Machines ยท 11 Nov 2026
The sensor that never blinked
One morning, a cartoning machine made almost 3,000 cartons and Fyr counted none of them. Finding out why led us to a quieter problem: on ordinary days, the count had been running short as well.
6 min read
A count is the simplest thing a monitoring system reports, and the easiest to take on trust. If the screen says a machine made 3,000 cartons today, nobody goes and counts the cartons. So when a count goes wrong, the question that matters is not only what happened, but how long it had been wrong before anyone noticed.
On the morning of 11 November, Fyr showed a cartoning machine running at about 49 cartons a minute, with zero cartons counted for the day. Both could not be true. We set out to find which number was wrong, for how long, and whether it had happened before. To answer that, we went back through ten days of readings from the machine and compared Fyr's count, minute by minute, with the machine's own.
We found two problems. The first was loud: a sensor had been stuck on since the evening before, and almost 3,000 cartons went uncounted. The second was quiet, and in some ways worse. On ordinary days, with the sensor working, Fyr's count still ran short of the machine's, by close to 30 percent on the worst day.
How we checked
Fyr counted this machine's cartons from a product sensor at its check station. The sensor flashes once for each pocket that holds the right product, and Fyr counted the flashes.
The machine also keeps its own record. Every carton it closes gets exactly one shot of glue, and the machine totals the shots every minute. That tally comes from the machine's own controls, not from anything Fyr adds, so we treated it as the reference: the number Fyr's count should match.
We compared the two over the working days from 2 to 10 November. For each minute the machine recorded, we counted the flashes Fyr saw between that minute's total and the one before it, and left out any minute with a gap in the readings. That left 1,527 minutes of production, about 25 hours. For the stuck sensor, we went through every reading from the evening of the changeover to the moment it was fixed.
The sensor that stuck
At 6:39 PM on Tuesday 10 November, after a changeover, the sensor came on and stayed on. The machine ran for 12 seconds that evening before it stopped. When it started again the next morning, the sensor was still lit. Through the whole morning, it went dark only twice, each time for about a fifth of a second, and both times while the machine was stopped.
A lit sensor looks like a pocket with product in it. To the machine, every pocket passed its check. To Fyr, which counted flashes, nothing was happening at all. The machine ran for a little more than three hours with the sensor stuck. Until 10:44 AM, when we changed how Fyr counts, the machine made 2,947 cartons, and Fyr counted none of them. By the time the sensor was fixed, the machine had made more than 5,500.
A count of zero is not missing data. It says nothing was made.
When someone on site looked, the changeover had gone wrong. Once it was put right, at 1:21 PM, the sensor went straight back to flashing. Fyr had kept the machine's minute totals in its readings, so those 2,947 cartons are back in the record.
Cartons a minute, 11 November
Wednesday, in ten-minute averages
Short even when it worked
The stuck sensor explained the zero. It did not explain something else in the comparison: on days when the sensor was plainly working, its count was still lower than the machine's.
Across the 1,527 minutes we compared, the sensor counted 48,838 cartons and the machine 56,957. Fyr's count was 14 percent short: about one carton in seven was never counted. The gap was not steady. It was 30 percent on 2 November, and it narrowed through the week until, on 9 and 10 November, the two counts agreed to within 2 percent.
How far short the sensor counted
Against the machine's own count
The readings suggest why. On the days with the widest gaps, almost every flash showed up in exactly one reading. A flash is shorter than the gap between two readings, so one that falls between them is never seen at all. On 9 and 10 November, when the counts agreed, about 3 percent of flashes lasted long enough to show up in two readings in a row, against almost none earlier in the week. Longer flashes are harder to miss. That is the best evidence we have for the explanation, but it is not proof: the readings cannot measure how long a flash lasts.
A count that is 14 percent low is, in a way, worse than a count of zero. Zero gets noticed the same morning. A count 14 percent low still looks like a busy day, and nothing about it invites a second look.
What we changed
Fyr now counts this machine's cartons from the machine's own tally of glued cartons. A covered, dirty or misread sensor can no longer change the count, and no flash is too short to see, because the machine has already done the counting.
The sensor is still watched, because it still does a job: it is the machine's check that each pocket holds the right product. If it has been lit for a full minute while the machine runs, Fyr raises a warning on the machine's page saying what to check, and clears it once the sensor goes dark.
To see how that warning would have behaved, we replayed all ten days of readings against it. It fires during both stuck episodes and at no other time. On the Tuesday, it would have fired at 6:40 PM, a minute after the sensor stuck. It also catches a shorter episode on the Friday before: the sensor stayed lit for more than ten minutes early in a shift, and in the minute or so the machine ran, it made 31 cartons that Fyr never counted. Those are back in the record now too.
Limitations
- Rejects. The machine's tally counts every carton it glues, including any it later rejects. A count of good cartons should leave those out, and we have not measured how many there were.
- Product. The tally counts closed cartons, not cartons with the right product inside. Checking the product was the sensor's job, and while it was stuck, no pocket was checked at all.
- Why the gap moved. Longer flashes explain the narrowing gap better than anything else we found, but we do not know what made them longer later in the week, or why the gap opened to 36 percent again on the Wednesday afternoon. The product on the line is our best guess.
- Coverage. We compared only minutes with unbroken readings, about 25 hours across six days. The machine ran for longer than that, and some days are covered far better than others.
- The Friday episode. What held the sensor on that Friday is still unknown.
- Scale. This is one machine over ten days. The warning has been live for hours, not weeks, so the replay is the evidence so far.
What we took from it
The plain lesson is to count the work, not a signal that usually comes with it. A shot of glue is a carton closed. A flash is only a sign that a carton probably went past.
The less obvious lesson is about checking. The stuck sensor was found within a morning, because zero is plainly wrong. The shortfall was there on most of the days we looked at, and it had passed for normal, because a number that is a little low still looks like a number. Any count is worth checking against an independent one before it is trusted.