Historical data

All times UTC

Saved hourDownload hour

Live data

10-minute window · 3-minute processing buffer

Station status

Results

187 hours · 26 catalogue events

Single-station STA/LTA → Standard multi-station trigger → Travel-time-constrained association. Both multi-station methods use the same station network.

Continuous week · main evaluation

Validation budget candidates/dayCSV (all budgets)

168 continuous hours · 7 catalogue events

Model	Recall	Precision proxy	Candidates/day

Event-focused dates · supplementary evaluation

19 selected hours · 19 catalogue events. Event-enriched sampling; not representative of continuous operation.

Model	Recall	Precision proxy	Candidates/day

Each table uses the same hours and one-to-one catalogue matching for all metrics: recall = matches / catalogue events; precision proxy = matches / detections; candidates/day = detections / hours × 24. Unmatched candidates remain unverified.

Evaluation setup

Station identity was selected on training dates; thresholds on separate validation dates. Final results are retrospective: missed events informed station expansion. The improved system changes both coverage and association. Unmatched candidates need review. The notebook separates these effects.

Method

Full illustrated notebook

One shared signal-processing pipeline. Three rules for turning station triggers into candidate events. Settings below are the default demonstration settings.

Shared input: clean each segment

BGS miniSEED · vertical HHZ · 100 Hz. Split at gaps without interpolation; exclude conflicting overlaps. Remove the linear trend and mean, taper edges, then apply a fourth-order 2–15 Hz zero-phase band-pass filter.

Actual waveform before and after the implemented preprocessing

Shared score: classic STA/LTA

Divide the mean squared signal over the last 1 second by its mean over the last 30 seconds. A sudden energy increase raises this ratio. All three methods use classic STA/LTA, not recursive STA/LTA.

Actual signal, 1-second and 30-second mean squared energy, and STA/LTA ratio; threshold 8 example

1. Single-station STA/LTA

Use EAB, selected on training dates. Start a trigger above ratio 8; end below 1.2. Each retained EAB trigger becomes a candidate.

2. Standard multi-station trigger

Apply the same 8 / 1.2 thresholds across the network. ObsPy coincidence_trigger combines overlapping trigger intervals with at least two stations. No extra overlap extension is added.

3. Travel-time-constrained association

Collect station intervals at 2 / 1.2; merge gaps up to 8 seconds, retaining the first onset.
Seed groups with the strongest remaining trigger. Visit nearby stations first and choose at most one trigger from each.
Every pair must satisfy |Δt| ≤ min(65 s, distance / 3 km/s + 8 s). This is a distance-based timing bound, not an estimated epicentre.
After grouping, retain members with peak ratio ≥12. Keep the candidate only if at least two members remain.

What changes on real data?

Silverdale · 7 June 2026 · ML 1.1. EAB does not trigger around the event. The association retains AR09, ESK and GVIE at different onset times; two onsets match predicted catalogue arrivals. Vertical lines show retained onsets, not fitted source locations.

Real Silverdale STA/LTA traces: EAB misses the event while AR09, ESK and GVIE form an association

Catalogue matching is a separate evaluation step: retained onsets are compared with predicted P/S arrivals within ±8 seconds. Catalogue times do not create the detections. Parameter-search and earlier-model experiments are in the notebook.