You don't pick a sensor, you study it

Seen from the outside, a connected hive looks like a shopping list. A scale underneath, a temperature and humidity sensor inside, a microphone, maybe an accelerometer, a radio to send the data out. You look at the list, add up the prices, and think you have understood what the project costs.

I found out early enough that the shopping list is the last thing you write, and that it is not even the expensive part. The expensive part comes first, and it is understanding what the number that sensor produces actually means.

In this article I take a single question, perhaps the one every beekeeper wants answered most, and follow it all the way down: will this colony swarm? Not to explain how bjtOS answers it. To show how many decisions you have to make before you can even attempt it, and how many of those decisions have already been made by someone else, over thirty years of work published in scientific journals. Anyone who puts a sensor in a hive is inheriting that work, whether they know it or not.

Why swarming, of all things

Because it is common and it costs money. In a 1957 study of English production apiaries, James Simpson concluded that, left to themselves, between 10 and 40% of colonies would swarm in an average year. It is not a disease: it is how a colony reproduces, which is to say, divides. But for whoever keeps it, it means losing half the population at the worst possible moment, just before the harvest.

And then because it is the perfect textbook case: swarming is not an instant, it is a state that matures over weeks, and that changes everything for anyone who has to measure it.

First decision: what exactly am I measuring

No sensor measures “swarming”. A sensor measures a physical quantity: degrees, pascals, grams, hertz. The first question is not which sensor to buy, it is which physical quantity changes reliably when a colony prepares to swarm. That is a biology question, and the answer is in a library, not in a component catalogue.

Biology says that quite a lot changes in a colony that is getting ready. The workers start feeding the queen less, with fewer nurses and more young bees, from the moment the queen cells are started (Allen, 1960). The queen, put on a diet, ends up losing about a third of her own weight, with the minimum reached when the queen cells are capped, and she regains it within five days of the swarm leaving (Morse, Dyce and Young, 1966). Even inside the workers’ bodies something is different: in a 2025 study, in pre-swarming colonies vitellogenin stayed significantly higher in ten and fourteen day old bees, whereas normally at that age it drops as they shift to the forager role (Klett and colleagues, 2025).

Here a small language trap opens up, and it is worth telling because it is typical of the trade. Beekeepers call that state “swarm fever”. In the scientific literature that term does not exist: people talk about swarm preparation, about colonies in pre-swarming. And the word fever, in the bee literature, means something else entirely, that is, raising the temperature of the comb against a heat sensitive pathogen. Anyone designing has to watch out for these overlaps: if you take the field jargon literally, you risk going off to look for a fever that is not there, and missing the signal that actually is.

The thermometer knows the answer, but it knows it late

There is a fever, in a sense. But it lasts a few minutes.

Aleksejs Zacepins and colleagues monitored ten colonies continuously, from 1 May to 31 August 2015, recording nine swarming events. Before takeoff the temperature measured above the upper hive body rises by 1.5-3.4 °C, from the 34-35 °C typical of the brood nest up to 37-38 °C, and the colony takes from eight to twenty minutes to warm up before leaving (Zacepins and colleagues, 2016).

That paper also contains two sentences that are, in effect, design specifications already written: a single measurement point is enough to detect the event, and one measurement per minute is sufficient. Those are exactly the two questions faced by anyone designing the firmware of a battery powered node, where every extra reading is energy you no longer have. Someone has already measured them, and published them.

But the most important conclusion is another one: the warning time is not a property of the sensor, it is a property of the phenomenon you chose to watch. The thermometer is not imprecise, not slow, not cheap: it does its job very well. It is simply that the warm up is the engine starting, not the decision to leave. With temperature you know they are leaving while they are leaving. For an anti-theft system that is perfect. For preventing a swarm, twenty minutes are worth nothing.

Where you put the sensor is a result too, not a hunch

A 2019 study on Apis cerana, the Asian honey bee, followed six swarming events measuring temperature at several points in the nest, and found that the most sensitive point is not where you would instinctively put it. It is low down, in the space between the hive wall and the first frame, towards the entrance. There the warm up is more legible than elsewhere, with average rates going from around 0.07 °C per minute to over 0.3 °C per minute in the final phase.

