1. Why a banknote counter is really a weighing instrument

Optical and magnetic counting is fast and it is good. A modern friction-fed counter runs at over 1,500 notes per minute and can tell you a note is a note, which denomination it is, and whether it is a forgery, all in the time it takes to pass a contact image sensor. What it cannot tell you is whether the pile it just processed is the pile it started with. Notes stick. Notes fold. Notes tear. A note that passes the head twice is counted twice, and a note that never leaves the hopper is never counted at all. Every operator in every cash room has a story about the strap that was one short.

Weight is the only second opinion that is completely independent of the transport. It does not care whether the note passed the sensor, whether two notes moved together, or whether the stack was fanned. It cares about total mass on the tray — and that makes it the natural cross-check for the optical count in exactly the situations where the optical count is fragile.

The category is not small, and it is not shrinking. One market study sizes the global banknote counter market at USD 2.64 billion in 2025, rising to USD 2.78 billion in 2026 and USD 4.16 billion by 2034, a 5.17% CAGR, with Asia-Pacific holding the largest regional share at roughly 41.5% and the fastest growth rate at 8.83%. A second study, looking specifically at currency counting machines, puts growth at 8.7% in 2026, driven by central bank fitness-sorting mandates that force banks to replace mechanical counters with electronic ones on a four-year rather than seven-year cycle. The reasons are mundane and durable: cash usage in the UK actually rose for a second consecutive year to 19.9% of transactions, and in many emerging markets it exceeds half of all retail payments.

Everything below is written from the supply side of the table. When I say “we” I mean load cells, force sensors, junction boxes, transmitters and indicators — nothing with a roller in it.

2. Where weight enters a cash-handling machine

I quote weighing into cash equipment at four different stations, and they have almost nothing in common technically. Getting a customer to name the station before naming the capacity is the single most useful thing I do on these projects, because the error budget is completely different at each one.

Notice the shape of that table: the resolution you need is roughly one part in ten thousand of the working span at every station. That is not a coincidence — it is the threshold at which a single note, or a single coin, becomes visible in the reading. We will come back to that number, because it is what decides which load cell and which A/D you can get away with.

3. The uncomfortable truth about note weights

Here is where most “count your cash by weight” marketing quietly falls apart. Two different things are going on with cash mass, and only one of them is trustworthy.

Coins are controlled by law. A mint strikes to a statutory mass and the tolerance is tight, so a circulated coin is reliable to roughly ±1–2% — wear removes a little metal, which is the only drift that matters. That is why coin counting by weight works so well, and why bagged-coin recognition (weigh the bag, divide by the unit mass) is a legitimate technique rather than a gimmick.

Banknotes are not. There is no statutory mass tolerance for a note. The published figures are nominal averages derived from substrate grammage and sheet area, and the substrate is a cotton-linen blend that absorbs and releases moisture with the room. One hundred United States notes weigh about 100 g not because a rule says so, but because they are all printed on the same size sheet of the same stock at about 1.0 g each.

There is one more signal hiding in the mass, and it is worth knowing about. Banknote substrate sits around 85–90 g/m². Ordinary copier paper sits around 80 g/m². A stack of notes printed on office stock is therefore measurably lighter than a stack of the real thing — a few percent, well outside the noise floor of a decent tray cell. I would never let a customer sell that as counterfeit detection, because it is not; a photocopy is not a forgery and a forgery will not necessarily be printed on 80 g/m² stock. But as an independent second signal that the optical head knows nothing about, it is genuinely useful, and cheap, and it costs you nothing to log it.

4. The project: a six-currency bench counter

The enquiry came in as a weighing question, which is unusual. An OEM building a portable bench-top counter for retail and banking had already specified the optical path — contact image sensor, magnetic head, UV lamp — and had run into a wall on the commercial side. Their target buyers were small retailers, hotel front desks, and bank branches that handle more than one currency. Those customers did not want a counterfeit detector. They wanted something that would settle a till dispute: count the drawer, count the strap, count the coin bag, print the result, and be fast enough that nobody avoided using it.

