Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Flight-log readings

See which readings HPR Sim takes from a flight log alone, the bound each states, and when it withholds one instead of guessing.

In short

  • What it models: the readings taken from a flight log on its own, with no design file and no simulation. They are liftoff, apogee and its time, the top speed in the climb, and landing. The rest are the flight time, the mean rate of descent, and each canopy’s descent rate with the main’s opening. They come from any logger’s height track. The logger’s own barometric altitude will do. So will a pressure turned into a height by the 1976 standard atmosphere, or an altitude kept in a column of its own. Two logs of one flight go on one timeline. An accelerometer along the rocket’s axis adds the top acceleration and burnout. Without one, burnout is a floor: the top speed’s time less its bound. Each reading states a bound on its error and where it came from, or is withheld with one of 28 reasons. It never guesses. A GNSS (GPS) height gives no reading.
  • Sources: the running median is the Hampel filter at threshold zero, as Pearson and colleagues define it. An accelerometer’s trace is cleaned by the same filter at threshold five. The landing’s thresholds and the least climb are those of Debrief, the project owner’s earlier flight-log analyzer. The flat top of a clipped accelerometer and the marks that tell the main’s descent from the drogue’s are Debrief’s too, each with HPR Sim’s changes. The bounds’ terms come from the noise model of the synthetic flights and the hypsometric equation. They also come from the safety codes’ 3:1 thrust to weight and 20° rail. Another source is the National Weather Service’s rate for a rapid change of pressure. Another is the drag that sets a canopy’s rate and slows its fall after it opens. The last is the 1976 standard atmosphere’s density scale height (References). The rest are HPR Sim’s choices, recorded in ADR-281, readings on every logger, and ADR-283, the rest of the readings.
  • How well it is validated: on the synthetic flights’ tuning set, every reading given stayed within its bound. The set is 171 flights logged by a barometer. Each was logged again at 36 rates and steps and cut 6 ways. The 71 of them with an accelerometer were read again with it at 500, 100 and 20 samples a second. Thirty-eight were read with it set to ±8 g. The largest share of its bound any error took was 0.737, on a top acceleration, and 0.717 on an apogee (Checked against). The bounds were tuned on those same flights, so this shows they are wide enough there, not on flights held apart. On an invented PerfectFlite log, every reading lands within its bound. Each but burnout’s floor also lands within the tighter one its test works out (evidence row G10). A real log has no truth beside it. The 1,058 records of the project’s sample logs show which readings each logger gives (an apogee from 592 of them), not how right they are. All twelve readings’ 90% intervals pass the two-sided test of ADR-257, calibrated intervals on the synthetic sets A and B, after their points were fitted on 2,000 flights of their own, and a floor’s lower end passes that test’s lower tail, a clipped top’s on tops written near an imposed clip (evidence row G34).
  • What it leaves out: the Mach number and dynamic pressure; a lower end on an apogee floor, a top speed floor or a top whose burn the log ends in (#1071). Loggers noisier than the tuning set’s six simulated logger profiles and rockets slower off the pad than 2 g can be read outside their bound. Air more than 15 K from standard is out too, and so is any correction for the day’s air, which sits in the bound instead. Where the log records the pad’s temperature and its height above sea level (from its pressure, an altitude above sea level, or a stated ground elevation), a day past 15 K is noted with its size (A day past 15 K). A canopy’s filling time isn’t in the opening’s bound (#779). A fall under no canopy is read as one canopy’s (#780). A canopy far above the tropopause is out. Accelerometer noise that runs from sample to sample is out; it would narrow burnout’s bound (#787). A top clipped at the high-g part’s range is out; no tuning flight reaches one. A board’s three axes give no top acceleration (#781; What it leaves out).

The method

Why these rules

Code: hpr_flightdata::readings (API reference); on the command line hpr analyze; from Python, FlightLog.readings() (Reading your own logger). The readings came with M4.2d, the milestone that added hpr analyze, on PerfectFlite’s logs alone (ADR-108, the decision on how a log is read). M7.2b, readings on every logger, takes them from every logger’s height track and gives each a bound held on the synthetic flights (ADR-281). M7.2c, the rest of the readings, adds the top acceleration and burnout from an accelerometer along the rocket’s axis. It adds the flight time as a reading of its own, and the descent under each canopy (ADR-283).

A barometric altimeter’s height track is a good record of a flight’s heights, with flaws a reading must survive:

  • Steps. It moves in steps of its resolution, a foot for a PerfectFlite, so the top of the climb sits flat for several samples.
  • Pulses. When the ejection charge fires, the pressure inside the electronics bay jumps. For a tenth of a second the height swings tens of feet away from the truth. Take the highest sample as the apogee and you read that swing.
  • Noise and drift. The sensor’s noise is correlated from one sample to the next, so it doesn’t average away quickly, and its zero drifts.
  • The day’s air. A logger turns pressure into height by the standard atmosphere. So on a day warmer or colder than standard, every height is off by a share of itself.

So every height and time comes from the track after a running median, which removes the pulse. Each reading states a bound that covers the rest. Each has a rule that says when the log can’t support it.

The height track

Every reading of a height or a time comes from one track, the first of these the log holds. The code each reading names as its source says which:

sourcethe trackits zero
barometerthe logger’s own barometric altitude, the record’s primary trackthe logger’s zero, which it sets on the pad
pressurea pressure channel, turned into a height by the 1976 standard atmosphere, as every logger documented does it (the height formulas note)the pad, found from the samples
altitude_channela barometric altitude kept in a channel, above the ground or above sea levelthe ground for the first; the pad, found from the samples, for the second
gnssa GNSS receiver’s altitudenone: no reading is taken
accelerometernot a height: the accelerometer along the rocket’s axis, for the top acceleration and burnout alone (The accelerometer)a fall in vacuum reads 0 and the pad +1 g, gravity put back where the logger took it out

A GNSS height gives no reading. A receiver’s navigation filter lags a climb and overshoots its top by more than a barometer’s bounds allow (receiver_lag). Its fix drops out or lags at liftoff (receiver_dropout). An altitude whose sensor the log doesn’t state, such as a table’s Altitude column read by its header, gives none either (unstated_sensor). A barometer’s bounds can’t be stated for it. A column mapping can mark such a column as the barometer’s own track (Reading your own logger).

A track’s step is the logger’s resolution where its reader knows it, a foot for a PerfectFlite. Otherwise it is the smallest gap between two of the track’s values. A pressure’s step becomes a step in height that grows as the air thins, Δh=Δp/(ρg), with ρ the standard atmosphere’s density at that height. At sea level, 10 Pa is 0.83 m (2.7 ft).

The running median

The running median replaces each sample with the median of the samples within K places of it. The window is cut short at either end of the log. The code sets K from the log’s own sample interval Δt so that the window spans 0.3 s. At a PerfectFlite’s 20 samples a second, K=3, seven samples.

Pearson and colleagues define the Hampel filter by the window’s median mk and its scale

Sk=1.4826×median|xk−j−mk|

A sample more than t⋅Sk from mk is replaced by mk. At t=0 every sample is replaced, and the filter is the running median (§2, eqs. 1 and 2).

Two properties make it the right tool here:

  • It removes any pulse up to K samples wide. Such a pulse holds at most K of the window’s 2K+1 samples, so the median is never one of them. On the public log below, the ejection pulse’s high side is two samples wide.
  • It reads a clean peak a little low, and never high. At 20 samples a second, the apogee of a noise-free trace reads at most 0.077 m (0.25 ft) low. The steps behind that number:
    1. At the highest sample, K+1 of the window’s 2K+1 samples lie within ⌈K/2⌉ places of it (⌈K/2⌉ is K/2 rounded up). So the median is no lower than the altitude that far away.
    2. The true peak lies within half a sample of the highest sample, so that altitude is at most (⌈K/2⌉+12)Δt from the true peak in time.
    3. At apogee the vertical speed is zero, so drag has no vertical part and only gravity curves the path. In that time the height falls at most g((⌈K/2⌉+12)Δt)2/2.
    4. With K=3 and Δt=0.05s, that is 9.80665×0.1252/2, or 0.077m.

Debrief uses the Hampel filter at t=4 over the same window. FusionSpace HPR doesn’t, because of what the public Pnut log shows. Around its ejection pulse the trace dips below itself just before the pulse and stays lower after it. Those samples sit in the pulse’s own window and widen its scale Sk. So the Hampel filter keeps the pulse. The result’s highest value is the pulse’s 1,028 ft above the altimeter’s zero, against the 1,009 ft apogee the altimeter states. After the running median, the apogee reads 1,010 ft above the zero, which is 1,009 ft above the pad.

The pad and the climb

The readings measure heights from the pad, so the pad must be found first, and only from samples taken before the rocket moved. Three rules do it, each with a number set from the noise:

  1. The allowance. The filtered track may sit up to 2 m (6.6 ft) from the pad and still be on it. If the pad’s own filtered samples stray farther, the allowance is that far instead. One step is added to either. The 2 m (6.6 ft) is four times the synthetic flights’ fast correlated barometer noise, 0.5 m (1.6 ft), HPR Sim’s choice. Half a foot, the rule before M7.2b, read liftoff up to 6 s early on a noisy 500 Hz barometer. The noise lifted the pad’s last samples above it. An allowance of 1 m (3.3 ft) read one tuning flight 6.8 s early, against a 4.5 s bound. One of 2 m (6.6 ft) never did.

  2. The climb begins at the first filtered sample twice the allowance and a step above where the log starts. That is 4.61 m (15.1 ft) on a PerfectFlite, whose step is a foot. The pad’s noise lifts the filtered track by at most the allowance, from below at the start or from above at the climb. So twice that is out of its reach. Bad samples at the log’s very start would lift that start, so the first three are judged first. One or two of them are bad, and the readings leave them out, if they stand more than this rise and 9 m (30 ft, three times the barometer’s noise) above the second lowest of the samples after those three, over the median’s window and at least five. Two cases are left alone:

    • All three stand that far above. That is a fall of the track’s own, which reads like a pressure pulse on the pad at ignition.
    • The track stays at their level again, for at least as long as a rocket at 2 g takes to leave that band, before a top further above it than a coast through the band reaches. That level is the pad. A low stretch far below it counts in the pad’s reach, so the apogee is withheld as no_climb.

    Eight sample logs have good first samples and then fall 31 to 654 m (102 to 2,146 ft) into the climb: a disturbed pad, below.

    Samples low at the start are not left out, as a logger’s zero before the climb looks the same. At 5 and 2.5 samples a second they read the apogee high. At those rates on a log from launch, and more so the faster the rocket climbs, two high samples at a level the climb itself soon reaches can read it low, past its bound (an open issue).

  3. The pad is the median of the samples up to the last one within half that rise of the start, before the climb. Half the rise is 2.30 m (7.6 ft) on a PerfectFlite. The median means that no one sample’s jitter sets the pad, and the half rise means the climb lends it few samples. A track that rests at two levels before the climb is read as two levels, below.

The pad’s error. A log that starts just before liftoff holds little of the pad, and a rocket may already be moving when its first sample is taken. Take a rocket that leaves the pad at A=2g or more and is moving at v when the log starts. It climbs vt+12At2 in the time t the track stays within half the rise of its start. That climb can’t pass the rise r, the noise allowed at either end. So its samples bound v, and with it how far up the rocket may already be:

v≤(r−12At2)/t,epad=v2/(2A)

That much enters the liftoff, apogee and landing bounds. The acceleration is the least the safety codes allow. The NAR’s high-power code asks that a rocket weigh no more than a third of its motor’s certified average thrust. Tripoli’s allows less than 3:1 only with an active stability system. Thrust three times the weight leaves twice gravity to climb with. The tuning set’s slowest rocket averages 30 m/s² over its first 5 m (16 ft).

On the public Pnut log, for example, the filtered track stays within 2.30 m (7.6 ft) of its start for 0.40 s. Suppose a rocket at A=19.6m/s2 climbed no more than the 4.61 m (15.1 ft) rise in that time. It was moving at no more than 7.60 m/s (24.9 ft/s) when the log started, from (4.61−19.6×0.402/2)/0.40. So it can’t have been more than 7.602/(2×19.6), or 1.47 m (4.8 ft), up. That is within the allowance, so the pad stands, and the 1.47 m (4.8 ft) is part of the apogee’s 27.4 m (90 ft) bound.

A short pad. Where the pad’s error could pass the 2 m (6.6 ft) allowance, the pad is short (short_pad). Either the log starts after liftoff, as a logger that records from launch writes its zero and then the climb. Or it holds too little of the pad to tell liftoff by. What happens next depends on the track’s zero:

  • A zero set on the pad, the logger’s own altitude as its format’s reader reads it: the heights are read from that zero. Liftoff is withheld unless a sample before apogee sits within the allowance of the zero. The top speed is withheld too if it peaks within 0.5 s of the log’s start. A faster climb may have come before the first sample there.
  • A zero only a word vouches for. A mapped table’s altitude, or a channel whose header calls it above the ground (#646, a wrong unit or a wrong zero in a mapping). Such a zero may sit anywhere within 19.9 m (65 ft) of the pad (below). With no pad of samples, nothing places it better than the first sample does. That sample reads s above the zero after a climb of up to the pad’s error C. So the zero stands between s below the pad and C − s above it. The heights are read from it, as from a zero set on the pad. The larger distance, max(|s|, min(19.9 m, C − s)), enters the liftoff, apogee and landing bounds as the pad’s error. A log too sparse to bound the climb before it takes the whole 19.9 m (65 ft).
  • Any other track, a pressure or an altitude above sea level: no zero places the pad, so every reading is withheld. The top acceleration is withheld too, as it is searched for up to apogee.

A disturbed pad. A rocket on the pad doesn’t sink below it. Yet some logs fall tens to hundreds of meters straight into the climb, by its look a pressure pulse at ignition (#1101). If the fall’s samples counted as pad, every height would read high. Where the filtered track falls below its start by more than the rise plus 0.015 m/s (0.05 ft/s) of drift since the log began, then climbs without resting, the top speed is withheld (disturbed_pad). To rest is to stay within half the rise of one height for as long as a rocket at A takes to cross the rise, 0.69 s on a PerfectFlite; a fall to a level the track holds is no disturbance. The apogee is still read, from the samples resting before the fall, else the logger’s zero; with no zero, nothing is. Liftoff’s bound reaches back over the fall to the last sample on the pad, with no 90% interval; with none, liftoff is withheld. On the two such records that state their apogee, the reading comes within about 0.3 m (1 ft) of it; the first sample as the pad reads up to 268.8 m (882 ft) high, the fall’s end up to 339.9 m (1,115 ft) (the survey’s report).

