🌎 All English Articles  |  🇯🇵 Japanese Version

Accelerated Life Testing: Simulating Years of Wear in Weeks

Accelerated Life Testing: Simulating Years of Wear in Weeks English

One of the strangest constraints of mechanical design is that customers expect a product to last ten years, but nobody involved in the development schedule is willing to wait ten years to find out if it will. I have worked as a contract mechanical designer inside major manufacturers in Japan for two decades, and accelerated life testing (ALT) is the tool that resolves this contradiction — a way of compressing years of real-world wear into weeks of lab time by deliberately increasing stress levels, cycle rates, or environmental severity beyond what the product sees in normal service. Done well, ALT gives you genuine confidence in a design’s field life. Done carelessly, it gives you a false sense of confidence backed by a number that does not actually correspond to anything real. This article covers how I have approached ALT across several projects, including a couple of cases where I got the acceleration model wrong and paid for it later.

The Core Idea: Acceleration Factors and Their Limits

Every ALT plan rests on an acceleration factor — a multiplier that relates stress under test conditions to stress under field conditions, letting you convert test hours into equivalent field years. The most common models I use are the Arrhenius model for temperature-driven degradation, inverse power law models for mechanical fatigue under increased load or cycle rate, and combined models like Eyring for cases involving both temperature and another stressor such as humidity or voltage. The mathematics behind these models is well established and has been for decades, but the part that actually determines whether your test means anything is not the math — it is whether the failure mechanism under accelerated stress is the same failure mechanism the product would experience in the field.

This is where I have seen ALT plans go wrong more often than the math itself ever does. On a gearbox project, we accelerated a fatigue life test by increasing rotational speed well beyond the rated operating speed, using an inverse power law model to justify the acceleration factor. The test ran faster than real time, produced gear tooth failures on schedule, and the calculated acceleration factor suggested we had validated roughly eight years of field life in about three weeks of bench time. The problem, which we did not catch until a field failure came back for analysis eighteen months after launch, was that the failure mode in the field was pitting fatigue driven by lubricant film breakdown under sustained moderate load, while the failure mode we had accelerated in the lab was bending fatigue at the tooth root driven by the elevated speed and inertial loading. We had accelerated a real failure mode, just not the dominant one the product actually experienced in service. The acceleration factor number was mathematically correct and practically misleading, because it described a different physics than the one that mattered.

Choosing What to Accelerate: Temperature, Cycles, or Load

Every accelerated test has to pick a stressor to increase, and the choice has to be grounded in the actual dominant failure mechanism, which means failure mode analysis has to come before test planning, not after. For products where degradation is chemically driven — seal hardening, lubricant breakdown, adhesive creep, plastic embrittlement — temperature acceleration using an Arrhenius-based model tends to be appropriate, because these mechanisms genuinely do speed up in a predictable way as temperature rises, within a bounded range where the underlying chemistry does not change character.

For products where degradation is mechanically driven — fatigue cracking, wear at a sliding or rolling interface, fastener loosening under vibration — cycle rate acceleration or load acceleration tends to be more appropriate, though load acceleration carries a real risk of pushing the mechanism into a different regime entirely, as the gearbox case above demonstrated. On a linear actuator project, we accelerated wear life by increasing cycle rate from a field-typical 20 cycles per hour to a test rate of 400 cycles per hour, while holding load and stroke length constant at field-representative values. Because we kept load and stroke unchanged and only compressed the time between cycles, the wear mechanism at the lead screw interface stayed consistent with what we expected in the field, and a post-test teardown comparing wear patterns against a field-returned unit with known service hours showed good correlation. That correlation check — comparing accelerated test wear patterns against actual field wear patterns — is a step I now insist on for every ALT program where a returned field unit is available, because it is the only real evidence that the acceleration is representative rather than assumed.

Weibull Analysis and the Trap of Small Sample Sizes

Once failure data comes in from an accelerated test, Weibull analysis is the standard tool for characterizing the failure distribution and projecting field reliability, typically expressed as B10 life — the time by which 10% of a population is expected to fail. I use Weibull plots on essentially every ALT program, but I have learned to be cautious about how much confidence a small sample size actually supports. A Weibull shape parameter estimated from four or five failures carries wide confidence intervals, wide enough that two different engineers analyzing the same small dataset can walk away with meaningfully different projected B10 life numbers depending on how the data points happen to fall.