That brief turned it into a weighing problem in three places at once: a note tray for loose and strapped notes, a coin cup that had to recognise bagged and rolled coin, and a reconciliation platform. The requirements I worked from, in the customer’s own terms:

We quoted a precision single-point cell for the note tray, a larger single-point cell for the coin cup, a heavy single-point cell for the reconciliation platform, and a high-resolution ADC front end with an RS-485 link so the machine’s own controller could poll the weight whenever it wanted. Straightforward on paper. Three things then went wrong, and none of them was the load cell.

5. Three problems that were not the load cell’s fault

5.1 The hundred-gram strap that weighed 103 grams

The first field trial went badly and produced the best lesson of the project. The tray had been calibrated in the customer’s engineering office with a nominal 100-note strap — 100.0 g, dead on. Three weeks later, a trial unit in a humid coastal branch reported the same strap at 102.6 g and flagged a surplus note on every single bundle. The plant manager was ready to ship the units back.

Two things were wrong, and only one of them was in our scope. First, the tray had been calibrated against notes. That is always a mistake: you have calibrated a scale using a transfer standard whose mass is a function of the weather. Second, the firmware compared total mass against a hard nominal — 100 notes × 1.0 g — with a window so narrow that ordinary moisture variation in cotton-linen paper fell outside it. At 75% relative humidity that variance is real, repeatable, and nobody’s fault.

The fix was procedural plus one firmware change. Calibration now uses an external certified calibration weight — a real mass standard, which is also why a $6 calibration weight is a line item on the accessory list of every machine in this class. And the note check became a tolerance band, not a value: for US notes we shipped a default of ±1.5% per bundle with a three-consecutive-stable-readings requirement before the result is announced, and the band is operator-adjustable per currency. A 100-note strap is now checked against a window, not a number, and the phantom surplus disappeared.

If you take one design rule from this article, take this one: when the artefact you are weighing is hygroscopic, calibrate with an inert mass and judge with a window. Both halves matter, and the second half is the one that gets forgotten.

5.2 Paper dust, and the zero that drifted while the tray was loaded

Second problem: by the end of a long shift in a busy branch, the tray read 0.6 g high with nothing on it. Over eight hours of counting, that is enough to manufacture a partition error out of thin air. The customer suspected the cell. The cell was fine.

The cause was mechanical. Every note that passes through a friction counter sheds a little fibre and a little surface coating, and over a few thousand notes that becomes a fine grey dust layer. Some of it lands on the pivots and the mounting bosses. A load cell measures strain in its own body; if the load path picks up a compliant film of paper dust at the pivot, the zero shifts. Add thermal expansion of an aluminium bracket warming up next to a solenoid and you get the slow creep the customer was seeing.

Three fixes, in increasing order of how much I like them:

• Seal the cell properly. We specified an IP65 cell with the strain-gauge cavity potted rather than merely covered, which matters in a dust environment — the potting is what stops the debris reaching the gauge, not the housing.
• Shield the pivot. A simple lip above the load pivot, added by the OEM in their tooling, keeps the dust film off the working surface at zero cost.
• Bound the zero tracking. This is the one that turns a mechanical problem into a software one and mostly makes it go away. Automatic zero tracking is essential in a machine that runs for eight hours, but an unbounded zero tracker is dangerous: if it runs while the tray holds a load, it will quietly eat part of the load. We set it to operate only below 5 g and only when no counting cycle is in progress, with a maximum correction rate of 0.05 g per second. That is fast enough to follow dust, slow enough that it cannot absorb a note.

5.3 The transport motor that sang into a millivolt cable

Third problem, and the most instructive: with the transport idle, the tray read a beautiful stable number. The moment the feed motor started, the reading jumped ±4 g. The customer’s first instinct — and mine, briefly — was a faulty cell. It was not. It was a 1 mV/V signal running down an unshielded cable in the same loom as a PWM-driven motor and standing next to a high-voltage static eliminator bar. The cell was faithfully reporting the electromagnetic environment it had been installed in.

