Load Cells for Self-Service Weighing Stations in Smart Restaurants: The Reading Now Issues an Invoice

1. Four questions a serving line has to answer

A self-service weighing station in a restaurant is not a scale with a display bolted on. It is a pricing instrument that happens to use a load cell. The guest takes a plate, serves himself, and walks away; there is no cashier to check anything; and the number that appeared on the screen while he was still reaching for the ladle has already been written into his account. That single change — the price is computed from the mass — moves the whole design into a different legal world.

Before we get to load cells, it is worth being explicit about what the machine is actually being asked to do, because only one of these four jobs belongs to the sensor.

I make load cells, junction boxes and indicators. I do not make the station, and I do not certify it. But roughly every second enquiry I get about a canteen or buffet weighing line arrives with a specification that cannot be approved, and the sender does not know it. So this article is written from the position I actually occupy: the person who has to explain, before the purchase order, that the accuracy you asked for and the platform you asked for cannot exist in the same instrument.

2. The reading issues an invoice

I have written before about scales that measure things nobody charges for: the coffee shot scale, the 3D-printer filament spool, the bath mat, the pet bowl. All of those live in a comfortable place, legally speaking. In the European Union a weighing instrument used for a purpose outside the six regulated categories carries only the manufacturer’s name, Max and Min — and it must not carry a conformity marking. Nobody inspects it. The worst that happens when it drifts is that you get a bad cup of coffee.

A restaurant self-service weighing station is on the other side of that line, and it is worth naming the exact clause, because it is the whole reason this article has a standards section at all. Directive 2014/31/EU — the Non-Automatic Weighing Instruments Directive — lists six regulated uses in Article 1(2). The relevant one is (f):

The reason this matters so much commercially is that the demand side is no longer voluntary. On 31 December 2030 every EU member state has to reduce food waste at retail and consumption — and that bracket includes restaurants and food services — by 30 % per capita against a 2021–2023 baseline, with a separate 10 % target for processing and manufacturing. That is written into the revised Waste Framework Directive (Directive (EU) 2025/1892), and the member states had to nominate a competent authority for it by 17 January 2026. The Union throws away roughly 59 million tonnes of food a year, about 131 kg per person, with an estimated market value near €132 billion, and 11 % of that arises in food service and restaurants.

You cannot cut waste you cannot measure. That is the mechanical link between a sustainability target agreed in Brussels and a strain gauge bonded to an aluminium beam. The category was born in China, where the smart-canteen market passed ¥41 billion in 2026 and is still growing above 28 % a year, where canteen penetration went from 3–5 % in 2021 to 12–15 % in 2026, and where more than 42 % of state-enterprise, university and campus canteens have already moved to pay-by-weight self-service. The reported results there are consistent: labour cost down 15–25 %, waste down 30–50 %, payback in 1.5 to 2.5 years, and about ¥120,000 a year saved in a 1,000 m² canteen. The instrument behind every one of those numbers is a scale with a legal obligation attached.

One last boundary, because it trips up almost every project I see. A weighing instrument is non-automatic when a person intervenes in the weighing — places the load, decides the reading is acceptable, completes the transaction. That is what a guest serving himself does. It is also what a member of staff does at a manual checkweighing bench. Automatic catchweighers, checkweighers, belt weighers and gravimetric fillers are covered by the MID instead, in Annex MI-006. So a self-service station, however much software is inside it, is a NAWI. If a supplier tells you their canteen scale is “MID MI-006 approved”, they are describing a different machine.

3. e is a legal integer, not an accuracy knob

Here is the trap. Every project brief I receive says something like “we need 1 gram accuracy”, and the sender is thinking of it the way you think about a ruler: a finer mark is a better ruler. In legal metrology that is backwards. Under OIML R76 an instrument is classified by e, the verification scale interval, and the classification table does not let you pick e on its own. It pairs e with two other numbers and then hands back the largest capacity you are allowed to have.

