Ask an engineer what a 10,000-hour rating means and most will answer, correctly for bearings, that ninety percent of units survive it. That is the L10 convention, and it is so deeply embedded in rotating-machinery practice that nobody checks whether it applies.
For strain wave gears it frequently does not. Harmonic Drive LLC’s published engineering data states that its rated torque at rated speed is based on mean wave-generator bearing life — L50. Half the population. If you sized a fleet of a hundred joints against that figure expecting ten failures, the honest expectation is fifty.
The part that wears out is not the one you are watching
The second surprise is which component sets the limit. The flexspline looks like the vulnerable part — it is a thin steel cup being deformed elliptically millions of times, which is a textbook fatigue problem. Manufacturers design it for effectively infinite fatigue life at rated torque, and in normal service it is not what ends the drive’s life.
The wave generator bearing is. It runs at input speed, which at 101:1 is a hundred times your output speed, and it is doing so while being held in an elliptical shape. That bearing sets the published life, which is why the life figure moves with input speed rather than output speed — a detail that catches people sizing a slow, high-torque axis and assuming the drive will last forever because the output barely turns.
Flexspline fatigue does end drives, but as a consequence of overload and ratcheting rather than of normal duty. We covered those paths separately in how harmonic drives fail.
One number cannot describe your duty
Published life assumes a specific torque at a specific speed, continuously. Real machines index, hold, dwell and reverse. The correct input to a life calculation is not your peak torque and not the arithmetic mean of your cycle: it is a cube-weighted average, because bearing life goes with roughly the cube of load.
That weighting has a consequence worth internalising. A short, hard transient contributes far more to wear than its share of the clock suggests, and a long idle contributes almost nothing. Two duty cycles with the same average torque and the same peak can differ by a factor of several in real life, purely in how the torque is distributed.
What we publish, and why it is hedged
Across our range, fifteen part numbers carry a service lifetime of 10,000 h (reference value) and eight carry a design life of 15,000 h. The words “reference value” are doing real work there, and we would rather explain them than quietly drop them.
They mean the figure describes the drive under the reference conditions it was rated at, not under yours. We are not in a position to publish an L10 figure for your duty cycle, because we do not know your duty cycle. Any supplier who quotes you a confident hours number without asking for one is quoting a catalogue, not an answer.
Three questions worth asking any supplier
- What survival fraction is this referenced to? L10 and L50 are both legitimate. Undeclared is not.
- At what input speed and torque? Life tracks input speed. A figure without both is not usable.
- Is it a bearing life, a fatigue life, or a grease life? These have different failure signatures and different maintenance answers, and the shortest of the three is the one you actually get.
None of this makes published life figures useless. It makes them a starting point that has to be adjusted for your duty before it means anything about your machine. Send us the duty cycle — torque, speed and time fractions — and we will do that arithmetic against a real part number rather than a reference condition.
Related: The cube rule · Nobody publishes a relubrication interval.