Fixing Accelerometer Clipping in High-Impact Plyometrics

Five reps into a 45 cm drop jump set, the raw acceleration trace reads the exact same peak every time: 16.0g, 16.0g, 16.0g, flat as a shelf for two or three samples before it drops back down. No athlete decelerates identically five times in a row, least of all on a fatiguing stretch-shortening protocol, but the sensor reports it that way. Meanwhile ground contact time for those same five reps swings from 187 ms to 241 ms, and reactive strength index bounces from 1.66 to 2.14 within one set. That reads like a technique problem, or fatigue, or a noisy athlete. It is neither. It is a sensor pinned against the ceiling of its own measurement range, and the timing numbers built on top of that ceiling are not describing the landing anymore - they are describing wherever the algorithm happened to land on a plateau it was never built to see.

Accelerometer clipping, also called saturation or railing, happens when the true acceleration during an event exceeds the maximum value the sensor's analog-to-digital converter can register. Instead of reporting the real peak, the device reports its own ceiling value for every sample the true signal stays above it. A ±16g accelerometer sampling a landing that actually peaks at 22g does not report 22g - it reports 16.0g, repeated, for however many samples the true impact stayed above 16g.

Visually this is easy to spot once you know the shape. A genuine peak rises, holds for a sample or two at its true maximum, and falls away in a smooth, slightly asymmetric curve. A clipped peak rises just as sharply, then goes dead flat - identical values, sample after sample, with essentially no noise on top - before dropping. That dead-flat run, sitting at exactly the device's rated maximum, is the fingerprint. If your platform hides raw traces, a proxy tell is a peak-acceleration column that returns the same round number rep after rep across a training block, matching a spec you can find in the device manual.

High-impact plyometrics push a sensor past its range for two reasons that compound each other: the movement is genuinely violent, and the segment where contact-detection sensors usually sit is where that violence peaks.

Landing force scales sharply with drop height and stiffness. McNitt-Gray (1991, Journal of Biomechanics) measured landing kinetics from drop heights of 0.32 m, 0.72 m and 1.03 m and found peak vertical ground reaction force climbed from roughly 3-4 times body weight at the lowest height toward 8-11 times body weight at the highest, an increase only partly offset by greater knee and ankle flexion at the larger heights. A stiffer landing - the kind a reactive-strength protocol actually wants, since minimizing ground contact time is the point - dissipates that force over a shorter window, which pushes peak acceleration higher still at the same drop height.

Where the sensor sits multiplies the problem. A unit at the sacrum sees an attenuated version of the impact, damped by the leg's mass and the joints above it, which is why sacral-mounted trackers commonly run fine on a ±16g range. A unit at the shank or shoe sees the impact close to where it is generated and routinely registers two to three times the peak value the same landing produces at the sacrum. Fong and Chan (2010, Sensors), reviewing wearable inertial sensors across lower-limb biomechanics studies, catalogued accelerometer ranges spanning roughly ±2g to ±20g in the literature and flagged that ranges under about ±16g were frequently inadequate for high-impact jumping and landing tasks - guidance that is necessarily general, aggregated across many devices and protocols rather than isolating clipping's effect on one metric, but the direction matches what shows up in shank and foot-mounted plyometric data in the field.

Ground contact time and reactive strength index are not corrupted by clipping the way you might expect, which is exactly what makes it easy to miss. Flight time, the phase most algorithms use to derive jump height, happens in near free-fall well under 1g, so a clipped landing peak never touches it. What clipping corrupts is the part of the algorithm that has to locate the landing event itself.

Most contact-detection logic anchors the exact timestamp of touchdown, midstance, or toe-off to the acceleration peak or the shape of the curve around it - some methods take the peak's sample index as the reference point for splitting the contact window into braking and propulsive halves, others use a slope-change method that assumes one clean, unimodal peak. A clipped signal breaks that assumption. Instead of a single maximum, the algorithm sees a plateau of identical samples, and depending on where it locks on within that plateau, the computed contact boundary shifts by a handful of samples in either direction, rep to rep, with no consistent bias. That is the detail worth sitting with: clipping does not reliably make ground contact time read short, or reliably read long. It makes it read inconsistently, in a way that looks exactly like biological variability in an RSI trend line but is actually noise from a sensor pinned at its ceiling.