For Class III — medium accuracy, which is where every bench scale, platform scale and retail scale lives — the relevant lines of the table are these:

Read those three lines together and something uncomfortable falls out. The capacity is not a free variable. If you want a 1 g division, the largest instrument you may build is 10 kg — and the smallest portion you may charge for is 20 g. If you want a 2 g division, the ceiling is 20 kg and the floor is 40 g. If you want to put a whole serving bay on the platform and you need 30 kg of capacity, then e must be at least 5 g, and the smallest bowl you may price is 100 g.

But the class table is only half of it, and this is the half almost nobody quotes. The load cell underneath has its own classification, in OIML R60, and its own ceiling on the number of intervals it may carry. A Class C load cell — which is essentially the entire commercial strain-gauge family, and every part in this article — is certified between 500 and 10 000 intervals, and the “C3” that appears on nearly every datasheet means exactly 3 000. Class B exists on paper, from 5 000 to 100 000 intervals. I have never quoted one, because it is not a thing the aluminium parallel-beam family is made in.

I want to point out something about that table rather than let you find it. It has no row where the minute column and the capacity column are both generous. The division is bought with the platform. This is the same truth as the one I keep repeating about capacity and absolute error, expressed in the currency of legal metrology instead of the currency of physics — and it is the reason I now ask every canteen enquiry the same first question, which is not about accuracy at all. It is: what is the smallest bowl you sell, and how much does that bowl cost?

Because if the smallest legal bill is €0.40, nobody at the counter will ever notice a 1 g division. And if the smallest bill is a bowl of soup priced at €0.80, then a 2 g division is already five times finer than the smallest price change the guest can see. Which raises a question that takes this article somewhere I did not expect when I started writing it.

Half the “1 g” specifications in this industry are d, not e

There are two symbols in legal metrology that look the same on a datasheet and mean completely different things. d is the scale interval the display shows — the smallest step you can read. e is the verification scale interval, the unit the inspector uses to calculate the tolerance and the class. They are frequently different, and when they are different, both have to be marked on the instrument. Tolerances always follow e, never d.

Now consider the two documents that land on my desk most often. The first is a competitor’s solution sheet for a canteen weighing line — a parallel-beam digital cell, 50 to 100 kg capacity, dividing value 1 to 2 g, platform from 250 × 350 mm up to 400 × 500 mm, RS485 or RS232. The second is my own datasheet for the digital planar cell in the card list at the end of this article, which offers 1 g up to 50 kg and 2 g at 60 to 100 kg, with a 300,000-count converter.

Both numbers are real. Neither one is e. Watch what happens if you treat them as e. A 100 kg platform with a 1 g division has n = 100 000. That is not Class III — Class III stops at 10 000. It is not even in the same family; it is Class II, whose interval count runs from 5 000 to 100 000 and whose minimum capacity is 50 e, so 50 g. And it needs a load cell certified to 100 000 intervals, which is the very top of Class B, a class that does not exist in the parallel-beam family at all. On a 50 kg cell with a 2 g division you get n = 25 000, which is the same story one size down. And if you put a 100 kg cell and a 2 g division together, n = 50 000. All of these are Class II numbers stamped on Class C parts.

I am deliberately writing this section against my own product page, not just the competitor’s. When you buy load cells for self-service weighing stations from me and the datasheet says 1 g, you are buying a cell that can resolve 1 g. Whether the station may legally price in 1 g is decided by the table two screens up, and by what platform you bolted on top. A supplier who will not draw that distinction for you is either new to the segment or hoping you are.

4. Two tolerances, and 70 % of one of them is mine

You now have e. The next thing you need is the tolerance that goes with it, and there are two of them, and they are not the same number. On initial verification, Class III instruments must hold ±0.5 e from zero up to 500 e, ±1.0 e from 500 e to 2 000 e, and ±1.5 e above that. In service, all three are exactly doubled: ±1.0 e, ±2.0 e and ±3.0 e.

So a 2 g station has to hold ±1.0 g / ±2.0 g / ±3.0 g when it is new, and double that in service. That is the budget for the whole machine: the load cell, the junction box, the cable, the indicator, the A/D, and the platform. How is it divided?

