Manufacturing teams talk a lot about average cycle time and mean time to changeover, then scratch their heads when the line keeps missing takt, or the schedule slips every Friday afternoon. Averages hide the real story. If you dig into the distribution of changeover durations, you often discover two distinct peaks. That’s bimodality, and it explains why an operation behaves like two different factories living under the same roof.
Understanding and acting on bimodal data changes how you set targets, staff your line, and plan your product sequence. It shortens downtime, lifts throughput, and reduces finger pointing. It also demands a different kind of thinking, because the variation is not random noise to be shaved down with generic efficiencies. It’s a signal that two processes are running under one label.
I have seen this pattern in packaging lines, extrusion, mixing, SMT, injection molding, and even software deployment. The details vary, but the strategy remains consistent. You find the second peak, name it, and then rewrite your changeover playbook to prevent getting dragged into it.
What a bimodal distribution looks like on the shop floor
Imagine a line change from Product A to Product B. Over a month, you log 60 changeovers. You plot the durations. Instead of one bell-shaped mound centered around, say, 28 minutes, you get six sigma tools two humps: one near 18 to 22 minutes and a second near 45 to 55 minutes. Averages might land at 32 minutes, which sounds fine on paper, but almost no single changeover actually takes 32 minutes. Half are quick. The other half are painful.
The practical consequence is predictable misery. Schedulers plan to a mean that rarely occurs. Supervisors nag to “hit the average.” Operators feel gaslit, because they know some setups just take longer, and the current process design doesn’t address that reality. The organization chases phantom efficiency while ignoring the structural split in the data.
Bimodality almost always signals distinct modes in the work:
- A low-friction mode where parts, tools, and recipes line up, and the team flows. A high-friction mode where one or more conditions trigger searching, rework, extra checks, or awkward coordination.
On the lines I’ve supported, the triggers ranged from SKU tiering and allergen constraints to tool index drift and changeover timing across shifts. On an SMT line, peak one occurred when the pick-and-place program and stencil were already validated. Peak two appeared when engineering had to touch the program or swap nozzles due to unexpected component substitutions. On a beverage filler, the fast cohort matched size-to-size neck finishes. The slow cohort bundled in sanitation and valve tuning. The data did not lie. Two modes, one metric.
The trap of averages and the value of a bimodal chart
The simplest way to see it is to build a bimodal chart of observed changeover times. Histogram bins of 5 minutes often work for this purpose. If you lack volume, visualize with kernel density estimation to avoid overfitting to a few bins. The litmus test is not academic. Ask: does the data form two clear mounds, or at least a shoulder that suggests a second cluster? If yes, resist the urge to collapse it with averages.
I like to overlay the chart with categorical labels. Color bars by shift, weekday vs weekend, product family, or whether the change required sanitation. Often you’ll immediately see that the slow peak clusters with a specific attribute, which gives you a handle to pull. If the pattern is subtle, compute medians per attribute and compare. You are not trying to win a statistics prize. You are trying to isolate the conditions that shove a changeover into the slow lane.
One practical rule: never set a KPI target to the mean of a bimodal distribution. Either define separate targets for each mode, or better, re-engineer the process so the slow mode becomes rare and explicit. If you must use a single target during the transition, pick the 75th percentile of the fast cohort. That encourages maintaining fast conditions without setting the team up for failure when slow-mode triggers occur.
Where the second peak comes from
Bimodality emerges because changeovers are not homogeneous. They bundle several steps that can vary in different combinations. The conditions that create the slow cohort typically fall into one or two of these buckets:
- Structural deltas between SKUs. Going from 500 ml PET to 2 L glass may require parts changes, sanitation, torque recalibration, and labeler rethreading. Those are not the same species of changeover as 500 ml to 600 ml PET on the same capper. Hidden dependencies. If the lab’s release of the next SKU lags by 20 minutes, the crew waits. Schedule looks late, and the line takes the blame for a problem it did not create. Tooling ambiguity. Shared tools, missing gauges, or wear that forces extra adjustment passes. You see this when the first bag or bottle is out of spec, then the team circles through “just a hair more” three times. Cross-functional handoffs. Engineering needs to tweak a recipe. Quality wants double checks on ingredients after a recent nonconformance. Maintenance jumps in to reset a sensor that drifted. These are all legitimate safeguards, yet they belong in a defined path, not as surprise add-ons that inflate time. Timing and staffing. Night shift with two fewer people or less experience can push a borderline setup into the slow peak, especially when changeovers happen near breaks or shift changes.