Two levels. A logger switched on above the pad and carried down to it rests at two levels, and so does a reading the pressure moves before liftoff. The track can’t tell the two apart. Where it falls as above and the medians of its stay at the start and of its last rest before the climb lie more than the rise apart, the pad is that last rest, read back from the rest’s end while its samples stay within the rise of each other. The gap between the two levels enters the pad’s error, as either may be the pad, so the bounds hold whichever of the two it was, while the pad’s noise stays within the rise and bar the short stay below. The pad reported is the last rest’s level alone (#1158). A synthetic log resting 30 m (98 ft) up for 5 s, then 3 s on the pad, reads its 98.1 m (322 ft) apogee as 97.96 m (321 ft) on the logger’s zero, with a bound of 40.9 m (134 ft) (the_pad_is_the_last_rest_before_the_climb). The climb is still found from the log’s start. A fall from the last rest into the climb is a pulse, as one from the start is.

Each reading

Each reading names its method and its provenance. The provenance is measured when the reading is read straight off the track, and derived when it is worked out from it. It is clipped for a top acceleration at its part’s range or held flat at its top. Such a top is a floor, its clip saying which (The accelerometer). There are nine methods:

readingrulemethodprovenance
liftoffthe last sample before apogee whose filtered height is within the pad’s allowancelast_on_the_padmeasured
apogeethe filtered track’s highest value; its time the middle of the run of samples that hold ithighest_filteredmeasured
time to apogeeapogee’s time less liftoff’s
top speedthe steepest least-squares slope of the filtered track over 0.5 s, from liftoff to apogeesteepest_slopederived
landingthe first sample after apogee back within the pad’s allowance that stays below it, 3 m (9.8 ft) and a step for a second, and no sooner than a fall from rest in vacuum could lose that heightback_on_the_padmeasured
flight timelanding’s time less liftoff’sliftoff_to_landingmeasured
mean descent ratethe filtered height lost from apogee to landing, over the time taken, part of landing’s reading
top accelerationthe axial accelerometer’s highest reading over the burns, searched from liftoff’s earliest time (the log’s start if liftoff is withheld), gravity in; a floor where it clipped, or where the thrust doesn’t end on the channelhighest_readingmeasured, or clipped
burnoutfrom the accelerometer: where its reading is least after it falls through zero, the end of the thrust’s fallleast_after_thrustmeasured
without it: a floor, the top speed’s time less its bound; the reading is the top speed’s timespeed_peakderived
drogue descentthe filtered height lost from apogee to the main’s opening, over the time takenleg_chordderived
main descentthe same from the main’s opening, found back from landing, to landing; from apogee where the descent shows one canopyleg_chordderived

Liftoff is the last sample before the track shows the rocket climbing out of the pad’s reach. So it comes after the true liftoff by the time that climb took.

A floor holds one way only, and has up to three numbers. The reading is where the rule lands. For burnout without an accelerometer, that is the top speed’s time, 2.20 s on the invented PerfectFlite flight below. The floor is that reading less its bound: 2.20 s less 1.32 s, 0.88 s, which hpr analyze prints as “no sooner than 0.88 s”. The truth is at or after the floor, or for a top acceleration at or above it. It may be later, or higher, by any amount. The third, the lower end, is tighter: the figure the truth stayed above in about 19 of 20 synthetic flights (A floor’s lower end). The floor is the limit nothing fell below on the tuning set. The errors the tuning set reports for a floor are the reading’s, not the floor’s (Checked against; the glossary).

The top speed is HPR Sim’s own, worked out from the heights rather than read from a speed column a logger writes. Every logger’s track can give one, and its bound can be stated. A 0.5 s fit averages the barometer’s correlated noise down to a few meters a second. It is short enough to keep most of the peak. It needs at least 5 samples in the fit, so a log sampled slower than every 0.125 s has none (sampled_too_slowly). It is withheld at 227.7 m/s (747 ft/s) and above (transonic). That is Mach 0.8 in the coldest air the bounds allow below 20 km (12 mi). The standard atmosphere’s 216.65 K less 15 K gives that air. From Mach 0.9 the pressure behind an ogive’s shoulder jumps on its way to Mach 1.0 (the Mach error note §1–2). So a speed derived from the heights there can be far off. Debrief’s two refusals stay (implausible_speed and noisy_speed; When a reading is withheld).

Landing is the first sample within the pad’s allowance, at least Debrief’s 2 m (6.6 ft), that meets one more test. From it, the track stays under the allowance and 3 m (9.8 ft) more, and a step, for Debrief’s second. So it comes before touchdown by the time the last few meters took. That is about 0.4 s under a main at 6 m/s (20 ft/s) on a PerfectFlite’s log. A barometer drifts during a flight, so the ground can read a little above or below the pad. The mean descent rate covers the drogue and the main together. Each canopy’s own rate is read from the shape of the descent (The canopies).

The flight time is a reading of its own, flight_time: landing’s time less liftoff’s, its bound the sum of theirs, so it needs both. Landing’s reading doesn’t carry it.

The top acceleration comes only from an accelerometer, never from the heights. Differencing twice turns a track’s steps into spikes: at 20 samples a second, a one-foot step differenced twice is 0.3048/0.052=122m/s2, over 12 g. So a log with no accelerometer gives none (no_accelerometer), and its burnout is a floor: the top speed’s time less its bound.

The accelerometer

The top acceleration and burnout come from an accelerometer along the rocket’s axis, and only from one the log names as that axis (ADR-283 §1–4). A table in the code lists each logger family’s axial column, such as AltOS’s acceleration or the Raven’s Axial Accel (Gs). It also lists whether the family says its readings hold gravity, and the part’s range where the family states one. A column whose name holds the word axial, as a table’s or a mapping’s may, is read with neither known. A board’s three axes don’t say which lies along the rocket, and their magnitude never falls below zero at burnout. So they give neither reading (axis_unstated).

AltosUI writes acceleration for every board, but an EasyMini or a TeleMini has a barometer and no accelerometer (the Altus Metrum manual’s specifications). So on one the column is the board’s estimate from its barometer, not a sensor’s reading. It gives no reading here. Each AltOS board with an axial accelerometer adds columns to the file. They are an inertial unit’s (accel_x and on: TeleMega, EasyMega) or a GPS receiver’s (latitude and on: TeleMetrum). So an AltOS CSV with neither is taken for a barometer-only board’s and gives no top acceleration (no_accelerometer). Its burnout comes from the top speed. An export that lost those columns loses its reading too. A TeleGPS, a GPS receiver with no accelerometer, would pass if its export wrote the column. On the sample logs 17 records are so withheld, 16 AltOS CSV and one AltOS table.

What an accelerometer reads. It doesn’t feel gravity’s pull. It feels every other force on the rocket, over its mass: the specific force. Along the axis that is the thrust less the drag, over the mass. Standing on the pad, the part reads +1 g, the pad’s push. Coasting, it reads the drag’s pull back, below zero. Falling in vacuum, it would read 0. Some loggers write what the part reads; others take gravity’s 1 g out on the pad, so the pad reads 0. HPR Sim puts the 1 g back where it was taken out, so every reading is the specific force, +1 g on the pad. So is the top acceleration: it holds the 1 g, where hpr sim’s max_acceleration_m_s2 is the rocket’s own acceleration, from the equations of motion. Less 1 g, the top acceleration is the rocket’s own acceleration up on a vertical climb.

Which way up, by three witnesses. A part mounted nose down writes the thrust negative, and its readings are turned over. Which way up is settled as gravity is (below), by three witnesses. Each is read on the trace as written, before it is turned. Each rests on physics that holds whether or not the logger takes gravity out. Each reads within 0.25 g of its level:

  • the pad, the steady level before the trace first moves 1 g from it, either way. That counts if it comes no later than liftoff’s latest time by the heights. At rest a part that holds gravity reads +1 g nose up and −1 g nose down. One that takes it out reads 0 either way. So about 1 g says nose up, about −1 g nose down, and about 0 says nothing;
  • the coast, the last 1 s before apogee’s earliest time, read only where a coast was flown. The window is after liftoff’s latest time (the log’s start, where liftoff is withheld). It is also after the trace first moves from its pad, and before an apogee the log doesn’t end at. Coasting, a part reads the drag’s pull back, −D/m, or −D/m−1g where gravity is taken out. It is never above 0 nose up while it coasts, its thrust ended and no canopy open, and the mirror nose down. So 0.75 g or more below 0 says nose up, and 0.75 g or more above it says nose down. A drag only pushes the reading further. Nearer 0 says nothing;
  • the thrust, on the cleaned trace, found as the burns are (both below), from liftoff’s earliest time. The first span, in either sign, that holds past 2 g for 0.5 s says its sign. The exception is a span of the other sign that comes first. A short burn’s span does so before a heavy drag’s, and a knock’s rebound before the thrust. The thrust then says nothing. Where no span holds that long, the sign every span shares says, a short burn’s; spans both ways, or none, say nothing. The span that says must rise by liftoff’s latest time, as the thrust is what lifts the rocket. Where liftoff is withheld because the log starts in the climb, it must rise within 0.5 s of the log’s start. One that rises later is a drag after a thrust too weak to read past 2 g, or a second burn, and says nothing.

The way is settled when the pad or the thrust says and no witness disagrees. A thrust read at the start of a log whose liftoff is withheld settles only beside a coast that says the same way. A log that starts after burnout has the drag there instead, past 2 g the other way. With no pad to tell by, nothing else could catch it. The coast alone settles nothing. Its second can hold a weak thrust still burning or a canopy opened early, either of which reads the other way. So it can only disagree. Otherwise the top acceleration and burnout from the accelerometer are withheld (direction_unstated), and burnout comes from the top speed. So are they where the trace, turned and with gravity put back, never passes 2 g, as a rocket leaving the pad does. Such a trace shows no thrust. The thrust’s witness reads the trace as written. So on a log that takes gravity out, a thrust under about 3 g reads under 2 g and says nothing, and the pad decides. Where the pad says nothing too, as on a log that takes gravity out, both readings are withheld. Three AltOS CSV records of the sample logs are withheld so: their short burn comes before a heavy drag. A knock on the pad before liftoff’s earliest time isn’t searched, and a canopy’s shock, however large, comes after the thrust.

On the tests’ short burn, gravity in, the part reads 40 g for 0.3 s. Then a drag of −8 g eases to nothing by 6 s. The drag’s span holds past 2 g for about 1.7 s and the burn’s for 0.3 s, but the burn’s comes first. So the thrust says nothing. The pad’s 1 g says nose up, and the top reads 40 g. Turned by the drag’s span alone, as an earlier rule turned it, the top read 7.9 g. Six sample logs are withheld as direction_unstated; Checked against says why.

A sample that isn’t a number or has no time is left out. So is one that repeats the time of the last kept, the first kept, as the height track keeps AltOS’s rows that share a time. A part sampled faster than every 0.002 s, 500 times a second, the fastest the bounds were checked at, gives neither reading (sampled_too_fast). Burnout then comes from the top speed.

A cleaned trace. A charge’s shock or one bad sample could turn the readings over, end the thrust early or pass for burnout. So those are found on a cleaned trace. A Hampel filter over 0.1 s puts each sample back to the window’s median if it is more than five robust standard deviations from it. A robust standard deviation is 1.4826 times the window’s median absolute deviation, as Pearson and colleagues define it. The filter takes out a spike narrower than half its window and keeps a step or a steady run as it was. The 0.1 s and the five are HPR Sim’s choice. The top acceleration is read off the samples themselves, not the cleaned trace.

Gravity, by three witnesses. Whether the 1 g is in the readings is settled by three witnesses, each within 0.25 g of the level it reads:

  • the family’s word, where the logger’s maker says;
  • the pad, read off the accelerometer itself: the steady level before the thrust lifts it by 1 g. That counts if it comes no later than liftoff’s latest time by the heights. About 1 g holds gravity; about 0 doesn’t;
  • the coast, the last 1 s before apogee’s earliest time, where a rocket falls freely but for a small drag. About 0 holds gravity; about −1 g doesn’t. It is read only where a coast was flown: after the trace first moves from the pad and before an apogee the log doesn’t end at. Unlike the direction’s coast, it isn’t also held to liftoff’s latest time: coarse heights put that time late, past a coast that was flown, and a family’s word the coast would contradict then settled alone.

A witness that reads neither level, such as the pad of a log that starts in the thrust, says nothing. Gravity is settled when at least one witness says and none disagrees, with one exception: a coast at about −1 g doesn’t settle it alone. A part that holds gravity reads its drag there, and a rocket still fast at apogee, arcing over, can drag about 1 g, which reads as gravity out would. Putting 1 g back on that log would read its top 1 g high (#806). A coast at about 0 still settles it alone: without gravity a coasting part reads −1 g or below, drag only pushing it further down, so reading within 0.25 g of 0 takes a net push forward, thrust less drag, of between about 0.75 and 1.25 g. A motor still burning in the coast’s second gives such a push. On a part that takes gravity out, with no family word and no pad, that reads gravity in and the top 1 g low, outside its bound: a risk left open, and no test pins it (#1131). Otherwise the top acceleration is withheld (gravity_unstated), and burnout comes from the top speed. The 0.25 g is HPR Sim’s choice, set out in What each bound holds. The pad is read off the accelerometer rather than as the samples before liftoff’s earliest time by the heights. Most Raven logs start about 0.5 s before liftoff, against a liftoff bound near 0.9 s. That rule, tried when this one was chosen, withheld 69 of the 80 Raven CSV records. The coast settles the Raven’s logs that start in the thrust, whose family doesn’t say (ADR-283). On the sample logs no record’s gravity rested on a coast at −1 g alone, or on a coast read off the pad: with both rules the committed report reproduces, its per-family counts of readings given and withheld unchanged.

The top acceleration is the highest specific force while the motor burns. A burn runs from a cleaned reading past 2 g to where the cleaned trace, averaged over 0.05 s, next falls through zero. Where it doesn’t, the burn runs to apogee. A span counts as a burn only if the cleaned trace holds past 2 g for 0.5 s in all. A canopy’s or a charge’s shock or a knock doesn’t. Where none does, the first span is the burn, a short motor’s. This was measured once, when the rule was chosen, on the sample logs. Of the spans after the first, every one that started within 3 s of apogee held past 2 g for at most 0.47 s. Every longer one started more than 3 s before it, though nothing beside those logs says which of the longer ones are burns. The 0.5 s is HPR Sim’s choice. The burns are searched from liftoff’s earliest time by the heights, or from the log’s start where liftoff is withheld. A rocket at rest reads 1 g, so anything more before it moves is a knock or a bump on the pad. A charge or a canopy’s shock in a coast lies outside the burns, so it isn’t taken for the top. A second burn, an air start’s or a sustainer’s, is searched too. No top or burnout earlier than liftoff’s earliest time is given. Its bound adds the part’s own errors, the noise around the top and what a peak between samples can hide:

ba=sAa+β+p+g(1−cos20∘)+5σ+Δ+j

Here sAa is a tenth of the reading a, for the part’s scale. β is 0.05 g of bias. p is how far the pad strays from the level it witnessed. g(1−cos20∘), 0.06 g, is the share of gravity a part that zeroes on a tilted rail takes out. It is counted only where the logger took gravity out. σ is the noise, measured within 0.5 s either side of the top. Δ is a step, and j is the jump to the sample either side. On the tests’ trace, a thrust rises to a top of 6 g, 58.8 m/s². The scale alone is 5.9 m/s² of the bound, and the bias 0.49 m/s².

