M10. Finding Cycles

What It Is

The period search looks for periods the readings repeat at, scores how sure it is of each, and names the wheel each one matches. Its scores are what Periods draws (chapter 11), its lines are what the headline names, and its cycle is what Cycle folds (chapter 12).

How It Is Worked Out

  1. The watch’s own clock. Each beat is placed at its number times the nominal beat period, not at the time the sound card heard it. A wheel turns once in a fixed number of beats: the fourth wheel once a minute is 480 beats at 28,800 bph, and the escape wheel’s turn is twice its teeth in beats:

    escape wheel period = teeth × 2 ÷ beats per second
    20 teeth at 28,800 bph (8 beats a second):  40 ÷ 8 = 5 s
    15 teeth at 28,800 bph:                     30 ÷ 8 = 3.75 s
    15 teeth at 21,600 bph (6 beats a second):  30 ÷ 6 = 5 s
    

    On this clock a wheel’s period is exact, and its cycle doesn’t smear when the rate changes.

  2. The series. For the rate, the timing offset of every clean beat from a steady rate, in 0.25 s bins (0.5 s for takes over 12 hours, 4 s over 48 hours). For the amplitude and the beat error, their 2 s windows.

  3. Slow change off. A running median over 30 minutes is taken off (a quadratic on a take under 20 minutes), so the mainspring running down isn’t scored as a long cycle.

  4. Score. The spectrum is divided by the shape the series’ own noise gives (AR(1), a first-order autoregressive model in which each reading carries a share ρ of the one before, for readings; a random walk plus white jitter for the beat timing) and by its local median, and each trial period sums its first 1, 2, 4 or 8 harmonics:

    score(P, h) = Σₖ₌₁ʰ (normalised power at period P/k)
    

    Each sum is scored against the spread the same steps give on simulated noise of the take’s length and step, and the best score against how high the best score over the whole search runs on that simulated noise: the false-alarm chance for the whole search. Under 0.0003, past what the simulation reaches, the text says “beyond the simulated range”.

  5. Listed when the false-alarm chance is under 1% across the whole search (the dashed line in Periods), the cycle explains at least 0.5% of the series’ variation, and it passes Holm’s correction with every other test of the take (M9).

  6. Fold and take off. The strongest period is refined, the series is folded at it (the median at each point of the turn), that average cycle is taken off, and the search runs again on what is left.

  7. Name. A line matches a wheel when their periods agree to within 1%, or to the take’s resolution, whichever is wider:

    resolution = P² ÷ take length
    

    At 60 s in a 2-hour take that is 0.5 s. Otherwise the nearest wheel and how far off it is are given.

Why This Way

  • Harmonics: a smooth swing puts all its power at its period. A short event once a turn (a tooth catching, a dip for a few seconds) spreads it over many harmonics, and a search at one period would miss it. Summing 1, 2, 4 or 8 catches both; twice the true period scores worse, since half its harmonics are noise.
  • Folding and taking off: wheels’ periods are whole-number ratios of each other (60 s, 30 s, 20 s…), so one fault shows at several periods. Taking off each cycle found keeps it from being listed again.
  • Dividing by the noise shape: watch readings wander, so a raw spectrum climbs towards long periods; without this, every long period would look like a line.
  • The beat timing for the rate: a 10 s rate reading can’t show any cycle faster than 20 s. The timing of each beat can: the escape wheel’s swing is a fraction of a millisecond, beat by beat. A cycle under 30 s is given as the timing swing in ms; a slower one as the rate’s swing, in s/d.
  • The resolution: a take holding N turns of a period can only place it to about one part in N.

When You’d Use This

  • Matching a line to a wheel whose period is off by a little.
  • Telling a smooth swing from a once-a-turn knock: which harmonic count scored best.
  • Planning a take: a cycle needs several turns inside the take.

What It Helps Find

  • A wheel that catches or rubs once a turn, at that wheel’s period or its harmonics.
  • An off-centre wheel or pinion, at half the period (twice a turn).
  • A disturbance from outside the watch (the room, the stand) at a period no wheel has.

What It Can’t Tell You

  • Which part. A line at a wheel’s period says the readings repeat with that wheel’s turn: the wheel, its pinion, its arbor or the hand it carries.
  • Long periods. Periods over about 15 minutes, half the running median, are weakened by it. A barrel’s turn of hours is beyond it.
  • Periods under the shortest searched: 3 s for the rate’s timing up to 12 hours, 4 s from 12 to 48 hours and 32 s beyond (the longer of 3 s and eight bins of the series), and 16 s for the amplitude and beat error. A wheel faster than that can fold into slower cycles that aren’t there.
  • A guarantee. The 1% is calibrated on simulated noise of those two kinds. Noise of another kind (bursts, flicker) can pass or fail differently; the timing-misfit caveat says when the beat timing doesn’t fit its model.

The project’s long-run notes give the search in full.


Previous: M9. Steadiness Statistics · Next: M11. Time and Frequency Maps

This page comes from the app's user guide at v1.0.1. See it on GitHub.