I find this the most underrated detail in the whole trade. In practice, a sensor’s position is almost always decided by mechanical constraints: where there is room, where the cable is not in the way, where the bees will not build over it. In the literature, position is an experimental result, with its own uncertainty and its own confidence interval. They are two worlds that talk to each other very little, and every time they touch you learn something.

The microphone: you choose the band you listen to before you listen

The idea that a colony announces the swarm with a sound is very old: Aristotle described a peculiar sound made by bees a few days beforehand. Testing it with instruments, on the other hand, is recent.

In 2008 Stefano Ferrari and colleagues recorded three hives for 270 hours, capturing nine swarming events. They found an increase in power spectral density around 110 Hz, which as the event approaches grows in amplitude and climbs towards 300 Hz, with occasional rapid transitions from 150 to 500 Hz.

Hidden in that paper there is another parameter that looks bureaucratic and is instead an irreversible design choice: they sampled at 2 kHz. The sampling rate decides the highest frequency you will ever be able to see in your data: what lies above it is not badly heard, it simply no longer exists, and no algorithm brings it back. Whoever designs a node has to choose that number before recording a single bee, and chooses it by reading what others have found. Choosing it too high means throwing away energy and memory to record noise; too low means erasing forever the very signal you were looking for.

The accelerometer, and the twenty hertz surprise

The turning point comes when someone stops listening to the air inside the hive and starts feeling the wood and the wax. Martin Bencsik and colleagues, in 2011, used a transducer fixed to the hive to follow the colony’s vibrations over time, automatically and non invasively, and showed that the swarming process can be identified days in advance.

Along the same road, in 2020, came the most solid piece of work I have read on the subject. Michael-Thomas Ramsey and colleagues instrumented 25 colonies over two seasons, most of them at INRA in Avignon, with piezoelectric accelerometers embedded in the wax at the centre of the brood frames. A recording every three minutes, band up to 5500 Hz, three hertz resolution, over thirty million data points, 11 primary and 19 secondary swarms observed. Two algorithms tell the colonies that will swarm from those that will not with over 90% accuracy, with a warning time reaching up to 22 days in one configuration and around 10 days in the other.

And then the result that turns intuition upside down: the largest contribution to the discrimination comes from the low frequencies, around 20 Hz, consistent with a vibrational signal the bees exchange among themselves. Above 200 Hz the contribution is modest, towards 800 Hz less still.

Twenty hertz is at the bottom edge of human hearing. It is a vibration more than a sound. It means that the most informative signal for this question is not where a microphone would look for it, and that an ambient microphone and an accelerometer sunk into the wax are not at all the same thing, however much they both end up on the shopping list under “acoustic sensor”. Whoever writes the list without having read that paper buys a plausible, wrong object.

The limits, which never make it into the brochures

The part that convinced me of the quality of that work is not the percentages, it is the limits the authors declare themselves.

The method based on instantaneous spectra works reliably only if applied within the swarming season: out of season it loses meaning. The second method needs at least ten days of history before it can compute an alarm, and much more memory. Both produce false positives during daylight hours, because of the foragers coming and going, so much so that the authors recommend averaging the data between midnight and five in the morning.

Let me stop a second on that last one, because it is the heart of what I wanted to tell. “Average the data between midnight and five” is one line of code. Whoever reads it in a firmware takes it for a programmer’s arbitrary choice. It is instead the conclusion of an experiment on twenty five colonies: there is an entire study inside that line.

The last limit is the finest. Primary swarms, the ones with the old queen, remain according to the authors remarkably hard to predict, and the reason is biological: there is no singing from the virgin queens, that piping which in secondary swarms offers an obvious acoustic handle. Thomas Seeley and Jürgen Tautz had also described, in 2001, worker piping, pulses of about 0.8 seconds rising from 100-200 to 200-250 hertz, which starts about an hour before takeoff and grows in step with the swarm warming up. An hour before. Not weeks.