Clipped. A part reads no higher than its range. A top is clipped, for one of two causes the JSON’s clip keeps apart. It is range where the top reaches the part’s range as its family states it, within two steps and 0.25 g. The part read no more. Otherwise it is flat where the top holds flat for 4 samples or 0.05 s, whichever is more samples. That is how Debrief tells a part that clipped. A part that clipped at a range other than the one stated, or with none stated, writes such a top. Flat means within Debrief’s 0.3% of the top or 0.25 m/s². That is narrowed to half the part’s noise where that is less, and never below half a step. A part at its range writes one value, while a thrust’s plateau read by a quiet part still scatters by its noise. Debrief’s tolerance alone was tried when the rule was chosen, before the tuning set held a log whose tops reach its part’s range. It called 82 of its 284 tops clipped (ADR-283). A clipped top is a floor: the truth is no lower than the reading less its bound, and may be higher by any amount. The JSON marks it is_floor, its provenance clipped and its cause in clip, and gives the part’s range as range_m_s2 where the family states one. hpr analyze says “clipped at the part’s range” only for range. For flat it says the top held flat at its top, so it may have been clipped. Either way the top may have been higher.

A thrust held in part. A top is a floor too where the accelerometer holds the thrust only in part. With no end to it, the last burn never falls through zero before the accelerometer’s samples end or apogee. That happens when the log, or the accelerometer alone, stops while the motor still pushes. The top may lie past the samples. Burnout isn’t read off such a trace, and comes from the top speed instead. With no start, the trace is already past 2 g at its first sample searched, as when a log starts in the climb. The top may have come before it, as a motor’s ignition spike can. Either way it is measured, its clip null, and hpr analyze says the accelerometer holds the thrust only in part. The synthetic report counts this cause apart, as thrust_in_part; no tuning-set top reads so, as every tuning log starts on the pad.

Burnout from the accelerometer. In the climb the reading is (T−D)/m, the thrust T less the drag D over the mass m. It is above zero while the thrust beats the drag, below zero once it doesn’t, and after burnout −D/m, which eases as the coast slows. So the reading falls through zero as the thrust tails off. It is least at burnout, where the thrust’s fall ends and the drag’s easing begins. Debrief takes the fall through zero as burnout, which comes earlier, by as long as the thrust takes to tail off below the drag. The cleaned trace is first averaged over about 0.05 s, an odd number 2h+1 of whole samples. That is 27 at 500 samples a second, 7 at 100 and 3 at 20. Their first and last lie 2h intervals apart, 0.052, 0.06 and 0.10 s. That cuts its noise by the square root of the n samples averaged. With N five times the averaged noise and a step,

N=5σ/n+Δ,

an averaged sample lies within N of its truth but about once in 1.7 million samples. So the samples next to burnout read within 2N of the least. The thrust falls steeply and the coast’s drag eases slowly, so burnout lies near the first of them. The reading is that first sample, less half the window, h intervals, since the average reads the least late by that much. Its bound is the span of those samples, an interval, and the window’s 2h intervals, at least two intervals. If the averaged trace never climbs 4N back above its least before apogee, the accelerometer gives no burnout. It gives none either if the least lies within 3N of zero, too little drag to see. Nor does it after more than one burn. The tuning set’s ten designs each fly one motor in one stage, so it holds no bound for the end of a second.

Take the same trace, sampled 100 times a second with a noise σ of 0.087 m/s². The thrust falls through zero to −0.6 g at burnout, 4.1 s. The drag eases to nothing by apogee, 8 s. Averaged over seven samples, N=5×0.087/7, 0.16 m/s². The drag eases by 0.6 g over 3.9 s, 1.5 m/s² a second. So the samples within 2N of the least span about 0.22 s. With an interval and the window’s 0.06 s, the bound is about 0.3 s. The reading falls within a sample, 0.01 s, of 4.1 s.

Burnout from the top speed. A rocket gains speed only while the thrust beats the drag and its weight. So the top speed comes at or before burnout, the last burn’s end. Where the accelerometer gives no burnout, or the log has none, burnout is a floor. Its reading is the top speed’s time, and the floor is that time less its bound. Burnout came no sooner than the floor, and may have come later by any amount. A motor whose thrust tails off below the drag and the weight burns on after the rocket stops gaining speed. The bound adds three times. The first is the time a coast slowing at g takes to lose twice the speed’s 5 m/s (16 ft/s) noise. The fit can place the peak that much late. The others are half the 0.5 s fit and an interval. On the invented PerfectFlite flight, which has no accelerometer, the first is 2×5/9.81=1.02s. That, half the fit’s 0.25 s and the 0.05 s interval add to 1.32 s. So hpr analyze reads burnout no sooner than 0.88 s, the top speed’s time, 2.20 s, less that bound. The true burnout is 2.10 s, 0.10 s before the top speed’s time. The JSON gives such a burnout’s reading as time_s and its bound as bound_s, and marks it is_floor. A top speed withheld, as transonic on a fast flight, withholds it too.

The canopies

Each canopy’s descent rate comes from the filtered height track, its legs split where the main opens. The opening starts from Debrief’s walk back from landing and its marks (ADR-283 §6). HPR Sim compares rates at the ground’s density, smooths them with a running median, and since ADR-366 takes the opening at the likeliest single change of the descent, not the walk’s first stop.

The main’s opening. Near the ground the rocket hangs under its main. The descent’s rate is the least-squares slope of the filtered track over 1 s, widened where needed to hold four of the track’s steps. A fit that would need more than 10 s is refused (step_too_coarse). The main’s rate is the median of that rate over the last 2 s before landing, or the fit’s span where that is longer. The last 0.25 s is left out. Walking back from landing, the run of samples whose rate stays within 1.6 times the main’s is the main’s descent. The limit is 3 m/s (9.8 ft/s) more than the main’s rate if that is more. Its first sample is the opening. These marks are Debrief’s, set on its own logs.

That walk stops at the first rate past the limit, however short. A real descent passes it more than once: the rocket swings under its main, and a coarse log’s rates scatter. So HPR Sim asks where one change in the descent is likeliest (ADR-366). Call a rate slow if it is within the limit and fast if it is past it. Split the descent’s n rates, from apogee to landing, at s, with a slow ones among the s before the split and c among the n−s after it. If the rates were independent, slow with one chance before the split and another after, the split’s log-likelihood at the best two chances would be

L(s)=ℓ(a,s)+ℓ(c,n−s),ℓ(k,m)=klnkm+(m−k)lnm−km

Here ℓ(k,m) is for a side of m rates with k slow ones, and a term with nothing in it counts as zero. This is the change-point of a two-valued sequence (D. V. Hinkley and E. A. Hinkley, “Inference about the change-point in a sequence of binomial variables”, Biometrika 57, 1970, pp. 477–488). A side that is all slow or all fast has ℓ=0, the most it can be. A side that mixes them has less, and the more evenly it mixes, the less. The main’s descent starts at the split with the greatest L. HPR Sim adds one condition: the later side must hold the larger share of slow rates, since the main is the slower canopy. Of two splits as likely, the one nearer landing is taken.

  • Where the rates go from fast to slow once, after the slow start from rest at apogee, the split is the walk’s. A test tries every such descent of its sizes.
  • A fast stretch under the main is weighed against the slow rates before it. Moving the split past it would put those slow rates into the drogue’s leg. Which costs more depends on the two lengths and on how long the drogue’s leg is: the worked case below shows one balance.
  • The split is never later than the walk’s. Every rate after the walk’s start is slow, and adding slow rates to an earlier side that holds a fast one only lowers its ℓ.

A worked case: 100 fast rates of a drogue, then 8 slow, a bounce of 10 fast, and 30 slow to landing. The walk stops after the bounce, after 118 rates. There L=ℓ(8,118)+0=−29.25: eight slow rates sit in the drogue’s leg. After 100 rates, the first slowing, L=0+ℓ(38,48)=−24.56: ten fast rates sit in the main’s. The split is there. With a drogue’s leg of 20 rates and 4 slow ones before a bounce of 6, the balance falls the other way, and the split is after the bounce.

The rates aren’t independent, since they are a running median of overlapping fits. So the likelihood is the rule’s reason and gives no confidence of its own. The opening’s bound, under The opening’s bound below, is derived for a descent whose rates go from fast to slow once. Where they change more often, the tests here hold the split’s opening to that same bound, and nothing proves it. One shape is known to break it: a drogue falling near the limit, slow for a few seconds and then fast just before the main opens, reads its opening seconds early (issue #1593, the bound where the split is earlier than the walk).

Three tests hold the rule to the cases that showed the walk wrong (#903). Two invent a flight with a drogue at 18 m/s (59 ft/s) and a main opened 750 m (2,461 ft) up.

  • A swing. Under a main at 5 m/s (16 ft/s), 40 s after it opened, the rocket holds its height for 2.25 s and then falls 4.5 m/s (14.8 ft/s) faster than the main’s rate for 2.5 s. The walk stopped at the swing. The split reads the opening and each leg’s rate within its bound, on a smooth log at 20 Hz, one in 0.3048 m (1 ft) steps, and one at 5 Hz.
  • A coarse log. 732 logs at 5 Hz in 0.93 m (3.1 ft) steps, with noise within 1.7 m (5.6 ft) either way: the main at 4 to 7 m/s (13 to 23 ft/s) in steps of 0.05 m/s (0.16 ft/s), each under 12 draws of the noise, 61 rates by 12 draws. The split reads two canopies on all 732, the opening and each leg’s rate within its bound.
  • The real flight that showed the fault, a StratoLogger CF’s log from a public forum thread (cited in the issue, not committed; the test reads it where it is cached, and reads nothing elsewhere). Its rocket swings under the main. The issue measures the drogue’s descent at about 17.3 m/s (56.8 ft/s) and puts the opening between 87 and 92 s, the log’s header giving the main charge at 87.15 s. The walk read the opening at 108.45 s and the drogue’s descent at 14.6 m/s (48 ft/s). The split reads 90.00 s, with a bound of 0.76 s, and 17.3 m/s (57 ft/s), within 1.4 m/s (5 ft/s). The charge fired 2.85 s before the reading: the bound leaves out the time a canopy takes to come out and fill.

That is one real flight with a logged charge time (evidence row G10). A second, the public Genesis flight, has a simulation beside its log: the walk read the log as one canopy, and the split reads the main’s opening 806.18 m (2,645 ft) up, where the simulation opens it at 797.33 m (2,616 ft). No bound is stated for that height, and the simulation’s drogue rate is outside the log’s interval, so this shows the two agree roughly, no more (the residuals report). How the split’s opening compares with a logged charge time on others hasn’t been measured, and a canopy’s filling time is still outside the bound (#779). The survey’s real logs have no truth to hold a reading to. On them the rule reads a drogue’s descent on more logs than the walk did, and on fewer in no family of loggers; ADR-366 has the counts, and the survey’s report the readings as they stand.

The rates it splits are their running median over 2 s, or twice the fit where that is longer. So a spike of the fit’s noise on a long leg is taken out, while a step, the opening, stays where it is. The raw rate was tried when the rule was chosen. On 2 of the 87 tuning flights whose main’s opening their own logs give, the walk missed the drogue. A noise spike on a 125 s main leg stopped it 125 s after the opening.

At the ground’s density. A canopy falls at the speed where its drag balances the weight, v=2mg/(ρCDS) (the descent rate). So its rate goes as ρ−1/2. Up high, in thinner air, one canopy falls faster than it does near the ground. From a high apogee its early descent would read as a drogue’s. So HPR Sim scales each rate, read at the height h above the pad, to the ground’s density:

v0=vexp(−h/2H)

H is the troposphere’s least density scale height, 7,000 m (22,966 ft). In the 1976 standard atmosphere, whose temperature falls by L, 6.5 K/km, the density falls as exp(−h/H) near any height. Here H=RT/(g−RL) for air’s gas constant R and the temperature T. At the coldest the bounds allow, the tropopause’s 216.65 K less 15 K, H is 7,289 m (23,914 ft). It is taken down to 7,000 m (22,966 ft). Where the air cools with height, as through the standard troposphere, a canopy’s rate grows by no more than exp(h/2H) at the least H. So the scaling can only slow a rate read high up, never quicken it, and one canopy doesn’t read as two. A surface inversion, air that warms with height, is the exception, as L is negative there and H shrinks. At 260 K, warming by 20 K/km, H is about 4,800 m (15,748 ft). Over an inversion 300 m (984 ft) deep, a canopy’s rate then grows by 3.2% from the ground to its top. The scaling, with H at 7,000 m (22,966 ft), takes out 2.1%. So a rate read there is about 1% quicker than near the ground, which is small beside the mark’s 1.6. When the rule was chosen, comparing each rate at its own density named a main on a tuning flight that opened one canopy at apogee.

A drogue, or one canopy. A drogue came before the main only where the descent shows its leg. That needs two things. First, the 2 s just before the run, at the ground’s density, fell faster than the run’s mark. This is HPR Sim’s addition: a drogue whose rate crosses the mark leaves the run’s start inside the drogue’s leg. Second, the descent from apogee to the opening fell faster than 1.4 times the main’s rate, by its median, as Debrief asks. Otherwise the descent shows one canopy (one_canopy): the drogue’s descent is withheld, and the main’s runs from apogee to landing. So a drogue no faster than the mark reads as one canopy, whatever its ratio to the main. The main’s descent is then the whole descent’s. In the tests, a drogue at 15 m/s (49 ft/s) over a main at 10 m/s (33 ft/s) reads as one canopy. The main’s mark is 16 m/s (52 ft/s). Debrief’s ratio of 1.4 alone would tell them apart. A rocket that came down under none reads the same way.

Each leg’s rate is the filtered height it loses over the time it takes. The drogue’s leg runs from apogee to the main’s opening, and the main’s from there to landing. A leg that loses no more than 3 m (9.8 ft) or a tenth of its top’s height, Debrief’s least, is withheld (short_leg). Its bound is the error two heights of the same track can hold between them, over the leg’s time (What each bound holds).