On a hinge mechanism project, we ran an ALT with six units and observed failures at what looked like a reasonably tight cluster, producing a Weibull shape parameter suggesting predictable wear-out behavior with a comfortable B10 life margin over the ten-year design target. When a follow-up test with fourteen additional units was run later in the program for an unrelated reason, the expanded dataset revealed a bimodal failure distribution — most units following the wear-out pattern we had originally characterized, but a distinct minority failing much earlier due to what turned out to be a separate mechanism, a manufacturing-induced stress riser from a stamping die wear issue that only affected a subset of production lots. The original six-unit test had, by chance, sampled zero units from the early-failure population, and the Weibull analysis built on that data was confidently wrong. I now push for a minimum of eight to ten units on any ALT program feeding a reliability claim that matters for a warranty commitment, specifically to reduce the chance of a small sample missing a secondary failure population entirely.

Handling Censored Data and Suspended Units Honestly

Not every unit in an ALT program fails during the test window, and how you treat units that survive to the end of the test — censored data, in reliability terminology — has a real effect on the projected life number. I have seen engineers under schedule pressure simply drop surviving units from the Weibull analysis entirely, treating only the failed units as data, which systematically biases the projected life shorter than it should be and, worse, throws away genuinely useful information about the units that made it through.

On a compressor bearing ALT program, we ran twelve units and stopped the test at a fixed time budget with eight failures observed and four units still running. Early analysis by a junior engineer on the team excluded the four survivors and computed Weibull parameters from the eight failures alone, producing a B10 life estimate that looked concerning relative to the design target. When I re-ran the analysis treating the four survivors as right-censored data points rather than discarding them, the corrected B10 life came out meaningfully higher and comfortably above target, because the censored analysis properly credited those four units for the runtime they had already accumulated without failing. The units had not failed; excluding them from the dataset was mathematically equivalent to pretending they did not exist, and it produced a needlessly pessimistic and technically incorrect result. I now make it a standing rule on every ALT program to define upfront how censored data will be handled in the analysis, before the test starts, so nobody has to make that judgment call under the pressure of a disappointing early result.

Building the Test Matrix

Stressor Product Type Where It Fits Typical Acceleration Model Key Risk if Misapplied
Temperature Seals, adhesives, lubricants, plastics Arrhenius Chemistry changes character above threshold temp
Cycle rate Mechanisms with time-independent wear Direct time compression Heat buildup from faster cycling adds unintended stress
Load / stress amplitude Fatigue-critical structural components Inverse power law Can shift dominant failure mode entirely
Humidity Corrosion-prone assemblies, electronics Peck / Eyring Condensation behavior nonlinear near saturation
Combined (temp + humidity + vibration) Complex electromechanical assemblies Multi-stress models Interaction effects hard to isolate and validate

I use a matrix roughly like this early in ALT planning to force an explicit conversation about which stressor is actually appropriate for which failure mechanism, rather than defaulting to whatever test chamber happens to be available that quarter — a mistake I discussed in a different context around validation test sequencing, and one that shows up just as often in ALT planning.

Validating the Acceleration Model Against Real Field Data

The single best thing I have done to improve ALT credibility on any project is building in a deliberate feedback loop between accelerated test predictions and actual field return data once the product ships. On a valve actuator product line, we predicted a B10 life of roughly 12 years from ALT data using a combined thermal and cyclic model. Three years into field deployment, warranty return data showed an early failure rate meaningfully higher than the ALT projection would have predicted for that point in the product’s life. Rather than dismissing the discrepancy, we pulled a sample of the returned units and found that field ambient temperatures in one particular installation region ran consistently 8-10°C higher than the design assumption used in the original Arrhenius acceleration factor calculation, because units were being installed in enclosed outdoor cabinets with poor ventilation that the original environmental assumptions had not accounted for.

This did not mean the ALT method itself was flawed — the acceleration model was mathematically sound — but the environmental input assumption feeding the model was wrong, and no amount of careful Weibull analysis in the lab could have caught an error in an assumption about field conditions that originated outside the test entirely. Since that project, I treat every ALT prediction as provisional until at least some field data exists to check it against, and I push for a formal comparison step, typically at the one-year and three-year field marks, between predicted and actual failure rates. When they diverge, the divergence itself is valuable information, often pointing not to a flaw in the test method but to a wrong assumption about the actual conditions the product experiences in the world, which is a distinction that determines whether the fix belongs in the design or in the installation guidelines.

コメント

タイトルとURLをコピーしました