Load Cells for Sorting Scales: The Law Caps the Error Spread

1. What a sorting scale actually does

A sorting scale — most of the world calls it a checkweigher — is not a scale with a conveyor bolted to it. It is a decision machine that happens to weigh. Every pack on the line crosses a short weigh belt, the instrument compares the reading against limits it was given, and something physical happens at the other end: a pusher fires, a flap drops, an air jet pops, or the pack simply continues into one of several bins. Nobody looks at the number. The number looks at the pack.

That makes it the third kind of weighing I have written about on this blog. On a trade-approved self-service weighing station the reading issues an invoice — the mass becomes money while the guest is still holding the spoon, which is why that application drags the entire legal chain of non-automatic weighing behind it. At an AI checkout the reading only confirms what a camera already recognized, and the weighing is a supporting actor. Here, the reading fires an actuator. Three different legal worlds, three different sensor conversations — and because the sorting scale usually stands two metres from the first of those three, I drew the boundary between them in my article on load cells for packaging machinery.

Before we get anywhere near a datasheet, it is worth separating the jobs people lump under “sorting scale”, because they are legally different instruments — and because load cells for sorting scales are chosen differently for each one:

One of the oldest arguments for the machine — and my favorite slide in any customer meeting — is the sampling math. A line running 100 packs a minute produces 6,000 packs an hour. A diligent QA team pulling 15 manual samples an hour checks 0.25 percent of production. The other 99.75 percent leave the plant on trust. A checkweigher is the machine that replaces a statistic with a census: every pack is weighed, every deviation is logged, every reject has a timestamp. When a retailer or a weights-and-measures inspector later asks how you know your net content is legal, the answer is either “I sampled” or “I measured all of it.”

2. The three Packer’s Rules behind every checkweigher

If you sell prepacked goods into the EU — and most of my customers do — the document that shapes your packaging line is not a checkweigher brochure. It is Council Directive 76/211/EEC, the average-quantity system that governs prepackages from 5 g to 10 kg, plus its sampling update 2009/23/EC. Pack to its rules and you may mark the package with the “e” sign, which acts as a metrological passport across the EU and EEA: your packs are checked at origin and circulate freely afterward.

The directive compresses into three rules that everyone on a packaging line should be able to recite. Inspectors call them the Packer’s Rules:

The tolerable negative error (TNE) scales with pack size:

Run the 500 g example through the three rules and the structure appears. TNE is 15 g, so T1 = 485 g and T2 = 470 g. Your batch mean must sit at or above 500 g. At most one pack in forty may land below 485 g. Not a single pack may land below 470 g — ever. For a 20 g stick pack the numbers get uncomfortable fast, because the TNE there is 1.8 g: T1 is 18.2 g and T2 is 16.4 g. You are managing a legal boundary that is two grams wide at eighteen hundred packs an hour.

One more number from the same legal family, because it silently sizes the sensor: the checking equipment used to demonstrate compliance — which on most e-marked lines is the checkweigher itself — must have a total uncertainty of no more than one fifth of the TNE. For a 500 g pack that is 3 g. For the 20 g stick pack it is 0.36 g. The machine you put at the end of the line is not just a filter; it is your measuring standard for the whole average-quantity system.

3. OIML R51: the rule book that names the sensor’s error

Here is where most buyers pick the wrong standard, and it costs them a redesign six months later. A checkweigher is not a non-automatic weighing instrument. Nobody places the pack on the platform and waits for a stable reading; the machine weighs and decides without an operator in the loop. The instrument family for that behaviour is OIML R51, automatic catchweighing instruments — and R51 is the document that, uniquely among weighing standards, writes the load cell’s error into the law as a statistical quantity.

R51 splits catchweighers into two accuracy classes:

If you only remember one thing about R51, remember that class X is judged on dispersion. The table below is the maximum permissible standard deviation for class X(1) — the tightest common checkweigher class — and the right-hand column is the number I put in front of buyers of load cells for sorting scales before I quote anything:

Run the division and something tidy appears: across the whole band from 100 g to 5 kg, that standard deviation is the TNE divided by 18.75 — 15 ÷ 0.8, 9 ÷ 0.48, 4.5 ÷ 0.24, 75 ÷ 4.0. In service the divisor loosens to roughly 15. So R51 does not say “your readings must be accurate”. It says your readings must cluster inside a window that is about one eighteenth of the legal tolerance you are trying to defend. That is a completely different engineering target, and it is the reason a load cell that looks superb on a datasheet can still fail a checkweigher type test.

