WORK ORDER — Maintenance and Improvement Record
Work order no: WO-A-2026-03 Date issued: 2026-09-11 Category: Vision station lighting
Machine / cell: Inspection Station IS-1, 12-megapixel camera with a 100 mm white ring light
Status: Closed Engineer: M. Okafor (vision) Total labour: 7.5 hours Parts cost: EUR 96.00
Reason for work: intermittent false rejects on the machined part surface check
1. Fault description as logged by the shift team
Thursday 06:40, operator record: the IS-1 station rejects scattered good parts through the day, the reject rate jumps from the usual 0.4 percent to anywhere between 1 and 6 percent, and there is no pattern in which part, which batch, or which hour, which drives quality mad because the same good part passes one shift and fails the next. The station reads a 0.4 mm machined mark on the flange face of a pump housing and must not call a clean part defective. The vision team ran their own checks, the camera, the lens and the part handling all behave, so the complaint spent three days pointing at the measurement recipe until I got the call on Tuesday morning.
The station uses a strobing ring light that fires a short pulse synchronised with the camera shutter to freeze the part and kill ambient light, and the pulse width is 90 microseconds on a 120 Hz trigger. That design choice is the single most common source of a statistical reject problem, because a strobe light that flickers in brightness from one trigger to the next changes every measured grey level, and a grey level change near the threshold looks exactly like a defect the part does not have. I brought the light meter and the oscilloscope rather than another recipe revision, because a reject pattern that ignores parts, batches and hours usually lives in the strobe, not in the algorithm.
Evening note to the file: a vision station thinks in grey levels, and a strobe that varies its pulse energy writes fake defects into every frame it serves, so the light source is the first witness to interview, not the last.
2. On-site inspection record (Day 1)
I clamped the light meter into the inspection fixture in the exact part position and logged the integrated pulse energy over a thousand consecutive strobe trigger events, and the record was a rollercoaster: the pulse energy held a steady 42.0 lux-microseconds for long stretches, then dropped to 31.5 for a few frames, then recovered, with no pattern in the dips. The histogram showed two clean populations, the design value around 42 and a shadow population around 32, and a station that switches between two light levels is literally a different measuring machine every time it swaps, so the grey level of the same clean surface swung by the equivalent of the 0.4 mm mark we are trying to find. The oscilloscope on the strobe driver told me why: the trigger-to-light delay drifted by up to 35 microseconds from one pulse to the next, so the 90 microsecond pulse sometimes clipped against the camera exposure window, and each clip shaved pulse energy off the frame exactly the way the two populations in the histogram said.
I isolated the cause to the strobe driver module, a compact constant-current pulse driver that switches the ring light LED current at a nominal 2 A peak, and traced the drift to its control loop: the driver’s pulse timing comes from an RC-oscillator stage whose timing capacitor sits next to the LED power stage on the same small board, and the LED stage heating shifts the capacitor value, so the pulse width and the trigger delay breathe with the board temperature. Cold board, full pulse, hot board, clipped pulse, and the station changed its mind about good parts as the driver warmed and cooled through the day. The camera exposure window itself was stable, the software exposure was fixed at 180 microseconds, so the whole fault was the strobe driver behaving like a slow heart that skips beats whenever it gets warm.
Evening note: the two-population histogram is the fingerprint, the vision team saw two grey levels on one clean part and called it part variation, but the part never changed, only the strobe pulse energy did, and the scope found the timing drift that hid the two light levels in the same room.
3. Corrective action taken
I replaced the strobe driver module with the temperature-stable version, a proper constant-current driver with a quartz-timed pulse stage and the timing circuit moved off the LED power plane, which the maker specifies for rough duty in a machine cell rather than for a bench, and I added a 3 mm thermal gap pad under the new board so the LED heat no longer launders into the timing network. Before I fitted it I logged the trigger-to-light delay on the old driver over a two-hour warm-up and saw the same 35 microsecond drift grow and shrink with the board temperature, which confirmed the diagnosis before the new part went in. The new driver holds the trigger-to-light delay at 12 microseconds with a 2 microsecond jitter across the whole warm shift, and the pulse energy histogram collapsed from two populations to one tight spike at 42.4 lux-microseconds with a 0.4 percent spread, so the camera now shades every clean part the same way at 07:00 and at 17:00.
