High-Volume Manufacturing Scrap Reduction

A high-volume production area running multiple parallel lines, with automated process steps, in-line inspection, and MES-based defect reporting by station and defect code. Scrap was reported daily and reviewed weekly.

The area was not in crisis. That is actually the harder version of this problem: scrap sat at a level everyone had learned to plan around. It appeared in the budget as a fixed rate rather than as a loss with a cause, and the standing explanation was operator technique and material variation — the two explanations that are least actionable and least falsifiable.

Scrap performance showed recurring defect modes that could not be resolved through operator adjustment alone. Corrective actions had been attempted before and had produced short-lived improvements that decayed within weeks — the signature of containment being mistaken for corrective action.

Stated as a question the work could actually answer: are these defects the output of a stable process running at its natural capability, or the output of an unstable process with assignable causes that have never been isolated?

• Production could not be stopped for experimentation; analysis had to run on live production data and short, scheduled trials.
• No capital was available at the outset, so the first round of solutions had to be process and standard-work changes.
• Defect coding was inherited and partially ambiguous — multiple codes described overlapping failure modes.
• Any change had to survive across three shifts and a rotating operator population.

I owned the improvement agenda for the area and led a nine-person team. I ran the analysis personally, set the trial plan with production leadership, and owned the PFMEA, control plans, work instructions and inspection criteria that the changes were written into. Maintenance, tooling and quality inspection each owned specific actions.

Correcting the defect coding changed the answer. Once the two overlapping codes were separated, the recombined Pareto showed that the largest true contributor was not the mode the area had been working on. It was concentrated on specific tool positions and correlated with time since the last tooling intervention — which pointed away from operator technique and toward a process window that was drifting between maintenance events.

Control charting the associated process characteristic made the mechanism visible: the process was not unstable in the sense of random upsets. It was drifting predictably, spending progressively more time near the specification limit until it crossed. The mean looked acceptable in the weekly summary the whole time, which is exactly why a monthly average review had never detected it. Capability computed over the full interval was materially worse than capability computed just after an intervention.

That reframed the problem from “why do operators produce this defect” to “why is the process allowed to drift to the edge of the window before anything triggers.” The second question has engineering answers.

• Defect codes separated and the coding standard rewritten, with inspector re-training so the data stayed trustworthy.
• Process window tightened and the intervention trigger changed from calendar-based to condition-based, using the characteristic that predicted the drift.
• Control charts placed on that characteristic at the station, with reaction rules written into standard work — what to do, by whom, at which signal.
• PFMEA updated for the mode, with the detection rating re-scored against the new in-process control rather than end-of-line inspection.
• Control plan and work instructions revised; inspection criteria aligned so the station-level check and the final check were measuring the same thing.
• Weekly scrap review changed from reporting the average to reviewing the control charts, which is the only version of that meeting that can catch a drift early.

Scrap fell 15% across the area. A parallel value-stream mapping program run with the same team captured an additional $180K. Gross utilization on the area rose 20% — roughly $250K annually — through time studies, line balancing and standard work, and changeover time fell 18%.

The result I care about more is that the improvement did not decay. The gain was held by a condition-based trigger and a control chart with reaction rules, not by sustained attention from the engineer who found it.

Verify the measurement system before trusting the Pareto. The single highest-leverage hour of this project was spent discovering that two defect codes meant the same thing to some inspectors and different things to others. Every analysis downstream of bad categorization is confidently wrong.

Averages hide drift. A mean inside specification is perfectly compatible with a process that is spending half its life at the edge of the window. Reviewing summary statistics instead of control charts is how a problem survives years of attention.

Containment is not corrective action. The previous attempts had all worked — briefly. They worked because someone was watching. The test of a corrective action is whether it still works when nobody is.

Defects attributed to people are usually attributable to process windows. “Operator technique” is rarely a root cause. It is usually a description of what happens when a process window is too narrow to run reliably without exceptional care.