HPU Data Collection & Validation

Collect and validate hours-per-unit data for learning curves: touch labor vs span time, rate normalization, disruptions and common data problems.

19 min read 5 sections Guide 3 of 6, HPU Forecasting Updated

New to this topic?

Read these first to get the most out of this guide:

Touch
Labor Hours Only
20+
Units for Reliable Curve
±5%
Clean Data Accuracy
3 to 5
Common Data Problems

Touch Labor vs. Span: What to Measure

Learning curves model the improvement in direct labor effort. The fundamental unit of measurement is touch labor hours: the time a technician’s hands are on the product, performing value-added work. This is distinct from several other time measures that are frequently confused with it.

90% CURVEHPU
Learning curveEach time quantity doubles, hours per unit fall to the rateThe Learning curve guide
MetricIncludesUse for Learning Curves?
Touch laborDrilling, fastening, routing, wiring, bonding, testingYes: this is the correct input
Direct laborTouch labor + direct support (material handling, kit staging)Sometimes: if support ratio is stable
Earned hoursPlanned hours credited upon task completionNo: reflects the plan, not actuals
Span timeCalendar duration including wait, queue, and idle timeNo: non-labor factors dominate
Charged hoursHours booked to a work order (may include overhead, meetings)Only if overhead is stripped out

Charge Number Contamination

In many factories, technicians charge time to the unit they are physically near, not the unit they are working on. Cross-charging, support labor booked to production charge numbers, and “bucket” work orders that pool hours across multiple units are the most common sources of data corruption. Always verify that hours are truly unit-specific before using them in a learning curve.

Normalizing for Rate Changes

Production rate affects hours per unit independent of learning. When a factory doubles its rate, it hires new workers, adds shifts, and stretches supervision thinner. These rate effects look like learning disruptions if you do not account for them. Conversely, a rate reduction can artificially improve hours per unit because experienced workers have more time per task.

LOT 3MIDPOINT
Lot midpointA lot's average hours sit at its midpoint unit on the learning curve
Rate Normalization Example Adjustment Method

Scenario: Production rate changes from 2 units/month to 4 units/month at unit 50.

Observed data: Units 45 to 49 average 4,800 hours. Units 51 to 55 average 5,200 hours. The learning curve predicts 4,650 hours at the midpoint.

Rate penalty: 5,200 − 4,650 = 550 hours attributable to rate change (workforce dilution, overtime, new hire inefficiency).

Normalization options:

  • Option A: Subtract 550 hours from units 51 to 55 and taper the adjustment over the next 20 units as the workforce stabilizes
  • Option B: Segment the curve: fit one curve to units 1 to 49 and a new curve starting at unit 50 with an adjusted T1
  • Option C: Include a rate variable in a multivariate regression (requires sufficient data at multiple rates)

Option A is most common for isolated rate changes. Option B is better when the rate change is permanent and large. Option C requires 30+ data points across multiple rate regimes.

Accounting for Disruptions

Disruptions break the learning pattern. A strike, pandemic, supply chain crisis, or major engineering change does not just add hours to the affected units. It can partially reset the learning curve for units that follow. The key is distinguishing between disrupted hours (which should be removed from the data) and learning loss (which is a real change to the curve going forward).

BREAK
Rate disruptionThe learning curve is running down when a break bumps hours back up, then it recoversThe Rate disruption guide
Disruption TypeTypical ImpactData Treatment
Production halt (< 3 months)5 to 15% hours increase for 3 to 8 units after restartFlag affected units, add disruption factor, do not exclude
Extended shutdown (3 to 12 months)15 to 30% increase, partial workforce turnoverSegment the curve; treat post-restart as a partial reset
Supply chain disruptionRework hours spike due to out-of-sequence workSeparate rework hours, normalize base hours to pre-disruption trend
Major engineering changeAffected operations reset to near-first-unit levelsDecompose by operation; reset only changed operations
Workforce turnover (> 30%)10 to 20% increase, gradual recovery over 10 to 20 unitsModel as a learning curve reset with a lower T1 than original

What Makes “Good” HPU Data

Good HPU data has five characteristics. If any of these are missing, the learning curve analysis will be unreliable and the resulting forecasts will carry hidden bias.

VISIBLEHIDDEN
Cost of qualityThe cost of quality is an iceberg: scrap and rework above, the larger hidden part belowThe Cost of quality guide
Five Attributes of Quality HPU Data Checklist
AttributeDefinitionHow to Verify
Unit-specificHours are traceable to a single serial numberCheck charge number structure; confirm no pooled work orders
CompleteAll operations for the unit are closed and finalVerify no open work orders, no pending rework, all buy-off complete
ConsistentSame scope of work across all units in the datasetConfirm configuration, charge number definitions, and labor categories are stable
AccurateHours reflect actual work performed, not planned or estimatedCross-check with timekeeping system, compare to supervisor logs
TimelyData is current, with no retroactive adjustments pendingConfirm accounting period is closed, no pending cost transfers

Common Data Quality Problems

After reviewing HPU datasets from dozens of aerospace programs, these are the problems that appear most frequently. Each one biases the learning curve in a specific direction, and the biases do not cancel out: they compound.