I also reviewed the light mounting while the ring was off: the ring bracket allowed the light to sit 2 mm high on one side, enough to tilt the illumination plane by 1.1 degrees and cast a soft gradient that had been adding a slow drift to the grey level, so I shimmed the bracket flat to the fixture face and re-checked the tilt with the dial indicator at 0.05 mm across the ring diameter. I set the exposure back up and ran the station reliability test, and the false reject rate on clean reference parts dropped from the wandering 1 to 6 percent to a steady 0.1 percent.
| Check item | Before | After | Spec |
|---|---|---|---|
| Pulse energy (histogram) | 42.0 / 31.5 two populations | 42.4 single spike | ±2% |
| Trigger-to-light delay | 12 to 47 μs drifting | 12 ±2 μs stable | ±5 μs |
| Ring tilt across diameter | 1.1° | 0.05 mm / <0.1° | <0.2° |
| False reject rate, clean parts | 1 to 6% | 0.1% | <0.5% |
The maintenance follow-up sheet now lists the strobe driver as a logging item, image brightness statistics recorded hourly by the station controller, so a pulse energy drift shows up as a trend in the reject log instead of a mystery after three days.
4. Acceptance result after the repair
I ran the released station across two full production shifts, roughly 4,800 parts through the surface check, and the false reject count over that run was 6 parts, half of them borderline cases on actual machining burrs rather than light faults, which puts the measured reject rate at 0.12 percent against the 0.5 percent acceptance limit. Quality took the station back immediately, and the process log shows the hourly brightness statistic flat at 42.4 lux-microseconds plus or minus 0.4 percent through both shifts, so the station is no longer two different measuring machines depending on the weather inside the cell. The 0.4 mm machined mark reads with a sharp 18 grey level contrast on the clean background, and the previously cursed clean parts pass every time, so I closed the order with the station released for production.
5. Hours and parts list
- Labour — 7.5 h total: pulse energy and scope inspection 3.0 h, driver replacement and thermal pad 1.5 h, ring tilt re-mount and shim 1.0 h, two-shift release verification 2.0 h
- Parts — temperature-stable strobe driver module EUR 62.00, thermal gap pads EUR 9.00, shim set EUR 5.00, light meter floor stock EUR 20.00 usage; total EUR 96.00
- Scrapped in the fault window — none, the false rejects went to the re-check rack and passed under the repaired light
6. Retrospective notes for the next engineer
A vision false-reject problem that jumps between one and six percent and ignores parts, batches and shift hours is a lighting problem until proven otherwise, because the algorithm and the optics do not change with the temperature, but a strobe timing stage sitting on the LED power plane does. The two-population pulse energy histogram was the clue the team had been printing for three days without reading it, both grey levels on one clean part were two states of the same light, not two states of the part. Put the hourly brightness log on the station, because a trend line is worth more than a surprise, and when you replace a strobe driver in a hot cell, buy the quartz-timed version and keep the timing circuit away from the LED heat, it is the whole difference between a stable inspection light and a slow heart that skips beats.
Attachments: pulse energy histogram, scope capture of trigger-to-light delay, ring tilt measurement sheet, two-shift release record (filed with WO-A-2026-03).
Glossary — strobe ring light: circular light pulsed in sync with the camera shutter; pulse energy: the integrated light delivered in one strobe pulse; lux-microsecond: unit of the light × time product; trigger-to-light delay: time from camera trigger to the light pulse peak; false reject: a good part the station calls defective.
Field note — why a strobe light was right and what the selection rule should be
It is fair to ask why the station strobes at all instead of running the ring light continuously, and the answer is the arithmetic of exposure and motion. The part stops in the fixture with a residual 0.3 mm positioning settle, and the camera sits 190 mm away with a 25 mm lens, giving a depth of field of about 6 mm and a working field of 64 by 48 mm, so the part needs to be frozen during the 180 microsecond exposure or the motion blur smears the 0.4 mm mark into a ghost that the recipe reads as a shallow defect. A continuous light is fine for a slow inspection, but the takeaway for the next engineer is that the strobe was the right choice and the driver was the wrong part, the timing drift, not the concept, was the failure. When selecting a strobe source for a measuring station, check three numbers on the data sheet before buying: the trigger-to-light delay stability across operating temperature, usually quoted as a jitter figure, the pulse energy repeatability frame to frame at full current, and whether the timing stage shares a board with the LED power stage, because shared boards are exactly where this drift came from. A source that delivers a constant pulse energy at both ends of a hot shift keeps the grey level fixed, and a fixed grey level is the only honest foundation for a 0.4 mm defect decision. The ring tilt rule is the second quiet lesson: a 1.1 degree tilt made a gradient that the histogram averaged away most of the time but that leaked through on borderline parts, so set the light plane parallel to the part plane within 0.2 degrees and shim the bracket, not the recipe, because the bracket is a quarter hour of shimming and the recipe is a week of tuning a machine that was never the problem.
