerprint is a discrepancy between the fuel-card transaction value and the tank-level rise after a refuel event. Losses per event are higher (50 to 200 litres) and tied to specific stations or shifts.

The third pattern is hijacked-truck dump, where the entire load of diesel is drained at a holding yard after a vehicle theft. The fingerprint is a single, very large tank drop (often the entire usable volume) outside any operational geofence. This pattern overlaps with stolen vehicle recovery and is usually addressed through the SVR layer rather than through fuel-management alone.

What the fuel-management stack looks like in hardware

A working anti-siphoning deployment uses four sensor sources. The combined feed is what makes the algorithm trustworthy. Any one source on its own produces too many false positives.

• Capacitive or ultrasonic tank-level sensor. Fitted into the tank from the top, measures fuel volume directly to roughly 0.5 percent resolution. The source of truth for "how much is in the tank right now".
• CAN bus engine feed. The vehicle's electronic control unit reports fuel-rail consumption, engine load, RPM and odometer. Tells you how much fuel the engine has consumed since the last reading.
• GPS unit. Position, speed and ignition state. Places every fuel event in context.
• Fuel-card transaction feed. Most SA fleet cards push refuel transactions to an API. Closes the loop on whether a tank-level rise was real or phantom.

Wialon's fuel-control module is the global benchmark for the algorithmic layer. Geotab provides equivalent depth on the CAN bus side. TG Tracking integrates these sources into TG Online, so the dispatcher sees one fuel timeline per truck rather than four separate widgets.

How fuel algorithms detect siphoning events

The core detection logic is a comparison between expected and actual. Expected fuel consumption is a function of distance, engine load, gradient and idle time. Actual fuel consumption is the difference between two tank-level readings, adjusted for refuels. When the actual figure exceeds the expected figure by more than a tunable threshold (typically 8 to 12 percent), the platform flags a fuel event.

Three refinements matter for South African conditions. First, gradient correction: a truck climbing the N3 escarpment burns more diesel than the same truck on a flat run. Second, idle correction: a driver sitting in N1 traffic burns fuel without moving, which is expected behaviour, not a flag. Third, terrain correction: gravel roads and stop-start delivery in Soweto consume more fuel per kilometre than long-haul runs. The algorithm calibrates against the specific truck's historical curve.

The output is not an accusation. It is an event log with a confidence score, a litre-loss estimate and a location pin. The dispatcher decides what to do with each event.

A 7-step protocol for setting up fuel alerts in TG Online

The setup below is the version TG Tracking uses with new Gauteng fleets. Every step has a verification check before moving to the next.

1. Baseline the truck. Record three months of normal operations, with no alerts firing. The baseline produces the expected litres-per-100km curve, the expected idle-hours-per-shift, and the expected refuel pattern.
2. Fit the tank-level sensor. Calibrate against ten known fill points (10 percent, 20 percent, up to 100 percent). The sensor's reading at each point is matched against a measured volume. Any non-linearity is captured in a calibration table.
3. Wire in the CAN bus feed. Confirm the vehicle's CAN protocol and pull at least four parameters: instantaneous fuel rate, total fuel consumed, engine RPM and odometer.
4. Synchronise the fuel-card API. Ensure refuel transactions arrive in TG Online within ten minutes of the swipe. Time-sync errors are the single largest source of false positives.
5. Configure the algorithm thresholds. Start at 12 percent variance for siphoning detection and 5 percent variance for refuel discrepancies. Tune downwards over the first month.
6. Define the alert routing. Decide who sees a fuel event in the first instance. The dispatcher or the fuel manager. Most fleets route to a single fuel manager to avoid alert fatigue across the dispatch desk.
7. Run a 30-day calibration window. Review every event in week one and week four. Adjust thresholds based on which events were real losses and which were noise.

After the calibration window, the system runs in standing operations and only the events that exceed thresholds reach a human.

What good looks like: a fuel scorecard table

The table below is an example of the fuel scorecard a Gauteng fleet might publish weekly. It is the headline output of the stack, not the detailed event log.

The board-level question is whether the lost rand value trends down over twelve weeks. If it does, the stack is working. If it does not, the threshold tuning or the depot security has a gap that needs investigation.

