As an Amazon Associate, Akermin earns from qualifying purchases. Product links on this page are affiliate links.
Choose the interval from the shortest feature you need to see, then check that the memory survives the deployment at that rate, and if the two do not fit, shorten the deployment rather than lengthening the interval. A record that spans the right period at the wrong rate looks complete and is quietly useless, because the feature you came for has been averaged or stepped over and nothing in the file announces that it is missing. A span of one second to eighteen hours covers four and a half orders of magnitude, and almost the whole of that span is the wrong answer for any particular job.
The interval is a question about the fastest thing you care about
Every quantity worth recording has a characteristic time over which it changes. A room responds to an opened window over tens of minutes. A heating circuit cycles over something like a quarter of an hour. Solid material gaining or losing moisture moves over days. A motor drawing starting current does its interesting work inside a second. Those characteristic times, not the length of the deployment, are what set the interval.
The rule that follows is simple to state and routinely ignored. Divide the duration of the shortest feature you need to describe by five, and that is your upper bound on the interval. Five samples across a feature give it a recognisable shape, ten give you a gradient you can trust, and two give you nothing but a straight line between two arbitrary points that happens to connect them. If you only need to know that something happened rather than what shape it had, you can relax towards two samples, but you have given up any ability to describe it.
Memory depth and event duration set the floor and the ceiling
The event sets the ceiling on the interval and the memory sets the floor. Those two bounds can easily cross, and when they do the honest response is to change the plan rather than to split the difference. Either shorten the deployment and return for a second one, use a unit with deeper memory, or accept explicitly in the write up that the fast feature was not captured.
There is a third constraint that only appears on long deployments, which is power. On units that wake a sensing element for each sample, the consumption scales with the sample rate rather than with elapsed time, so a fast interval on a long deployment can exhaust the cell before the memory fills. On mains powered units this disappears entirely, which is part of why a permanently powered installation can afford a rate that a battery unit cannot. The guidance on setting up long term environmental logging covers the deployment side of these trade offs.
One second to eighteen hours, worked into spans
Assume a memory that holds ten thousand readings, which is a convenient round figure for working out the shape of the problem. The span is simply the interval multiplied by that count, and the shortest describable feature is roughly five times the interval.
| Interval | Span at ten thousand readings | Shortest feature you can describe | Typical subject at that rate |
|---|---|---|---|
| 1 second | About 2 hours 47 minutes | Around 5 seconds | A motor start, a drive ramp, a valve stroke |
| 10 seconds | About 1 day 4 hours | Around 1 minute | A kettle, a shower, a lift car cycle |
| 1 minute | About 7 days | Around 5 minutes | Occupancy, a heating call, a window opening |
| 5 minutes | About 35 days | Around 25 minutes | Boiler cycling, cooking, a ventilation recovery |
| 30 minutes | About 208 days | Around 2 hours 30 minutes | The daily cycle, weather driven swings |
| 6 hours | About 2,500 days | Around 30 hours | Slow drying, long term drift |
| 18 hours | About 7,500 days | Around 4 days | A sealed space nobody will revisit soon |
The middle of that table is where most building work belongs, and the two ends exist for genuinely different kinds of question. Note how quickly the span grows: moving from one minute to five minutes multiplies the deployment length by five while only coarsening the describable feature from five minutes to twenty five, which is an excellent trade for occupancy work and a terrible one if you were hunting a short extract fan run.
Where a fast interval buys nothing but a larger file
Sampling faster than the quantity moves adds rows and no information. A wall drying out over weeks, recorded every second, produces a file in which consecutive values are identical and the only variation is sensor noise. Worse, that noise now dominates the apparent short term behaviour of the record, which invites people to read structure into it that is not there.
Some instruments offer a smoothing option precisely for this situation. The Fluke 289 logging multimeter states a selectable filter described as a smoothing mode, intended to steady the display when the input is changing rapidly or is noisy, and separately a graphical review of a logged session with the ability to zoom in up to fourteen times. That zoom capability is the other half of the interval decision: an instrument that lets you examine a dense record in detail after the fact makes a fine interval more defensible, because you are not forced to read the whole file at one scale.