This is where a document most integrators have never seen comes in. When a NAWI is certified from separately approved modules — which is how nearly every scale in the world is actually built — each module is assigned a fraction of the maximum permissible error, and the fractions are combined in quadrature. On real declarations of conformity for retail platforms the indicator gets 0.5, the load receptor — platform, frame, structure — gets 0.5, and the load cell gets 0.7. Add the squares and you get 0.25 + 0.25 + 0.49 = 0.99. The fractions are designed to fill the budget exactly once, with nothing left over and no component generously funded.

And now the elegant part, which took me two years to appreciate. Under OIML R60 a Class C load cell’s own error limits are p_LC × 0.5 v, p_LC × 1.0 v and p_LC × 1.5 v, where v is the cell’s own minimum verification interval, equal to Emax divided by its interval count. A 6 kg C3 cell therefore has v = 6 000 g ÷ 3 000 = 2 g, and its own limits come out at ±0.7 g, ±1.4 g and ±2.1 g. Those are not similar to the numbers in the table above. They are the same numbers. Two recommendations written decades apart, by different committees, add up to the same place — and the reason is that v = e is the design intent, not a coincidence. One cell, three thousand divisions, one station.

Which gives us the first genuinely useful rule of this article, and it is a rule about paperwork rather than electronics. A %F.S combined-error figure does not tell you whether a load cell is legal for trade. It tells you one number at one load. The document that answers the question is the R60 certificate, because it carries the staircase: the interval count, the minimum verification interval, and the error limits at each step. When an integrator sends me a competitor’s datasheet and asks whether they can price food with it, the correct answer is almost always “show me the certificate” — and about half the time there is no certificate to show.

5. A tare is a material property

Everything so far has been about the food. Now consider the thing the food is sitting on, because in a self-service station the plate is not a container. It is a deduction, and the deduction is stored in a database as a number. That number was written by somebody, in an office, from a supplier’s catalogue, and it describes a dry plate.

Meanwhile the plate on the platform has just come out of a commercial dishwasher at 65 to 85 °C wash and 82 to 90 °C rinse. It is warm, and in the case of anything porous, it is not the same object it was in the catalogue.

Read the third row again. Earthenware at 5 to 15 % uptake is a 35 to 105 g tare error on a single plate, and it is not exotic — it is the cheap buffets and the outdoor terrace sets. But even the middle row is enough to cause a legal problem, so let us do that one properly. Stoneware, 700 g plate, 1.6 % uptake in the middle of the range: the plate that leaves the dishwasher weighs 711 g, and the database says 700.

Eleven grams. At €20 per kilogram that is €0.22 per tray. Small enough that nobody will ever complain about it at the till, which is exactly what makes it dangerous. Run the year: a hotel or campus restaurant doing 900 trays a day, 360 days, is 324,000 transactions, and 11.2 g each is €72,576. Switch the same site to vitrified porcelain at 0.5 % uptake, where the plate gains 3.5 g, and the same 324,000 trays come to €22,680. The difference between two plate materials is €49,896 a year — and that is not a tableware budget, it is a tare convention.

Why “just press zero with the plate on” makes it worse

The obvious field fix — have the guest put the empty plate on and press zero, so the wet weight disappears — works perfectly, once, for one plate type. It also moves the plate out of the tare path and into the zero path, and those two devices have completely different legal limits, which is the subject of the next section. And it quietly breaks every other plate in the building, because there is no longer a tare value at all: the plate has become part of the zero. A soup bowl, a side plate and a dessert dish now all weigh zero, and every guest who picks a different dish is mis-priced by the difference in the tableware.

The legal metrology community has a name for the pattern that does work, and it is worth quoting. Tare values must correspond to the scale interval of the instrument, so your 700 g tare on a 2 g station has to be rounded to a multiple of 2 g — mine was 700, which is fine, but a 703 g average plate is not. A tare value may be increased by repeated operations of the tare device, but cannot be reduced by the same method — before a lower tare can be established, the instrument must be returned to zero. A pre-set tare cannot be modified or cancelled while any subsequent tare operation is still in use. And the pattern the guidance actually recommends for direct sales to the public is the one-shot tare: an empty container is tared; the full container replaces it without the instrument ever seeing zero; and when the transaction is complete and the pan is clear, the tare drops out and the instrument returns to zero.