In one plant, simply separating allergen to allergen changeovers into their own standard with an explicit sanitation window removed a third of the “late” flags. The average changeover got worse on paper, but the plan became accurate, and the crew stopped firefighting.
Make the modes explicit in your standards
You cannot erase a slow mode until you recognize it as a different process. That begins with how you write standards. Many shops have a generic changeover SOP that tries to cover every possibility. It becomes fifteen pages long, part instruction and part encyclopedia. No one references it during the heat of the setup.
A better pattern is to codify modes:
- Mode A: like-to-like or low-delta changeover, under 20 minutes, no sanitation, no program edits, same container family. Mode B: cross-family changeover with sanitation and torque recalibration, 45 to 60 minutes, additional verifications by QA. Mode C: engineering-required program or tooling change, time budget 60 to 90 minutes, with pre-release signoff.
Once your standards reflect these paths, the team stops pretending that all changeovers fit one clock. You also gain clean data. When you log times, tag each changeover with its mode. The next chart you draw will sharpen. Often, the observed second peak fractures into two narrower mounds, each with repeatable characteristics.
Scheduling to avoid the second peak
The best way to cut downtime is not to execute faster, but to avoid entering the slow mode in the first place. That is a planning problem. The schedule should stitch a run sequence that minimizes cross-family transitions, waterfall allergens in a logical order, and aligns changeovers with skill availability. I’ve seen 10 to 25 percent throughput gains from nothing more than reordering the queue.
There’s a trap here. In a high-mix environment, the sales promise can force your hand. You must make a glass run between two PET runs, or you must satisfy a rush order that triggers a complex tooling swap. Acceptance of reality does not mean surrender. It means you pull forward the prerequisites that keep that forced slow changeover from getting even slower. That can include pre staging tooling, dry running recipes during lulls, or roping in maintenance for a quick calibration while the machine is still warm.
Schedulers benefit from a simple rule set backed by data, not folklore. For example, each adjacent pair of SKUs in your portfolio can carry a switching cost, driven by historical times by mode. A planning tool that treats the sequence as a traveling path with switching penalties can often find a run order that is both feasible and friendlier to uptime. You do not need a perfect optimizer. Even a spreadsheet that tallies penalties will steer you away from unnecessary slow-mode hops.
Designing changeovers for flow, not heroics
Changeovers go long when the team must think, search, or wait. If you can cut those, you cut downtime. The mechanics are familiar, but the way you apply them depends on which mode you’re targeting.
Start with externalizing steps. In the fast mode, much of the prep can happen while the line runs: preset clamps, stage parts kits, preheat tools, verify the next recipe on a shadow PLC. In the slow mode, externalization requires more planning. If sanitation is involved, stage verified sanitation kits and seals. If you need a torque recalibration, pre stage the calibrated wrench and the torque card with the new spec visible. When you audit a setup, track time spent moving feet more than hands. The feet tell you where the flow breaks.
Then attack ambiguity. Standardize clamp positions with fixed stops or positive locators so you don’t tune by feel. Color code similar parts to avoid a mid-setup swap. Embed photos at the point of use. In regulated environments, tie changes to checklists that live with the machine, not in a binder across the aisle. The slow mode often includes extra verifications. That’s fine, keep them, but remove choice from the sequence so they do not require debate.
I’ve watched a crew cut a 48 minute slow-mode setup to 32 minutes by packaging it as a kit in a lightweight cart with foam cutouts. Nothing glamorous, just no more wandering to find a 3 mm hex or the right spacer set. Five minutes here, four minutes there, plus the confidence that comes with a known path.
Capability before speed
There is a temptation to time every step and chase seconds. That helps, but only after you lock capability. Capability here means the first run after a changeover holds spec at the target rate, without rework or fiddling. Many lines carry a hidden tail of micro stops during the first 15 minutes of production. The changeover “ends” on the report, but the line is still hurting. When you see a bimodal chart of changeover times, consider charting the first hour after restart as well. You may find that the slow mode not only takes longer to switch, it also degrades the next hour’s yield.