What we changed, in the order that mattered:

1. Shielded four-wire cable, and the shield earthed at the indicator end only. Earthing a shield at both ends creates a ground loop and makes the noise worse, not better. This is the single most common wiring error I see in customer prototypes.
2. Physical separation. The cell cable was routed down the opposite side of the chassis from the motor harness and the ioniser lead, with a minimum 30 mm gap and no parallel runs longer than 50 mm. Free, and worth more than any filter.
3. Ferrite sleeves on the motor leads, at the motor end, which suppressed the commutator brush noise that was the dominant term.
4. Measure in the settle window. The most effective fix was also the cheapest: the firmware now takes the weight only during a 250 ms pause between feed cycles, when the motor is off and the tray is mechanically quiet, and requires the reading to be stable within ±0.1 g across that window before accepting it. Weight verification and high-speed counting do not have to happen at the same instant.

Resolution, incidentally, was never the problem, and this is worth stating plainly because it is where most spec sheets mislead. The tray cell we used runs at 1.0 mV/V, not the 2.0 mV/V that dominates commercial weighing. It did not need more signal, because the front end resolves 5,000,000 internal counts across a ±10 mV input — roughly 2 nanovolts per count. More output from the cell would not have bought us anything; better shielding and a quieter measurement instant bought us everything.

6. What weight can and cannot do

I would rather lose a sale than have a customer design a feature that cannot work. So here is the honest capability map for the weighing layer in a cash machine.

The architecture that falls out of this table is a two-out-of-three arbitration, and it is what I now recommend by default. The optical head produces a count. The weighing layer produces a mass-derived count. The operator, or the strap label, produces an expected count. If the mass-derived count agrees with the optical count, the machine reports and prints with confidence. If they disagree beyond the tolerance band, the machine does not guess — it re-passes the bundle and, if the disagreement survives the second pass, prompts for a manual check. That is a much better product than one that claims to count cash by weight, because a machine that says “I am not sure” is worth more in a cash room than a machine that says “one hundred” and is occasionally wrong.

7. The arithmetic that actually picks your load cell

Before you argue about sensitivity class, accuracy class or which A/D to buy, do the matching arithmetic. It is one line, and it settles the argument in about two minutes. The standard matching relationship between a load cell and an indicator is:

Worked through the coin cup on that six-currency machine — a 30 kg single-point cell at 2.0 mV/V, 5 V excitation, one cell, 2 g divisions (15,000 divisions across the span):

(2.0 × 5 × 30,000 × 1000) ÷ (15,000 × 1 × 30,000) = 0.67 μV per division

Now compare that with the converter. Our RS-485 front end accepts ±10 mV and resolves 5,000,000 internal counts across that range — about 2 nV per internal count. One 2 g division therefore carries roughly 335 internal counts. That is the margin that decides whether your coin cup is usable: not the headline accuracy class of the cell, but how much of the A/D’s resolution each division occupies. Fewer than about ten counts per division and you are gambling with noise; a few hundred and the job is comfortable.

Run the same arithmetic on the note tray and the picture changes, which is the interesting part. A 3 kg cell at 1.0 mV/V, 5 V, one cell, 0.2 g divisions (15,000 divisions) gives you 0.33 μV per division — half the coin cup’s signal, because the sensitivity is half. Against a 2 nV converter that is still about 167 counts per division, which is comfortable, and it is exactly why we accepted a 1.0 mV/V cell on the tray and did not chase a higher-output cell. The resolution comes from the converter, not from the cell — and the cell you can actually buy in the right physical size matters more than the one with the prettiest number on the datasheet.

8. Where the weighing core ends

I am a weighing-component exporter, and pointing at the boundary is a service rather than a limitation. Here is where the line falls on a banknote counter, so that nobody builds a schedule around a supplier who was never going to deliver half of it.

