Fixing Bar Velocity Spikes From Bouncing: How to Spot and Filter the Artifact

Rep four of a 100 kg bench set reads 1.38 m/s. Reps one, two, three and five are all sitting between 0.71 and 0.79 m/s. Nothing about the lift looked different from the platform - same bar path, same pace, maybe a slightly harder chest bounce at the bottom. But that one number is now dragging the set average up by almost 15%, and if you are autoregulating off mean velocity, it just told you the athlete had capacity for another rep when in fact rep four was identical effort to the others.

This is not a sensor malfunction and it is not a breakthrough in bar speed. It is what happens when a stretch-shortening rebound, a cable going briefly slack, or a bar bouncing off safety pins gets read by a velocity sensor exactly like an intentional, controlled concentric rep. The device did its job correctly - it measured what happened to the bar. The problem is that what happened to the bar was not a clean concentric contraction, and treating the reading as one corrupts every downstream decision built on it.

A velocity sensor - whether it is an IMU strapped to the bar sleeve or a tethered linear position transducer - has no way to distinguish muscular force production from momentum carried over by an external rebound. Three mechanisms produce the spike, and they show up in different lifts.

Stretch-shortening rebound at the bottom of the lift. On a bench press or squat, a hard, uncontrolled bounce off the chest or out of the hole converts elastic energy stored in tendon and connective tissue into an immediate reversal of bar direction. Padulo, Laffaye, Chaouachi and Chamari (2013, Journal of Sports Sciences) compared bench press execution with a bounce technique against a controlled, paused technique at matched loads and found the bounce condition produced measurably higher peak barbell velocity and a shorter time to reach it, despite no difference in the load lifted. The extra speed is not new force output from the athlete; it is stored elastic energy being released over a very short window, which a sensor reads as an unusually steep, unusually brief velocity peak.

Cable slack rebound on tethered devices. Linear position transducers measure the payout of a cable, and if the cable goes slack for even a fraction of a second - common when a lifter drops fast into the bottom of a squat or lets a deadlift settle hard before the pull - the string can briefly overshoot as it re-tensions, registering a velocity value the bar itself never actually reached.

Bar drop at the top or during rerack. A bar dropped onto safety pins, or released and caught by j-hooks with some downward travel first, produces a sharp deceleration-then-rebound signature that some logging software fails to trim from the rep window, especially if the device is still sampling when the set is technically over.

Not every high reading is an artifact, and treating every outlier as noise is its own mistake - you will eventually throw out a genuinely fast rep. Four checks separate a bounce spike from a real one.

• Duration of the peak. A genuine concentric peak velocity is reached gradually and holds for a meaningful fraction of the rep. A bounce spike is a narrow, near-instantaneous transient - if your device exports a velocity-time curve, a bounce artifact looks like a thin needle rather than a rounded hill.
• Where in the rep it occurs. Real peak velocity in most lifts occurs in the first third to half of the concentric range of motion. A spike that appears right at the bottom turnaround, or right at lockout as the bar is being racked, is positioned exactly where a rebound would occur, not where genuine peak velocity does.
• Consistency with the rest of the set. One rep at 1.38 m/s surrounded by four reps at 0.71-0.79 m/s in the same set, same load, same athlete, same day is a statistical outlier by any reasonable measure - more than 2.5 standard deviations from the set mean in most working sets. A genuine improvement in speed under fatigue almost never looks like a spike on one rep flanked by normal reps on both sides; fatigue-driven change is gradual across a set, not a single-rep event.
• Match against video or a mean propulsive velocity value if your device reports both. If peak velocity spiked but mean propulsive velocity for the same rep looks unremarkable, the spike lives specifically in the portion of the rep sensitive to rebound, which is a strong tell.

Not every velocity metric is equally vulnerable. García-Ramos, Pestaña-Melero, Pérez-Castilla, Rojas and Haff (2018, Journal of Strength and Conditioning Research) compared mean velocity, mean propulsive velocity and peak velocity in the bench press for reliability and load-prediction accuracy, and found mean propulsive velocity - which excludes the portion of the rep where the bar is decelerating - produced more consistent load-velocity relationships than peak velocity across loads and lifters. Peak velocity, by definition, captures the single fastest instant of the rep, which is exactly the instant a bounce artifact occupies. Mean propulsive velocity, by only integrating the phase where acceleration remains positive relative to gravity, structurally excludes most of what a rebound contributes, because a bounce-driven transient sits at the boundary between the eccentric and concentric phases rather than inside the true propulsive window.

