Files
aimsight-labeling/auto/analyses/NORDEND_2026-11-04T2112_2026-11-07T1412/report.md
T

85 lines
10 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# NORDEND: a spider web on the reference CTR03 makes all the curves jump
Selection: from 04/11 at 21:12 to 07/11 at 14:12 (UTC). Hint from the user: the detectors do not see the disturbances of `corr_x` and `corr_y`. These disturbances are a comb of peaks, mostly at night: at each peak, almost all targets jump together, for one measurement, then come back.
## What the data show
### 1. The peaks: the reference CTR03 rejected, the compensation falls to two references
- **When**: 38 measurements, all isolated instants or pairs: on 05/11 at 01:00, 01:15 and 03:30; the night of 05 to 06/11, from 23:15 to 09:45 (13 measurements); the night of 06 to 07/11, from 21:30 to 10:45 (22 measurements).
- **What moves**: at each of these instants, and only there, the reference CTR03 goes above its error threshold (0.40 to 0.58 for a threshold of 0.4) and the station discards it. The compensation then runs only on CTR02 and CTR04: their reprojection error falls to zero and `sigma0` is missing. The base changes and **15 to 18 targets out of 18** jump together: `corr_y` downward (median of the targets about −0.07 to −0.08, i.e. 20 to 30 times the noise for PR17 and PR18), `corr_x` a little upward (about +0.01 to +0.02). The most touched targets are PR17, PR18, PR11 and PR13. Everything comes back at the next measurement.
- Commands: `python3 tools/scripts/inventory.py data/NORDEND --start … --end …` (rejections of CTR03 and "0 kept; outside: CTR02 (empty), CTR03 (r), CTR04 (empty)" at the 38 instants); `scripts/pics_ctr03.py` (for each instant, the error of CTR03 and the number of targets more than 4 noises from their local level: 15 to 18 in `corr_y` at the 38 rejections, never more than 3 elsewhere, except at two instants of 05/11 read below).
- **What does not move**: the raw positions. Against the common movement of the targets, no target jumps at the rejections (`overview.py --rigid 2026-11-05T18:00Z`, residual of the targets of the order of a hundredth of a pixel). CTR02 and CTR04 stay under their threshold. No other target is rejected at these instants.
- **Level of proof of the mechanism**: **confirmed** (reference rejected, reprojection at zero, common jumps at the same instants, nothing in the raw positions).
![CTR03 against three targets: error_value at the top, corr_y at the bottom; grey zones: the labels of CTR03](fig1_ctr03_corr_y.png)
At the top, the error of CTR03 (blue) rises at once on 05/11 around 19:45 and lives right at the threshold until 07/11 around 12:30; at the bottom, each time it passes the threshold, PR17, PR18 and PR11 dive together for one measurement. Between two rejections, their curves are at their level.
![Reprojection and residuals of the references against the targets (overview.py --rigid)](fig4_rigide_reprojection.png)
Bottom panel: the reprojection of CTR02 and CTR04 falls to zero at each rejection of CTR03 (green crosses at the top); the compensation then runs on two references. Top panels: CTR03 moves slightly away from the targets during its two nights; the targets stay a block.
### 2. The cause: a thread of spider web in front of the lens, on CTR03
- A web stretched very close to the lens crosses the image in wide blurred bands, lit at night by the infrared, beaded with drops. It is there the whole batch. The wind moves it from one night to the next.
- On 05/11 at 01:00, a big bright blurred point of the thread sits on CTR03: this is the rejected measurement. At 04:00, it has moved toward CTR02; the error of CTR02 then rises until the morning, without rejection or effect on the curves.
- On 05/11 from 19:00-21:00, a band of the thread comes onto CTR03 and stays there two nights (thumbnails of 21:00, 23:00, 06/11 06:00, 07/11 04:00). The error of CTR03 goes from about 0.1 to 0.3-0.4 and stays right at the threshold, also by day (thread not lit, invisible). On 07/11 at 12:30, the error drops back; in the evening, the thread has moved beside the spot (image of 07/11 at 20:00).
- The same web bothers a few targets, without touching the other curves: PR09 on 05/11 in the morning (halo on its spot at 06:00, rejected several times from 03:15 to 09:30, then at 15:00), PR10 on 06/11 at 09:30, PR04 on 07/11 at 11:30, PR07 on 07/11 at 14:00 (bright band very close, image of 14:00). Only their rejected points are missing.
- Commands: `thumbnails.py --sensors CTR03,CTR02,CTR04` and `--zone 1550,700,700,600`; whole images by `ls data/NORDEND/images/ | grep …` then reading.
- **Level of proof**: **confirmed** for the thread on CTR03 the nights of 05 and 06/11 and at 01:00 on 05/11 (seen on the images of those hours); **strongly supported** for 03:30 on 05/11 and for the daytime rejections (no image at the instant, or thread invisible by day; same reference, same signature, thread seen the night before and after).
![Whole image of 05/11 at 23:00: the web in front of the lens](fig2_image_20261105T2300Z.jpg)
The bright blurred bands are the threads of the web, very close to the lens. The group of targets is at the bottom right; CTR03 is the highest point to the left of the group, under the band that descends toward the bottom right.