In Europe, for example, the ECB publishes a database of successfully tested banknote handling machines by manufacturer and country. In India, BIS certification reshapes procurement specifications for cash-handling equipment. Both of those attach to the machine, tested as a whole, by the party who builds and sells it. Our job is to make sure the weighing channel inside that machine is the part that never causes a test failure — which means a cell whose calibration is traceable, a mounting scheme that does not load the cell sideways, and a wiring layout that does not put millivolts in the same loom as a motor.

9. Recommended load cells and weighing electronics

Four components cover essentially every cash-handling machine I quote: a precision tray cell, a mid-capacity cell for the coin station, a heavy single-point cell for reconciliation, and a high-resolution control front end. Where the machine is slim or battery-powered we additionally use the WST112 flat mini cell — its 1,000 Ω bridge draws far less excitation current than a 350 Ω cell, which matters when the whole machine runs from a USB charger — and where the customer wants a printed slip straight from the instrument rather than from their own controller, the WST1503 indicator with built-in thermal printer removes a PCB from the design.

WST209 — Precision Note-Tray Cell

A laboratory-balance-grade single-point cell in a 110 × 33 × 10 mm aluminium body. This is the cell for the note tray and the strap check, where one note has to be visible in the reading.

• Capacity: 0.3, 0.6, 1, 1.5, 3 kg
• Class: C3 · comprehensive error ±0.02% FS
• Repeatability: ±0.015% FS · hysteresis ±0.015% FS
• Creep: 0.03% FS / 30 min
• Output: 1.0 ± 0.1 mV/V · 405 Ω in / 350 Ω out
• Temperature: compensated -10 to +40 °C, operating -20 to +60 °C
• Sealing: IP65, potted gauge cavity · 4-wire shielded, φ4 × 250 mm

Best for: the note tray, the 100-note strap check, and any station where 0.1 g resolution on a 1–2 kg span is the requirement.

View WST209 Specs →

WST216 — Coin Cup / Bulk Tray Cell

A 200 × 200 mm platform cell that suits a coin cup, a rolled-coin tray or a bagged-coin station. Its 2.0 mV/V output gives the front end twice the signal of the tray cell, which is what you want on a heavier span.

• Capacity: 4, 8, 10, 15, 20, 30 kg
• Class: C3 · comprehensive error ±0.02% FS
• Repeatability: ±0.02% FS · hysteresis ±0.02% FS
• Creep: 0.02% FS / 30 min
• Output: 2.0 ± 0.2 mV/V · 406 Ω in / 350 Ω out
• Temperature: compensated -10 to +40 °C, operating -35 to +70 °C
• Material: anodised aluminium · IP65 · 4-wire shielded, φ3.8 × 150 mm

Best for: loose and bagged coin counting, where the statutory coin mass spec makes weight a genuinely primary measurement.

View WST216 Specs →

WST207 — Reconciliation Platform Cell

Designed for counting scales and 250 × 350 mm platforms, which is exactly the size of a till drawer or a coin bin. Use it where the number only has to be right to the nearest coin.

• Capacity: 3, 5, 6, 10, 15, 20, 30, 40, 50, 60, 100 kg
• Class: C2 · comprehensive error ±0.03% FS
• Repeatability: ±0.03% FS · hysteresis ±0.03% FS
• Creep: 0.03% FS / 30 min
• Output: 2.0 ± 0.2 mV/V · 405 Ω in / 350 Ω out
• Temperature: -10 to +40 °C operating range
• Platform: 250 × 350 mm · IP65 · 4-wire shielded, φ4 × 400 mm

Best for: end-of-day till reconciliation, bagged coin and mixed cash by weight, on a bench-top machine.

View WST207 Specs →

WST1502 — High-Resolution Weighing Controller

The front end that makes 0.1 g resolution on a kilogram tray practical. Isolated RS-485 Modbus RTU means the machine’s own controller polls the weight instead of the operator reading a display.

• Conversion: Σ-Δ A/D · input range ±10 mV
• Resolution: 5,000,000 internal counts
• Speed: 10 / 20 / 80 conversions per second, selectable
• Excitation: 5 V DC, drives up to 4 × 350 Ω or 8 × 700 Ω cells
• Interface: isolated RS-485 Modbus RTU · optional Modbus TCP, 4–20 mA, set-points
• Functions: counting function, programmable digital filter, programmable zero tracking, accumulation
• Power: battery operation up to 30 hours · stainless steel housing

Best for: embedded weighing inside the machine — the programmable filter and bounded zero tracking are what let you run an eight-hour shift without drift.