Two practical consequences follow. First, a buyer who asks me for “a 0.01 percent cell so the checkweigher passes” has mis-identified which term dominates the sum; I will come back to that in the first case below. Second, class Y graders are judged in MPE bands (typically 1.5 e to 2.5 e depending on class and load), which is the easier regime — if your machine only sorts fish into three sizes and never feeds a net-content declaration, you are not fighting the 18.75 divisor at all.

4. A decision is not a price

I have spent several articles on this blog drawing one line through the weighing world: is a price computed from this reading? If yes, the whole legal machine of trade measurement wakes up. If no, you are solving a mechanical and process problem instead. A sorting scale sits on the “no” side — but on its own terms, which is why it is worth a chapter.

The verdict is a strange object. It has no invoice attached, so on some lines the instrument needs no approval at all: a plant that uses a checkweigher purely to catch its own filler drift can run an unapproved machine, and many do, because the machine never makes a claim to a customer or an inspector. But the moment the same machine is the evidence you present for average-quantity compliance, it becomes the measuring standard for your legal net content — and R51 plus 76/211/EEC apply whether or not anybody stamped a plate on it.

One asymmetry is worth stating plainly, because it is the mirror image of what I wrote about enforcement on the road. On a highway, the inspector’s tolerance is generous and your legal limit is brutal. On a packaging line it is the opposite: the legal tolerance for a single pack is generous (a 500 g pack may legally weigh 470 g), while the process capability you must demonstrate is harsh (a standard deviation of 0.8 g). The law has already forgiven your worst pack. Your retailer has not.

5. Speed is the third variable

Capacity and accuracy are the two numbers a customer brings to me. Speed is the one they forget, and it is the one that changes the answer — which is why load cells for sorting scales cannot be specified from a static-scale checklist. A checkweigher gives the cell a fixed amount of time to produce a trustworthy reading, and that time is not negotiable: it is the length of the weigh zone divided by the belt speed.

The industry publishes accuracy as a function of rate, and the shape of that curve is the single most useful commercial fact in this article. Machine builders who publish it honestly give figures in this neighbourhood: about 0.15 g at a leisurely 60 packs per minute, 0.45 g at 180 per minute, 0.85 g at 300, and 1.3 g at 450. Notice what that is: the tolerance widens by roughly a factor of nine while the rate widens by seven and a half. You are not paying for speed with money alone — you are paying with accuracy, and the exchange rate is bad.

Two adjustments have to be made to every brochure number before it reaches a real line. First, the quoted rate is a nominal ceiling; on real lines with real product, builders and integrators quietly derate it by 15 to 25 percent depending on how the product behaves — loose powder and light bags cost more than rigid cartons. Second, the quoted accuracy is a two-sigma laboratory figure measured with a well-behaved test pack on a warm machine; in production, expect the real spread to be 1.5 to 2 times worse. Both of those adjustments apply to my own case studies below, which is why I show the arithmetic rather than the brochure.

6. Case one: the stick-pack line that rejected a tenth of its good packs

The arithmetic below is the first thing I run on any enquiry for load cells for sorting scales, and it is the reason the enquiry usually changes shape before I quote. A stick-pack machine builder in Southeast Asia came to me with a sentence I hear in some form almost every month: “Our checkweigher is rejecting too many good packs. We think the load cell is not accurate enough. Please quote your best 0.02 percent cell.” They build the filler and the checkweigher together; the product is instant coffee, 20 g per stick, running three shifts.

I asked for two things before quoting anything: the reject band width they had set in the instrument, and one shift of reject logs with timestamps. Asking for the log before the quotation is the habit that has saved me the most wasted samples over the years, and on a sorting line it is the only place where the difference between a bias problem and a spread problem is visible. My article on load cells for packaging machinery covers the sibling weighing stations on the same machine; here I stay on the verifier itself.

