How Shift and Operator Differences Move Your Thresholds

The same vision system, five false alarms in a day shift and thirty in a night shift: a very common complaint after go-live. The first instinct is to loosen the threshold, at the cost of more misses — the problem moves rather than resolves. The more effective sequence is to separate the causes first. Did the images themselves change, through ambient light at shift changeover, daylight through windows, or focus drift as equipment warms up? Did the incoming material genuinely change overnight, with a different supplier or a different storage age? Or do the crews really work differently, with different placement, pace and cleaning habits? The three have entirely different remedies, and only after separating them does tuning make sense.

Split the data three ways first

A single overall false alarm count tells you nothing about cause. Split it at least three ways: by shift, by check item — whether every item rose or only one or two — and by hour, whether the increase is uniform across the shift or concentrated after start-up or before handover. Once split, conclusions often surface on their own. If the rise concentrates on one check item, that item's imaging condition is probably the most light-sensitive. If it concentrates in the first hour, the usual explanation is a cold machine, or state left behind by the previous shift. If it is uniform across the shift, incoming material or operator difference is more likely. None of this needs any investment beyond putting shift and check item into the write-back fields, which is exactly why field design insists on them.

Per-shift thresholds are allowed, but must leave a trail

If the cause is environmental and cannot be removed at source in the short term, giving different shifts different thresholds is an acceptable engineering compromise. It needs two companions. First, threshold versions must be recorded, so that every change to per-shift thresholds carries a version number written into each verdict record. Second, there must be a review path, so that two standards are never running unnoticed — in a quality system that is not acceptable, because a part passing on days and failing on nights becomes a quality incident the moment it is traced. The safer position is therefore to remove the difference at source where possible, treat per-shift thresholds as a transitional measure, and state clearly in the process documentation which of the two you are currently in.

The other half: operator rotation

Beyond shift timing, rotation has an easily missed effect. The same station in different hands sees different placement attitude, different seating depth, and even different surface condition. A system sensitive to attitude shows this as more false alarms after a crew change. The right fix is usually not to loosen the criteria but to add a little constraint in the fixture — a locating block, a datum edge — which suppresses the variation far more reliably than teaching the algorithm to tolerate every pose. How camera and fixture work together is covered in camera mounts and fixtures. One more thing worth stating in advance: new operators typically run both a higher defect rate and a higher false alarm rate for their first weeks. That is normal, and if the acceptance period happens to cover it, both sides should mark those weeks out when reading the numbers.

Periodic review beats one-off tuning

Lamps age, lenses collect dust, fixtures wear, incoming material changes. All of these accumulate slowly, none of them triggers an alarm, and so none of them announces itself. By the time the line notices more false alarms, the drift is usually weeks old. The countermeasure is a cheap periodic review: every so often, run a fixed set of standard samples or standard parts through the system and check whether the verdicts still match the previous run. If they do not, something in the system state changed — go and look at lighting, cleanliness or thresholds. Equipment or quality staff can do this in a few minutes, and it catches most drift early. Model-level drift and when to retrain are covered in model drift and recalibration.