"However, based on our knowledge about the machinery or environment in question, we can probably make intelligent decisions about what to save and what to toss."
Some practical problems with that approach.
1) you are assuming you know what level of data will be useful in a future case.
2) smoothed data may remove outlier values that are later found to have been important.
3) pre-processed data may introduce aliasing effects - you get artefacts that were not actually there.
4) assumptions about cause and effect are not valid if underlying constraints are breached.
A system needs to be able to switch on more detailed archiving when the root cause of a recurring situation needs to be identified.