Simons and Bradshaw (2016, Sports Biomechanics) compared accelerometer-derived load measures against a video-based criterion across movements ranging from fine motor tasks to gross, high-impact athletic movements, and reported agreement weakening substantially at the high-impact end - correlations above 0.90 for lower-intensity movements dropped to roughly 0.60-0.70 for jumping and change-of-direction tasks, with part of that divergence attributed to the sensor's fixed measurement ceiling relative to the accelerations those movements produce. Their comparison used a single accelerometer model and mounting configuration, so the exact correlation values will not transfer to every device, but the pattern - accuracy degrading as impact magnitude rises toward a sensor's rated limit - is the same mechanism at work in a clipped drop-jump trace.

Matching sensor range to mounting location and drop height means picking a setting with headroom above the impacts you actually expect, rather than leaving whatever range shipped by default. The table below is a starting point from typical field values, not a hard ceiling - a stiff lander or a box height above what is listed can still exceed the low end of a band, which is why the range-check protocol in the next section matters more than any single number here.

Two notes go with it. A wider range is not free - it trades away resolution, since the same bit depth from the analog-to-digital converter now covers a larger span of possible values, so a ±100g setting reports coarser increments than a ±16g setting at the same sample rate. For timing-based metrics like ground contact time and RSI this trade-off rarely matters in practice; it matters more for protocols using the raw peak value directly, such as estimating leg stiffness from impact force. And if the device has no selectable range at all, the workaround is a lower-impact mounting location, a lower test drop height until a wider-range unit is available, or a dedicated high-g sensor for that specific protocol.

Run this once on any drop jump or depth jump protocol feeding ground contact time or RSI into a training or return-to-play decision, and repeat it whenever the device, mounting location, or drop height changes meaningfully.

1. Confirm the sensor's rated accelerometer range from the device manual or app settings, and note whether it is user-selectable or fixed in hardware.
2. Have the athlete perform 5 reps at the drop height and mounting location you actually intend to use in testing.
3. Export or view the raw acceleration-time trace for each rep and inspect the top of the landing peak for a run of two or more consecutive samples at an identical value near the device's rated maximum - that flat plateau is the clipping fingerprint, not noise.
4. If clipping is present and the range is user-selectable, step up to the next available setting (±16g to ±32g, for example) and repeat the same 5 reps.
5. Compare rep-to-rep consistency of ground contact time and RSI before and after, not just the average - a coefficient of variation that drops sharply after widening the range confirms the original values were clipping noise rather than genuine variability.
6. For an independent check, film one rep in side-view slow-motion video at 240 fps or higher and count frames from first ground contact to toe-off, multiplying by frame duration to get a contact time to compare against the sensor's post-fix value.

A U18 volleyball athlete performing a 45 cm drop jump on a shank-mounted IMU fixed at ±16g showed the flat-plateau signature on every rep of a 5-rep set. Raw peak acceleration, ground contact time and RSI for that set are shown below, alongside the same 5 reps after switching the sensor to ±32g.

The averages alone tell a partial story - clipped ground contact time averaged 211 ms against a corrected 212.2 ms, close enough that a glance at means alone might miss anything wrong. The rep-to-rep spread is where the artifact shows: contact time swung across a 54 ms range while clipped and a 9 ms range once corrected, and RSI's coefficient of variation dropped from roughly 10% to under 2% after the fix - the signature of removing measurement noise, not of an athlete suddenly becoming more consistent. True peak values after the fix ranged from 18.9g to 24.6g, every one of them above the ±16g ceiling that had been clipping the set, confirming the flat readings were hard saturation rather than a coincidence of technique. Test-retest RSI variability in clean drop-jump protocols typically sits in the mid-single-digit-to-low-double-digit percentage range, which means the clipped data here was sitting right at the edge of passing for normal biological noise instead of an obvious equipment problem - the exact trap that makes this artifact easy to overlook.