- 32 analyses: each label kept, rewritten or removed by its analyst (section Requalification of the report; mm magnitudes where the targets have them) - SNCF01 x3, WILDSTRUBEL 2026-09-01 -> 2026-10-06 x3: verified from their existing labels under the new rules - long labels (no analysis): auto/analyses/long-labels/<STATION>/; NORDEND-169 and -195 run to 07/11, so NORDEND-283 and -284 (their continuation) are removed - reading-rules.md, stations.md: the citations of removed or rewritten labels point to their successor, or to [expert] Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
101 lines
11 KiB
Markdown
101 lines
11 KiB
Markdown
# Labels: WILDSTRUBEL, 2026-10-20T00:00Z → 2026-11-01T00:00Z (UTC)
|
||
|
||
Task (first pass): verify the 2 existing labels of the range (`work/existing.jsonl`), and read the 2 spanning labels (`work/spanning.jsonl`) within the range.
|
||
Task (2026-11-17): requalify these labels under the new rules (`labeling-rules.md` §1: only what stands out from the noise of the corrected curves, beyond 5 own noises; No data only when the station delivers nothing; missing points not labelled for themselves). Both labels are kept unchanged, with their ids and review fields (see "Requalification" below).
|
||
|
||
Verdicts per old id:
|
||
- **WILDSTRUBEL-200**: kept. It is the only measurement of the range (31/10 20:15). Raw positions of all sensors shift together (x by about +3.3 px, y by about +10.4 px against 09/10). The tilt goes from about 0.20 (09/10) to 0.39. Seen target by target with `--rigid`, the targets do not move against each other. corr_y is shifted on PB1_2, PB2_2, PB_5, PB4 and PA2_2, and corr_x on PA_9. R2 is rejected. The reprojection error is high on PA_8, PB3_2 and R1.
|
||
- **WILDSTRUBEL-201**: kept. No measurement after 20:15 until 13/11 02:15. The metrics stop after the 20:15 synchronisation, with the voltage normal. The other stations (NORDEND, JUNGFRAU, CLARIDEN…) have full exports from 01/11 to 05/11, while WILDSTRUBEL's are empty.
|
||
|
||
## 1. My labels
|
||
|
||
### 2026-10-31T20:00Z → 2026-10-31T20:30Z (WILDSTRUBEL-200, kept)
|
||
- **Phenomenon of the catalogue**: "A step taken during a measurement outage" (Displacement), with the camera tilting ("The camera tilts little by little").
|
||
- **Precedents**: none useful. The `precedents.py` query returned Visibility cases from LENZSPITZE and NORDEND, which do not fit. The neighbouring WILDSTRUBEL labels 197–199 give the context of the outages.
|
||
- **Family**: Displacement. The cause is not established: a camera movement that the compensation does not fully remove. It is not Thermal, because a single point shows no cycle at equal temperature (housing at about 25 °C at 20:00). It is not Target issue, because nothing in the configuration changed.
|
||
- **Measured**: there are 13 sensors at 20:15, and R2 is rejected (error_value 0.98). The compensation runs on 5 references.
|
||
- Raw positions, against 09/10 07:00, through `overview.py --rigid`:
|
||
- common translation tx 4.2 px, ty 9.1 px;
|
||
- target residual: median 0.09 px, max 0.5 px (PA_9);
|
||
- reference residuals: up to 0.24 px (PA_8 dx).
|
||
- Pose: pan goes from −0.01 (08–09/10 nights) to +0.08, and tilt from 0.18–0.20 to 0.39. Rotation is −0.048, within its usual range. sigma0 is 0.0064, against about 0.003–0.004 before.
|
||
- corr_y daily medians, 09/10 → 31/10:
|
||
- PB1_2: −0.24 → +0.41
|
||
- PB2_2: −0.19 → +0.24
|
||
- PB_5: 0.10 → 0.19
|
||
- PB4: 0.04 → 0.10
|
||
- PA2_2: 0.04 → 0.08
|
||
- PA_8 (ref): 0.11 → 0.00
|
||
- corr_x: PA_9 goes from 0.16 to 0.51.
|
||
- reproj_err: PB3_2 0.39, PA_8 0.35, R1 0.17, against about 0.25 / 0.25 / 0.09 on 09/10.
|
||
- Outside the range: the same shifted levels are still there on 13/11 and 15/11, so the step lasts. The label correctly says that a single point within the range cannot tell.
