NexTier, Pearsall · MSU2
Event review · Oct 24–25, 2026
MSU2 red zone review
Event Review - C.Lobozzo - Tuesday November 3rd 2026. Post DEMOD - Archived Video Only
Oct 24, 11:36 to 11:48: people moved in the red zone while armed, and nothing was captured
All five MSU2 views in sync. Spreadsheet: "Other camera angles miss recording of individuals in the red zone when pop-off blew" (11:41:37).
11:36:00
Armed
Armed (red overlay on the quad)Not armedMoment noted belowClick any video to enlarge it.
Source: Deep Vision recording, MSU 2 (rntl-scl186). Arm state read from the "● ARMED" overlay on Chan4, one sample per second.
What the video shows
| Time | View | Observed |
|---|
Click a row to jump there. Observations are from the recording only; no detections or alerts from that day are on record.
Why nothing was captured
- The zones were in place. Drawn on the Oct 24 views, the camera's zones still cover the pumps (Chan1), the pumps and manifold (Chan2) and the wellhead (Chan3), close to the customer's target drawings.
- The people seen at 11:40 to 11:41 were outside the zones. The two workers by the forklift (Chan1) and the worker left of the pumps (Chan2) are in the open yard, not in a red zone, so no alarm is the expected result for them.
- Chan4 has no zone, so movement there cannot alarm.
- The system was disarmed twice in this window (11:41:45 and 11:43:20). Anything after 11:43:20 could not alarm.
- Not confirmed: the zones were last read from the camera on Oct 16. We have no record of a change after that, but nobody read them on Oct 24.
ZonesWhat the zone should be, what the camera had, and where it fell on Oct 24
Target: blue outlines drawn on the Oct 15 zone-violation report. Camera zone: red = alarm area, blue boxes = areas excluded from alarming, as read from the camera on Oct 16. Hover over any picture to enlarge it.
Spraying: six events in the spreadsheet, all on the MSU2 PTZ
Pick an event in "Watch the events"; the up-close and stills sections below switch with it. Green boxes mark spray wherever the detector found it: the pop-off, 15:39 and the unconfirmed 06:31 haze. The 16:19 and night events have no boxes because the detector did not find them. A white spray or mist plume over the pump manifold. Clear in daylight; a low-contrast haze at night under the site lights.
Watch the eventsAll five MSU2 views in sync, one event at a time
Listed time
Time listed in the spreadsheetMoment noted belowGreen box on the PTZ: probable spray found by the detector (held 1.5 s)Click any video to enlarge it.
| Time | View | Moment |
|---|
Source: Deep Vision recording, MSU 2 (rntl-scl186), full 5-minute window around each listed time (12 minutes for Oct 25 01:29). Review copies at 640 px, 10 fps.
Up close
Stills around the listed time
| Listed time | Spreadsheet note | Seen in the stills |
|---|---|---|
Source: Deep Vision, MSU2-Cam1 (PTZ), full frame, one still at each time shown. The PTZ sat on the same view for every event.
Spray can be detected by mutating the video
Offline tests on the MSU2 PTZ recordings from the spreadsheet events. Early results from six events, not proven accuracy.
1 · ViewsCan a different picture make spray stand out
Which view shows spray best
Green box: the spray, found automatically from the processed video and drawn in the same place on all eight views. The dawn frame has no box because nothing stands out.
| View | Result |
|---|---|
| High contrast | Slightly easier to see the plume by eye. Also brightens everything else, so it adds no separation. |
| Low contrast | Worse. Spray is already low contrast; this washes it out. |
| Black and white | Worse. Removing colour makes spray harder to see. |
| Processed view A | Easier for a person to spot the plume than the original. |
| Processed view B | Shows the spray clearly in daylight, with some noise from glare. |
| Processed view C | Too noisy to use; the whole frame lights up. |
| Processed view D | The cleanest result. Dark everywhere except the spray. |
Camera settings: no contrast setting on the camera makes spray stand out; the processed views are done on the video, not in the camera. What may help on the camera, to test live: keep colour on, and reduce heavy image processing such as WDR. A thermal camera would show spray far better at night.
2 · Grid detectorTested on the recorded events
Grid test: catches daytime spray, misses both night events
The PTZ view is split into a grid of tiles, and each tile is checked against its own recent history. Red tiles are changing in the way spray changes them. An alert needs several neighbouring red tiles for a short time. Changes across a large part of the frame at once are ignored, because that is the camera refocusing, not spray.