Instruments with no interval at all
A large part of the measurement toolkit has no scheduler whatsoever, and understanding that is part of understanding intervals. A thermal camera records a frame when a person decides to record one, so the effective sampling interval is however often somebody is willing to return and press the shutter. The FLIR One Pro states an output of 480 by 360 upscaled from a 160 by 120 sensor with a visual and thermal blend, and the FLIR One states 240 by 180 upscaled from 80 by 60. A timed or scheduled capture facility is Not stated in the listing bullets for either one, and neither needs one for the inspections they are built around.
What that costs you is comparability. Two thermal frames taken a fortnight apart differ because of the fabric, because of the weather, because of the time of day and because of where you were standing, and separating those is the whole difficulty of repeat thermal work. A logged record has no such problem because the instrument did not move and the clock did not care. The reviews of the FLIR One and the FLIR One Pro cover what each captures, and the comparison against the Edge Pro covers the handling difference between a phone attachment and a standalone body.
Choosing a number before you leave the site
Write down three things before you set the interval: the shortest feature you need to describe, the length of the deployment, and the memory depth of the unit. Divide the first by five to get the upper bound, divide the memory by the deployment in samples to get the lower bound, and if the bounds cross, change the deployment rather than quietly accepting a coarse record. Then set the clock, confirm the unit is actually logging, and note the launch time on paper as well as in the software.
Finally, be honest in the record about what the interval cannot show. A published listing that gives nothing at all is a useful reminder of how little some product pages carry: the Bacharach maintenance kit associated with refrigerant leak detection publishes a part number and nothing else, so interval, memory, accuracy and range are each Not stated in the listing bullets. When a listing is that sparse, the specification has to come from the manual or not at all, and writing down which figures you could not confirm is part of doing the job properly.
Frequently Asked Questions
How many samples do I need across an event to call it resolved?
Two per cycle is the absolute theoretical floor for detecting that a repeating thing exists at all, and it is nowhere near enough to describe it. Five samples across the shortest feature gives you a recognisable shape, and ten gives you something you can measure a gradient from. Work backwards from the feature you expect: if a ventilation recovery takes twenty minutes, a four minute interval is a sensible target and a twenty minute interval will show you a straight line between two unrelated points.
What is aliasing and can it really happen on building data?
It is the appearance of a slow cycle in the record that does not exist in reality, produced when a fast repeating process is sampled slower than twice its own rate. It happens on building data regularly. A plant item that cycles every twenty minutes, sampled every twenty one minutes, produces a long slow wave in the file with a period of several hours. Nothing in the data marks it as false, which is why the interval should always be chosen against the suspected mechanism rather than against convenience.
Does a stored sample represent an instant or an average?
It depends on the instrument and the mode, and the difference matters more than people expect. Some loggers store the instantaneous reading at the tick of the interval, some store the mean over the interval just elapsed, and some store a minimum and maximum pair alongside it. An instantaneous sample can land on a spike and exaggerate, while an averaged sample can smooth a genuine excursion out of existence. Find out which one you are getting before you interpret a peak.
Is there any harm in simply choosing the fastest interval available?
Several kinds. The memory fills before the deployment ends, which is the worst outcome because you lose the later part of the record rather than the earlier. Battery consumption rises on units that wake the sensor for every sample. The file becomes large enough that the analysis step gets skipped. And a fast record of a slow quantity contains no more information than a slow one, just more rows of the same number.
Should the interval change between a reconnaissance visit and the real deployment?
That is usually the most efficient approach. Go out first with a short, fast deployment of a day or two at a fine interval, purely to find out what the shortest meaningful feature is. Then set the real deployment at an interval chosen from what you saw, with enough memory margin to cover the full period. Two visits with the right settings beat one long visit at a guessed interval.
Why would any instrument offer an eighteen hour interval?
For deployments where the quantity genuinely moves that slowly and nobody will return for a very long time. Slow material drying, a sealed archive space, a long term check on a structure settling into a new condition. At that interval the instrument is recording a trend rather than events, and the design assumption is that the battery and the memory both have to survive unattended for an extremely long stretch.