View WST1502 Specs →

10. FAQ

Do banknote counters actually use load cells?

Yes, but not in the way most people assume. The load cell is not the counter. It is the verification layer sitting behind the optical count, plus the primary counter for coins. Machines in this class typically carry a precision platform cell for the note tray and bundled notes, a mid-capacity cell for the coin cup, and sometimes a larger platform cell for bulk reconciliation. The optical head counts notes as they pass; the weighing channel asks whether the pile weighs what that count implies.

Can you count banknotes by weight alone?

On a single known denomination, approximately — a 100-note strap of one denomination is a legitimate and very useful weight check, and the useful window is around ±1.5%. Across mixed denominations, generally no. Every United States note weighs about 1.0 g regardless of value, so weight cannot distinguish a dollar from a hundred. Euro and sterling series are size-graded and therefore give a usable mass separation between denominations, but it is a corroborating signal rather than a certified answer. Treat weight as the second opinion, never the sole count.

Why does a stack of 100 notes not weigh exactly 100 grams?

Because banknote substrate is a cotton-linen blend and it is hygroscopic. The nominal 1.0 g per note for US currency is an average of a paper that absorbs and releases moisture with the room; the same strap can differ by 2–3% between a dry office and a humid branch. Coins do not have this problem — their mass is a statutory specification controlled at the mint — which is precisely why coins count well by weight and notes need a tolerance window.

What resolution does the weighing tray actually need?

As a rule of thumb, about one part in ten thousand of the working span — the point at which a single note or coin becomes visible in the reading. On a 1–2 kg note span that is 0.1 g; on a 30 kg coin span, 2 g. Then check that your A/D can carry it: divide the signal per division by the converter’s resolution per count and make sure you have a comfortable margin, not just a passing one. Our own working figure runs to a few hundred A/D counts per division.

How do you stop paper dust from drifting the zero?

Three layers. Seal the cell properly, with a potted strain-gauge cavity rather than a housing that only covers it. Shield the load pivot with a lip in the tooling so the dust film never reaches the working surface. And give the firmware a bounded automatic zero tracker — active only below a small threshold and only outside a counting cycle, with a capped correction rate. An unbounded zero tracker on a loaded tray will quietly eat part of the load, which is a worse failure than the drift you were trying to fix.

The reading jumps when the transport motor runs. Is the cell faulty?

Almost never. A millivolt signal sharing a cable loom with a PWM motor and a high-voltage static eliminator will report exactly that noise, faithfully. Fix the path before you replace the transducer: shielded four-wire cable with the shield earthed at the indicator end only, physical separation from the motor harness, ferrite sleeves at the motor, and — usually the most effective change — take the weight only in the settle window between feed cycles, requiring three consecutive stable readings before accepting the result.

Which load cell class do I need for a counting machine?

For the tray, C3 with repeatability around ±0.015% FS and creep no worse than 0.03% FS per 30 minutes. For the coin station, C3 with creep of 0.02% FS per 30 minutes where available. For a reconciliation platform, C2 is usually adequate because the requirement is only to the nearest coin. Note that creep is specified as a percentage of full scale per half hour — multiply it out across an eight-hour shift before you decide it does not matter, because on a light span it very much does.

Does VektorForce supply the complete money counting machine?

No, and we are deliberate about it. We supply the weighing core — load cells, single-point cells, high-resolution controllers and indicators with RS-232 / RS-485 outputs, junction boxes, and a calibration protocol with a traceable record per unit. The note transport, the authentication sensors, the count-arbitration firmware, the enclosure, the EMC testing and any machine-level approval or central-bank machine listing belong to the OEM who builds and sells the machine. We would rather be excellent at the measurement layer than mediocre across the whole chain.