UNITCUM AVG
Unit and cumulative averageTwo curves: unit hours fall faster than the cumulative average that trails themThe Unit and cumulative average guide
ProblemSymptomBias DirectionFix
Cross-chargingHours scatter increases, some units inexplicably high or lowRandom: inflates variance, weakens R²Audit charge practices, reconcile with supervisor records
Incomplete units in datasetRecent units appear to have fewer hours (still in work)Understates recent hours, overstates learningOnly include fully completed units with all work orders closed
Configuration changes not trackedStep change in hours at a specific unit numberOverstates or understates learning depending on directionMap configuration changes to unit effectivity, normalize or segment
Overhead in touch laborAll units appear higher than expected, flat learningUnderstates learning rate (flatter curve)Strip non-touch hours (meetings, training, general support)
Retroactive cost transfersHistorical unit hours change between reporting periodsUnpredictable: corrupts trend analysisLock data after accounting close, use snapshot date

The Incomplete Unit Trap

This is the single most common and most damaging data error. If your dataset includes units that are still in production, those units will show fewer hours than they will ultimately accumulate. The learning curve regression will interpret this as faster-than-actual learning, producing an optimistic forecast. Always verify that every unit in your dataset has completed all operations and all work orders are closed.

The Bottom Line

The quality of your learning curve analysis is limited by the quality of your data. Use touch labor hours, not span or charged hours. Normalize for rate changes before fitting a curve. Identify and treat disruptions explicitly rather than letting them corrupt the regression. Verify all five data quality attributes before starting analysis. The most common error (including incomplete units) produces systematically optimistic forecasts that can undermine proposals and EACs. Next: Unit & Cumulative Average Curves, how to plot, regress, and build confidence intervals on clean data.

Frequently asked questions

What is the difference between touch labor and span time?

Touch labor is the actual hands-on hours a technician spends working on a unit: drilling, fastening, routing, testing. Span time is the calendar duration from start to finish of a unit, including wait time, material delays, shift gaps, and queue time. Learning curves apply to touch labor, not span. A unit might have 5,000 touch labor hours but a span of 60 working days. Using span-based metrics for learning curve analysis will produce misleading results because span includes non-labor factors that do not follow learning patterns.

How do you handle disruptions like COVID in the data?

First, identify the disruption window precisely: which units were in flow during the disruption, and which operations were affected. Then choose a treatment: (1) Exclude those units from regression entirely and note the gap. (2) Flag them as outliers and include a disruption factor in the model. (3) Normalize the hours by removing quantified disruption hours (e.g., rework from supply chain issues). The worst approach is to include disrupted units without adjustment, which will corrupt your learning rate estimate.

How many data points do you need for a reliable learning curve?

Minimum 8 to 10 completed units for a preliminary curve, but 20+ units spanning at least one doubling of cumulative production for a reliable regression. The key is not just quantity but coverage: you need data across a range of unit numbers. Ten data points clustered between units 40 to 50 tell you less than five points spread across units 10, 20, 40, 80, and 160.

Share this guide: LinkedIn X Follow SymplProcess on LinkedIn

Take this to a room

The running order

For anyone assembling HPU data. They should leave able to check five data quality attributes and to spot the incomplete-unit trap.

7 beats · 12 min
  1. 1

    The analysis cannot beat the data

    Everything downstream (the curve, the forecast, the EAC) is limited by what goes in.

    • A perfect regression on contaminated hours produces a confident wrong answer.
    • Most learning curve disputes are actually data disputes.
    • Clean data gets you around plus or minus five per cent. Dirty data gets you anything.

    Ask the room Where do our hours per unit actually come from?

  2. 2

    Touch labour, not span, not charged

    Three different numbers that get used interchangeably, and only one belongs in the curve.

    • Touch labour: hours actually worked on that unit.
    • Span: calendar time from start to finish, including waiting.
    • Charged: whatever landed on the charge number, which is not the same thing.
  3. 3

    Charge number contamination

    Technicians often charge to the unit they are near, not the unit they are working on.

    • Cross-charging, support labour on production numbers, bucket work orders pooling hours across units.
    • These are the most common sources of corruption.
    • Verify hours are genuinely unit-specific before using them at all.

    Ask the room Would we detect cross-charging here?

  4. 4

    The incomplete unit trap

    The single most damaging error, and it is easy to make silently.

    • Units still in production show fewer hours than they will finally accumulate.
    • The regression reads that as faster learning than is real.
    • The result is a systematically optimistic forecast, in a proposal or an EAC.
  5. 5

    Normalise before you fit

    Rate changes and disruptions have to be handled explicitly, not left in the data.

    • Normalise for rate changes first; a rate change moves hours for reasons unrelated to learning.
    • Identify disruptions and treat them as disruptions, not as noise.
    • Otherwise the curve absorbs them and the slope is meaningless.
  6. 6

    Five attributes of good data

    Check all five before starting, every time. It takes an hour and saves a month.

    • Complete units only. Unit-specific hours. Touch labour.
    • Normalised for rate. Disruptions identified and flagged.
    • Twenty or more units for a reliable fit.
  7. 7

    What we do next

    Audit one dataset against those five before we use it for anything.

    • Pull the dataset and mark which units are genuinely complete.
    • Sample charge numbers and check for cross-charging.
    • Flag every known disruption on the timeline.
    • Only then plot it.

    Ask the room Which dataset do we audit first, and who owns the charge-number question?