Supervised learning wants labelled failures. For a rail bridge that means data recorded from that structure, with that damage, under that traffic — and the only way to get it is for the damage to already have happened. The training set you need is the outcome you are paid to prevent.
Threshold alarms fail for a quieter reason. A single channel crossing a limit is a late signal; by then the change is large. The early signal is not in any one channel, it's in the relationship between them — how strain at one point moves with strain at another, how the modes shift together as the train passes. Damage breaks the relationship long before it breaks a limit.
Novelty detection works from the one dataset you always have: the structure behaving normally. Build the statistical description of that, and anything that doesn't fit it is, by definition, new. You never have to specify what you're looking for.