Then the sentence that explains why this whole article exists. In the same guidance: “It is the responsibility of the operator of a weighing instrument to ensure that it is at zero before each transaction.” A self-service station has no operator. So who holds that responsibility, and what have you installed in the machine to discharge it? That is not a rhetorical question — it is the design review, and if your answer is “the firmware does it,” you need to know what the firmware is legally allowed to do.

6. The zero device is not a tare device

Every integrator I have met treats “zero” as a convenience function. In legal metrology it is a device with a name, a range and a rate, and all three are fixed by the recommendation rather than by your firmware team. Three numbers matter.

Also fixed by the recommendation, and more important than any of the ranges: zero-tracking may only run when the indication is at zero — or the negative net is equivalent to gross zero — and only when the equilibrium is stable. And when the equilibrium is disturbed, the instrument must not print, store, zero or tare at all. That last sentence is the one I would put on the wall of every canteen station. The law already tells you that you may not take a zero while the guest’s arm is still on the counter. You do not need to invent that rule; you need to make sure your firmware obeys it.

7. Three ways to mount a cell in a serving line

Everything I have ever been asked to quote for a self-service station turns out to be a variation of one of three arrangements, and the choice is almost always made by whoever drew the steelwork, months before anyone asked a metrologist. The three arrangements are not variations in quality. They are variations in how much of the tray a single cell is asked to be responsible for, and that single decision sets the capacity, which sets v, which sets e. The division is decided at the drawing stage and merely discovered later.

Arrangement C is the one in the reference designs, and for a simple reason: it is the only one where the cell never sees the plate. If the guest fills a bowl on a pad, the pad weighs the food and the bowl, and the plate stays on the tray where it belongs. Arrangement B is what you need when a whole loaded tray has to be priced in one go, which is what most canteens actually ask for. Arrangement A looks like the cheap version of B and is almost always the expensive one, because the plate has moved onto the platform and you have bought yourself the tare problem from section 5 with no way to solve it mechanically.

There is one more number in arrangement A and B that almost nobody reads properly, and it is on every single-point datasheet I sell: the recommended platform size. WST1201 says 250 × 350 mm. WST202 and WST1204 say 400 × 400 mm. WST1205 says up to 800 × 800 mm. That number is not a suggestion about counter space. It is the size of lever the cell has been characterised for, and it is the reason a single-point cell can hold its specification when a guest drops a plate on the far corner. Fit a 600 × 500 mm top to a 400 × 400 mm cell and the cell has not become inaccurate — it has become unpredictable at the corners, which the eccentric loading test will find and which you cannot fix in firmware.

And if you do go to four corners, the summing device stops being a convenience. On a NAWI, the junction box and every component that sums the load cells has to be protected against unauthorised access, because a trim potentiometer that anyone can turn is a way to change what the instrument measures without leaving a trace. This is where the choice of cell starts to change the paperwork. WST1101 is a planar beam cell designed to be used three or four to a platform, and it is not corner-adjusted — the corner matching has to happen in your summing box, which means your summing box has trim pots, which means your summing box has to be sealable. WST1201 and WST1204 arrive with the off-centre compensation already installed at the factory to OIML R60, so the summing box only has to add numbers. The digital cell did not make the scale more accurate. It removed a screwdriver from the counter.

So the rule I now use, after nine of these projects: count the places a guest can put something down. One small pad means arrangement C with an analogue cell, which is the cheapest and the best. A tray crossing the counter means arrangement B, and prefer cells that already carry their corner compensation. Arrangement A survives only if the platform is at or under the cell’s rated size, which usually means a platform too small to be useful. The arrangement you draw decides the division, and the division is the thing the law cares about.

8. Plate-level: what the smallest legal bill is

Here is the consequence of section 3 that catches people out, and it is about the bottom of the range rather than the top. You can always print a big number. You cannot print a bill smaller than 20 e. The minimum capacity is not a suggestion or a warm-up zone; below Min the instrument is not required to be accurate at all, and a legal-for-trade instrument that prices below its own Min is selling something it does not stand behind.