The opening’s bound. A fit centered half a window before the opening sees only the drogue’s descent, faster than the mark. So the run starts no sooner than half a window and a sample before the opening. After the opening the main slows the fall as quadratic drag slows any falling body (Taylor, Classical Mechanics, §2.4),

dvdt=−g(v2vm2−1),

and from any speed it reaches the mark v1 within

τ=vm2glnv1+vmv1−vm,

where vm is the main’s rate at the opening’s height, the rate near the ground grown by exp(h/2H). So the run starts no later than τ, half a window and a sample after the opening. The bound is τ, half the 1 s fit and two sample intervals. One interval is for the sample the run starts on, and one more is HPR Sim’s margin. A canopy’s filling time isn’t in it: the synthetic flights open each canopy at once (#779, the opening’s bound and a canopy’s filling time).

On the invented PerfectFlite flight, the main’s rate near the ground is 6.0 m/s (20 ft/s), so the run’s mark is 9.6 m/s (31.5 ft/s). That is 1.6 times the rate, more than 3 m/s (9.8 ft/s) above it. The run back from landing starts at 21.40 s, 148.4 m (487 ft) up. There the main’s rate is 6.06 m/s (19.9 ft/s), the rate near the ground grown by exp(148.4/14,000). From there the main slows any fall to the mark within

τ=6.0619.6ln15.663.54=0.46s,

so the opening’s bound is 0.46+0.50+0.10=1.06s, as hpr analyze prints it. The main truly opened at 21.14 s, 150 m (492 ft) up. The reading is 0.26 s late, as the 1 s fit still sees some of the drogue’s faster fall just after the opening. The drogue’s leg loses 241.7 m (793 ft) in 11.125 s, 21.7 m/s (71 ft/s). That is below the drogue’s true 25 m/s (82 ft/s), as the leg starts at apogee, where the rocket falls from rest. Over the same span the true rate is 21.74 m/s (71.3 ft/s).

Bounds and refusals

What each bound holds

Each reading states the bound its error stays within. A bound is a sum of terms, each the size of one error the synthetic flights draw, or measured on the log itself. Most of an apogee’s bound is the day’s air.

  • Apogee height. The day’s air and the barometer’s gain as shares of the height h. The noise, the pad’s error and the drift since the log began. Last, a step at the pad and one at the peak, and the median’s peak bound:

    bh=(sT+sG)h+n+epad+dta+Δ0+Δa+bmed
  • Apogee time. The time the top of a coast takes to fall by the noise and a step, 2(n+Δa)/g. Plus the longer interval either side of the peak’s flat run.

  • Liftoff. The time a rocket at A=2g takes to climb the allowance and the pad’s error, once above the pad and once below, 4(a+epad)/A. Plus the longer interval either side of the sample, twice.

  • Top speed. The day’s air and the gain as shares of the speed v. The port’s error (the port is the hole that lets the outside air in to the barometer) while the rocket accelerates at a1, 0.02va1/g. The fit cutting the peak, (a1+a2)w/4. Here a1 and a2 are the climb’s change of speed before and after it, measured on the fits. Last, the noise in a speed.

  • Landing. The time the descent takes to fall the allowance, the noise, the pad’s error, the logger’s height above the ground, a step and the drift. It takes the descent at its rate near the ground, taken slow by a step and twice the noise. Plus the longer interval either side of the sample, twice.

  • Flight time. Liftoff’s bound and landing’s, added.

  • Mean descent rate. The apogee’s height bound and the height landing is read early by, then the rate times the apogee’s and landing’s time bounds. All of it is over the descent time.

  • Top acceleration. The part’s scale as a tenth of the reading, its bias and how far the pad strays. A tilted rail’s share for a part that took gravity out. Last, five times the noise, a step and the jump to a neighboring sample (The accelerometer). A clipped top’s bound is one-sided: the truth is no lower than the reading less it.

  • Burnout. From the accelerometer, the span of the averaged samples within 2N of the least. Add an interval and the averaging window, or two intervals if that is more. From the top speed, a floor: 2×5m/s/g, half the fit and an interval.

  • Each canopy’s rate. The day’s air and the barometer’s gain as shares of the height the leg loses. The noise at either end, the drift over the leg and a step at either end. For a leg from apogee, add the median’s peak bound. All of it is over the leg’s time.

  • The main’s opening. The time the main takes to slow the fall to the run’s mark, half the 1 s fit and two intervals (The canopies).

termsizesource
sT, the day’s air6.92%a height read by the standard atmosphere is off by ΔT/T of itself over the layer it spans (the hypsometric equation); ΔT is the synthetic flights’ 15 K range of air (ADR-280 §2), and T no colder than the standard atmosphere’s 216.65 K below 20 km (12 mi). Near the ground it is about 0.35% a kelvin
sG, the barometer’s gain0.25%an MS5607 barometer chip’s ±1.0 mbar over the 400 mbar span its datasheet states it on, the widest of the synthetic flights’ parts; reading that error as a gain is an inference (ADR-280 §3)
n, the noise3 m (9.8 ft)the fast and slow correlated noise, the part’s white noise, the pad’s median and the port’s error at apogee’s sideways speed, added and tuned on the tuning set
d, the drift0.015 m/s (0.049 ft/s)the synthetic flights’ fastest drift, 3 Pa in 30 s (ADR-280 §3), at the density of air 5 km (3.1 mi) up, 0.74 kg/m³
Δ, a stepthe track’s stepthe rounding of the logger’s numbers; a pressure’s step at the height it is read at
bmed, the median’s peak bound0.077 m (0.25 ft) at 20 samples a secondgravity’s bend over the median’s reach (The running median)
epad, the pad’s errorup to 2 m (6.6 ft)how far up the rocket may be when the log starts (The pad and the climb)
a, the allowanceat least 2 m (6.6 ft), and a stepfour times the fast correlated noise, HPR Sim’s choice
the logger above the ground3 m (9.8 ft)a tall rocket’s electronics bay on a rail, HPR Sim’s choice: it lands lower than it started by as much
the port, while acceleratinga pressure coefficient of 0.02the ogive-cylinder’s measured −0.02 one caliber behind its shoulder below Mach 0.9 (the Mach error note §1–2), the largest the synthetic flights draw
w, the fit’s width0.5 sHPR Sim’s choice; a fit cuts a peak by about (a1+a2)w/8, and twice that is allowed
the noise in a speed5 m/s (16 ft/s)tuned on the tuning set
sA, the accelerometer’s scale10%the ADXL375 accelerometer chip’s ±10% sensitivity (its datasheet, p. 3, in the sensor noise note, read once and not re-read against the sheet), the widest of the synthetic flights’ parts (the acceleration)
β, the accelerometer’s bias0.05 gthe bias the synthetic flights leave after the pad’s zero, the top of the 0.01 to 0.05 g the sensor noise note calls a guess, as no source states one
a tilted rail0.06 g, g(1−cos20∘)the 20° from vertical the NAR’s and Tripoli’s high-power safety codes allow a rail (the NAR’s model rocket code allows 30°, 0.13 g): a part that zeroes on the pad takes out gcosθ
the accelerometer’s noise5σmeasured on the log, within 0.5 s of the top, or from the top to a second past the fall through zero for burnout; a normal noise passes it, either way, about once in 1.7 million samples
gravity’s witnesses0.25 ga 20° rail’s 0.06 g, a part’s zero offset and the small drag near apogee, HPR Sim’s choice
the descent’s fit1 s, widened to hold four steps, at most 10 sa canopy’s rate changes slowly, so a fit twice the climb’s averages the noise down; four steps keep a step from reading as a spike, HPR Sim’s choice
H, the density scale height7,000 m (22,966 ft)the 1976 standard atmosphere’s 7,289 m (23,914 ft) at its coldest less 15 K, taken down (The canopies)

The two tuned terms are the 3 m (9.8 ft) and the 5 m/s (16 ft/s). They were set on the same flights the bounds are held on (Checked against).

A worked example

The invented PerfectFlite log is the one the command line’s own example reads (hpr analyze prints it). It reads an apogee of 390.144 m (1,280 ft), 10.275 s into the log, on a track in one-foot steps sampled 20 times a second. Its filtered track stays within 2.30 m (7.6 ft) of its start for 0.80 s. That is longer than the 0.69 s a rocket at rest would need to climb the 4.61 m (15.1 ft) rise at 2 g. So the pad’s error is zero. The bound adds up so:

termworked out, in m, m/s or sbound term, m
the day’s air0.06924 × 390.144 m27.01
the barometer’s gain0.0025 × 390.144 m0.98
the noise3.00
the pad’s error0.00
the drift0.015 m/s × 10.275 s0.15
a step at the pad, and one at the peak2 × 0.3048 m0.61
the median’s peak bound0.08
the bound31.83

hpr analyze prints it rounded up, as it prints every bound: “within 31.9 m (105 ft)”. The flight’s true apogee is 390.31 m (1,280.5 ft), so the reading is 0.17 m (0.56 ft) low. That is about a two-hundredth of its bound, as it should be on a log written from the true heights with no noise. About 85% of the bound is the day’s air, which a real log can’t rule out.

The 90% intervals

A bound is the size every error stayed within on the tuning set. A 90% interval is narrower: it says where the truth usually lies, over the days and loggers the synthetic flights draw. Twelve readings state one. hpr analyze prints each after its bound, rounded outward, such as a 90% interval of 369.8 to 410.3 m (1,213 to 1,346 ft), and a floor’s lower end as “95% lower end”. hpr analyze --json and Python’s readings() carry each as low and high, under its reading’s key: liftoff.interval_s, apogee.height_interval_m, apogee.time_interval_s, max_speed.interval_m_s, landing.interval_s, flight_time.interval_s, landing.mean_descent_rate_interval_m_s, max_acceleration.interval_m_s2, burnout.interval_s, drogue_descent.interval_m_s, main_descent.interval_m_s and main_descent.opening.interval_s. A floor has no interval: a clipped top acceleration states lower_end_m_s2 and a burnout floor lower_end_s, as A floor’s lower end sets out; an apogee or top speed floor states neither.

An interval need not hold its reading. Liftoff is read late, so its interval lies before the reading: the worked example’s liftoff reads 0.80 s, with an interval of 0.41 to 0.64 s.

Each interval comes from the reading’s own bound. The synthetic flights’ calibration set is 2,000 flights on a seed no other set shares (the sets). Each of its flights’ own logs was read; 362 of them are the GNSS-only tracker’s, which give none of these readings, so 1,638 are left for the height track’s. The top speed is also withheld at or above 227.7 m/s (747 ft/s), where the barometer’s port may read the transonic flow over the nose (Each reading), which leaves 576 top speeds. The rest have fewer for reasons of their own: 641 logs hold an accelerometer, 564 of them read a burnout that isn’t a floor, and 781 flights opened a drogue, so 781 give its rate and the main’s opening. Each reading’s error was then taken as a share of its bound: the reading r less the truth, over the bound b. The interval runs from the reading less the 95% point of those shares times the bound, to the reading less the 5% point times the bound:

[r−q95b,r−q05b]

Each point is the ⌊0.05(n+1)⌋-th from its end of the n sorted shares, the split-conformal rank (Lei and colleagues, 2018): the 81st of 1,638, the 28th of the 576 top speeds. Each tail then misses about 5% on average over flights drawn as those are.

reading5% point q05, share of the bound95% point q95, share of the bound
liftoff0.2030.482
apogee height−0.6310.637
apogee time−0.3310.227
top speed−0.2180.265
landing−0.3350.156
flight time−0.379−0.132
descent rate−0.4750.495
top acceleration−0.4950.568
burnout−0.0650.0689
drogue rate−0.6220.612
main opening0.1380.567
main rate−0.5740.586

Some points lean. Liftoff is the last sample within the pad’s allowance, which the rocket leaves only after climbing it, so it is read late, and both its points are above zero: its interval lies wholly before the reading. The flight time runs from that late liftoff to a landing read early, so it reads short, and its interval lies wholly after the reading. The apogee time’s errors reach further early than late, so its interval reaches further after the reading than before it. The main’s opening is read after the canopy has slowed the descent, so it is read late, and its interval lies wholly before the reading. cargo xtask synthetic fits the points again and fails unless they equal the library’s.

A worked example. The invented PerfectFlite log’s apogee of 390.144 m (1,280 ft) has a bound of 31.83 m (104.4 ft) (above). Its interval runs from 390.144 − 0.637 × 31.83 = 369.87 m to 390.144 + 0.631 × 31.83 = 410.23 m (1,213 to 1,346 ft). The true apogee, 390.31 m (1,280.5 ft), lies inside. Most of the width is the day’s air, as most of the bound is. The same log’s flight time reads 45.00 s with a bound of 2.91 s, so its interval runs from 45.00 + 0.132 × 2.91 = 45.38 s to 45.00 + 0.379 × 2.91 = 46.10 s. It doesn’t hold the reading: the flight time reads short, so the interval says the flight most likely lasted a little longer.

Checked against sets A and B. The two test sets, which the points never saw, each pass the test of ADR-257, calibrated intervals (the report). The truths below an interval and those above it are each counted, so an interval that misses on one side only fails. Each count gets its 99.9% Clopper-Pearson band, the range of shares that count rules in. The test fails unless both sides’ bands hold 5% and the inside count’s holds 90%:

setreadingreadings comparedtruths belowtruths insidetruths above
Aliftoff8203572956
Aapogee height8205572045
Aapogee time8203474640
Atop speed2861225816
Alanding8203874537
Aflight time8204074535
Adescent rate8204772845
Bliftoff8455273954
Bapogee height8453875948
Bapogee time8453576446
Btop speed286926215
Blanding8452478833
Bflight time8453276647
Bdescent rate8454075550
Atop acceleration3561232222
Aburnout3071028611
Adrogue rate4283237917
Amain opening4282238323
Amain rate8204473739
Btop acceleration3681833020
Bburnout3252428120
Bdrogue rate4362539219
Bmain opening4361439230
Bmain rate8454176044

The GNSS-only logs, 180 of A’s and 155 of B’s, give none of these readings, and the transonic mark withholds 534 more top speeds on A and 559 on B. So the top speed is tested on fewer than 300 readings a set, where a tail passes with a wider share of misses either way: 4 to 28 of 286 (1.4% to 9.8%), against 22 to 63 of 820. Every interval narrowed to 0.8 of its width fails the test, and so does every one widened to 1.25 or shifted down one standard error, the top speed’s included. Past the mark, though, the rule lets through only the top speeds that came out low. The two Calisto designs mostly fly past it, and of the 17 top speeds they gave on A and B, 8 left the truth above the interval (the report’s By design).