The log answered the question in about ten minutes of arithmetic. Start from the law. A 20 g pack falls in the 5 to 50 g band of 76/211/EEC, where the tolerable negative error is 9 percent of the nominal quantity:

Now the production side. Their line was rated 180 sticks per minute, which in practice with a light, slightly sticky stick and a hot machine was running at roughly 140 per minute on a sixteen-hour day — about 134,400 sticks a day. The cell they had was quoted at plus or minus 0.3 g, which is a two-sigma figure, so the ideal standard deviation was 0.15 g; allowing the usual production penalty of about 1.7 times, their real spread was nearer 0.25 g.

They had set the reject band at plus or minus 0.5 g. Divide the band by the spread: 0.5 ÷ 0.25 = 2.0. Two sigma, two-sided, is a false-reject rate of about 4.55 percent. Applied to 134,400 sticks, that is roughly 6,100 rejected sticks a day — and from the log, most of them were perfectly legal product. The cell was not the culprit. The band was.

There is a second lesson in the same log, and it is the one I actually spend my time explaining. The reject band should not be symmetric. The law has a hard floor (T2) and a soft warning line (T1) beneath nominal, and no ceiling at all above it. Your filler, meanwhile, has a cost that rises monotonically with overfill. So the rational band is tight below and loose above: reject aggressively on the low side, where a legal violation lives, and let the high side be governed by filler trim rather than by the rejector. A symmetric band rejects your most profitable packs.

Then came the part where I had to disappoint them. They asked for a 0.02 percent cell, so I did the sum on our own parts. Our WST201 in a 3 kg capacity is specified at plus or minus 0.02 %F.S comprehensive error, which is 0.6 g — and the class X(1) standard deviation budget at 20 g is 0.144 g, with the checking-equipment limit at 0.36 g. The cell’s own comprehensive error is already larger than the whole legal budget. Buying a tighter cell would not fix that, because comprehensive error is a static number while the budget is a statistical one, and neither of them is where their false rejects came from.

What I did recommend: keep the analog cell class they had, take the capacity down rather than the price up (a smaller spring element raises the millivolt-per-gram output, which is what actually buys resolution at 20 g), tighten the mechanics and the sticker knife so the stick does not jitter on the belt, and fix the environment first. Their plant had a loading door open at the end of the line and a measured swing of 8 to 10 degrees Celsius across the day. At 3 kg, a zero-temperature effect of 0.02 %F.S per 10 degrees moves the zero by 0.6 g; the pessimistic figure our own product page also lists — 0.1 %F.S per 10 degrees, and yes, both values appear on the same page, which I have flagged for years — would move it by 3 g. Either way it is several times the entire X(1) spread budget. I asked them to close the door and add a draught shield before I sold them anything at all.

Six weeks later the false-reject rate was down to a figure they were happy with, achieved with the original cell class. The purchase they had planned — a premium cell to solve a control problem — never happened. I count that as a good outcome, because they still call me for the second line and for the smaller-capacity cells, and because the alternative would have been a returned shipment and a broken relationship.

7. What I supply, and what the machine builder owns

I run an export business for weighing components: load cells, force sensors, junction boxes and weighing indicators. I am not a checkweigher manufacturer, and I have never sold a reject mechanism. Since the whole value of this article depends on that line being clear, here it is drawn explicitly.

Why insist on this? Because on a sorting line the failure modes are mostly between the components. A cell with a superb datasheet bolted to a frame with a resonance at the belt frequency produces a worse verdict than a mediocre cell on a stiff frame. If I let a customer believe I sold the machine, I would be quietly accepting blame for the belt, the filter and the rejecter — three things I do not make and cannot see. What I can do is give you the weighing solution: the part of the problem that is genuinely mine, specified with numbers, and tested before it leaves.

8. Capacity: the belt is tare, the surge is real

On a static platform scale, capacity is chosen for the object. On a checkweigher it is chosen for the conveyor, and this single reversal decides load cells for sorting scales more often than accuracy class does. The pack is a rounding error in the mechanical load; the weigh belt, the rollers, the side guides and the drive all sit on the cell, and every one of them moves.