Read the third and fourth columns together and a strange inversion appears. A station with a finer division cannot sell a smaller portion — it can sell the same portion at a better resolution, but its floor is high. The €0.40 station and the €8.00 station are not better and worse versions of each other; they are two different products. If your menu includes a 30 g garnish, a shaving of parmesan, a spoon of sauce sold by weight, then a 6 kg station at e = 2 g cannot legally sell it, because 30 g is below the 40 g floor. No amount of engineering fixes that. Either the item is sold by piece, or the station’s Max has to come down, or the law changes, and only one of those three is in your hands.

And the division itself is nearly free once you are above the floor. At €20/kg, one 2 g division is four cents. The entire argument between a 1 g station and a 2 g station, the argument that drives tender documents and doubles the price of a load cell, is a two-cent argument per division — and because a legal instrument only has to be accurate to half a division at rest, the amount of money that actually hinges on the choice is one cent. I have never once been asked to justify the accuracy of a station at that level, but I have been asked to justify the price. When a customer tells me they need 1 g, I now ask one question: what is the lightest thing on your menu that anyone buys by weight? If the answer is above 40 g, the 1 g is decoration, and I say so, which occasionally costs me the cell.

There is a mechanical half to this section as well, and it is the reason I now ask for the plate before I ask for the platform. (The same arithmetic runs the other way on a baby scale, where the cradle is the tare and the patient is the live load — the cradle is lighter, so the division survives, which is exactly the trade this section describes.) The reference product in this market — the machine nearly every European canteen integrator benchmarks against — uses a platform of 390 × 286 mm with two divisions, 2 g and 5 g, and minimum capacities of 40 g and 100 g. Those are not arbitrary numbers. They are exactly the C3 line from section 3: 6 kg at e = 2 g is 3 000 intervals; 15 kg at e = 5 g is 3 000 intervals. Somebody at that company sat down with the same table I have been quoting and picked the two points on the line that fit a dinner plate. A 270 mm plate on a 286 mm platform leaves 8 mm a side. That is the whole design envelope of a plate-level station: not “how much can it weigh”, but “does the plate go on, and does it go on nearly centred every time”.

9. Counter-level: why the steel sets the division

Now put the same exercise on a counter instead of a plate, and watch the division collapse. This is the part of the conversation where a customer who has asked for 2 g finds out that he is going to get 50 g, and that nothing about my cell is at fault. It is worth doing in full, because the chain is short and every link is checkable.

The station in question is a full service counter: a stainless top 1 200 × 700 mm, 3 mm thick, with two basin openings and a sub-frame carrying the load cells. The live load — the food — is at most 8 kg, and the customer asked for a 15 kg platform on that basis, which is a sensible way to size a scale and a completely wrong way to size a counter. Here is what the cells actually see.

I have written about the same mechanism before, in leg-loaded silos and feed towers, where a tank holding tonnes is weighed by the steel underneath it and the live load is almost a rounding error. A canteen counter is the same problem at a twentieth of the scale, and it fails for the same reason: somebody sized the sensor for what goes in rather than for what it is bolted to. The last row of the table above is the one that should be on the wall of every canteen kitchen. Twenty-eight per cent of the load cell’s capacity is permanently occupied by the counter it is bolted to. The food the station exists to weigh is a thin layer on top of 41.8 kg of steel, and every gram of it moves inside the top sixth of the cell’s range, exactly where a %F.S specification is least informative. What a customer means when he says “medium accuracy” is that the food is medium-sized. What a load cell hears is that it is going to spend its life at 28 % of full scale, at a division of 50 g, no matter how good it is.

Compare that with what the same station would have done on a 60 kg cell. Sixty kilograms on the C3 line is exactly e = 20 g, n = 3 000, Min = 400 g, and a smallest legal bill of €8.00 instead of €20.00. The 41.8 kg of steel turned a 20 g instrument into a 50 g instrument and a €8.00 floor into a €20.00 floor. Nobody chose that. It happened between the drawing and the quotation, and it is invisible because the number that caused it — the bare mass of the counter — usually does not appear anywhere on the specification.