Six of the readings have a flattering side: a reading too low makes the flight look safer than it was. A low apogee height flatters a waiver, and a low top speed a fin’s margin. The top speed is vertical; a fin’s airspeed is higher on an angled climb, by 6% on a 20° one. A low top acceleration flatters the load the airframe took. A low mean descent rate, or a low rate under either canopy, says the descent was gentler than it was. The mean rate runs from apogee to landing, drogue and main together, so it is no landing speed: that is the main’s rate. The report gives each one’s mean error, the reading less the truth, with its sign and at the figures of its largest error, as its later digits differ from one computer to another. The apogee height’s is +1 m (3 ft) on A and, on B, less than half a meter and not below zero, against errors up to 237 m (778 ft). The top speed’s is +1.2 m/s (3.9 ft/s) on A and +1.3 m/s (4.3 ft/s) on B, against errors up to 26.8 m/s (88 ft/s). Neither is low on average. The descent rate’s is −0.01 m/s (−0.033 ft/s) on both, a hundredth of a meter per second low against errors up to 6.81 m/s (22.3 ft/s). The other three are as small: under 0.5 m/s² (1.6 ft/s²) on the top acceleration, against errors up to 15.7 m/s² (51.5 ft/s²), and 0.02 m/s (0.066 ft/s) or less on each canopy’s rate, against errors up to 1.51 m/s (4.95 ft/s) on the drogue’s and 5.87 m/s (19.3 ft/s) on the main’s.

The test pools the five loggers that give these readings and the ten designs. The 90% is over the days the flights draw, spread evenly within 15 K of the standard atmosphere, and the misses aren’t spread over those days. A barometric height is off by about the day’s offset from standard over its temperature, and so are the speed and the rates taken from it. On days within 10 K of standard, no height and no rate of descent missed, and 4 of 360 top speeds did. On days warmer than that, above about 25 °C (77 °F) at a sea-level field, a warm day reads low, so the truth lay above the interval: for the height, 45 of 137 on A and 48 of 156 on B (#1055); for the descent rate, 45 of 137 and 50 of 156; for the main’s rate, 39 of 137 and 44 of 156; for the drogue’s, 17 of 69 and 19 of 80; for the top speed, 16 of 57 and 15 of 69 (#1061). On a hot day, then, expect each of them above its interval’s top far more often than 1 time in 20. So for a waiver, take the apogee plus its bound; for a fin’s margin, the top speed plus its bound; for a landing, the main’s rate plus its bound; for a load, the top acceleration plus its bound. None of these holds if the reading is a floor, which has no upper figure. Each bound holds the day’s air to 15 K. The bounds, too, were measured only on synthetic flights.

The times lean by logger. On the AltOS logs, whose rate drops from 100 to 10 samples a second at apogee, 31 of 185 apogee times on A and 30 of 186 on B lie later than the interval (#1056). Liftoff on the PerfectFlite and EggTimer logs leaves its truth above the interval (18 of 144 and 23 of 181 on A), and on the AltOS and 500 Hz IMU logs below it. On the 500 Hz IMU logs, landing and the flight time leave theirs above it, 18 of 171 and 24 of 171 on A (#1062). The top acceleration leans most. Its one pair of points serves two accelerometers: on the AltOS logs 34 of 185 truths on A and 38 of 186 on B lie outside the interval, 22 and 20 above it, the flattering side, and on the 500 Hz IMU’s none of 171 and 182 do (#1070). The report’s bold rows mark each group whose counts alone fail. The report gives each logger’s and each band’s counts. A day more than 15 K from standard, or a logger noisier than these, can see more misses still. No real flight has checked an interval (evidence row G34).

A floor’s lower end

A floor’s truth may lie any amount above it, so it can’t state an interval. It states a lower end instead, the figure the truth stayed above in about 19 of 20 synthetic flights: the reading less its own 95% point times its bound, r−q95b. That end is tested alone. The band on the share of truths below it must hold 5%, the lower tail of the same test. As that test is two-sided, too few truths below fail it as too many do. The probes are deliberately wrong lower ends: raised by one standard error, or lowered by two, each must fail. The standard error here sizes those probes. It is the 95% point less the median, times the bound, over 1.645, as an interval’s is its width over twice 1.645. Lowering by one isn’t enough to fail on a few hundred floors: of 211, a band still holds 5% with 2 truths below.

floorfloors fitted95% point q95, share of the boundmedian, share of the boundwhere the lower end falls
burnout, from the top speed’s peak437−0.0335−0.194just after the reading
top acceleration, clipped3480.259−0.124just below the reading
setfloorfloors comparedtruths belowtruths below, raised one standard errortruths below, lowered two
Atop acceleration1813520
Aburnout21111550
Btop acceleration1873440
Bburnout2137481

Every floor as stated passes, and every raised or lowered one fails.

Why a clipped top is fitted near the clip. No synthetic flight’s accelerometer reaches its own range, so the top acceleration’s floors come from each flight’s log written again with the accelerometer clipping at 8 g. The AltOS part has no such setting; its log clips there too, to reach a count the test can fail on, and its reader tells the clip by its flat top.

How far past a clip a top goes depends on the design, and no log says. A first fit, on tops as flown, found most of them far past it, so its 95% point came out negative and put the lower end above the reading. It held on A (8 of 252 below) and failed on B (31 of 285): one of the ten synthetic designs, the Bella Lui, whose tops only just pass 8 g under B’s model, left 21 of 22 truths below it.

So the fit takes the one place where the log’s own error can put a truth below a clipped reading: a top under the clip that reads as clipped. A part’s scale error, its gain, can be off by up to s, ±3% on the 500 Hz IMU and ±10% on the AltOS part. The reader calls a top clipped from 0.25 g under the range, and the part’s bias of up to 0.05 g helps it there, so a truth from (1−0.3/8)/(1+s) of the clip can be read as clipped: 0.934 of it on the IMU and 0.875 on the AltOS part. Past 1/(1−s) of the clip, 1.03 and 1.11, every such part clips.

Each flight’s true top is moved to a point of its own in that window, a different point on the calibration set and on each of A and B. A top that then reads whole isn’t a floor and isn’t counted; the report gives how many. Of the rest, only a top under the clip that the gain carried up to it can miss. A top truly past the clip lies above the reading: on the tops as flown, none of A’s 252 and none of B’s 285 lie below the lower end. That check is close to certain by construction, so the test on the moved tops is what the result rests on.

So read the lower end as how far the part’s own gain can carry a reading above the truth, not as how far past the clip the top went. The result rests on the window as well as on the flights. The window’s edges come from the parts’ stated scale errors and the reader’s own threshold, but they were set after A and B had been read, so these two sets don’t hold it apart as they do the other points (ADR-308). Probes not in the report show how far it moves: with both parts’ window at 1.00 to 1.10 of the clip, no top lies under it and the 95% point goes to −0.0496, a lower end above the reading that the test no longer supports; at 0.90 to 1.10 it is 0.236 against this window’s 0.259. On A and B the lower end misses less than its 5%, 3 of 181 and 3 of 187, all on the AltOS logs: an error to the safe side, and the test’s band still holds 5%. It is stated only on a clipped top: a floor whose burn the log ends in has none, as an apogee floor has none, since no set reads one to test (#1071).

A burnout floor’s lower end lies a little after the reading, by 0.0335 of the bound, as burnout usually comes after the top speed. How far after depends on the motor’s tail-off, so it can lean by design: on the Bella Lui, 8 of 61 truths lie below it on A (#1070).

A worked example. With an invented clipped top reading 78.5 m/s² (8.00 g) and a bound of 9.00 m/s², the lower end is 78.5 − 0.259 × 9.00 = 76.2 m/s² (7.77 g). The floor, the reading less the whole bound, is 69.5 m/s² (7.09 g). The true top is most likely above the first and, on what the tuning set held, no lower than the second. Both are lower limits: neither sizes a load, as the true top may be much higher than either. Use the reading plus its bound only for a top that isn’t a floor (is_floor false).

  • J. Lei, M. G’Sell, A. Rinaldo, R. J. Tibshirani and L. Wasserman, “Distribution-free predictive inference for regression”, Journal of the American Statistical Association 113(523), 2018, §2.2, for the split-conformal rank that sets each interval’s points. V. Vovk, A. Gammerman and G. Shafer, Algorithmic Learning in a Random World, Springer, 2005, ch. 2, for its origin.

A day past 15 K

Each bound holds the day’s air to 15 K from standard. This section covers what the readings do on a day further off. It is a note beside the readings, not a correction of them, and it assumes the offset measured at the pad holds up the whole climb. It is tested only on invented flights in air offset from standard by the same amount at every height (Ussa76::with_offset), not on a real log.

Standard air is 15 °C at sea level, less 6.5 °C a kilometer of height. So 15 K over standard is reached at these temperatures at the pad:

Pad’s height above sea levelStandard there15 K over standard
sea level15 °C (59 °F)30 °C (86 °F)
1,000 m (3,281 ft)8.5 °C (47.3 °F)23.5 °C (74.3 °F)
1,500 m (4,921 ft)5.25 °C (41.45 °F)20.25 °C (68.45 °F)

Past that, the bound may not hold. A barometric height then can read low by more than the bound allows for the day’s air, and so can a speed or a rate of descent worked out from the heights. That is the flattering side: an apogee too low for a waiver, a descent too slow. Where the log records the pad’s temperature and its height above sea level, the readings check the day against the 15 K (#1259).

What it compares. The day’s offset ΔT is the log’s temperature at the pad less the 1976 standard atmosphere’s at the pad’s height above sea level. The standard’s temperature falls 6.5 K a kilometer from 288.15 K at sea level, to 216.65 K at 11 km (6.8 mi), and holds that to 20 km (12.4 mi). The code holds it above 20 km (12.4 mi) too, where the standard warms again, which errs toward a larger share:

ΔT=Tlog−Tstd(zpad)
Tstd(z)=288.15K−0.0065K/m⋅z

Tlog is the median of the log’s temperature samples over the span the pad is read from. The pad’s height above sea level, zpad, is the first of three. A pressure channel over the same span, turned into a height by the standard atmosphere, gives the pad’s pressure altitude, the height whose standard temperature the logger’s conversion assumes. Next comes an altitude channel above sea level, then the ground elevation the logger states.

When it notes. When |ΔT| is more than 15 K. hpr analyze then prints a line above the readings with the log’s temperature, the offset and where the pad’s height came from. Below the apogee, the top speed, landing’s mean rate and each canopy’s rate, it prints a note with the size. On a low flight a day just past 15 K can move the readings by less than the bound’s own 6.92%, as the size below shows. The note then says the day is past 15 K but within the bound’s share at this flight’s height. Its JSON carries the comparison as day_air whenever the log has one, with offset_k, covered and share.

The size. By the hypsometric equation, dz=−(RT/g)dp/p, the height between two pressures goes as the air’s mean temperature between them. So a height read by the standard atmosphere is off by ΔT/T¯ of itself, where T¯ is the standard’s mean over the layer (the height formulas note). The note takes the standard’s temperature at the apogee’s height above sea level, ztop, the pad’s plus the apogee’s reading: the coldest air the climb spans. So its share is the largest the layer allows. Where the apogee is only a floor, as when the log ends on the climb, the top may lie higher, so the note takes the standard’s coldest, 216.65 K, instead:

sday=|ΔT|Tstd(ztop)

On a hot day the true value may lie up to sday of the reading above it; on a cold day, below. The bound’s own term is sT, 15/216.65 = 6.924%. A speed or a rate is a height over a time, so it takes the same share.

A worked example, with invented numbers. The pad’s pressure altitude is 1,190 m (3,904 ft) and the log reads 308.15 K (35 °C) there. The standard there is 288.15 − 0.0065 × 1,190 = 280.415 K, so ΔT is 27.735 K, past the 15 K. The apogee reads 2,738 m (8,983 ft) above the pad, 3,928 m (12,887 ft) above sea level, where the standard is 262.618 K. So sday is 27.735/262.618 = 10.561%, 289.2 m (949 ft) of the reading. The bound allows 6.924% of it for the day’s air, 189.6 m (622 ft). The difference, 3.637% of the reading, is 99.6 m (327 ft): the day’s air alone may put the true apogee that far past the day’s-air part of the bound.

A note, not a wider bound. A logger’s temperature channel is often the board’s or the sensor chip’s, in the electronics bay. The sun and the electronics warm it, so on a hot pad it can read above the air. Widening every bound to it would move the bounds, and the 90% intervals calibrated on the synthetic flights, by a figure that may be the bay’s. So the bounds stay as they are, and the note gives the size for the flyer to weigh against the field’s own weather reading.

Its limits.

  • A log with no temperature, or with no pressure, altitude above sea level or stated ground elevation, gets no check. Nothing then says whether the day was within 15 K.
  • A temperature channel whose samples are all one value over the whole log, and that value a unit’s zero read as kelvin within 0.01 K, isn’t taken for the air: 0 K, 273.15 K (0 °C) or 255.372 K (0 °F, (0−32)×5/9+273.15). A logger with no sensor writes zeros, and the readers turn them into kelvin first, so zeros in °C and °F read as plausible air. A channel of one other value is read: a logger quantized to a degree can read one value over a short log. A real sensor that reads exactly 0 °C or 0 °F throughout is refused too, and gets no check. A temperature outside −100 °C to 100 °C isn’t taken for the air either.
  • A stated ground elevation may be the field’s surveyed height, not its pressure altitude. The two differ by up to a few hundred meters with the weather, a kelvin or two of Tstd. An altitude channel above sea level has the same catch. A logger may set it by an altimeter setting, the day’s sea-level pressure, so it reads nearer the true height than the pressure altitude.
  • The offset is taken at the pad and held up the climb. Where the true offset shrinks with height, the note overstates; where it grows, it understates. A hot afternoon’s air often cools faster than the standard’s 6.5 K/km. It tends toward the dry adiabatic lapse rate, g/cp≈ 9.8 K/km, found in any meteorology text (References). So a hot pad’s note leans high. Under a morning inversion, where the air near the ground is cooler than the air above it, a hot pad’s offset grows with height and the note leans low. On a cold pad the two swap: an inversion makes the note lean high, and air cooling faster than standard makes it lean low.
  • hpr_flightdata::readings::day_air pins it. On an invented flight from a 1,190 m (3,904 ft) field, in air 30 K over standard at every height, the note fires and its share matches the rule. The true apogee lies past the reading plus its bound, and within the reading plus the note’s share. The same flight 30 K under standard notes the other way, and 10 K over is covered. An offset of exactly 15 K is covered, and 15.01 K isn’t. A log ending on the climb takes its share at 216.65 K. Columns of one value, zeros in kelvin, °C or °F, give no check.

