Most launch teams measure reliability by counting failures and hoping the count falls. Hope is not a growth model. What separates a managed launch from a lucky one is whether you can demonstrate, with plotted data, that the design is getting more reliable week by week and that you understand why.

Test-analyse-fix (TAF) cycles, growth modelling and launch readiness evidence are three parts of one system. The cycles generate the data, the model turns the data into a projection, and the evidence pack turns the projection into a decision. Across two decades in automotive and aerospace, I have seen programmes that ran all three well launch on time with defensible risk positions, and I have seen programmes that only tested, and their launch reviews were guesswork with charts.

This article sets out how the pieces fit: the discipline inside each TAF cycle, how to choose stresses that reflect the customer's duty cycle, what the Crow-AMSA growth model actually tells you, how to verify fixes and their interactions, and what a launch readiness board should demand before it signs.

What Test-Analyse-Fix Means in Practice

A TAF cycle is not test until it breaks, fix it, carry on. Each cycle has three disciplined phases: a defined test exposure, a root-cause analysis of every failure against defined failure modes, and a verified corrective action implemented into the hardware or software under test. The critical discipline is that the fix must be introduced into the test article deliberately, tracked by configuration, and confirmed by continued exposure. Fix failures but never reintroduce corrected units to the same stresses, and you have testing, not growth.

Failure analysis is where most programmes go soft. A solder joint that fractures under thermal cycling is not a solder failure; it is fatigue driven by coefficient of thermal expansion mismatch, joint geometry, or voiding in the reflow profile. Each failure must be classified: design deficiency, manufacturing defect, part quality, test artefact, or genuine wear-out. Only design and part-quality findings feed the growth curve meaningfully; manufacturing defects belong on a separate containment track, otherwise you flatter the design.

Configuration control during TAF is non-negotiable. Every unit in the chamber must be traceable to a fix level. I have audited growth curves that looked magnificent until we discovered the last eight weeks of testing ran on units carrying an unreviewed firmware change. The moment you lose traceability between failure, cause, fix and retest, the curve becomes decoration. Keep a simple matrix: failure ID, mode, root cause, fix, fix level introduced, verification result. That matrix is your growth evidence.

Choosing Test Stresses and Durations

Growth testing only works if the stresses reflect the customer's duty cycle. For an electronic control unit, that means combined thermal cycling and vibration per the applicable environmental profile, humidity where the installation sees it, and voltage transients matching the vehicle network: load dump, cold crank, reverse polarity. For a mechanical assembly, it means duty-cycle fatigue: door slam counts, latch endurance, bracket vibration at the package's natural frequency. Testing at rated conditions only tells you the product works on a good day.

Accelerated stress raises the failure rate deliberately so failures arrive within the launch window. Thermal cycling between extended temperature limits accelerates solder fatigue through the Coffin-Manson relationship; elevated temperature accelerates electrolytic capacitor degradation and polymer creep through Arrhenius behaviour. The point is not precision on the acceleration factor. It is ensuring the failure modes you precipitate are the modes the field will see, not artefacts of overstress such as melting or dielectric breakdown that never occur in service.

Sample size and duration must be planned against a target, not against the calendar. If the reliability objective is expressed as a mission life or a warranty-period survival probability, work backwards to required test hours using an agreed acceleration model and confidence level. Run multiple units in parallel so failure spacing gives you resolution on the curve. A single unit tested to destruction yields one failure mode at a time and very little statistical confidence.

Reliability decisions are made at the process, not in the report that describes it afterwards.
Reliability decisions are made at the process, not in the report that describes it afterwards.

The TAF discipline loop

  1. 01ExposeDefined stress profile matching the customer duty cycle
  2. 02AnalyseRoot cause and classification of every failure against the mode taxonomy
  3. 03FixCorrective action introduced at a tracked configuration level
  4. 04Re-verifyCorrected units returned to the same stresses to confirm the fix
A cycle only counts as growth when every phase closes with traceable evidence; skipping retest converts the loop into ordinary testing.

Growth Models: Making the Curve Say Something

The workhorse of reliability growth is the Crow-AMSA model, an extension of the Duane approach, which treats cumulative failures as a non-homogeneous Poisson process with a power-law intensity. In plain terms: plot cumulative failures against cumulative test time on log-log axes, and a genuinely improving programme produces a straight line with a slope below one. A slope near one means the failure rate is not falling; you are finding failures as fast as ever. A slope above one means you are going backwards, often because a design change introduced new modes.

The management value of the model is projection. Given the fitted intensity function, you can estimate how many test hours remain before the design reaches its target failure intensity, which tells you whether the launch date is achievable or whether you need more test capacity, more samples, or more aggressive fixes. That is a far better conversation than testing until the deadline and seeing. I have used such projections to justify a launch delay, and on other occasions to defend an on-time launch with evidence; both decisions rested on the same curve.