Depot security: where the fuel stack meets the fence line

Algorithmic detection catches the event after the fact. Depot security stops the event happening in the first place. A working Gauteng depot has four overlapping controls, each tied to the same fuel data.

• Pump-side cameras that capture every fuel transaction with a clear shot of the truck registration, the pump display and the operator. Footage is keyed to the fuel-card transaction.
• Geofenced refuel zones. A truck that refuels outside the geofence triggers an alert. A truck inside the geofence with no fuel-card transaction within a ten-minute window also triggers an alert.
• Driver biometrics on the pump. The pump only dispenses on a verified fingerprint or facial scan, which prevents driver-card swapping.
• Reconciliation routine. A weekly reconciliation between fuel-card transactions, dispenser totalisers and tank-level rises. Discrepancies above 1 percent are investigated.

These controls work without telematics, but their value compounds when the data joins up. A depot collusion event found by the reconciliation routine is corroborated by pump-side video and geofence logs, and the investigation closes in hours rather than weeks.

Where TG Tracking sits in the stack

TG Tracking is not a fuel-sensor manufacturer. The sensors come from specialists and the algorithms come from platforms like Wialon and Geotab. What TG Tracking provides is the local layer: the install team that calibrates the tank sensor against the fuel card, the alert thresholds tuned for South African corridors, the fuel scorecard fleet directors recognise, and the data hand-off into SARS-compliant logbook reporting.

A typical 40-truck deployment runs over four to six weeks: week one is hardware fitment, weeks two and three are calibration, and weeks four to six are algorithm tuning. After that, the stack runs in the background and the fleet director reviews the weekly scorecard.

Next step: a one-week fuel audit

Three numbers any Gauteng fleet director should know within seven days of reading this article:

• The litres-per-100km figure for the worst-performing truck in the fleet, and the gap to the best.
• The total rand value of refuel discrepancies in the last quarter.
• The number of trucks in the fleet that do not have a tank-level sensor fitted.

If the numbers are unclear, the data is not joining up. TG Tracking runs a one-week fuel audit for Gauteng fleets that produces these three numbers and a costed plan for the missing layers. Book through the TG Online portal.

Frequently asked questions

How accurate is a tank-level sensor compared to the truck's own fuel gauge?

A capacitive or ultrasonic tank-level sensor reports volume to roughly 0.5 percent resolution after calibration. The truck's stock gauge is a coarse float-based system with 5 to 10 percent resolution. The fitted sensor is the source for fuel-control algorithms; the dashboard gauge is for the driver only.

Will the CAN bus feed work on older trucks?

Most heavy trucks built since 2007 carry a J1939 CAN bus. Older trucks may need an aftermarket reader or a workaround using only the tank sensor and GPS. Accuracy drops without the CAN feed, but it is still useful for catching large siphoning events.

How long does it take to calibrate the algorithm?

Plan for four weeks. The first two weeks build the expected-consumption curve, the third tunes thresholds against real events, and the fourth validates the tuning by replaying historical data. After calibration, the algorithm runs unattended.

What about driver privacy under POPIA?

Fuel-management data is tied to the vehicle, not the driver, and is treated as legitimate business processing under POPIA. Driver-specific scoring requires written notice in the contract and a documented retention schedule. TG Online includes both as standard for SA fleet customers.

Can the algorithm detect theft from a parked truck?

Yes. The tank-level sensor reports continuously, and the algorithm flags any drop while the engine is off and the vehicle stationary. This is the most common siphoning pattern and one of the easier events to catch.

How does fuel data feed into SARS logbook reporting?

TG Online generates SARS-compliant logbooks that include kilometres, business-versus-personal categorisation and fuel costs per trip. The fuel-card transactions and tank-level data combine to give an auditable per-trip cost, which removes manual reconciliation.

What is the typical payback period?

Most Gauteng fleets see a payback of three to nine months on the hardware and subscription cost. Fleets with weak depot security and no prior tank sensors typically pay back inside a quarter. Fleets with strong depot controls already in place pay back closer to nine months but still see a measurable monthly saving thereafter.