Case studiesMetal fabrication · Carbon steel · United States
Carbon steel fabrication
A measured baseline for a plant that needs to triple its capacity.
- 75,000 sq ft
- Plant
- 3 days
- Camera install
- 2 weeks
- Kickoff to first results
“We've done more in a couple of weeks than we would have in six months of trying to figure this out by hand.”
Plant Production Lead
A US metal fabricator working in carbon steel operates a plant of 75,000 square feet. Jobs are long and welding is manual, with small crews working in cells. The plant has limited industrial engineering staff.
Revenue is growing more than 50 percent year over year, and the plant needs to roughly triple its capacity to keep pace. It is weighing three routes to get there: a redesigned layout, collaborative robots (cobots) for welding, and changes to how the work is done.
What it needed was a measured baseline against which to judge them. Leadership's view was that welding set the pace of the plant, and it wanted data to test that. As in most plants of its size, standard times predated the current product mix, and how the work varied from cell to cell was understood from experience rather than from measurement.
The first five weeks
What was delivered, and when
Week 0
Install
- Three days on site
- Cameras, network and on-site storage
- Every target view confirmed
Week 1
First readout
- Time use by weld cell
- Weld and grind time
- Shift profile
- Cell utilization
Weeks 2 to 4
Full measures
- Cycle times and throughput by station
- Work in process, queues, lead time
- SOPs for the core product
- Value stream map
- Ranked changes
Week 5 on
Continuous
- Measures keep updating
- Adoption of each change tracked
- Layout and throughput models
What Cobalt built
Cobalt installed cameras in three days, with complete coverage of the target processes and no interruption to production. The system observes those processes continuously, across every shift.
Generic vision models do not recognize the specific work of a fabrication cell, so Cobalt built tooling for this plant. An image classifier, trained and validated on human-labeled footage from the plant's own floor, tells its activities apart: welding, grinding, fit-up and handling, walking, and waiting. A vision-language model is configured to answer open-ended questions about a station, such as how long an operation took.
How the video becomes measures
Two models, built for this plant's work, watch the same video
Continuous video
Every covered station
Every shift
Image classifier
Trained and validated on human-labeled footage from this floor
Eight activity classes
Vision-language model
Configured for open questions
“How long did this operation take?”
Measures
Time use
Cycle time
Throughput
Flow
The activity classes
Defined with the plant, so that each one answers a question it asked
- Welding
- Welding or tack-welding in the cell, from short positioning tacks to long seams.
- Grinding and cutting
- Grinding surfaces or edges, removing excess weld, abrasive cutting or trimming.
- Tool work
- A tool is clearly in use, but the operation cannot be told apart reliably.
- Other hands-on work
- Fit-up, measuring and checking, handling and rigging, setup and housekeeping.
- Walking
- Moving within, into or out of the cell.
- Waiting
- No hands-on work under way. The class describes what is visible. It does not say whether the time was necessary.
- No one visible
- Nobody can be seen in the cell. Someone hidden behind equipment cannot be ruled out.
- Unresolved
- No usable classification was obtained for that moment. Reported, not hidden.
What the plant has now
About two weeks after kickoff, the plant had time-use measures for every weld cell, a profile of activity across the shift, and a view of how many cells were being worked at any moment. Waiting and walking appear as measured categories for the first time. Variation between cells doing the same work is now visible, which is the starting point for a shared standard method.
The plant can now see where its welding hours go, which is where raising throughput starts. Throughput, cycle times, a value stream map, SOPs for the core product and a working model of the plant follow. The same measures keep being reported, so any change the plant makes is judged on the same basis.
Time use by weld cell
Share of shift
- Value-added work
- Support work
- Waiting
- Walking
- Breaks and idle
| Station | Value-added work | Support work | Waiting | Walking | Breaks and idle |
|---|---|---|---|---|---|
| Weld cell A | 21% | 41% | 15% | 8% | 15% |
| Weld cell B | 17% | 43% | 18% | 9% | 13% |
| Weld cell C | 14% | 38% | 16% | 12% | 20% |
| Weld cell D | 24% | 42% | 13% | 6% | 15% |
Shift profile
Share of covered stations with hands-on work under way, by half hour
| Time | Hands-on |
|---|---|
| 06:30 | 31% |
| 07:00 | 58% |
| 07:30 | 71% |
| 08:00 | 74% |
| 08:30 | 72% |
| 09:00 | 38% |
| 09:30 | 61% |
| 10:00 | 73% |
| 10:30 | 75% |
| 11:00 | 69% |
| 11:30 | 22% |
| 12:00 | 49% |
| 12:30 | 68% |
| 13:00 | 70% |
| 13:30 | 41% |
| 14:00 | 57% |
| 14:30 | 29% |
Cell utilization
Weld cells with hands-on work under way, of 4, by half hour
Three routes to more capacity
The same measures and models inform all three routes, so each is judged against one baseline.
Layout redesign
A 3D model of the floor, built from the facility drawing and the video, compares layout options by walking distance, crane moves and expected throughput.
Cobots
The throughput model carries a cobot lane beside the manual cells. It shows how many units cobots could take, what that does to the constraint, and what the lane needs upstream to stay fed.
Process changes
Ranked changes to staging, sequencing and method come out of the time-use and cycle-time measures. Each one is linked to the video behind it.
20 to 30%
Expected improvement in throughput from the three together.
Start with one operation
Tell us which operation you want measured. Get in touch.