Take the arithmetic from case one, but with a modest weigh conveyor. Suppose the moving assembly on the cell weighs 3.1 kg and the stick pack weighs 20 g. A digital single-point cell in an 11 kg capacity sits at 28 percent of Emax while the line runs — a comfortable static position. Now look at what dynamic service means for the same cell: our WST1206 and WST1209 digital cells are specified with a safe dynamic load of 70 percent of Emax, which is 7.7 kg in the 11 kg size. Against the 3.1 kg of live load, that is a headroom factor of about 2.5. In practice, the pack is irrelevant to that calculation: the same cell would carry a 500 g pack with an 18-fold margin on the pack alone, and would still be limited by the belt.

The second capacity number is the one almost nobody quotes, and it is the reason I put it in writing: Vmin, the minimum load cell verification interval, given by Emax divided by the Y value from the OIML R60 certificate. It is the smallest span in which the cell is allowed to make a legally meaningful statement. This is the same arithmetic that governs any weighing station fed by a vessel — the hopper scales I wrote about in load cells for hopper scales live with exactly the same floor — except that here the tare is a moving belt instead of a static vessel.

Read that table against the class X(1) spread limits from section three and the design conversation changes shape. At a 500 g pack the legal spread budget is 0.8 g; the X(1) cell in that row delivers a 0.73 g verification interval, which is just barely adequate — and adequacy is not comfort. At a 20 g stick the budget is 0.36 g for the checking equipment, which no 11 kg cell can reach, at any price. This is why I keep saying that the answer to a small pack is a smaller spring element, not a more expensive certificate.

9. Case two: the grader where I did not sell the digital cell

A seafood equipment builder in Europe was designing a three-lane weight grader for fish fillets: 300 to 800 g pieces, sorted into three size classes, with a recheck lane at the end. They had already decided they wanted digital cells on every lane. The reasoning was the one I hear constantly: “Digital is more accurate, and we get rid of the junction box and the trimming.” This is the class Y branch of the family — the same machine type as the egg grading lines I wrote about earlier, where the verdict is a size class rather than a pass or fail.

Both halves of that sentence are worth examining. The junction box half is true and valuable — a digital cell network genuinely removes the summing box, the trimming pots and the long analog cable run. The accuracy half is where the trouble starts. A digital cell does not measure better than a good analog cell; it transmits better. And transmission has a rate, which becomes a hard limit when the weighing window shrinks.

So I asked for the line drawing and did the dwell arithmetic for each lane. Everything depends on two numbers: the length of the weigh zone, and the belt speed.

A digital cell on an RS485 bus with one device and a 9,600 baud link reports at about 20 hertz — one fresh reading every 50 milliseconds. In the slow lane’s 0.7 second window that is fourteen readings, plenty for a filter to work with. In the fast lane’s 0.25 second window it is five. Five samples is not a measurement; it is a guess with a confidence interval. Switching to CANopen at 125 kilobaud raises the rate to about 50 hertz, which gives 12.5 samples in the same window — workable, but with no margin for the addressing overhead that appears the moment you put four cells on the same bus.

There is a quieter problem too. Bus latency adds directly to the decision time: a Modbus RTU round trip in a real installation can take 17 to 23 milliseconds, and if the reject decision is issued after the reading arrives, that latency pushes the reject window past the point where the piece is still under the pusher. Above roughly 200 pieces per minute, latency stops being a detail and starts producing misfires.

This is the moment where a supplier earns or loses trust. I had a clean, higher-priced, easier-to-install product to sell for all three lanes, and I told them not to buy it on one of them. That is not charity: a customer who discovers in commissioning that a cell cannot keep up will not distinguish between the cell and the supplier who chose it. I also showed them the alternative honestly — if they could slow the fast lane to 40 metres per minute or lengthen the weigh zone by 100 mm, digital would work there too. Line geometry and cell type are one decision, not two.

10. What a digital cell really buys on a sorting line

Since I argued against digital cells on one lane, it is only fair to say precisely what they do buy, because for load cells for sorting scales the list is longer than it is on most machines.