In short: the very swarming the beekeeper wants to prevent is the one that leaves the fewest traces.

The most expensive problem is not the sensor, it is knowing what really happened

There is one last cost nobody ever talks about. To train or validate any method you need the right answers: which colonies swarmed, and when, to the minute. In the 2020 study the events were confirmed by examining the daily spectrograms by eye, looking for the vibrational signature and the drop in signal at the moment of departure.

Labels cost more than sensors. That is why work exists whose only purpose is to produce well annotated data: a recently published dataset gathers eleven months of continuous recordings from 53 hives in Quebec, with audio, temperature and humidity set alongside measurements made by hand by experts, population, brood, varroa infestation, honey yield, winter mortality.

And here you also understand why I am cautious about certain recent results announcing 98 or 99% accuracy in recognising colony states. Three hives and nine events (2008). Twenty five colonies and thirty swarms over two seasons (2020). Fifty three hives for eleven months (the dataset). These are three different orders of solidity, and a very high number obtained on the first is worth less than an honest number obtained on the third.

The sensor is the easy part

What this reading has already changed for me, concretely, is the shopping list. I know that an ambient microphone is no substitute for an accelerometer in contact with the wax. I know that the thermometer, for this specific question, is a detector and not a predictor, and that it should be judged on what it does well. I know that whatever I build will have to be validated within the season, at night, and on a number of colonies that is not three.

That is not nothing. They are all things I did not have to discover by getting them wrong, because someone had already discovered them and written them down.

The sensor, in the end, is the easy part. It costs little and it goes on in half an hour. What costs is knowing what the number it produces means, and that knowledge cannot be bought: it is read, it is checked, and every so often you discover it says the opposite of what you were hoping.

The sources, for anyone who wants to go and look

• Simpson, J. (1957). The incidence of swarming among colonies of honey-bees in England. Journal of Agricultural Science 49(4), 387-393.
• Allen, M. D. (1960). The honeybee queen and her attendants. Animal Behaviour.
• Morse, R. A., Dyce, E. J., Young, R. G. (1966). Weight changes by the queen honey bee during swarming. Annals of the Entomological Society of America 59(4), 772-774.
• Seeley, T. D., Tautz, J. (2001). Worker piping in honey bee swarms and its role in preparing for liftoff. Journal of Comparative Physiology A 187, 667-676.
• Ferrari, S., Silva, M., Guarino, M., Berckmans, D. (2008). Monitoring of swarming sounds in bee hives for early detection of the swarming period. Computers and Electronics in Agriculture 64(1), 72-77.
• Bencsik, M. et al. (2011). Identification of the honey bee swarming process by analysing the time course of hive vibrations. Computers and Electronics in Agriculture 76, 44-50.
• Grozinger, C. M., Richards, J., Mattila, H. R. (2014). From molecules to societies: mechanisms regulating swarming behavior in honey bees (Apis spp.). Apidologie 45(3), 327-346.
• Meikle, W. G., Holst, N. (2015). Application of continuous monitoring of honeybee colonies. Apidologie 46, 10-22.
• Zacepins, A. et al. (2016). Remote detection of the swarming of honey bee colonies by single-point temperature monitoring. Biosystems Engineering 148, 76-85.
• (2019). The temperature increase at one position in the colony can predict honey bee swarming (Apis cerana). Journal of Apicultural Research 58(4).
• Ramsey, M.-T., Bencsik, M., Newton, M. I., Reyes, M., Pioz, M., Crauser, D., Simon Delso, N., Le Conte, Y. (2020). The prediction of swarming in honeybee colonies using vibrational spectra. Scientific Reports 10, 9798.
• Zhu, Y. et al. (2023). MSPB: a longitudinal multi-sensor dataset with phenotypic trait measurements from honey bees.
• Klett, K., Ihle, K. E. et al. (2025). Vitellogenin plays a role in regulating honey bee swarming. Scientific Reports 15, 36569.