Grid alert (probable spray)Combined: processed view D under the grid; orange = rising, red = over the alert levelMoment noted below
| Time | Moment |
|---|
| Listed time | Note | Grid result | Status |
|---|---|---|---|
| Oct 24 11:41:37 | Pop-off blew | Alert 11:41:38 | Caught |
| Oct 24 15:39:00 | Alert 15:39:07, repeating until 15:39:52 | Caught | |
| Oct 24 06:29:40 | Alert 06:31:02 to 06:31:16 over the manifold, 90 s after the listed time | Unconfirmed | |
| Oct 24 16:19:56 | Cycled for 47 sec | No alert. Nothing visible in the video either | Not seen |
| Oct 24 11:30–11:50 | 20 min, daylight | One false alarm at 11:32 (camera refocus), removed by a rule that ignores frame-wide changes | 0 false alarms |
| Oct 25 01:31:00 | Ran till 1:39:20 with spraying | No alert in 12 minutes. The score came close once but never held | Missed |
| Oct 25 02:27:22 | Night | No alert in 5 minutes | Missed |
Source: Deep Vision recordings, MSU2-Cam1 (PTZ). Line = detector score each second; dashed line = alert level; red ticks = alert; blue marker = time listed in the spreadsheet.
3 · PlanHow we would get to a working alert
Detection plan
| Step | What | Gives us |
|---|---|---|
| 1. Prove it on more footage | Run the grid over full days of MSU2 PTZ recording (day and night), and over the night events once downloaded. | Real catch rate and false alarms per hour, day and night. The current numbers come from 45 minutes. |
| 2. Review database | Every alert saved with time, camera, tiles, a 10 s clip and a still. Reviewers mark it spray, not spray, or unsure. | The false-alert and accuracy log asked for, and labelled examples for training. |
| 3. "Probable spray" check | Each grid alert's clip goes to a vision model: spray or mist yes or no, and where. Only "yes" goes out as an alert. | Fewer false alarms. Adds a few seconds. |
| 4. All cameras, PTZ first | Same grid per camera view. The PTZ needs a separate baseline per preset, and pauses while moving. | Coverage on the quad channels too. Spray is faint on the quad at this distance, so the PTZ stays the main sensor. |
| 5. Add pressure | Pump pressure from the frac van as a trigger; video confirms and records. | The most reliable pop-off signal. Needs their data feed. |
| 6. Trained model | Once there are hundreds of reviewed examples, train a spray model; later, on the camera itself. | Better night performance and fewer rules to tune. |
Expectation to set: this is a "probable spray, review it" alert within seconds. It is not a shutdown interlock. At night the current method misses spray entirely (0 of 2 on Oct 25); night needs a different approach before it can be offered.
A thermal camera should see spray day and night; how far depends on the lens
We have no thermal footage from the site, so this tab is an illustration plus arithmetic. The sensor sizes, lens options and spray sizes below are typical values for planning, not a product quote.
What is and is not simulated. Simulated: how many thermal pixels would cover this view, and where the spray sits (taken from our processed video). Not simulated: real temperatures. The background shading is the visible image turned grey, standing in for "warm equipment". Only a real thermal camera on site can confirm the contrast.
1 · AppearanceIllustration, not real thermal footage
What it would look like
Night panel: we could not find the spray in the night video, so the 15:39 plume shape is placed on the night frame to show the idea. In real thermal, site lights would not glow; they are bright here only because the visible image is the stand-in background.
2 · Can it detect sprayWhy it should work and what could go wrong
Why thermal should work
- Spray is cooler than what is behind it. Droplets cool by evaporation, and frac fluid is near ground temperature, while pumps, engines and exhaust run hot. That is a temperature difference, not a lighting difference.
- Night stops mattering. Thermal needs no light, so the glare and haze from site lights that beat the visible method go away.
- Wet stays visible. Equipment and ground that got sprayed stay cool for a while, so there is evidence after a 2-second pop-off.
- Liquid water blocks thermal light. A dense spray shows as a solid cool shape, not a faint haze.
What could still go wrong
- Fine, thin mist may only show weak contrast. Unknown until tested.
- Steam and exhaust look like plumes too, but they are hot, not cool. Cool-versus-hot is the main way to tell them apart.
- Rain and wash-down cool everything; they need their own rule.
- A thermal PTZ that moves loses its baseline each time. A fixed thermal camera on the manifold is simpler.
3 · RangeHow far from the camera spray can be seen
How far away can it detect spray
Range = spray width ÷ (pixels needed × size of one pixel at 1 m). "Spotted" = 4 pixels across, enough to trigger a check when contrast is strong. "Confident" = 10 pixels across, enough to recognise the shape. Pixel size at 1 m = sensor pixel pitch ÷ lens focal length. Air at these distances has little effect on thermal.
Where the spray was on Oct 24
| Event | Estimated distance from MSU2 | Note |
|---|---|---|
| Pop-off, 11:41:38 | about 18 m | Right side of the PTZ view |
| Plume, 15:39 | about 26 m | Upper middle, over the manifold |
| Far rows of the manifold | about 34 m or more | Top of the PTZ view |
Distances are estimates from a depth model run on one PTZ frame (Depth Anything V2), not measurements. This kind of model tends to underestimate far distances, so treat these as minimums. A site measurement of MSU2 to the manifold would replace them.
What this means: every spray in these examples was within roughly 20–40 m. At that range, even a mid-size thermal sensor with a standard lens covers a 1 m burst confidently, and a 30 cm leak is spottable. The open question is contrast, not distance. The next step is a short test with a thermal camera pointed at the manifold during pumping.