The costs are equally specific. Refresh rate is shared: our own WST1209, a four-cell network on a 125 kilobaud CAN bus, delivers 40 hertz, not 50 — the rate falls as you add devices. Latency sits between the physical weight and the verdict, which is harmless on a slow lane and dangerous on a fast one. And the cell no longer hands you a raw analog signal: if your machine’s filtering, autozero or diagnostics logic was written around an analog input, moving to digital means rewriting that logic, not just rewiring.

One further point deserves its own paragraph, because it is the most useful thing I learned from a checkweigher customer in five years. On a sorting scale, a failed channel does not always announce itself as a fault. A cell that has drifted low will reject perfectly good packs, and somebody will notice within the hour. A cell that has drifted high, or that has frozen at its last value, will pass everything — and the line will ship underweight packs to a retailer with a completely clean reject log. The failure that hurts you is the one whose symptom is “no faults today”. This is why I now recommend, on any multi-cell sorting machine, that the self-diagnostic path be treated as a functional requirement rather than a feature: an analog cell’s worst failure mode is a plausible number, a digital cell’s worst failure mode is no number — and the second is easy to alarm on, which is exactly what you want. If you want the general test procedure for deciding whether a cell has actually failed, I wrote it up in how to determine whether the load cell is damaged.

11. Case three: the missing leaflet weighs 0.7 grams

The third project is a pharmaceutical cartoner in Asia who wanted a completeness check: a folded patient leaflet weighing about 0.7 g must be inside every carton, and a missing leaflet has to be detected before shrink-wrapping. The carton itself weighs 45 g, and the packed carton plus the weight of the checkweigher’s own carrier fixture comes to about 1.45 kg on the cell.

The naive version of this problem looks easy. A 0.7 g step against a spread of 0.15 g is nearly five standard deviations, and the chance of missing such a step purely by noise is around one in a million. On paper, statistically, you cannot miss it. In the plant, they were missing about one carton a day, and the customer complaint that brought them to me was a recall.

Here is the diagnosis, and it took about a day of their reject log to confirm. A cardboard carton’s mass is not a constant. Paper takes up and gives off moisture with the ambient humidity, and in a tropical plant a 45 g carton can swing by close to a gram across a shift — on some days more. Their detection logic compared each carton’s gross weight against a fixed target. But the carton’s own tare drift was larger than the leaflet they were trying to detect. Some cartons dry out by 0.7 g and look exactly like a carton with a missing leaflet; others absorb 0.7 g and hide a genuinely missing leaflet completely. No load cell, at any resolution or any price, can separate those two cartons, because nothing in the gross weight distinguishes them.

The architecture I proposed was differential and sequential: weigh the empty carton immediately before filling, store that value against the carton’s position or a printed code, then weigh after the leaflet is inserted and compare the difference. Now the tare, including its humidity drift, cancels out of the subtraction, and the only thing left to detect is the 0.7 g step. Two further measures made it robust: a settling gate, so the carton must be motionless for a defined period before both readings are taken, and a dedicated leaflet-verification station upstream rather than asking one machine to do both the net-content check and the completeness check with one number.

On the sensor side, that architecture is undemanding, and this is the point I want readers to take away. Once the comparison is a difference between two readings taken on the same cell minutes apart, what matters is not comprehensive error but repeatability and zero stability — a specification that a well-made analog single-point cell satisfies easily in a 3 kg capacity, and which does not improve further when you spend four times as much on a certificate class. I recommended the analog part, and I have been shipping it to them since.

12. Giveaway: the money the checkweigher earns back

Machine builders sell a checkweigher on compliance. Plants buy it for something more mercenary, and it is worth doing the arithmetic openly, because a customer who understands the economics makes a better sensor decision — and buys the right load cells for sorting scales rather than the most expensive ones.

