well you can let 1000 of them run for a year and from the ratio of those that failed calculate (or rather guesstimate) for how many years can more than 500 be still running.
It's not perfect, but for a perfect solution you would need something like a time machine, I guess.
Some failures are at a random distribution - running 1000 for a year will help you estimate that.
But any moving part has a finite lifetime due to wear, and you want to know how long this is. Running lots of them for 1/1000 of the expected lifespan shouldn't produce any at all of this kind of failure.
The normal speed of the clock is very slow, so it should be easy to run it at an accelerated rate to test wear. (There are limits, though; apparently the Difference Engine jams frequently if you overclock it.)
It's not perfect, but for a perfect solution you would need something like a time machine, I guess.
edit: this article linked from wiki has nice explanations http://db.usenix.org/events/fast07/tech/schroeder/schroeder_...