Timekeepers.jl

Appearance

Spectral Views ​

Pick a spectral view in TKApp's View menu to spot noise, mains harmonics or a dead sensor that a trace hides.

ViewWhat you get
Time | Spectraa PSD panel next to each trace
Spectrathe same PSD panels at full width, no traces

Window and Scroll choose the samples, masks remove theirs, and the spectra update as soon as you mask or unmask. The line below the plots shows nfft, df = fs/nfft, Nyquist, segment length and segment count; the i badge explains each term.

How the spectra are computed ​

Welch's method on the visible window:

  1. Masked samples are removed, not zero-filled, which leaves good stretches. A gap between joined runs (e.g. high-rate Metronix) also ends a stretch, so no segment spans two recordings.

  2. Each stretch is cut into nfft-sample segments with 50% overlap. Leftovers shorter than nfft are dropped.

  3. Each segment: remove the mean, apply a Hann window w, FFT, scale |X(f)|² by 1 / (fs · Σ w²), double all bins except 0 Hz and Nyquist. The result is a one-sided PSD in amplitude²/Hz.

  4. Average all segments with equal weight.

Both axes are logarithmic. The 0 Hz bin is hidden, and the y axis shows at most eight decades below the peak.

How the transform length is selected ​

nfft = window length in samples ÷ 8, rounded down to a power of two, clamped to [256, 8192] (8192 for All). Widen the window for finer frequency resolution, narrow it for finer time resolution. If no stretch is as long as nfft, the panel says so: widen the window or remove a mask.

Performance ​

Workspaces (taper, FFT plan, buffers) are cached by (nfft, fs, noverlap, window, detrend), and a short timer delays recomputation while you scroll.

Internal API

Timekeepers._welch_psd, Timekeepers._welch_psd_segments and Timekeepers.SpectralWorkspace are internal and may change. For your own spectral analysis, get clean segments and use a dedicated package:

julia
segments = good_segments(ta, mask; min_samples = 4096)