To lift capability, codify startup criteria. That can be a short list: correct part-to-tool match confirmed, torque verified, first-piece inspection cleared, purge complete. Tie each to a physical indicator, like a green tag on a valve or a signed field on a digital traveler. If you have to choose between a 5 minute shorter changeover and a stable first hour, pick stability. Schedulers can handle a reliable 40 minute switch far better than a 30 minute switch that spawns 20 minutes of micro stops and scrap.
Data plumbing that pays off
Teams get stuck because they cannot see patterns in real time. They might log times in a spreadsheet, but no one slices by product family, shift, or sanitation requirement. Two simple practices make a difference.
First, collect start and end timestamps for each step in the changeover, not just the total. If you track tool removal, cleaning, tooling install, adjustments, QA checks, and ramp up separately, you can allocate your effort to the right place. In my experience, adjustments and QA waits dominate slow modes more often than tool swaps themselves.
Second, tag each changeover with attributes at the time of execution. Don’t rely on someone to remember on Friday what made Tuesday’s switch slow. A short digital form on a tablet at the machine works: SKU out, SKU in, mode, sanitation required, engineering involved yes/no, staff count, shift, and any upstream waits. Keep it fast. Drop-downs beat free text.
With even a month of tagged data, you can generate a clean bimodal chart by mode, and a table of median and 90th percentile times per transition category. That becomes your scheduling penalty matrix and your training plan.
Training to compress the slow mode
Most organizations train to the average changeover. That creates a gap. The crew learns the basics of the fast mode, then discovers the slow mode on a live line with a full schedule. A better pattern is to drill specifically on the slow mode. If cross-family swaps require additional torque checks, simulate it with a practice changeover on an off day. If the most painful transitions include a labeler thread and a filler bowl clean, practice those as a bundle. The goal is muscle memory under the conditions that typically cause drift and debate.
Cross-train deliberately. One plant cut their worst-case changeovers by 30 percent by adding a roving “setup lead” on the evening shift. This was a senior operator with deep knowledge of the slow mode who floated at the top of the hour to support any active changeover. They did not add headcount overall, they reshaped a position to address the long tail.
Finally, capture tribal knowledge in a format that survives turnover. A one-page photo guide for each major SKU family transition beats a thick binder. Put it at the machine, laminated and grease stained. The aim is not documentation theater. It is a working memory aid that lets an average operator perform like the best on their worst day.
Align incentives with modes, not averages
Metrics shape behavior. If you praise teams for hitting a blended average, they will cherry-pick easy runs and push the hard ones to the next shift. If you penalize every slow changeover, regardless of mode, you create cynicism and data games. Reframe targets in a way that respects the bimodality:
- Track and publish on-time performance to mode-specific budgets. A Mode A change targeted at 20 minutes, Mode B at 55, Mode C at 75. Hitting Mode A in 24 minutes triggers a root cause review, not an eye roll or a shrug. Measure total useful production time in the window after changeover. If the first hour is clean and meets rate, count it. If not, consider the changeover incomplete from a KPI view. This discourages rushing the endgame. Reward reductions in the frequency of slow-mode triggers. If the team finds a way to engineer a common tool across two families, celebrate the drop in Mode B frequency. Sometimes the biggest win is not running faster, but running in the fast lane more often.
These adjustments remove perverse incentives and surface the work that actually reduces downtime.
Engineering out the second peak
While much of the early work is procedural, the ceiling lives in engineering choices. You can’t schedule your way out of physics. If a cross-family swap requires five unique parts and a manual alignment, you will keep a second peak no matter how good your checklist is.
Engineering levers that consistently flatten the slow mode:
- Modular tooling with quick-connect interfaces and positive alignment features, so you replace assemblies rather than tune subcomponents. Pin-and-slot beats slotted holes every time. Commonization of parts across SKUs. If two bottle families can accept the same star wheel with a shim kit, do that. The payback often shows up in the changeover chart more than in the BOM. Recipe management that decouples operator-visible presets from engineering parameters. Let operators load verified sets without exposing every knob. This reduces accidental drift and the need for mid-setup adjustments. Predictive checks on wear-prone components tied to the schedule. If a nozzle or guide tends to cause a 10 minute chase every third changeover, replace it proactively when the plan calls for a slow-mode swap rather than gambling on condition.