It also matters at the load the station is actually used at. An empty receptor already reads 41.8 kg. Add a 340 g portion and the instrument is at 42.1 kg. On the 20 g instrument that is 2 107 e, in the 2 000—10 000 e band, where the permissible error is 1.5 e — 30 g, which is 8.8 % of the portion. On the 50 g instrument the same gross is 843 e, in the 500—2 000 e band, where the error is 1.0 e — 50 g, or 14.7 % of the portion. There is no configuration of the C3 line in which a portion measured on top of a 41.8 kg counter is better than one part in ten. That is not a load cell problem and it is not a tare problem. It is a “why are you weighing the counter” problem.

10. What a digital cell really buys you here

Half of a bowl-station conversion in Asia is specified with digital cells and half with analogue, and the argument for digital is usually stated as resolution. I want to separate the claims, because two of the three common reasons are wrong and the third is the one that matters.

Take the resolution claim first, because it is the one that gets repeated and it is the least useful. A 300 000-count digital cell at 60 kg resolves 0.2 g per count, so a 2 g legal division is ten raw counts wide. At 100 kg it is 0.333 g per count and the division is six counts wide. Six counts is thin but it is not marginal; the reason good indicators want five to ten counts under a division is to keep the last digit from flickering, and six to ten does that job. The counts were never the bottleneck. A 300 000-count converter does not make a station Class II, and it does not let you use a smaller division than the cell’s certificate allows. If you want the proof, look at my own datasheet: WST1201 lists a division of 0.1 g on its 5 kg capacity. Read that as a legal scale interval and it is n = 50 000 — an accuracy class well above anything a parallel-beam cell is certified for, and twice below the Class II ceiling only in the sense that it is half of it. It is not a division. It is a resolution, and it is honest to call it one.

The latency claim is also weaker than it looks. RS485 gives you the reading in 3 ms and RS232 in 20 ms. The busiest single-channel canteen line I have seen quoted does 1 200 people an hour, which is one transaction every three seconds. Seventeen milliseconds of difference is not a factor in whether a queue moves. It becomes a factor in a completely different design — a checkweigher rejecting a package on a moving belt — and if that is your application you should not be reading this article.

What the digital cell actually buys you is the cable and the corner trim pots, and both of those are legal problems rather than engineering ones. On a four-cell counter the analogue signal travels down four cables into a summing box, gets trimmed across four potentiometers that must be sealed, and then travels down a fifth cable to an indicator that has to resolve a 2 mV/V signal in a room with a dishwasher on the other side of the wall. That is a long way for two millivolts to travel while staying inside a ±0.5 e budget. The digital cell does the conversion where the physical quantity is, sends a number, and arrives with its corners already matched. It is not more accurate in the laboratory sense. It is more accurate at the end of the cable, which is where your station actually lives.

So: a four-cell service counter where the guest places a full tray should use digital cells, because the summing and the sealing are the real cost and the cells pay for themselves in the junction box and in the approval meeting. A single small pad where a bowl is placed should use an analogue cell with a short cable, because there is nothing to sum, nothing to trim, and nothing to protect except the cable itself. If you are choosing a cell for the first time rather than replacing one, the nine questions I use are set out in how to choose the right load cell for a project. In neither case does the choice change the two numbers the law cares about. Pick the digital cell because of what it does at the end of the cable. Do not pick it because of the counts, and never pick it because of the milliseconds.

11. The test that fails is eccentric loading

If you take one thing from this article into a design review, take this: of all the tests an approval laboratory runs on a self-service station, the one that finds the mistakes is the shifted-load test. The instrument has to hold the permissible error with the test load placed anywhere on the load receptor — at each corner, at each edge, off to one side. It is not a hard test to pass. It is a very hard test to pass by accident, because the errors it exposes are drawn into the steelwork months earlier and are invisible in every simulation.

The mechanism is a lever, and it is worth stating plainly because it explains why accuracy class does not help. A platform on a centre-mounted cell is a lever pivoting about the cell. A plate placed on the far corner produces the same vertical force as a plate placed in the centre, plus a moment the cell has to resist. A parallel-beam single-point cell exists precisely to absorb that moment without changing its output — that is the entire reason it looks different from a bending beam — but the absorption is only characterised within a certain platform envelope, and that envelope is the recommended platform size on the datasheet. Inside it, the corners agree. Outside it, they stop agreeing, gradually at first and then not gradually at all.

