Skip to content
Sign inGet Started

Refuelling or Theft? Why Flow Rate Alone Cannot Tell Them Apart in Tank Telemetry

By Saeed Baatiah · Published:

Short answer

The physical rates are far apart, retail dispensing is capped at 37.9 litres per minute and a bulk coupler reaches 2,270, but the telemetry layer destroys that information before it arrives. A CAN level is quantised and smoothed, so the discriminators that survive are ignition state, location, direction and temperature, not slope.

Key takeaways

  • Retail dispensing is legally capped at 10 gallons per minute (37.9 L/min), while a standard 4-inch API bulk coupler is rated to 600 gpm.
  • A CAN-derived fuel level is quantised at 0.4% per bit and broadcast once per second, about 1.6 litres per bit on a 400-litre tank.
  • Teltonika's own prescribed cure for level jumps is a moving average it warns "could delay the response to real changes".
  • Diesel expands about 0.085% per °C, so a 20 °C daily swing moves a 400-litre tank by roughly 6.8 litres with nobody touching it.
  • No standards body publishes a siphoning flow rate, so any rate-based rule is calibrated against a number that does not exist.

01The rates really are far apart

The intuition behind rate-based detection is sound, and the underlying physics supports it. Retail fuel dispensing has a legal ceiling: United States federal regulation states that the dispensing flow rate must not exceed a maximum value of 10 gallons per minute, which is 37.9 litres per minute. Bulk delivery operates in a different regime entirely. A standard four-inch API bottom-loading coupler, the connection a tanker uses, is rated by its manufacturer at 600 gallons per minute, roughly 2,270 litres per minute, about sixty times the retail ceiling. NIST Handbook 44 reinforces the separation by treating 30 gallons per minute as the metrological boundary between large-capacity bulk devices and small-capacity motor-fuel devices, assigning them different accuracy classes. So a bulk delivery, a retail fill, and a slow gravity draw genuinely are three different physical processes with three different signatures.

02And the telemetry layer throws the rate away

The problem is what survives the trip from the tank to the server. The FMS-Standard interface, which is how a truck exposes its fuel level over CAN, defines the fuel level signal SPN 96 with a gain of 0.4 percent per bit and a repetition rate of 1,000 milliseconds. On a 400-litre tank that quantisation step is about 1.6 litres, and the signal updates once per second. The same standard adds a warning that is easy to skip: the mentioned resolution is not related to the accuracy of the signal, so 0.4 percent per bit is the best case rather than a measurement guarantee. J1939 also defines fuel level as a ratio of fuel volume to total container volume rather than a volume, which means every litre figure downstream depends on a configured tank capacity and inherits any error in it. A coarse, once-per-second ratio is a poor instrument for measuring a slope.

03The filter that the manufacturer warns about

Raw CAN fuel level jumps while a vehicle moves, because the liquid moves. Teltonika documents the problem and prescribes the cure: an exponential moving average, with an averaging constant recommended in the range of 2,500 to 5,000 for moderate smoothing and 6,000 or more for very unstable signals. The manufacturer also states the cost of that cure in plain terms, advising that the value be adjusted gradually because setting it too high could delay the response to real changes in fuel consumption. That sentence is the crux of this article. The filter that makes the data usable is the same filter that flattens the very slope a rate-based theft rule wants to measure, and it cannot distinguish sloshing from a genuine fast drop, because both look like a rapid change. Smoothing is not optional on a moving vehicle, so the rate information is not recoverable by configuration.

04A record is only written when something triggers it

Even before filtering, the server does not receive a continuous curve. Teltonika acquires records by four simultaneous methods, time, distance, angle, and speed change, and entering zero disables each one individually. Whether a fuel drop is visible at all therefore depends on whether a record happened to be taken while it was in progress, and the time-based interval is the parameter that decides that. This is the mechanism behind the practical failure the owner describes: a two-minute siphoning event and a two-minute refuelling can each collapse into a single pair of samples, one before and one after, with an identical apparent slope between them. The slope in that case is not a measurement of the physical flow rate at all. It is an artefact of the sampling interval, and comparing it to a rate threshold compares the wrong quantity to a threshold that was never calibrated against it.