![The zone of the references CTR03 and CTR02, from 04/11 22:00 to 07/11 20:00](fig3_fil_sur_CTR03.png)
The thread moves: on CTR03 on 05/11 at 01:00, between the two at 04:00, on CTR03 at 21:00 (and all the following night), invisible by day (06/11 09:00, 07/11 10:00), then to the right of CTR03 on 07/11 at 20:00, when CTR03 has come back to its level.
![Whole image of 07/11 at 14:00: the web seen by day](fig5_image_20261107T1400Z.jpg)
By day, the web is seen against the light: a bright band descends through the group of targets, with a big bright point. It is at this hour that PR07 is rejected.
### 3. What also moves, unrelated to the web: the thermal cycle of the camera
- Every day, when the housing heats (housing temperature highest around 14:00 on 05/11, early afternoon on 06/11), the camera turns: the pan drops strongly, the tilt rises, the scale increases. It comes back slowly in the evening and all night, without rest. Morning turning points: 05/11 06:30, 06/11 07:00, 07/11 06:30.
- The compensation almost absorbs it, except on the right edge of the group: the `corr_x` of PR18 dips clearly in the middle of the day (7 to 10 times the noise), that of PR16 and PR13 to PR17 a little less.
- On 05/11 at 12:00, the camera makes a one-measurement jolt: the whole scene, references included, moves in the image. Many targets then drop a little in `corr_y` then come back. I put it in the cycle of the day.
- Commands: `metrics.py` (housing temperature, pose, `cib_corr_x`), `days.py --night 3-5 --bounds`, `scripts/cycle_jour.py`.
- **Level of proof**: **strongly supported** (the pose follows the housing temperature; the rock face does not move with the temperature, according to the station context).
### 4. What is missing: CTR01
CTR01 (local reference, read as a target, outside the compensation) is rejected a large part of the batch (124 measurements out of 260), as since 16/10. Its spot is small at night and almost invisible by day, when that of PR17 stays sharp. Its rejections do not make the other targets jump. Cause not established (**insufficient evidence**): state of the prism, or of the glass in front of it.
## Why the detectors see it badly (hypothesis)
Each peak lasts one measurement, and the next step is back. They also do not go through a rejected measurement of a target: the targets jump under their threshold, and only the reference is rejected. A detector that looks for ranges of rejections of the targets, or for deviations that last, does not see them. The signature to look for is a **reference rejected with the reprojection of the other two at zero**, and almost all targets jumping at the same instant. **possible**, not verified on the code of the detectors (outside what I am allowed to read).
## What remains not established
- Why CTR03 passes the threshold at these precise instants rather than at the neighboring ones: the thread covers it all night, its error oscillates around the threshold; the exact instant probably depends on the movement of the thread (wind), invisible at the rate of one image per hour.
- Between two rejections, CTR03 accepted but veiled is slightly shifted against the targets (of the order of a hundredth of a pixel): its effect on the pose is under the noise of the curves; I did not label it.
- The state of CTR01 (prism or glass); the nature of the rejections of PR04 at 11:30 on 07/11 (nothing on the thumbnail, cause deduced).
- After the batch: PR07 is rejected almost without stopping from 08 to 10/11, and the measurements are missing from 11/11 00:15 to 12/11 10:30. Outside the selection, not read.
## Labels
- 04/11 21:15 → 05/11 06:30, Thermal displacement: end of the cycle of 04/11, the camera comes back at night.
- 04/11 21:15 → 07/11 14:00, Target issue: CTR01 right at the threshold the whole batch, curve with holes.
- 05/11 00:45 → 01:30, Obstruction: thread of the web on CTR03 (seen at 01:00), CTR03 discarded, all targets jump.
- 05/11 03:00 → 09:45, Obstruction: thread of the web on PR09 (seen at 06:00), rejected points.
- 05/11 03:15 → 03:45, Obstruction: CTR03 discarded for one measurement (03:30), all targets jump; thread deduced.
- 05/11 06:30 → 06/11 07:00, Thermal displacement: camera cycle, with the jolt of 12:00.
- 05/11 14:45 → 15:15, Obstruction: PR09 rejected for one measurement, thread deduced.
- 05/11 23:00 → 06/11 10:00, Obstruction: thread on CTR03, first night, comb of peaks on all targets.
- 06/11 07:00 → 07/11 06:30, Thermal displacement: camera cycle.
- 06/11 09:15 → 09:45, Obstruction: PR10 rejected for one measurement, thread deduced.
- 06/11 21:15 → 07/11 11:00, Obstruction: thread on CTR03, second night, comb of peaks on all targets.
- 07/11 06:30 → 14:00, Thermal displacement: camera cycle, until the end of the batch.
- 07/11 11:15 → 11:45, Obstruction: PR04 rejected for one measurement, thread deduced.
- 07/11 13:45 → 14:00, Obstruction: PR07 rejected at 14:00, thread seen on the image.
For the client: the structure does not move. The peaks of the two nights are not valid measurements: the station had only two references left. The web in front of the lens must be removed (glass to clean); the problem will return as long as it stays.