And the guest works against you in a way a laboratory never does. Laboratory weights land vertically, in a marked position, on a settled platform. A guest slides a plate across the top, drops a bowl onto a corner, leans a forearm on the edge while reaching for the tongs, and puts an empty plate back on the platform it just left. The shifted-load test is not an exotic scenario. It is a description of lunch.

The four-corner arrangement changes the problem from one you cannot fix to one you can. With four cells, each one carries a quarter of the platform and a quarter of the lever, and the corner behaviour becomes a matching question between four channels rather than a property of a single envelope. That matching is trimmable, which is why wide platforms are built that way — but it also introduces a slow error that has nothing to do with corner geometry and everything to do with electricity. Both families of my cells are quoted at 2 mV/V, but the input impedance is very different: WST201 and WST202 have a 406 Ω bridge and WST1101 has 1 090 Ω. At the recommended 5 V excitation that is 12.3 mA and 4.6 mA per cell, and about 62 mW against 23 mW of heat dissipated inside the cell. Four of the low-impedance cells on one supply draw 49 mA; four of the high-impedance ones draw 18 mA.

Why that matters on a counter and not on a bench: the strain gauge grid heats itself, and it reaches its working temperature over minutes, not seconds. (If you already have a station in service and are trying to work out whether what you are seeing is a corner problem, a drift problem or a damaged cell, the checks in how to tell whether a load cell is damaged separate the three in about ten minutes with a multimeter.) On a four-cell platform, the four corners do not carry the same load — that is the definition of a leaning guest — so they do not heat identically, and the difference settles slowly and shows up as a corner drift that no trim pot can chase because the trim was set on a cold instrument thirty minutes earlier. It is the classic “it was perfect in the morning and eleven grams out after lunch” report, and it is one of the few places where a higher bridge impedance or a lower excitation is a metrological decision rather than a power-supply one.

12. Heat, steam, and one division per five degrees

There is a number in the temperature requirements that is easy to miss and expensive to ignore. Near zero, the indication on a Class III instrument must not change by more than 1 e for a 5 °C change. On a 2 g division that is 0.4 g per degree. Nothing about that sounds alarming until you watch a canteen through a service: the doors open at eleven, the hot line comes on, four hundred people arrive, and the room climbs 12 °C by half past one. Twelve degrees at 1 e per five is 4.8 g, or 2.4 divisions, or €0.10 a tray at €20/kg — €86.40 a day at 900 trays, which is €31 104 a year.

That number is a budget, not a verdict, and how it is spent depends entirely on the division you chose. The zero-tracking device from section 6 is exactly the mechanism designed to absorb slow thermal drift — but it can only absorb up to 4 % of Max in total, and it may only run while the indication is at zero or at the gross-zero equivalent, and stable. A busy station is loaded for most of the service and only sees an empty platform in the gaps between guests, which is precisely when the room is changing fastest. So on a real counter, a meaningful fraction of that 4.8 g survives to the till, and the question becomes how much of it my cell is responsible for.

The other half of the environment is not a number at all. A commercial dishwasher washes at 65 to 85 °C and rinses at 82 to 90 °C, and the trolley comes back through the dining room hot and wet. My cells are compensated over −10 to +40 °C and specified to operate from −20 to +60 °C. The compensated range is where the published temperature coefficients actually hold. Above +40 °C, the cell has not become unreliable, but it has left the range in which anything on its certificate is a promise. If the cell sits under a platform 400 mm from the dish-return lane, that is not a corner case; it is the standard layout, and it is the reason I ask for a photograph of the room rather than a plan drawing.

13. Twelve stations is twelve verifications

A canteen refurbishment is rarely one station. It is four counters, or a line of eight pads, or twelve identical units across three buildings, and this is where the arithmetic of everything above turns into a project schedule. Every station is a separate instrument. Twelve stations means twelve verifications, twelve sets of paperwork, twelve opportunities to fail, and one acceptance decision that depends on the worst of them. If the fleet is a series of small identical enclosures rather than weighing counters, I have argued elsewhere that the same fleet arithmetic applies to smart lockers, where one summing box and one seal settle the whole bank. Here the difference is that each station is priced separately by a member of the public, so each one has to be right on its own. The cost of the load cells is a line in the bill of materials. The cost of the verification is a day rate multiplied by twelve, and at that point the sensor is the cheap part of the programme.