Before any reading

A log passes four checks before any reading is taken from it. Each keeps the readings inside the range the bounds were checked over, or catches a track that isn’t what it says:

  • It is sampled no slower than every 2 s and no faster than every 2 ms, the slowest and fastest rates on the tuning set. Those are half a sample a second (sampled_too_slowly) and 500 a second (sampled_too_fast). A clock a part in a thousand faster than 500 a second still reads, as a logger’s rounding of its own clock may make it. One sample log, a Raven’s CSV at 800 samples a second, is withheld as too fast. The check is on the median interval, so a short burst of faster samples can still sit in a log that passes.
  • Its step is no coarser than 10 m (33 ft), or 100 Pa for a pressure (step_too_coarse), the coarsest checked on the tuning set. The coarsest the readers meet are a MARSA’s 8.8 m (29 ft) heights and an ARTS2’s 92 Pa raw counts.
  • Its clock increases (clock_out_of_order). A reader keeps a channel’s clock as the file writes it. That clock can repeat or run back, as a GNSS receiver’s fixes or a team’s merged table do.
  • Its climb slows as a coast does. A coast loses gt2/2 in the last t before its top, 78.5 m (257 ft) in 4 s, gravity slowing it and drag only adding. A track must lose more than twice its noise and step in the 4 s before its top, or it is withheld (slower_than_free_coast). The threshold is 6.6 m (22 ft) on a PerfectFlite. It isn’t a coast’s 78.5 m (257 ft), since a flight whose motor burns into those 4 s loses less.

A burst of samples. The rate check above is on the median interval, so a hostile or damaged file can hold a short burst of faster samples. Without a limit, a fit would walk the whole burst for each sample in it, and the time would grow as the square of the burst. So each least-squares slope reads at most twice its window’s steps at 500 samples a second, plus one: 501 samples for the 0.5 s top-speed fit, 1,001 for the descent rates’ usual 1 s fit, and 10,001 at the descent fit’s 10 s limit. A window holding more is read at every n-th sample, n the smallest whole number that brings the count within the cap. A log sampled no faster than 500 a second is read whole. The descent rates’ running median sizes its window from the descent’s median interval, but never from one shorter than the rate check allows. The landing’s search moves forward only.

Tests pin each of these. They count the samples a burst’s window is read at, and check the median’s window at a burst’s interval. A descent with a burst in its main leg still reads its drogue’s leg. Landing over a burst of 600,000 samples that dips to the pad and rises past the landing ceiling is held under 5 s; a walk over the burst for each sample took 47 s there once, on a fast desktop. The fits read at every n-th sample weren’t checked against the bounds, as the tuning set holds no log with a burst.

A disturbed top. A top that isn’t the coast’s is withheld as disturbed_top, by either of two checks. Which point is the apogee can’t be told, so the apogee, the top speed and landing are withheld; liftoff is still given.

  • A fall before the top. A coast doesn’t fall and rise again. Take a track that falls more than twice its noise and step below a peak in the 4 s before its top. If it then climbs to that top, it holds a top that a disturbance after apogee made. An example is a deployment’s pressure bump that lasts longer than the 0.3 s median takes out. A top such a bump holds after apogee can put the apogee’s time past its bound.
  • A spike at the top. A coast slows into its top at gravity and drag. Near the top drag is smaller than gravity. So it climbs no more than gΔ2 in its last Δ, as a body slowing at twice gravity would. And nothing falls from its top faster than gravity pulls it. A track holds a spike, such as a charge’s, if two things happen near its top. Near means within a second either side, or two samples on a slower track. It climbs into the top by more than gΔ2 in its last Δ. And it falls from it faster than a body at rest there can. Each exceeds its limit by more than twice the track’s noise and step. The fall allows for the filtered top trailing the true one. That allowance is the speed a fall from rest gathers within the median’s window and twice the noise and step. Then comes gravity’s fall from there.

The ballistic coefficient is the mass over the drag coefficient times the frontal area. Drag’s upward share stays under gravity over a last second for any ballistic coefficient over 150 kg/m², even at 120 m/s (394 ft/s) across. A rocket 29 mm (1.1 in) across, of 0.3 kg, has about 800. That covers a 1 s span. On a slower track the span is two samples, up to 4 s at half a sample a second. There the argument holds only over about 380 kg/m² on a vertical coast, and about 700 kg/m² at 120 m/s (394 ft/s) across. Below that a coast can climb into its top faster than the check allows. So a good top that a charge then drops fast may be withheld. The check only withholds, so it never reads such a top wrong.

On a PerfectFlite logging 20 times a second, twice the noise and step is 6.6 m (22 ft). So a coast climbs at most 16.42 m (54 ft) in its last second. In the tests, a step 20 m (66 ft) above a coast’s top is 0.6 s wide. It falls 29.0 m (95 ft) from the filtered top in one sample. A body at rest there falls at most 20.9 m (69 ft), so the track is withheld.

Either half of a spike alone is read, as a good top makes each. A parachute open before apogee slows the climb into the top harder than a coast does, but not the fall from it. The tests’ chute, opened at 40 m/s (131 ft/s), climbs 20 m (66 ft) in its last second, then falls at gravity. A charge that raises the bay’s pressure at the top drops the track fast, but only after the coast’s own climb. That climb is 50 m (164 ft) for half a second in the tests. Both read their apogee within its bound.

Two public logs of a charge’s transient show both outcomes (#702, a transient read as the peak; Public logs of faulty flights). A TeleMetrum log from a forum thread, whose raw peak was one sample of a pressure spike after the charge, now reads the coast’s peak. The median takes the one sample out. An MIT-licensed Blue Raven log of a breakup is withheld by the spike check. The tests rebuild both shapes with invented numbers.

Another check guards a height that is above the pad only by a word. A column mapping can call a column an altitude above the ground, and a column’s header can say so, but neither proves it. Such a track must start within 19.9 m (65 ft) of its zero (starts_off_the_pad). A logger zeroes on the pad, and its zero moves with the weather and the noise before launch. The 19.9 m (65 ft) is an hour on the pad plus the 3 m (9.8 ft) noise. The hour’s rate is the National Weather Service’s for pressure rising rapidly or falling rapidly, 0.06 inHg (203.2 Pa) an hour. At sea level’s density, 1.225 kg/m³, an hour’s 203.2 Pa is 16.9 m (55 ft) of height, and the noise makes up the rest. The hour is HPR Sim’s choice.

That check closes the first half of #646: an altitude above sea level mapped as one above the ground. Its second half, a clock in milliseconds read as seconds, is caught by the checks above. The samples come a thousand times too far apart, slower than every 2 s at any logger rate below 500 a second. And a track on such a clock loses less than a tenth of a millimeter in the 4 s before its top. A coast loses 78.5 m (257 ft) in the same time. A test holds both halves both ways: the wrong unit withheld by name, the same flight in its right unit read.

When a reading is withheld

A withheld reading has a code and a sentence with the log’s own numbers. Only the readings that need what is missing go. A log that starts in the air on the logger’s own altitude loses its liftoff, and with it the flight time. It keeps its apogee, top speed and landing. There are 28 reasons:

codewhenreadings withheld
too_shortthe track has fewer than 3 samples; or the accelerometer has fewer than 3 before apogeeall; or the top acceleration, and burnout from the accelerometer
no_climbthe filtered track never climbs twice the allowance and a step above its start, or 3 m (9.8 ft) above the pad, or the allowance and a step if that is higher; or, for landing, falls too little past its step and noise to time the descentall, or landing
starts_off_the_pada track above the pad only by a mapping’s or a column’s word starts more than 19.9 m (65 ft) from its zeroall
short_padthe first samples can’t rule out a rocket already more than 2 m (6.6 ft) upliftoff, and a top speed at the log’s start, on a track with a zero, the logger’s own or a word’s; all on any other track
ends_before_landingthe log ends before the track is back within the pad’s allowance and stays down for a secondlanding, the flight time and each canopy’s descent
faster_than_free_fallthe track comes down to the landing sample sooner after apogee than a fall from rest in vacuum could lose that height, 2h/g, allowing a step in h and a sample for the apogee’s timelanding, the flight time and each canopy’s descent
implausible_speedthe top speed is above 4,000 m/s (13,123 ft/s), Debrief’s ceilingthe top speed, and burnout from its time
noisy_speedfrom liftoff to apogee the speed swings below zero by more than 20% of its topthe top speed, and burnout from its time
transonicthe top speed reaches 227.7 m/s (747 ft/s)the top speed, and burnout from its time
sampled_too_slowlythe samples come further apart than 2 s; or, for the top speed, than 0.125 s; or, for the canopies, too far apart for 5 samples in the descent’s fit, every 0.25 s for a 1 s fitall; or the top speed and burnout from its time; or both canopies’ descents
sampled_too_fastthe samples come more often than every 2 ms, 500 a second, the fastest rate the bounds were checked atall, where the heights come that fast; where the accelerometer does, the top acceleration and burnout from it
step_too_coarsethe step is coarser than 10 m (33 ft) or 100 Pa; or so coarse for the descent near the ground that a fit holding four steps would span more than 10 sall; or both canopies’ descents
slower_than_free_coastthe track loses no more than twice its noise and step in the 4 s before its topall
disturbed_topthe track falls more than twice its noise and step below a peak in the 4 s before its top, then climbs to that top; or, within a second either side, it climbs into its top faster than a coast and falls from it faster than a body at restapogee, the top speed, landing, and the readings that need them
disturbed_padthe track falls before the climb and climbs without restingthe top speed; liftoff with no pad before the fall; all on a short pad with no zero
clock_out_of_orderthe height track’s clock doesn’t increase; or the accelerometer’s runs backall; or the top acceleration, and burnout from the accelerometer
no_height_trackthe log holds no altitude, pressure or GNSS trackall
unstated_sensorthe log’s only height doesn’t say whether a barometer or a GNSS receiver measured itall
receiver_lagthe only height is a GNSS receiver’s, whose filter lags a climb and overshoots its topapogee, the top speed, and burnout from its time
receiver_dropoutthe only height is a GNSS receiver’s, whose fix drops out or lags at liftoffliftoff, and the flight time
needsthe reading needs another that was withheld, such as landing on a GNSS trackas the sentence says
bad_recordthe record breaks what every reader guarantees: tracks as long as their clocks, on finite times that increase. No file reader produces one; only a record built by hand in code canall
no_accelerometerthe log has no accelerometerthe top acceleration; burnout comes from the top speed’s time
axis_unstatedthe log’s accelerometers don’t say which lies along the rocket’s axis, such as a board’s three axes or a table’s column read by its headerthe top acceleration; burnout comes from the top speed’s time
gravity_unstatedthe axial accelerometer’s readings may or may not hold gravity’s 1 g, and neither its family, nor its pad, nor its coast to apogee settles which, or they disagree, or only a coast at −1 g saysthe top acceleration; burnout comes from the top speed’s time
direction_unstatedwhich way up the axial accelerometer reads isn’t settled: of its pad (about 1 g nose up, about −1 g nose down), its coast to apogee (−0.75 g or less nose up, 0.75 g or more nose down) and its thrust (the first span past 2 g to hold there for 0.5 s, unless a span the other way comes first), none says or two disagree; or, turned as they settle and gravity put back, the trace never passes 2 g from liftoff’s earliest time to apogee, so shows no thrustthe top acceleration and burnout from the accelerometer; burnout comes from the top speed’s time
one_canopythe descent shows no drogue’s leg: at the ground’s density, its rate in the 2 s before the main’s leg never passes the main’s mark, 1.6 times the rate near the ground or 3 m/s (9.8 ft/s) more if that is more, or the descent before that leg doesn’t fall, on median, faster than 1.4 times that rate. A drogue no faster than the mark reads as one canopy, whatever its ratio to the main, and so does a rocket under one canopy or nonethe drogue’s descent; the main’s is read from apogee, the whole descent’s
short_lega leg of the descent loses no more than 3 m (9.8 ft) or a tenth of its top’s height; or the whole descent loses too little, or falls too slowly near the ground, to time a leg bythat leg’s descent; or both

“All” means liftoff, apogee, the top speed and landing, with the flight time, burnout and each canopy’s descent that need them. The top acceleration is searched for up to apogee, so a reason that withholds the apogee withholds it too. The exception is a log with no axial accelerometer, which is said first (no_accelerometer or axis_unstated). A burnout the accelerometer can’t give falls back to the top speed’s time, and carries the top speed’s reason where that is withheld too.

An apogee within half a window of the log’s end is read, but marked is_floor: the log may have stopped before the rocket did. The top speed and the smoothed apogee are then floors too, with no interval. Only a test’s cut flight has checked this: no synthetic log ends on the climb.

Two logs on one timeline

Each logger counts from a zero of its own, so two logs of one flight need their clocks lined up before they can be read together. Every board in a stack leaves the pad at the same instant, so each log’s liftoff reading is the same moment on its own clock, the method Debrief used (MIT; its reasoning and the rules it measured and deleted). The offset is the second’s liftoff less the first’s:

Δt=t2−t1,BΔt=B1+B2,

with ti each log’s liftoff and Bi its bound. BΔt is the most the two errors can add to. Its 90% interval runs from Δt−0.0515BΔt to Δt+0.0526BΔt, the two shares the 5% and 95% points of the error over the bound on two-logger flights of their own (ADR-257, calibrated intervals, §1, and ADR-312, two logs on one timeline).

A worked example, RocketPy’s Prometheus flight. The TeleMetrum reads liftoff at 0.23 s within 0.71 s, the TeleMega at 0.18 s within 0.83 s. So Δt = 0.18 − 0.23 = −0.05 s, within about 1.53 s (the command adds the unrounded bounds). The interval is −0.05 − 0.0515 × 1.53 = −0.13 s to −0.05 + 0.0526 × 1.53 = 0.03 s; the command, from the unrounded figures, prints −0.13 to 0.04 s. Each receiver’s UTC second, taken where it ticks in its log, puts the TeleMega’s clock less the TeleMetrum’s at −0.13 s. Over all 238 paired ticks that scatters from −0.24 to −0.03 s, and over the 72 at 100 samples a second from −0.15 to −0.11 s. A fixed difference between the receivers’ delays can’t be seen. So this offset is about 0.08 s away, and the interval’s low end, −0.129 s, is too close to that figure to say which side it falls. Both logs start at the boost, so an offset of zero would miss by 0.13 s, against this offset’s 0.08 s: a weak check, and one flight can’t show the interval’s coverage.

Checked against. On two-logger synthetic flights (their sets), every one of the 1,383 offsets given stayed within its bound, by 0.13 of it at most, the error’s median 0.03 s and its largest 0.21 s. A log with no barometer gives no liftoff, so its flight gives no offset: 617 of the 2,000. On the two-logger sets A and B the interval held the truth 622 of 703 and 612 of 680 times, passing the coverage test, and narrowed to 0.8, widened to 1.25 and shifted down it fails. All 135 offsets the 200 staged flights gave stayed within their bounds; there the interval missed none, so it is wider than it need be on those flights (the report).