I’ve rarely seen capital justification worksheets include “reduction in slow-mode frequency” as a line item, yet this is often where the ROI sits. A 5 minute improvement on 200 fast changeovers may equal a 30 minute cut on 40 slow changeovers. The latter often wins on net uptime and stress avoided.
Edge cases and judgment calls
Not every bimodal pattern deserves the same response. A few realities to consider:
- Low volume, high mix job shops sometimes accept a persistent slow mode. It is cheaper to plan for it honestly than to over invest in commonization that will not repeat. Regulated environments may require additional checks that you cannot compress without risk. In those cases, your best gains come from externalization and pre verification, not speed on the day. New product introductions create temporary second peaks as the team climbs the learning curve. Tag NPI runs separately and avoid letting their times pollute the stable modes. As learning accumulates, roll them into the standard cohorts.
Judgment matters. Don’t chase a perfect unimodal distribution if it requires brittle standards that crack under real-world variation. Aim for clarity, realism, and reliability first.
A brief case story: sorting the peaks on a filling line
A beverage plant I supported ran three bottle sizes across PET and glass, with allergen and non allergen formulations. The official changeover target was 30 minutes. Actuals came in two clusters: 18 to 22 minutes and 50 to 65 minutes. The team felt snakebit.
We rebuilt the data view. A bimodal chart tagged by container family showed the slow cluster almost entirely tied to PET to glass or glass to PET, with sanitation and star-wheel changes. A secondary factor was shift. Nights ran 10 minutes longer in the slow mode.
We split the standard into Mode A and Mode B. Mode A covered within-family changes with a 20 minute target. Mode B marked cross-family changes with sanitation, set to 55 minutes. Schedulers cut unnecessary cross-family jumps and stacked allergens to minimize flip flops. Maintenance supplied a pair of kitted star-wheel carts with positive locators and a torque card holder. QA created a laminated Mode B checklist with three photo checks at the labeler. We assigned a setup lead on nights during the expected Mode B windows.
In three weeks, Mode A held at 19 minutes median. Mode B fell from 57 to 44 minutes median, with the 90th percentile dropping from 65 to 52. The schedule stabilized, OEE rose by 6 points, and the tone on the floor improved because the targets matched reality.
How to get started this month
If you suspect bimodality in your changeovers, resist the urge to launch a broad initiative. Start with a tight loop:
- Build a simple bimodal chart of the last 4 to 8 weeks of changeover times. Tag by a few obvious attributes: product family, sanitation, shift. Define two or three modes that match the visible clusters. Write short, mode-specific standards. Set realistic time budgets for each. Adjust the schedule for the next two weeks to reduce slow-mode frequency. Where a slow mode is unavoidable, pre stage kits and assign a setup lead. Capture step-level timestamps for at least twenty changeovers across modes. Use that to target the top two time sinks in the slow mode. Review with the team and iterate. Celebrate time saved, but also celebrate avoided slow-mode entries.
This small loop erases the ambiguity that keeps averages in charge. In a month, you will know if your plant’s second peak is mostly planning, process, or engineering. Usually, it is a mix. The right sequence of moves becomes obvious once the data speaks the right language.
The deeper shift: from efficiency theater to operational truth
Bimodal data asks you to be honest about the work. It forces a move away from catchall SOPs, generic KPIs, and hope-based scheduling. In exchange, it offers a calmer plant and a playbook that fits the floor instead of the boardroom. When the team learns to recognize modes in their own data, the conversation changes. Operators flag a run as Mode B at noon and ask for the kit. Schedulers rearrange to keep the last hour of the shift out of a slow setup. Engineering picks a tooling standard that collapses two families into one mode next quarter.
The second peak does not disappear by pep talk. It shrinks as you remove the conditions that cause it and respect the ones you must keep. That is how downtime gets cut, not in theory, but shift after shift, product after product, chart after chart.