|
||
- **Images**: the camera takes no image at night (08/10: last image 16:00, first 04:35). On 31/10 the images run from 04:00 to 13:00. Thumbnails at 09/10 12:00 and 31/10 12:00/13:00, for PB1_2, PB2_2, PA_9, R1 and R2: the spots are sharp, with no obstacle. R2's measured frame has moved about 21 px in x, against about 3 px for the others, and its spot is no longer at the centre. It latched elsewhere, which matches its rejection.
|
||
- **Confidence**: high on the facts, low on the cause (as the label says).
|
||
- **Remarks, not enough to correct the label**:
|
||
- The description gives the tilt but not the pan rise (x +3.3 px), which is also clear.
|
||
- "No image at that hour" is true, but the camera never takes night images here.
|
||
- The rule "a lasting jump hidden in an outage is bounded on the outage and the first measurement after" could argue for a label from 15/10 08:00. The validated choice applies the single-measurement rule instead ("bounded one rate step on either side"). It does not double the No data label WILDSTRUBEL-199, which already covers the outage. I keep it.
|
||
|
||
### 2026-10-31T20:15Z → 2026-11-01T00:00Z (WILDSTRUBEL-201, kept)
|
||
- **Phenomenon of the catalogue**: "The station goes silent on its own, for a long time".
|
||
- **Precedents**: the previous gaps of WILDSTRUBEL. 182 and 192 are energy gaps (battery low before), and 197 and 199 are silent gaps. This one has no energy signature, like 197 and 199.
|
||
- **Family**: No data. Every sensor is missing.
|
||
- **Measured**:
|
||
- Metrics: the last `last_sync` is 20:15:13 (the measurement), and there is no metric from 21:00 onward.
|
||
- Battery at 20:00: BusVoltage 13.17 V, well above the knee (about 12.8 V); BatteryPercent 80.6, the usual post-charge night value since 30/10, while it was 100 % during the day.
|
||
- The next WILDSTRUBEL measurement is 13/11 02:15. NORDEND, JUNGFRAU and CLARIDEN have thousands of export rows on 01–05/11, while WILDSTRUBEL's files are empty. So this is not a common or end-of-file delay.
|
||
- **Images**: none after 31/10 13:00. The camera takes no night images, so that hour proves nothing either way.
|
||
- **Confidence**: high.
|
||
- **Remark**: "although the battery was full" is loose: the battery was at its usual well-charged night level, not 100 %. The point the label makes, that there is no energy signature, holds, so it is kept.
|
||
|
||
### Spanning labels, read within the range
|
||
- **WILDSTRUBEL-199** (No data, 15/10 08:00 → 31/10 20:15): it holds within the range. There is no measurement from 20/10 to 31/10 20:15.
|
||
- Metrics show only two values on 25/10 (01:00, 03:00), then come back on 30/10 at 18:00, with a sync every 15 min.
|
||
- There are 10 daytime images on 31/10, from 04:00 to 13:00.
|
||
- Inaccuracy: "come back … the battery full". On 30/10 at 18:00 the battery was 78 % at 13.09 V; it reached 100 % only during the day of 31/10.
|
||
- From 30/10 18:00 the station was running (syncs, images) without measuring: this is the phenomenon "The station runs, the measurement is missing", which the description does mention.
|
||
- **WILDSTRUBEL-130** (Target issue R3, to 01/11): within the range, the only measurement is a night one (20:15), where R3 is accepted, as usual at night. The label's "no daytime measurement after 15/10" is true. It is consistent.
|
||
|
||
## 2. What I left without a label
|
||
- Requalification: no label removed and none rewritten. WILDSTRUBEL-200 stands out from the noise (PB1_2 corr_y +26 noises, details below). WILDSTRUBEL-201 is a full outage: no sensor is measured.
|
||
- R2 rejected at 20:15: it is a missing point. Under the new rule it is not labelled for itself. Its effect on the other curves (the targets shifted beyond the noise) is what label 200 carries, together with the camera tilt.
|
||
- The range holds a single measurement instant (31/10 20:15). Nothing else touches corr_x or corr_y.
|
||
- There is no thermal label: a single night point gives no daily cycle of the pose.
|
||
- R2 rejected at 20:15 is part of label 200, which names it. Outside the range (13/11, 15/11), R2 stays rejected with a drifting raw position. To be looked at in the next batch, as a possible new Target issue or a moved reference.