The mechanism is simple. A filler cannot hit a target exactly, so it is set to overfill slightly; the overfill is giveaway that goes out of the factory door as free product. A checkweigher measures every pack, so the filler’s mean can be brought down toward the nominal quantity with a known, statistically defended margin — not with a guess. Published case studies show what that is worth. Take a line running 200 packs per minute, eight hours a day, 250 days a year: that is 24 million packs. Removing 2 g of average overfill saves 48 tonnes of product a year. On a slower 60 per minute line, 7.2 million packs a year, the same 2 g saves 14.4 tonnes. On a high-rate line — 3,000 packs an hour, sixteen hours, 300 days, 14.4 million packs — a 1.8 g reduction at two US cents per gram is over half a million dollars a year.

The most quoted numbers come from snack production, and they are consistent with the arithmetic above: a nuts line cutting average overfill from 4.1 percent to 1.0 percent — three percentage points of the pack — saved about 98,000 euros in eight months, roughly 12,250 euros a month. A smaller example: one million packs a month where the mean is pulled down by 0.5 g saves 500 kg of product, about 5,000 dollars a month at ten dollars a kilogram. None of these figures depend on the load cell; they depend on the machine being trusted enough to tighten the filler. That trust is the product.

Now the other side of the ledger, which almost nobody puts in a proposal. Every reject costs money. Take a 120 pack per minute line: 14.4 million packs a year. Run the reject band at three sigma with a spread of 1 g — plus or minus 3 g — and the false-reject rate from random variation alone is about 0.27 percent, or 38,880 good packs a year thrown away. At 3.00 dollars of product plus 8 cents of packaging each, that is about 119,750 dollars a year burned on product that was perfectly legal. Tighten the band to two and a half sigma, which is a tempting thing to do when a retailer complains, and the false-reject rate rises to 1.24 percent: 178,560 packs, nearly 550,000 dollars.

If you take one thing from this section, take the asymmetry. Giveaway is a mean problem and is solved by trusting the machine enough to move the filler’s target down. False rejects are a spread problem and are solved by widening the band or reducing noise. Buy a better load cell to attack the second, and you may gain a little; attack the first with a log and a filler trim, and you gain tonnes. The cell matters most in the third case — where a step must be detected against a variable tare — and there the spec you want is repeatability, not class.

13. Approvals by market

A sorting scale crosses more rule books than any other machine I write about, and the confusion is always the same: people assume the certificate on the instrument and the certificate on the cell are the same document. They are not, and the difference is where a project can stall for months.

Read the third column and the practical consequence is obvious: what I ship is evidence, not permission. I can give you a cell from a certified production line, with the class and the Y value documented, and a calibration certificate for the serial numbers on the pallet. I cannot approve your machine, and neither can any component supplier. If a builder tells you that fitting a certified cell makes the machine legal, that builder has the paperwork backwards.

14. Load cells for sorting scales: the four I recommend

The four models below cover the whole sorting-line conversation: an analog workhorse for slow lanes and heavier packs, a six-wire packing-machine cell for compact stations, a digital single-point cell for wet or washdown lines, and a digital bending beam for heavier packs and for the hoppers feeding the sorter. The capacity, class and IP rating differ, and so does what each one costs you in exchange — so I have written the trade-off into every card rather than leaving it out.

One thing I want on the record before the cards: our own WST219 and WST218 are the two models our catalogue names specifically for checkweigher systems, and neither appears in my four. That is not an oversight. Both are imperial-capacity cells (20 to 100 lb, and 50 to 500 lb) designed for the incumbent machine platforms, and on a metric-line project the four below are the cheaper, better-documented route. If you are building a replacement platform for an existing American-designed machine, ask me for the two imperial ones instead — that is a real difference in tooling, not a marketing preference.

And the honest gap in this list. If your packs are stick packs at 20 g, none of the four is small enough — on our own shelf the family that starts at 0.3 kg and runs to 6 kg, the WST205, is the answer, and a smaller spring element buys more millivolts per gram than any certificate can. Tell me the pack weight first and I will tell you whether the four above are even relevant, or whether you should be looking at a different capacity range entirely, as described in my selection guide.

15. Seven questions to answer before you ask for a quote

Every enquiry for load cells for sorting scales that later went wrong went wrong at one of these seven questions, and none of them is about the sensor. Answer them in writing, even roughly, and my reply stops being a price list and starts being a design.

16. Questions I get from checkweigher builders