05Temperature moves the level with nobody present

Before any threshold is applied, one correction is mandatory, because diesel changes volume with temperature and a tank standing untouched overnight will report a different level in the morning. Measurement Canada publishes the official volume correction factors for diesel per API MPMS Chapter 11.1: for 840 kg/m³ diesel the factor is 1.0000 at 15 °C, 0.9958 at 20 °C, and 0.9873 at 30 °C, approximately 0.00085 per °C, or 0.085 percent of volume per degree. On a 400-litre tank a 20 °C day-night swing is therefore roughly 6.8 litres of apparent change with no fuel added or removed. That figure matters because it is the same order of magnitude as a small real theft, so an uncorrected threshold set low enough to catch habitual small siphoning will fire on ordinary Gulf diurnal temperature variation instead, every single day.

06The sensor is five times less accurate while moving

Motion degrades the measurement itself, and one manufacturer publishes the split explicitly rather than quoting a single accuracy figure. Technoton states for its DUT-E fuel level sensor a measurement inaccuracy of not more than 0.2 percent for stationary fuel tanks and standing vehicles, and 1 percent for moving vehicles, a fivefold difference on the same hardware. Teltonika's UL202 ultrasonic sensor publishes a liquid measurement accuracy of ±0.5 percent. Those numbers bound the smallest level change worth acting on, and they say that the trustworthy window for detecting a small loss is when the vehicle is at rest. The standards world agrees that liquid motion corrupts measurement: OIML R117 requires the compartments of road tankers to be fitted with an anti-swirl device unless a compliant gas separator is present, which is a recognition at standards level rather than a sensor quirk.

07What the discriminators actually are

If slope is unreliable, what remains is context, and every element of it is available at better resolution than the fuel level itself. Movement state is one: Teltonika checks movement criteria every second, so ignition and motion are known at one-second granularity while the smoothed fuel level is not. Direction and magnitude are another, and OIML R117 supplies a floor for the magnitude by requiring dispensers with a maximum flowrate up to 60 litres per minute to have a minimum measured quantity not exceeding 5 litres, below which a dispensed volume is not metrologically acceptable in the first place. Location is the third: proximity to an authorised depot or fuel station separates an expected gain from an unexpected one. Temperature is the fourth, applied as a correction before anything else. None of these is a rate, and together they answer the question the rate was being asked to answer.

08A second, independent measurement channel

Where the stakes justify the installation cost, the strongest answer is to stop relying on level alone. A flow-based channel measures a different physical quantity and fails in different ways, so agreement between the two is meaningful evidence rather than a repeated opinion. Teltonika supports impulse counting from a flow meter, resetting the count with each record, and states that impulse-based fuel usage monitoring is much more accurate than other types of fuel metering sensor, recommending it for high fuel usage machines such as cranes, construction vehicles, mining machines and agricultural equipment. The input frequency limits are documented, 50 Hz on the first digital input and up to 30,000 Hz on the others. A level that falls while the flow channel records no dispensing is a far stronger signal than a level that falls quickly, because the two channels cannot be fooled by the same artefact.

09Detecting the slow, repeated loss

Habitual small-volume siphoning is a different detection problem from a single large event, and it is the one a per-event threshold is worst at. The regulatory world solves it by reconciliation over a period rather than by event detection. The US EPA requires statistical inventory reconciliation for underground storage tanks to detect a leak at least as small as 0.2 gallons per hour or a release of 150 gallons within a month, and its inventory-control threshold treats an overage or shortage greater than or equal to 1.0 percent of the tank's flow-through volume plus 130 gallons as an indication the tank may be leaking. The transferable idea is the structure, not the numbers: a small persistent discrepancy is found by comparing cumulative dispensed volume against cumulative level change over weeks, where the repeated pattern accumulates while random measurement error does not.

10The calibration error that manufactures a false drop

One source of apparent theft is not an event at all but a conversion fault, and it is worth eliminating before any investigation. Teltonika documents that analog fuel senders are non-linear and must be calibrated into a polynomial with coefficients a0 through a3, built from a table of litres against voltage. It also gives a diagnostic that costs nothing to check: the voltages must increase or decrease consistently, and if the values go up and down between different fuel levels, the underlying issue must be found before continuing. A non-monotonic calibration curve will produce level readings that fall while fuel is being added, which reaches the platform as a drop. Any theft alert on a vehicle whose calibration has not been verified should be treated as unproven, because the cheaper explanation has not been ruled out and confronting a driver on that evidence damages more than it recovers.

