If you wait that long, you're probably more concerned if the chain that's
holding the motor to the firewall is going to break after the motor cuts
loose from the mounts. All too often, the vibrations start to pick up
seconds or miliseconds before a catastrophic failure.

To do such a health-monitoring function properly, you really want some
seeded fault data to characterize what a "bad" engine spectrum looks like.
How many engines do you want to sacrifice to get the data? You can
approach
it from the "anything different from a healthy engine signature"
standpoint,
but that will likely result in a ton of false positive fault indications.


Are you suggesting that a bad engine will give clues to it's demise enough
in advance that you could actually do something about it? Clues that a
monitor could pick up on, but an experienced pilot wouldn't?