Cleaning the Timeline

Raw GWOSC segments include very short gaps, known GW events, and lock-loss artefacts. prune_segments() removes all of these in one call.

tq.prune_segments(
    rm_short_segments=True,
    rm_min_duration=22.0,   # drop segments shorter than 22 s
    rm_allevents=True,       # excise known GW events from GWOSC
    rm_window_length=30,     # ± 30 s around each event GPS time
)

Options

Parameter	Description
`rm_short_segments`	If `True`, drop segments whose duration is below `rm_min_duration`.
`rm_min_duration`	Minimum segment duration in seconds. Segments shorter than this are discarded. 22 s is the minimum needed to safely produce a 16 s analysis window with edge buffering.
`rm_allevents`	If `True`, fetch the GWOSC event catalogue and excise a window around each event GPS time. This prevents known signals from contaminating the noise training set.
`rm_window_length`	Half-width (in seconds) of the excision window around each event. `[event_gps − rm_window_length, event_gps + rm_window_length]` is removed.

Why remove events?

Sage needs a pure noise dataset — segments that are free from known gravitational-wave signals. The GWOSC catalogue lists GPS times of every confirmed detection. Leaving these segments in would contaminate the noise class with real events, undermining the assumption that the noise class contains only background noise.

rm_window_length=30 excises 30 s either side of each catalogued event GPS time. This is a conservative window that accounts for the full inspiral-merger-ringdown duration of the heaviest systems in the prior, ensuring no signal power leaks into the noise training set.

Inspecting the result

After pruning, tq.timeline is updated in-place. You can inspect durations and inter-segment gaps:

segs = tq.timeline[0]["segments"]

# Segment durations
durations = segs[:, 1] - segs[:, 0]

# Gaps between consecutive segments
gaps = segs[1:, 0] - segs[:-1, 1]