11Even a sealed meter is not exact

Reconciliation compares telemetry against a delivered or dispensed figure, and it is worth knowing that the reference side has its own tolerance. OIML places fuel dispensers and measuring systems on road tankers in accuracy class 0.5, with maximum permissible errors of 0.5 percent on one line and 0.3 percent on the other. A legally sealed, periodically verified meter is therefore permitted to disagree with true volume within those bounds, which sets a floor on how tight a reconciliation tolerance can usefully be: a rule that flags any discrepancy above a tenth of a percent is flagging approved metrology, not loss. NIST adds a sense of scale from the other direction, requiring a bulk-delivery test draft to be at least the volume delivered in one minute at the device's maximum rate and in no case less than 180 litres, which is roughly the magnitude a genuine tanker transaction occupies.

12A rule that uses what the data can actually support

The practical rule follows from the constraints rather than from a preferred threshold. Correct for temperature first, using a published volume correction factor rather than a local assumption. Evaluate level changes when the vehicle is stationary with the engine off, where the sensor's own published accuracy is roughly five times better. Classify by direction, magnitude against a minimum credible quantity, and location against authorised fuelling points, not by slope. Where the asset justifies it, add an independent flow channel and require the two to agree. Run a cumulative reconciliation over weeks to catch the slow repeated pattern that no single event threshold will see. And verify sensor calibration before treating any drop as evidence of anything, because a non-monotonic curve produces drops that never happened.

13What ENTEK.AI does not claim here

A number this article would seem to need is deliberately absent: the flow rate of siphoning. We looked for it in standards bodies and peer-reviewed sources and it is not published anywhere we could read. Calculations from first principles are possible, and we excluded them, because a derived figure presented next to a legally defined dispenser ceiling would read as an equally solid measurement and it is not. Vendor pages quoting gravity discharge rates were unreachable or secondary. This absence is itself the argument: a rule that fires when a drop is faster than siphoning is calibrated against a quantity nobody has published. We also publish no detection rate or recovered-volume figure from our own deployments, because we have not run that as a dated study with a stated method, and an estimate dressed as a measurement is worth less than the admission.

What each signal can and cannot establish
SignalResolution availableWhat it can establishWhy it is not enough alone
CAN fuel level (SPN 96)0.4 % per bit, 1 HzThat the level changed, and roughly by how muchQuantised, smoothed, and a ratio of a configured capacity
Apparent slope between recordsWhatever the record interval wasNothing reliableThe slope is an artefact of the sampling interval
Ignition and movement stateChecked every secondWhether the vehicle was at rest and offDoes not say what happened to the fuel
Geofence / locationGNSS, per recordWhether the change happened at an authorised pointA theft can occur inside a depot too
Temperature correctionPublished VCF per °CHow much change was thermal, not physicalA correction, not a detector
Flow meter impulse countUp to 30,000 Hz on DIN2-4That a dispensing event did or did not occurNeeds separate hardware installed
Cumulative reconciliationWeeks of accumulated dataSlow habitual loss no single event showsSays nothing about any individual event

Figures and sources

Questions

Can I detect fuel theft from the speed of the level drop?

Not reliably. A CAN level is quantised at 0.4% per bit, broadcast once a second, and smoothed by a filter the manufacturer warns delays real changes. Records are written only when a trigger fires, so the apparent slope often reflects the sampling interval rather than the physical flow rate.

How fast is siphoning compared with refuelling?

Retail dispensing is legally capped at 37.9 litres per minute and a bulk coupler is rated to about 2,270. We found no siphoning rate published by any standards body or peer-reviewed source, so we do not state one, which is precisely why a rate-based rule has nothing solid to be calibrated against.

My tank shows less fuel in the morning and nobody touched it. Why?

Thermal contraction. Diesel changes volume about 0.085% per °C per the published correction factors, so a 20 °C day-night swing moves a 400-litre tank by roughly 6.8 litres. Apply a volume correction factor before any threshold, or ordinary Gulf temperature variation will trigger alerts daily.

When is a fuel level reading most trustworthy?

When the vehicle is stationary with the engine off. Technoton publishes an inaccuracy of not more than 0.2% for standing vehicles against 1% for moving ones, a fivefold difference on the same sensor, which is why evaluating level changes at rest is the single cheapest accuracy improvement available.

How do I catch small repeated theft rather than one big event?

By reconciliation over a period rather than per-event thresholds. The US EPA approach for storage tanks compares cumulative dispensed volume against cumulative level change over a month; the repeated pattern accumulates while random measurement error does not, which is what makes a small persistent discrepancy visible.

Sources

Related insights