This does not mean peak velocity is a bad metric - for ballistic and jump-based testing it is often the metric you actually want, since a genuine explosive peak is the point. The issue is specific to bounced or rebound-assisted reps in strength lifts, where peak velocity conflates "fastest instant of true muscular output" with "fastest instant of stored elastic energy release," and a sensor cannot tell them apart from a single number alone.

Once you can identify a spike, you need a repeatable way to remove it without hand-editing every session. Three approaches, in order of how much technical setup they need.

1. Switch the primary metric to mean propulsive velocity. If your device or app supports it, this is the single highest-leverage change. It does not delete the artifact from the raw data, but it stops the artifact from reaching your training decisions, since the deceleration-phase transient it lives in is excluded from the calculation by definition.
2. Apply a statistical outlier filter to peak-velocity or mean-velocity data. A simple z-score cutoff works well for most home setups: flag any rep more than 2.5 standard deviations from the rolling set or session mean, and exclude it from the working average rather than deleting it outright - keep it visible but tagged. For larger datasets or automated pipelines, a Hampel filter (a median-based outlier detector using a rolling window of 5 reps and a threshold of roughly 3 times the median absolute deviation) rejects spikes without being thrown off by genuine gradual trends the way a simple standard-deviation cutoff can be.
3. Fix the mechanical source directly. For cable slack, add light pre-tension to the LPT line before the rep starts so there is no free travel to overshoot through. For bar drop artifacts, confirm your device's end-of-rep detection window closes before the bar contacts pins or hooks - most apps let you set a minimum velocity threshold or a maximum rep duration that trims trailing samples. For chest bounce specifically, a technique cue (a controlled 1-count pause at the chest) removes the artifact at the source rather than filtering it after the fact, and has the side benefit of making the strength stimulus more comparable to a paused competition-standard rep.

Run this once on any lift where you suspect bounce is inflating readings - bench press and box squat are the most common candidates.

1. Pull the last 5 sessions of rep-by-rep data at a consistent submaximal load (60-75% 1RM works well) for the lift in question.
2. For each session, calculate the mean and standard deviation of peak velocity across all logged reps.
3. Flag any individual rep more than 2.5 standard deviations above that session's mean.
4. For each flagged rep, check whether mean propulsive velocity for the same rep is also elevated, or whether it looks normal while only peak velocity spiked - the latter pattern is the bounce signature.
5. Count flagged reps as a percentage of total reps logged. Below 5% is typical background noise; above 15% suggests a technique or setup issue worth addressing directly rather than filtering around indefinitely.

Weakley, Morrison, García-Ramos, Johnston, Diefenbach and Banyard (2021, Sports Medicine), in a systematic review of commercially available velocity-monitoring devices, noted that between-device and between-technique variability in resistance training was frequently larger than the true biological change practitioners were trying to detect, and specifically flagged uncontrolled technique variation - including bounce and pause inconsistency - as a source of error that device accuracy alone cannot fix. That is the core limitation worth sitting with: no amount of filtering downstream substitutes for controlling the rep itself when consistency matters more than a single fast number.

An intermediate lifter (bench 1RM 100 kg) logged 6 reps at 75 kg with a moderately bouncy technique at the chest. Raw peak velocity data below, followed by the same set with a 2.5 SD outlier filter and a switch to mean propulsive velocity as the primary metric.

The raw peak-velocity average across all six reps is 0.94 m/s - a number with no physiological meaning, inflated entirely by reps 3 and 5. Excluding the two flagged reps drops the peak-velocity average to 0.725 m/s, closely matching what the mean propulsive velocity trend already showed across all six reps (roughly 0.60 m/s scaled to the same phase). Had this lifter autoregulated off the raw peak-velocity average, the session would have read as unusually fast and possibly triggered an unwarranted load increase the following week. The mean propulsive velocity column told the true story the entire time - it simply was not the metric being watched.