There is a clause in the verification procedure that rewards good engineering with something rare — less work. The standard test load is normally half of Max. If the instrument demonstrates a repeatability of 0.3 e or better, the verifier may drop to a third of Max; if it achieves 0.2 e, to a fifth. On a 6 kg station at e = 2 g those thresholds are 0.6 g and 0.4 g, and half Max is 3 kg against 1.2 kg at a fifth — a factor of 2.5 in the weight somebody has to carry up and down a service corridor. Now look at my own specification: a combined error of ±0.02 % F.S on a 6 kg cell is 1.2 g, which is three times the 0.2 e threshold. You are carrying 3 kg weights, and across twelve stations that is 36 kg of certified stainless steel being wheeled around a hospital.

That is not a criticism of the cell; it is the concession being designed for instruments that are far better than they need to be, and the honest reading is that the concession is a reward for repeatability, not for accuracy — two properties that are related but not the same, and only one of which is quoted on the datasheet. When a customer asks me whether he can claim the reduced test load, my answer is that he should ask his verifier, and that the way to earn it is to build the station so that it repeats, not so that it scores well: a stiff platform, a flat floor, no cable pulling on the cell, and a fixed mounting torque. Those are the four things I ask about when a client reports a repeatability problem, and about three times out of four the cause is the fourth one.

Two of those five were legal rather than technical, and they are worth naming. On a legal-for-trade instrument the calibration and adjustment must be sealable, mechanically or in software, and the junction box and every component that sums the cells must be protected against unauthorised access. A trim pot anyone can turn is a way of changing what the instrument measures without leaving a trace. So before a fleet goes into service: the summing box is sealed, the calibration switch or service mode is sealed or password-protected, the software version is identifiable and cannot be edited, the secondary display and any printed label agree with the main display to the same value, decimal place and unit, and the descriptive label is present with maker, model, serial number, certificate number, accuracy class, Max, Min, e and the temperature range. None of that is engineering. All of it is the difference between a station that is accurate and a station that is legal.

14. Standards: what applies, and what does not

I am a sensor supplier, not a notified body, so treat this table as a map of the conversation rather than as legal advice. It is the map I use, and the most useful thing about it is the number of rows where the answer is “no”. Most of the compliance noise around a self-service station comes from suppliers quoting documents that do not apply to it.

15. What I supply, and what I do not

The right-hand column is not a list of things I am apologising for. It is the reason the left-hand column is worth anything: if I claimed the right-hand column too, I would be a solutions company selling you a promise instead of a component I can measure.

16. Questions I get from canteen integrators

17. Six load cells for self-service weighing stations I would shortlist

These are the parts I would put in front of a canteen integrator, with the limitation of each stated next to it rather than hidden in a footnote. They are not six answers to the same question: two belong to a plate-level station, two to a wide counter, one is the four-corner option and one is for a whole counter bay. If you want a single load cell for a self-service weighing station rather than a set, the plate-level digital cell is what I would send first.

If you take one thing away from this article, take the constraint rather than the model. A self-service weighing station is a trade-approved scale with the operator deleted, which means every judgement that operator used to make — is this at zero, is this plate the same plate, is this weight worth arguing about — has to be answered by two integers, Max and e, and by the certificate of one load cell. Get those two right and the rest of the station is sheet metal and firmware. Get them wrong and no sensor on the market will rescue the design. Send me the drawing and the menu and I will do that arithmetic with you, including when the answer is that you do not need to buy anything from me.

Where I have quoted a standard rather than our own data, the source is the standard itself: the accuracy classes and the mpe staircase are from OIML R76-1 and the load cell classes and Y values from OIML R60; the direct-sales and zero-setting rules are from the same two documents; the food waste target is from the European Commission’s food waste pages; and the US marking and division rules are from NIST Handbook 44. The Dutch, Irish, German, Belgian, French and Polish cases are my own projects, and the names are omitted because the projects are theirs.