Hi @BuzzyQ, welcome to the forum 
That’s highly dependent on how it’s implemented, what kind(s) of faults it’s targeted at, and how it’s intended to be used.
As a trivial example, if the vehicles being used by the military and civilians are quite different then training a model to detect faults on one of them may not readily translate to the other, especially when it comes to differing dynamics, data streams, and integrations with things like control station software. The converse is also true - if they’re using the same vehicles and software then transferring could be very straightforward.
As a more nuanced consideration, if you develop some kind of module that gets connected to various data streams and trains itself per vehicle (or fleet) to identify when things seem abnormal, that may be readily transferable technologically, but may have stricter regulatory requirements and guarantees before it’s allowed to be used on military vehicles. If that kind of model doesn’t need to know what kind of data it’s processing then that would likely make it more broadly usable and easier to pass at least some acceptance requirements, but would then also likely require a more extensive training period at the start of each integration where it learns what “normal” looks like.
Sounds like lots of fun! Whether it’s conducive to starting a viable business depends a lot on you and the circumstances/environment you’re in, which I at least can’t meaningfully comment on or predict.
For some more general commentary, I do think fault detection is an area of underwater robotics that is likely to be quite impactful, and isn’t meaningfully widespread yet, but exactly what form(s) the successful implementations will take is dependent on market needs and forces that many of us are likely not well-versed in.
As a few off the cuff potential options, there could be live models that run in real time either on the vehicle (for self identification) and/or on the control station software (for supervisory identification to a pilot/operator), and there could also be cloud based services that do bulk, batch processing of logs to help identify potential issues across a fleet of vehicles (possibly across multiple organisations), and are used to tune maintenance and inspection schedules and the like. Live cloud-based options could also be possible, but seem a bit less likely to catch on because they require vehicles or operators to consistently be connected to the internet, which often isn’t practical (and opens some additional security and data protection risks).