Two cautions from experience. First, growth models assume the failure process is reasonably homogeneous; a major redesign or a supplier change resets the clock, and pre-change data must be handled accordingly. Second, fixes have delayed effect: a corrective action introduced at week ten does not benefit the test hours already accumulated. Plotting the curve without accounting for fix introduction times overstates maturity.

Some teams plot a projected curve that assumes all fixes were present from the start. Compare that against the demonstrated curve and the gap tells you how much of your reliability exists only on paper. Both curves belong in the review pack, with the demonstrated one carrying the decision.

The moment you lose traceability between failure, cause, fix and retest, the growth curve becomes decoration.

Tracking Corrective Actions and Their Effectiveness

Every fix carried into the growth programme needs an effectiveness check. A bracket stiffened to cure a resonance fatigue crack must be retested at the same vibration profile to confirm the natural frequency has moved out of the excitation band and that no new mode has appeared nearby. A conformal coating applied against condensation-induced leakage currents must be verified under humidity cycling with contamination present, because clean-lab humidity rarely reproduces field behaviour. Fixes that are not re-verified are hypotheses.

Watch for fix interactions, the curse of late-stage programmes. Addressing one weakness often loads another: a heavier heatsink shifts a board's resonant frequency; a firmware watchdog masks a hardware brown-out rather than curing it; a stronger adhesive introduces a stiffer bond line that cracks a plastic boss. After each significant fix level, review the full failure-mode list and ask explicitly which modes could have been worsened. Where the risk is credible, add targeted retest to the next cycle rather than waiting for the field to tell you.

Keep a running fix-effectiveness metric: of the corrective actions verified to date, how many fully eliminated their failure mode, how many reduced its frequency, and how many failed verification entirely. A programme where most fixes fully eliminate their modes behaves differently from one where fixes merely attenuate. The latter tends to plateau, and the growth curve flattens before target. Recognising that plateau early, while schedule margin still exists, is one of the most valuable things the curve does for you.

Launch Readiness Evidence: What a Review Board Should Demand

A launch readiness review should demand four artefacts. First, the demonstrated growth curve against the reliability target, with confidence bounds and the fix-introduction timeline visible. Second, the failure-mode register: every mode found, its classification, its root cause, its fix status, and critically the modes found but not fixed, with rationale. Third, verification evidence for each fix at the correct stress level. Fourth, an assessment of residual risk: which failure modes have never appeared in test but remain plausible from physics and field history on similar products.

The unfixed-modes list deserves particular scrutiny. Every programme ends with open items; the honest programme shows them. What matters is whether the exposure has been quantified: an engineering judgement, supported by stress-strength analysis or similar-product field data, that the residual risk is acceptable for launch. Running out of time is not a rationale. Three open modes with reasoned exposure assessments are worth more than an empty list produced by closing items on paper.

Maturity of the evidence matters as much as its content. A curve built from six failures with wide confidence bounds does not demonstrate a target has been met; it demonstrates the target has not been ruled out. Be candid about that distinction. Where statistical evidence is thin, say so and state the post-launch monitoring that will close the gap: early field returns analysed on the same failure-mode taxonomy as the test programme, so the growth story continues seamlessly from chamber to customer.

Demonstrated versus projected reliability

Projected curve

  • Assumes all fixes present from test start
  • Fit continues smoothly to target
  • Looks mature at review time
  • No accounting for fix introduction dates

Demonstrated curve

  • Fixes credited only from introduction onward
  • Slower, honest, statistically bounded
  • Shows the true plateau risk
  • The one a review board should trust
The gap between the two curves is the reliability that exists only on paper; the demonstrated curve must carry the launch decision.

Extending Growth Management into Early Production

Reliability growth does not stop at launch; the field simply becomes the test. Set up a closed loop in the first months of production: warranty returns and early failures are torn down, classified against the same mode taxonomy used in TAF, and fed back as new points on the growth curve. This is where the discipline of classification pays for itself. If field failures map onto modes you already understand, the model extends cleanly and you can predict warranty performance. If they map onto modes you never saw in test, you have discovered a coverage gap worth documenting for the next programme.

Manufacturing-induced failure modes deserve separate tracking. Design growth should be measured on design-caused failures; process escapes such as a missing torque, a misoriented diode, or contamination belong on a containment and process-improvement track. Mixing them muddies both. Over the first production months, the design failure intensity should continue the downward trend from the TAF programme while the process-related rate falls as the line stabilises. Two curves, two owners, one review.

Finally, capture the programme learning while it is fresh. Which test stresses found the modes that mattered? Which found nothing and could be trimmed? Where did the growth model's projections hold, and where did they mislead? A short, honest closure report on the growth programme is the cheapest reliability investment you will ever make. It sharpens the next launch's test plan, shortens its TAF cycles, and moves the organisation from hoping reliability improves toward proving that it does.