|
||
|
||
## Requalification
|
||
- Command: `python3 tools/scripts/ranges.py data/WILDSTRUBEL --columns corr_x,corr_y --k 5 --start 2026-10-01T00:00Z --end 2026-10-31T21:00Z --instants 2` (output in `work/requal_ranges_0110_3110.txt`).
|
||
- Why this window: the label's range holds only one measurement (31/10 20:15), so a level and a noise cannot be computed inside it. The window takes the level and noise of each sensor from the measurements of October before the outage (01/10 → 15/10 08:00).
|
||
- No `--rolling`: there is a single point after a three-week outage, so no local level exists around it.
|
||
- **WILDSTRUBEL-200** (Displacement, 31/10 20:00 → 20:30): **kept**. Targets beyond 5 own noises at 20:15:
|
||
- PB1_2 corr_y: +0.653 px, **+25.9 noises** (noise 0.025 px); about 3.6 mm (5.46 mm/px, 13.8 m).
|
||
- PB2_2 corr_y: +0.499 px, **+17.7 noises**; about 3.3 mm (6.57 mm/px, 16.5 m). PB2_2 corr_x: −0.160 px, −5.1 noises.
|
||
- PA_9 corr_x: +0.350 px, **+8.4 noises**; about 5.3 mm (15.06 mm/px, 39.1 m).
|
||
- PB_5 corr_y: +0.143 px, +6.4 noises; about 1.3 mm.
|
||
- Below 5 noises: PA2_2, PA3_2 and PB4.
|
||
- References, for context: PB3_2 cx −13.6 / cy +16.1, PA_8 cx +10.8 / cy −6.2, R1 cy −9.1, PA1_2 cy +6.3. R2 is rejected (error_value 0.985). R3 is accepted (0.43).
|
||
- The deviation is far beyond the threshold, so the label holds as it is. Its description rests on the shifted corrected curves, not on the rejection of R2, so it is not rewritten.
|
||
- **WILDSTRUBEL-201** (No data, 31/10 20:15 → 01/11 00:00): **kept**. `inventory.py --start 2026-10-31T00:00Z --end 2026-11-01T00:00Z` gives one instant present (20:15), and the next measurement is on 13/11 02:15. After 20:15 no sensor has any measurement, so the station delivers nothing, which fits the new No data rule.
|
||
- Counts: kept 2, rewritten 0, removed 0.
|
||
|
||
## 3. The documentation
|
||
- What helped:
|
||
- `families.md` and `labeling-rules.md`, for the No data rules (common gap against own gap, battery signature) and the bounds of a single measurement inside an outage.
|
||
- `reading-rules.md` §4, on the power-cut signature.
|
||
- `tools/DATA.md`, for the camera pose and its tilt sign.
|
||
- Tension between two bound rules: "a lasting jump hidden in an outage" and "a single measurement inside an outage" lead to different bounds for label 200.
|
||
- Precedents: none relevant from other stations.
|
||
- Requalification: the new §1 threshold was simple to apply. For a single measurement inside an outage, `ranges.py` needs a window that holds measurements on one side: this is not stated in the guide, but the `--start` margin solved it.
|
||
- No weather folder exists for WILDSTRUBEL (`data/WILDSTRUBEL/weather/` is absent), so there is no weather.
|
||
|
||
## Difficulties
|
||
- none
|
||
|
||
## Sources read
|
||
- `instructions/analyst.md`, `tools/GUIDE.md`, `tools/DATA.md`
|
||
- `labeling/families.md`, `labeling/labeling-rules.md`, `labeling/reading-rules.md`, `labeling/stations.md`
|
||
- `labeling/auto/WILDSTRUBEL.jsonl` (labels 130, 182–201), `work/existing.jsonl`, `work/spanning.jsonl`
|
||
- Station data through `inventory.py`, `metrics.py`, `summary.py` (daily, October to November), `overview.py --rigid`, `thumbnails.py`; the raw camera and BatteryPercent CSVs; image listings
|
||
- Requalification: `work/requalify_input.jsonl`, `ranges.py` (k 5, 01/10 → 31/10 21:00), `inventory.py` (31/10), `labeling/labeling-rules.md` §1–2 again
|
||
- Export row counts of NORDEND, JUNGFRAU and CLARIDEN for 01–05/11; `precedents.py` (one query)
|