Fixing IMU Temperature Drift That Corrupts Cold-Weather Sports Sensor Data

The team's outdoor testing block looks broken for the first quarter hour, every single winter. The countermovement jump numbers a biathlete puts up on the mat at 8am, warmed up in a 21°C locker room, sit at 46-48 cm, matching everything logged all autumn indoors. Twenty minutes later, standing on packed snow at -12°C wearing the same harness after the same warm-up routine, the device reads 41-42 cm on an identical jump, then climbs back toward 46 cm as the session goes on. Nobody got weaker walking outside, and nobody rebuilt their explosiveness jogging around cones for twenty minutes. The sensor's own silicon changed temperature faster than its correction model could track, and the bias that introduced looked exactly like a real drop in performance, right up until it quietly corrected itself.

A MEMS accelerometer or gyroscope measures motion by tracking how far a microscopic silicon proof mass deflects against a spring structure only a few micrometers wide, then converting that deflection into a voltage. Both the stiffness of that spring and the damping around it shift with temperature, and so does the electronics reading the signal, so identical motion produces a slightly different voltage at -10°C than it does at 20°C even though nothing about the athlete's movement changed. Manufacturers correct for this at the factory, running each chip through a thermal chamber and fitting a compensation curve, usually a polynomial, that maps temperature to an expected bias offset so the firmware can subtract it in real time.

That compensation curve is built from steady-state data: the chip sits at each test temperature until it stabilizes, then gets sampled. It was never validated against a device that starts at room temperature and loses 30°C in under a minute, because a factory thermal chamber does not operate that way. The chip's internal temperature sensor also lags the true silicon temperature by several seconds under a fast ramp, so the correction being applied is chasing a value the sensor already left behind. Layered on top of the ambient shift, the IMU's own processor and radio generate a few degrees of self-heating that only shows up a few minutes after power-on, adding a second, slower thermal transient on top of the outdoor one.

Two research groups working in inertial navigation, not sport, have measured pieces of exactly this problem. Aggarwal, Syed, Niu, and El-Sheimy (2008, Journal of Navigation) ran low-cost MEMS IMUs through thermal chambers across roughly -40°C to +75°C and found gyroscope bias swinging by several degrees per second across that full range, with the relationship clearly nonlinear rather than a straight-line offset, and showed a fitted temperature-compensation model removed most of that error against their chamber data. Their chamber ramps, however, were slow and controlled, on the order of a degree per minute, which is the steady-soak case a factory calibration curve is built to handle. They did not test a rapid, minute-scale transition, so how their model holds up under a fast indoor-to-outdoor shock is not something their data speaks to directly.

Woodman (2007, University of Cambridge Computer Laboratory Technical Report 696) documents the second piece: low-cost MEMS gyroscopes carry a self-heating transient after power-on that is independent of ambient temperature entirely, and the report recommends a warm-up period before trusting the output rather than reading it cold. The report is a survey drawing on general navigation-grade guidance rather than a sport-specific cold-exposure experiment, and the settling time it cites varies by sensor design, so it reads better as a reason to test your own device's warm-up behavior than as a number to copy directly. Stack the two transients together and a device carried from a warm locker room onto a snowy range is fighting an ambient shift its factory curve was never built for, plus a self-heating settle time on top of it, at the exact moment a coach wants clean numbers from the first rep.

The failure mode takes the same shape as any other slow sensor bias: smooth, one-directional, and easy to mistake for something physiological. A gyroscope bias of even 0.5°/s, integrated over a 2-second turn or cutting movement, adds a full degree of phantom rotation to that rep, and the error does not average out across a set because it points the same direction every time until the sensor finishes settling. An accelerometer bias does something similar to velocity: integrated over a 4-second sprint effort, a steady 5 mg offset alone adds roughly 0.2 m/s of phantom speed, enough to nudge a top-speed reading outside a typical day-to-day noise band and send a coach looking for an explanation involving spikes, fatigue, or an unusually good warm-up.

What makes this specific error easy to misdiagnose is that its shape matches genuine adaptation almost exactly. It sits furthest from the true value right after the temperature step, then converges smoothly toward the correct number as the sensor's internal temperature and the compensation curve catch up with each other, typically over five to twenty minutes depending on the device. A coach who checks jump height or sprint speed once per block, rather than watching the time series within a block, sees a clean improving trend and never gets the chance to notice that the improvement stops exactly where a stable-baseline device would have started.

Run this before the first outdoor session of the season, and again any time the swing between the warm-up space and the field or rink changes by more than about 10°C.

1. Check the device's spec sheet for a stated operating temperature range and confirm whether it applies factory temperature compensation per-sample or only at power-on; if that is not published, ask the manufacturer directly rather than assuming.
2. Capture a 15-second stationary bias sample indoors at the end of the athlete's normal warm-up, with the device flat and motionless, and log the raw output.
3. Carry the device outside and let it soak, powered on but stationary, for at least 10 minutes before the first working rep; a longer soak of 15-20 minutes is safer below -10°C or after an indoor-to-outdoor swing greater than 25°C.
4. Capture a second 15-second stationary bias sample outdoors at the end of the soak and compare it against the indoor sample; a shift larger than the device's stated steady-state residual points to an incomplete soak or an uncompensated sensor.
5. Run a diagnostic block of a fixed sub-maximal movement every 5 minutes for 30-40 minutes and log the time series rather than a single end-of-block number, watching specifically for a smooth one-directional creep in the first 10-20 minutes.
6. If a creep shows up, extend the pre-session soak time by 5-minute increments and repeat the diagnostic until the trend sits flat within the device's normal rep-to-rep noise band from the very first rep.

A cross-country ski team logged sprint split velocity every 5 minutes across a 40-minute outdoor interval session at -12°C, one week with no soak time and the next week with a 15-minute pre-session soak, both weeks against a radar-gun reference on the same hill segment.

Without the soak, the device opened the session 0.6 m/s under the radar reference and only closed the gap after roughly 20 minutes on snow, the exact window Woodman's warm-up guidance and the thermal-shock scenario above would predict. With the 15-minute soak completed before the first interval, every reading across the block sat within 0.2 m/s of the radar gun, which is inside the normal rep-to-rep noise for this measure. The soak time cost the session fifteen minutes; the misread trend it prevented would have cost a coach a false read on early-block fatigue or a false read on improving conditioning, depending on which direction the temperature happened to be moving that day.