What it leaves out. One number per pair: a clock running fast or slow isn’t corrected, by about 15 ms over a 300 s flight at a crystal’s 50 ppm. A board that wakes late or a log cut before liftoff has no liftoff to line up by. Nothing checks that the logs are of one flight. Two boards can agree and both be off by the same day’s air, so the spread says how far apart they lie and no more. On one real pair only (evidence row G35), and on no real staged flight (evidence row G36).

Validation and limits

Checked against

The synthetic flights’ tuning set. The synthetic flights are flights HPR Sim simulated, each with its truth and the log a simulated logger would write. Of the tuning set’s 200 flights, 171 carry a barometer; the other 29, a GNSS tracker’s, give no reading by rule. cargo xtask synthetic reads each of the 171 as its own logger wrote it. It then reads each again 36 ways in all. They cover every rate from 500 to half a sample a second. They cover every step from none to 10 m (33 ft) or 1 to 100 Pa. Last, it reads each cut 6 ways, as loggers that record from launch write them. Each is 20 samples a second, in feet or in 10 Pa steps. Each starts 0.3 s before liftoff, 0.1 s after, or 0.5 s after with the zero first. The 71 flights whose logger has an accelerometer are the AltOS-like and the IMU profiles. They are read again with it at 500, 100 and 20 samples a second. The 500 Hz IMU profile’s 38 are read once more with its part set to ±8 g (The acceleration). Some tops then clip at the part’s range and the rest don’t. Every reading given stayed within its bound, and the check fails the build if one doesn’t. A floor holds while it is no more than its bound above the truth. The largest share of its bound any error took, from the report and its JSON:

readinglargest share (of the bound)on the logits error, against its bound
liftoff0.603500 samples a second, heights with no step0.387 s within 0.643 s
apogee height0.71720 samples a second, heights with no step−151 m (−495 ft) within 210 m (689 ft)
apogee time0.56220 samples a second, pressure in 100 Pa steps−1.04 s within 1.85 s
top speed0.54420 samples a second, pressure in 100 Pa steps13.7 m/s (44.9 ft/s) within 25.1 m/s (82.3 ft/s)
landing0.5725 samples a second, pressure in 100 Pa steps−3.89 s within 6.8 s
flight time0.507500 samples a second, pressure in 100 Pa steps−1.89 s within 3.73 s
descent rate0.594100 samples a second, heights with no step−1.16 m/s (−3.8 ft/s) within 1.95 m/s (6.4 ft/s)
top acceleration0.737its accelerometer at 20 samples a second−13.7 m/s² within 18.6 m/s²
burnout0.54420 samples a second, pressure in 100 Pa steps, a floor from the top speed0.718 s within 1.32 s
drogue rate0.7025 samples a second, heights with no step−0.944 m/s (−3.10 ft/s) within 1.35 m/s (4.43 ft/s)
main opening0.641100 samples a second, pressure in 10 Pa steps0.617 s within 0.963 s
main rate0.7500 samples a second, heights with no step−1.17 m/s (−3.84 ft/s) within 1.68 m/s (5.51 ft/s)

The report also gives each reading’s errors over every tuning flight and every log together. It gives them beside the mean bound the readings stated (Errors on the tuning set). An error is the reading less the truth. Each error is held to its own reading’s bound. That bound differs from flight to flight and log to log, so the least or the most may pass the mean bound. The mean is for scale. A canopy’s rates are compared with the truth over the span the reader chose, its leg’s ends. So an opening read early or late shows in the opening’s error, not in the rates’. A floor is counted apart, as it may lie any amount below the truth. A floor is a burnout from the top speed, a top speed on a log cut on the climb (none in these sets), or a top acceleration that clipped or whose thrust has no end on the channel. The least, the most and the mean bound are to three figures, nearest. A mean is given to the decimal place of its largest error’s third figure, the digits every platform shares:

readinggiven readingsmean errorleastmostmean bound
flight time7,007−0.73 s−5.06 s3.48 s5.82 s
top acceleration2900.6 m/s²−13.7 m/s²13 m/s²16.3 m/s²
top acceleration, a floor32−48.1 m/s²−75.9 m/s²−1.58 m/s²12.9 m/s²
burnout, from the accelerometer296−0.0104 s−0.0943 s0.0476 s0.544 s
burnout, a floor1,237−0.24 s−1.03 s0.718 s1.3 s
drogue rate2,8420.04 m/s (0.13 ft/s)−1.61 m/s (−5.28 ft/s)1.35 m/s (4.43 ft/s)1.79 m/s (5.87 ft/s)
main opening2,8420.73 s−0.778 s5.3 s2.03 s
main rate5,4240.04 m/s (0.13 ft/s)−5.47 m/s (−17.9 ft/s)4.83 m/s (15.8 ft/s)1.71 m/s (5.61 ft/s)

Burnout from the accelerometer lands within 0.0943 s of the truth, to three figures, against a mean bound of 0.544 s. Without it, the top speed’s time comes 0.24 s before burnout on average, as a speed’s peak should. It ranges from 1.03 s before it to 0.718 s after, within its bound. The floor, that time less its bound, about 1.3 s on average, lies earlier still. The main’s opening is read late on average, by 0.73 s. Its bound holds the time a canopy takes to slow the fall and half the fit.

No tuning flight’s true top acceleration reaches the range its own part clips at: the nearest comes to 0.492 of it, the report says. So the 500 Hz IMU profile’s 38 flights are logged once more with the part set to ±8 g. The BMX160 and the MPU-6000 offer that full scale. Of the 38 flights’ true tops, 28 reach that range, and 10 don’t. Of the 32 tops read as floors, 28 are clipped at the range, all on that log, each a floor that holds. The floors’ mean error, −48.1 m/s², is mostly how far those clipped tops lie below their truth, as a floor may by any amount. The other 4 are held flat at their top, all on the AltOS-like profile’s ADXL375-class part read at 20 Hz. That is consistent with its noise there, 19.8 mg by the synthetic flights’ noise model, being under half its 49 mg step. So it writes most of a thrust’s plateau as one value. But no test pins a quiet part that reads as flat (ADR-283). No tuning-set top reads as a thrust with no end.

On their own logs, the main’s opening was found on 87 tuning flights with a barometer, each within its bound. On no log was a main named on a flight that opened one canopy or none. Those are counted as none to read. On logs sampled once a second or slower, the canopies are read only where coarse steps widen the descent’s fit enough to hold 5 samples. Elsewhere they are withheld, most as sampled_too_slowly, some as ends_before_landing (up to 11 of 171 at half a sample a second) and one as step_too_coarse.

The same report counts each reading withheld, by reason. At 20 samples a second and faster, the top speed is withheld as transonic on 117 or 118 of the 171 flights. At 5 samples a second and slower, it is withheld as sampled_too_slowly on all of them but the one withheld whole below. At half a sample a second, up to 11 flights end before their landing is seen. In 10 m (33 ft) steps one flight’s track is withheld whole as step_too_coarse. The logs written again are read as tables a column mapping states, so their heights are above the pad only by the mapping’s word. Of the cuts, heights from 0.1 s after liftoff withhold liftoff as short_pad on 36 flights. Heights from 0.5 s after, the zero first, withhold it on all 171. Both give the apogee on all 171, within its bound with the zero’s possible error in it, at up to 0.695 and 0.674 of it. A pressure has no zero set on the pad, so a pressure cut after liftoff withholds every reading on all 171. A pressure cut from 0.3 s before liftoff does so on 3. No tuning flight is withheld as disturbed_top or slower_than_free_coast.

The tuning set shows the bounds hold on the flights they were tuned on. It doesn’t show how often they would hold on others, nor on real logs. The flights are HPR Sim’s own, with a noise model taken from datasheets and a few public logs (Synthetic flights).

An invented flight. The file synthetic-pnut.pf2 holds a flight made up for the tests. It sits on the pad for half a second, then accelerates at 50 m/s² for 1.6 s and coasts with no drag. It falls from rest to 25 m/s (82 ft/s). It descends at 25 m/s (82 ft/s) to 150 m (492 ft) and at 6 m/s (20 ft/s) to the ground. It is written as a Pnut writes one, rounded to whole feet and feet per second. It has an ejection pulse of the public log’s shape a second after apogee. Every reading is known in closed form. The tests hold each within its stated bound, and within a tighter one worked out beside it for this log (evidence row G10):

readingtrueHPR Sim readsits boundthe test’s tighter bound
liftoff0.50 s0.80 s0.81 sfrom 0.50 s to 0.814 s, once the rocket has climbed 2 m (6.6 ft), a foot and the half foot that rounds it up
apogee390.31 m (1,280.5 ft)390.14 m (1,280 ft)31.9 m (105 ft)0.23 m (0.75 ft): half a foot of rounding and the median’s 0.077 m (0.25 ft)
apogee time10.258 s10.275 s0.88 sone sample, 0.05 s: the middle of the flat run at the top
top speed80.0 m/s (262 ft/s) at 2.10 s78.3 m/s (257 ft/s) at 2.20 s20.6 m/s (68 ft/s)no higher than the truth, and no lower by more than the fit’s cut across burnout, 3.7 m/s (12 ft/s), and half a foot a second
landing46.14 s, touchdown45.80 s2.11 sup to 0.38 s early: 2 m (6.6 ft) at 6 m/s (20 ft/s) is 0.33 s, and a sample more
flight time45.64 s45.00 s2.91 sfollows from liftoff and landing
mean descent rate10.93 m/s (35.9 ft/s) over HPR Sim’s span, 10.275 s to 45.80 s10.92 m/s (35.8 ft/s)2.1 m/s (6.9 ft/s)the two heights’ bounds over the span
burnout2.10 sno sooner than 0.88 s: the top speed’s time, 2.20 s, less its bound1.32 snone tighter: a floor, no more than its bound after the truth
drogue rate21.74 m/s (71.3 ft/s) over HPR Sim’s span, 10.275 s to 21.40 s, the fall from rest at apogee in it21.7 m/s (71 ft/s)2.2 m/s (7.2 ft/s)the two heights’ bounds over the span
main opening21.14 s, 150 m (492 ft) up21.40 s, 148.4 m (487 ft) up1.06 son this log, whose drogue falls far faster than the mark: no sooner than the truth, and no later by more than half the 1 s fit and a sample, 0.55 s
main rate6.0 m/s (20 ft/s)6.0 m/s (20 ft/s)0.8 m/s (2.6 ft/s)the two heights’ bounds over the span

The middle of the flat run at the top falls between two samples here, so the apogee time is 10.275 s. hpr analyze prints times to two decimals, as 10.28 s and 9.47 s after liftoff. The first sample of the flat run at the top would read the apogee over 0.2 s early. A test shows that too, so the rule of taking the run’s middle is pinned. The file’s highest sample is the pulse’s, 400.5 m (1,314 ft). The Hampel filter keeps it, and the median sets it aside; a test holds both.

A real flight. Debrief ships a public PerfectFlite Pnut log, trimmed from a publicly shared flight. Its terms upstream are unclear, so it isn’t committed here, and CI doesn’t have it. Where it has been fetched (cargo xtask refs fetch), a test reads it. The test holds the apogee above the altimeter’s zero, the highest sample, the pad, and each reading with its bound:

the altimeter statesHPR Sim reads
apogee1,009 ft above the pad1,009 ft (307.5 m) above the pad, within 27.4 m (90 ft); 1,010 ft above the altimeter’s zero, its pad a foot above it
highest sample1,028 ft above the zero, the ejection pulse, set aside
top speed264 ft/s (80.4 m/s), within 29.7 m/s (98 ft/s), derived from the heights
liftoff0.40 s, within 0.98 s
landing72.70 s, within 3.25 s

The altimeter’s apogee comes from the same pressure trace. So the agreement shows that HPR Sim reads the trace as the altimeter does, not that either is the true height. A Featherweight Raven flew on the same flight, and Debrief reports it agreeing at about 1,009 ft.

The sample logs. cargo xtask logs takes the readings from every record the project’s sample logs hold. Those are other flyers’ files kept outside the repository. The command counts each reading given or withheld, by logger family and by reason (the survey’s report). Of the 1,058 records, summed over the report’s families, an apogee is given from 592, liftoff from 549 and the top speed from 353. Landing is given from 471, the flight time from 404 and burnout from 426. The top acceleration is given from 160, the drogue’s descent from 277 and the main’s from 458. Each of the 28 reasons has a case. For 24, the case is the first file the survey withholds a reading of for it, by an anonymous id. For the four that none of the sample logs triggers (bad_record, faster_than_free_fall, starts_off_the_pad and step_too_coarse), the case is an invented file under validation/fixtures/logs/readings/. A test reads it to its reason. The invented fall, for one, drops at 15 m/s (49 ft/s) for 3 s after apogee, then at 300 m/s (984 ft/s) to the ground. These counts show what each logger can give. They don’t show how right a reading is: none of the sample logs has a truth beside it.

Four reasons stand out. The top speed is withheld as transonic on 197 records: many of these flights are fast. Liftoff is withheld as short_pad on 157, 71 of them from loggers whose files start close to liftoff. Those are 37 of 51 RRC3 CSV records and 34 of 44 Eggtimer CSV records. And the apogee, top speed and landing are withheld as disturbed_top on 75. The top speed is withheld as disturbed_pad on 10, liftoff on 5.

The new readings’ reasons follow the loggers. The top acceleration is withheld as no_accelerometer on 682 records, 17 of them the AltOS files above. It is withheld as axis_unstated on 136, whose boards or tables don’t say which accelerometer lies along the rocket. It is withheld as gravity_unstated on 10, and as direction_unstated on 6. On one, nothing says which way up the accelerometer reads. It never passes 2 g either way, and the log holds no pad or coast to tell by. On one, the coast reads near +1 g, nose down, against a pad and a thrust that say nose up. On three AltOS CSV records, gravity taken out, the pad says nothing and a short burn comes before a heavy drag. So the thrust says nothing either. On one, the accelerometer’s thrust rises after liftoff’s latest time by the heights, its clock running late against theirs. The drogue’s descent is withheld as short_leg on 12, and as one_canopy on 173, where the descent showed no drogue’s leg. With no truth beside the logs, that doesn’t say how each rocket came down.

