240 Rejected Units a Shift Eliminated by SPC Control Charts on a Mitsubishi Robot Weld Cell

Run the arithmetic and a shift losing 240 units works out to one bad part roughly every 105 seconds against a 42-second cycle time. The takt gap can hide that loss only while a downstream buffer keeps absorbing the miss. The cell builds a bracket subassembly with resistance-spot and MIG welds, with a six-axis arm feeding parts into a fixed weld gun. Nobody had changed the robot program. Rejects had climbed from a baseline near 30 per shift over about nine weeks, and the shop-floor explanation centered on electrode wear.

The response was to dress electrodes twice as often as the maintenance sheet required. The extra work showed up in the labor record, and the reject count did not move. The cell was spending more time on a suspected cause that gave back no quality improvement.

What the reject sheet showed

Every reject tag carries a defect code. Most bad parts here were coded as insufficient penetration, with the other codes taking a much smaller share.

That uneven mix pointed toward heat input. Insufficient penetration on a MIG weld follows heat input, and heat input is current times voltage over travel speed. Travel speed on a robot is fixed by the program. The power source holds voltage in a tight band. Current was the remaining variable, and it was not part of the daily review.

Before any chart went up at the cell, two weeks of the weld controller’s internal log were pulled. The older controller overwrites that file on a rolling window of about three days, so the export had to happen before the rollover erased more records. The file yielded several weeks of weld-current samples, each taken at the point of the weld. Plotted as individual samples against time, the data formed a cloud with no practical trend to act on.

How the charts caught the current drift

The first setup paired an X-bar chart with an R chart, both built from subgroups of five consecutive welds sampled once an hour. The centerline for average weld current sat where the process had been running well, and the control limits follow from that centerline plus and minus A2 times the average range. With subgroups of five, A2 is 0.577 from the standard Shewhart constants, which puts the limits a modest band above and below the centerline. Those bounds are what turns a shapeless cloud of readings into a line that either sits inside its expected spread or does not.

Hourly terminal monitoring carried an unavoidable lag. The signal existed only after the subgroup had been sampled, entered, calculated, and reviewed by someone with the authority to stop production.

During live use, the average-current chart signaled a few hours after startup. Three consecutive subgroups fell below the lower control limit. Under Western Electric rules, repeated points past a control limit indicate special-cause variation, and chance alone puts three points that far below the mean at under one percent.

That created a familiar procedural bind. The chart called for a stop while the robot was still cycling and routine production checks were still passing. Waiting another hour for confirmation meant continuing to build parts through a process already outside its own calculated limits. The rule on paper said stop; the parts coming off the fixture looked fine to the eye.

After the stop, the cause traced to a contact tip that had eroded enough to raise resistance while still passing routine checks. The arc at that station was starved of current. The tip was replaced during the stop, and the next sampled subgroup came back onto the centerline.

Weld-time range inside the five-part subgroups widened whenever the wire feeder liner packed with debris and caused an intermittent feed stutter. The range moved before the average did, because a stutter first spreads readings apart inside the subgroup. Watching the range chart gave roughly a shift and a half of warning before the X-bar chart reacted.

Those two fault patterns behave differently in the weld. Contact-tip erosion pulls all five current readings down together, so the subgroup average slides. Feeder stutter is less uniform. It widens the readings inside the subgroup first, and the average may hold near its centerline for a while.

Same range, different average

Take a constructed five-weld subgroup used only as an illustration, not as logged production data. The readings sit low and close together, only a handful of amps apart. The average lands well below the centerline, and the range stays small because the readings agree with one another.

A healthy subgroup can be scattered by that same handful of amps while staying centered where the process should be. The two subgroups can share the same range while their averages sit far apart. Internal variation is acceptable in each case, even though the current level has dropped in the first.

Level loss does not require wild scatter. Current can fall smoothly, with five tight readings at a time, while the R chart stays calm and the X-bar chart drops below the lower control limit. Sorting individual readings on the floor would bury the shift among ordinary sample noise. The paired charts track subgroup average and subgroup range independently.

A range-only display would have caught the feeder problem early and missed the eroding tip, since the tip dragged the whole subgroup down without pulling the five readings apart.

The gauge has to be trusted first

The penetration checks behind the reject codes are measured values, so the measurement system has to be reliable before any control chart built from those values is worth anything. In a setup like this, a subset of parts goes to a coordinate measuring machine for weld-bead cross-section verification. The calibration certificate is current, and repeatability is checked against a known reference artifact that sits well inside the tolerance band for bead geometry. A chart built on a drifting gauge sends maintenance after the measuring system instead of the weld process.

The maintenance rule that changed

The contact tip had been replaced on a fixed interval. The charts showed that tip degradation produced a current signal an average of two shifts ahead of the scheduled swap. They also showed that some tips ran well past the calendar interval with no current drift at all.

Moving the tip from calendar replacement to chart-triggered replacement did not go cleanly at first. Two early stops were false alarms traced to a data-entry transposition. For a stretch, chart signals were treated as advisory, so some drifting tips continued running until the next scheduled swap.

The rule became automatic only after the standard-work sheet was revised and a second round of buy-in was completed. Once below-limit runs triggered stops, tip consumption fell and starved-arc defects faded out. The scheme fits inside a total productive maintenance program because the chart supplies the condition signal for the one component whose wear moves the defect rate. Other scheduled tasks stayed on their existing intervals, since the data gave no reason to change them.

The number after four weeks

Real processes plateau. Insufficient-penetration rejects dropped sharply during the first full week with both charts running, then a residual band settled in. The remaining defects were positioning faults, tied to a different cause and a different charting problem, which kept total scrap from reaching zero.

Total scrap per shift landed in a band well below the original 240. The leftover defects belonged to a failure mode the two charts were never designed to detect.

Overall equipment effectiveness needs a narrow reading here. Availability barely moved. The improvement came through the quality factor, because more parts passed the gauge on the first pass. Throughput held steady, since the line had already been meeting takt on good parts until the buffer ran dry.

The history before the charts went live had to be reconstructed from reject sheets, because the earliest controller files had rolled over and disappeared. The climb from 30 rejects per shift to 240 survives in the reject history without any weld-current records behind it. Nine weeks of baseline drift happened inside a window the controller log no longer covers. What that missing stretch cannot tell anyone is how early the current would have crossed its limit, had a chart been watching from the first week the rejects began to climb.