What it leaves out

  • The day’s air. A logger’s height is its own conversion of the pressure, which assumes a standard atmosphere. On a warm day a barometric altimeter reads low, 6.5% on a day 20 K warmer than standard (Atmosphere). HPR Sim doesn’t correct it: it widens the bound by 6.92% of the height instead, which covers air up to 15 K from standard. On a day further off, a height can pass its bound. Where the log records the pad’s temperature and its height above sea level (from its pressure, an altitude above sea level, or a stated ground elevation), hpr analyze notes such a day with its size, and leaves the bound as it is (A day past 15 K).
  • Loggers unlike the tuning set’s. The bounds hold on the tuning set’s six simulated logger profiles, at their noise. A logger noisier than they are, or a rocket slower off the pad than 2 g, can be read outside its bound. A pad that drifts within the 2 m (6.6 ft) allowance, early one way and late the other, can read liftoff outside its bound. A review measured up to 1.22 times the bound at ±2 m (6.6 ft) on a synthetic climb. On a short pad, a zero only a word vouches for is placed by the first sample unfiltered. So its error can be understated by that sample’s noise, up to 3 m (9.8 ft) (#773, the pad error’s noise budget).
  • Rockets slow off the pad. The liftoff bound and the pad’s error assume at least 2 g off the pad. A rocket slower than that is read later than its liftoff bound. It may already be further up, when a log starts, than its pad’s error allows.
  • Fast flights’ top speeds. The top speed is withheld from 227.7 m/s (747 ft/s), and so is a burnout taken from its time where no accelerometer gives one. The heights aren’t checked for the port’s error near the speed of sound. The apogee, read near the top where the rocket is slow, stays within its bound on the tuning set, whose fastest flight reaches Mach 1.12.
  • A short stay on the pad. The climb’s first rise counts toward a rest, so a short stay on the pad after a carry down is read as a pulse from the higher level. After a 30 m (98 ft) carry and before a 4 g climb, that is up to 0.2 s from the stay’s first sample to its last at 20 samples a second (0.15 s on a pressure track), up to 0.35 s with the pad’s noise 2.3 m (7.5 ft) either side, two samples at 5 and at 3.3 a second, and one at 2.5. The apogee then reads low by the drop, past its bound on a 98 m (322 ft) flight at 20 samples a second for a drop of more than about 10 m (33 ft).
  • Wide pulses. A pulse wider than half the median’s window passes it. The bounds were held on the synthetic flights’ pulses, which are shaped like the public Pnut log’s and scaled by 0 to 2. A wider or taller one isn’t covered.
  • The deployments, the Mach number and dynamic pressure. The main’s opening is read from the descent’s rate, not from a deployment the log records, and the drogue’s leg is read from apogee. The Mach number and dynamic pressure aren’t read: they need the speed and the air all along the climb, which the readings don’t reconstruct.
  • A canopy’s filling time. The opening’s bound holds the time an open canopy takes to slow the fall, not the time it takes to fill. The synthetic flights open each canopy at once, so a main that fills slowly can be read later than its bound (#779).
  • A fall under no canopy. A rocket that came down under no canopy has its fall read as one canopy’s main_descent: the rate alone doesn’t tell a canopy from none. The tuning set’s flights of Cavour, a design with no parachute, read that way (#780).
  • A descent from far above the tropopause. The scaling to the ground’s density slows a rate read high up and never quickens it. The exception is about 1% in a surface inversion (The canopies). This holds while the density from the pad up falls no faster than exp(−h/H). Through the troposphere it falls slower, as its scale height is at least 7,289 m (23,914 ft). Above it, about 11 km (36,000 ft) up, the scale height is about 6,340 m (20,800 ft). On a day 15 K cold it is 5,903 m (19,367 ft). From a pad at sea level, the 1976 standard atmosphere’s density stays above exp(−h/H) to 34.6 km (113,500 ft) up. On a day 15 K cold it does so only to 21.3 km (69,900 ft). From a pad 1,500 m (4,921 ft) up, the limits are 28.3 km (92,800 ft) and 17.7 km (58,100 ft) above it, each rounded down. A canopy from higher could read faster than it falls near the ground, and one canopy as two. The heights are the standard’s geopotential heights, a little less than a height measured straight up, which puts the limits low, not high. The tuning set’s apogees reach 4,420 m (14,500 ft), so it can’t show this.
  • More than one burn. An air start’s or a sustainer’s second burn is searched for the top. But the tuning set’s ten designs each fly one motor in one stage, so its bound isn’t checked on one. Burnout then comes from the speed’s peak, a floor. A second burn that holds past 2 g for less than 0.5 s is taken for a shock. So its top is missed and burnout is the first burn’s end. A shock that holds there longer reads as a burn. Where liftoff is withheld, a knock on the pad wider than the cleaning takes out joins the burn. The knock can be taken for the top (#789). A part that reads above zero all through the coast, as a part that zeroed off its pad’s level can, never ends its burn. So its top is searched to apogee and given as a floor, and its burnout comes from the speed’s peak.
  • Noise that runs from sample to sample. Averaging n samples cuts the noise by n only where each sample’s noise is its own. A part whose noise runs on, as its own low-pass filter or a slow drift makes it, keeps more of it through the average. So N comes out too small and burnout’s bound too narrow. The synthetic flights draw each sample’s noise apart, so they can’t show this (#787).
  • Clipping at the high-g part’s range. Clipping at a part’s range is held on the tuning set only on the low-g IMU part set to ±8 g, where 28 tops clip. No tuning flight’s true top reaches its own part’s range, ±32 g or ±180 g (the loggers). So a high-g part’s top clipped at its range is held only by the unit tests. A clipped top is a floor either way.
  • The coast’s witness. It assumes the drag in the last second before apogee’s earliest time is under 0.25 g. A rocket still fast at apogee, arcing over, reads its drag there. The witness then says nothing or, past 0.75 g, can disagree with a family’s word and withhold the top acceleration (ADR-283). Alone at −1 g it settles nothing (#806). A net push forward of between about 0.75 and 1.25 g (thrust less drag) still acting in that second reads as gravity in, and the top then reads 1 g low (#1131, untested).
  • A board’s three axes. A log whose accelerometers don’t say which lies along the rocket gives no top acceleration (axis_unstated), and its burnout comes from the top speed’s time. The axes’ magnitude could give a top acceleration, but not burnout, as it never falls below zero (#781).
  • Intervals. A bound is a worst case held on the tuning set, not a 90% interval with a tested chance of holding the truth. Twelve readings state one too, a floor a lower end in its place (The 90% intervals).
  • A GNSS track. A log whose only height is a GNSS receiver’s gives no reading. The receiver’s lag and overshoot aren’t modeled, so no bound can be stated for it.
  • Disturbed tops, both ways. Some bumps come less than about a second after apogee, before the track has fallen past its noise and step. Such a bump can’t be told from the top unless it is a spike. The bump’s time may pass the apogee’s time bound. A transient that only climbs in fast, or only falls out fast, is read as the top. That is because an early parachute and a charge make those shapes on a good top. A spike that climbs in no faster than a coast is read as the apogee, however high it stands. The other way, a brief dip in the coast in the last 4 s before the top withholds a good apogee. How many of the sample logs’ 76 such tops are good isn’t measured (#762). The synthetic flights model no charge’s transient, so the bounds aren’t tested against one (#761).

Tests and references

Tests

  • hpr_flightdata::readings: a plain flight read within its bounds. The same flight read alike from a pressure channel, an altitude channel above sea level and the logger’s own track. A GNSS altitude withheld by name. Each withheld code on a log built to trigger it, only the readings that need it going. The pad’s rules, with a jitter in the first samples, noise up to the allowance and a slow noisy pad. A pad short by its samples and one read from the logger’s zero. A vacuum fall rounded to feet landing at 10 to 100 samples a second and one at 1.3 g refused. #646, a wrong unit or a wrong zero in a mapping, both ways. A mapped track’s short pad read from its zero within the allowance, its apogee’s bound grown by the distance, and withheld past it. A top held by a fall or a spike withheld, each by its own rule. A top with an early parachute or a later charge read. The shapes of the two public logs of #702, with invented numbers. The step and the clock at their edges. The top speed’s guards, each on its own. A property test of the median’s peak bound for every K from 1 to 6 and every place the peak can fall between samples.
  • hpr_flightdata::readings::accelerometer: a noisy axial trace’s top and burnout read within their bounds, burnout within a sample of the thrust’s end. A part mounted nose down, with gravity taken out, read alike. A dip, and a knock with its rebound first, on the pad not turning the trace over. A short burn before a heavy drag read nose up, gravity in and out, and nose down when mirrored. A pad that disagrees with the thrust withheld. A weak thrust on a log without gravity settled by its coast, or withheld where a heavy drag’s span disagrees. A coast before the trace moves, or in a log that ends in the thrust, saying nothing. Three traces that show no thrust withheld, each by its own words. The thrust’s witness and the pad’s and the coast’s levels, each way they say. A top held at the part’s range marked clipped at the range, the truth above the reading less its bound, and a quiet plateau not. An accelerometer that ends in the thrust giving its top as a floor and no burnout. Gravity left unsettled withholding the top acceleration and sending burnout to the top speed’s floor. The coast settling gravity where the log holds no pad. A repeated time keeping its first sample, and one after a sample with no time left out. A peak between two samples within the top’s bound. A trace that never falls below zero giving burnout as a floor from the top speed. The noise measured as the scatter about the curve, a straight line’s as 0.
  • hpr_flightdata::readings::canopies: a drogue to 150 m (492 ft) at 25 m/s (82 ft/s), then a main at 6 m/s (20 ft/s). Each leg’s rate and the opening read within their bounds. The opening’s bound holding the main’s slowing time to its 9.6 m/s (31.5 ft/s) mark beside half the fit. One canopy withholding the drogue and reading the main from apogee. A drogue no faster than the main’s mark read as one canopy. A drogue passing the mark for a moment, as a gust would make it, naming no main. One canopy from 8.6 km (28,215 ft) read as one at the ground’s density. Samples bunched at landing leaving the median its window. A short drogue leg withheld. A track too coarse for the main’s fit withholding both.
  • hpr_flightdata::readings::day_air: an invented flight in air 30 K over standard, its note, its share against the rule and the truth within it; 30 K under, the other way; 10 K over, covered. The edge, exactly 15 K, covered both ways and 15.01 K not. A log ending on the climb, its share at 216.65 K. No note with no height above sea level, no temperature, or one that can’t be the air: out of range, or one value throughout (zeros in kelvin, °C or °F). The standard’s temperature against its table. hpr analyze’s own tests: each note below each reading, in full, on a hot pad and a cold one; a day just past 15 K within the bound’s share; none at 70 °F; the pad’s pressure altitude named as such; and the drogue’s line calling its rate the mean from apogee.
  • hpr_flightdata::synthetic: the invented flight, its readings against the truth (burnout, the main’s opening and both legs among them). The Hampel filter keeping the pulse the median removes.
  • hpr_flightdata::filter: a pulse removed, a ramp untouched, the Hampel filter at zero equal to the median. Property tests that every output lies within its window and that the sliding median equals each window’s own, bit for bit.
  • cargo xtask synthetic --check: every reading on the tuning set within its bound, at every rate, step and cut. That includes each accelerometer at 500, 100 and 20 samples a second, and the IMU’s part set to ±8 g. A floor is no more than its bound above the truth. The report is reproduced on every platform.
  • cargo xtask logs and its tests: each invented sample read to its reason. And, where the sample logs are fetched, every record’s readings counted and the report reproduced.
  • crates/hpr-cli/tests/cli.rs: hpr analyze on a log alone in a folder, its JSON checked against the published schema; the public Pnut log, where fetched. crates/hpr-py/tests/test_logs.py: Python’s FlightLog.readings() equal to what hpr analyze prints, on each invented log.

References

  • J. Lei, M. G’Sell, A. Rinaldo, R. J. Tibshirani and L. Wasserman, “Distribution-free predictive inference for regression”, Journal of the American Statistical Association 113(523), 2018, §2.2, for the split-conformal rank that sets each interval’s points. V. Vovk, A. Gammerman and G. Shafer, Algorithmic Learning in a Random World, Springer, 2005, ch. 2, for its origin.
  • R. K. Pearson, Y. Neuvo, J. Astola and M. Gabbouj, The Class of Generalized Hampel Filters. 23rd European Signal Processing Conference (EUSIPCO), 2015, §2, eqs. 1 and 2. Pinned as pearson-2015-generalized-hampel in the reference lock file.
  • Debrief, lib/analyze/index.ts and lib/parsers/perfectflite.ts (MIT, the project owner’s own), for the landing’s thresholds, the least climb, the speed’s refusals and the format. The same files are the source for the flat top of a clipped accelerometer and the fall through zero it takes as burnout. They are also the source for the main’s opening found back from landing, the drogue’s ratio and a leg’s least drop. lib/flight/build.ts is for the first sample kept at a repeated time. Notes are in the readings note.
  • The accelerometer’s scale and bias: the ADXL375 datasheet’s ±10% sensitivity (p. 3), and the bias the synthetic flights assume, which no source states. Both are in the sensor noise note and Synthetic flights.
  • A canopy’s slowing after the opening: vertical motion under quadratic drag, v˙=−g(v2/vm2−1), J. R. Taylor, Classical Mechanics, University Science Books, 2005, §2.4. τ is its integral from any speed down to v1, HPR Sim’s step.
  • A canopy’s rate: drag balancing the weight, v=2mg/(ρCDS), so the rate goes as the density to the −½ (Recovery).
  • The day’s air past 15 K: the hypsometric equation, dz=−(RT/g)dp/p, and the temperatures of U.S. Standard Atmosphere, 1976, NOAA-S/T 76-1562, as the height formulas note lays them out. The range a logged temperature must fall in sits beyond the coldest and hottest air on record, −89.2 °C at Vostok and 56.7 °C at Death Valley, from the WMO’s archive of weather and climate extremes. The dry adiabatic lapse rate a hot afternoon’s air tends toward, g/cp≈ 9.8 K/km: J. M. Wallace and P. V. Hobbs, Atmospheric Science, 2nd ed., Academic Press, 2006.
  • The density scale height: U.S. Standard Atmosphere, 1976, NOAA-S/T 76-1562, its troposphere’s lapse of 6.5 K/km and its tropopause’s 216.65 K (Atmosphere). The formula H=RT/(g−RL) follows from its pressure and temperature.
  • The noise model and the range of air: ADR-280, the synthetic flights first §2–3, and Synthetic flights.
  • The port’s error: the Mach error note §1–2, from a wind-tunnel test of the static pressure along an ogive-cylinder through Mach 1.
  • How loggers turn pressure into height: the height formulas note.
  • The least acceleration off the pad: the NAR’s high-power safety code. Its limit is a liftoff weight no more than a third of the motor’s certified average thrust. Tripoli’s is the other (below 3:1 only with an active stability system). The most a rail stands from vertical, 20°: the same codes.
  • The rate of the day’s pressure change: the National Weather Service’s glossary, pressure rising (falling) rapidly. That is 0.06 inHg an hour or more, the threshold of METAR’s PRESRR and PRESFR remarks.
FusionSpace HPRVersion 0.1.0Issued 2026-11-21In preparationAbout HPR Sim