I can suggest a couple of things to check.
Opto couplers are complex mechanically, because they contain 2 different types of component, which are then immersed in a resin which has to be airtight, transparent to IR, have a thermal coefficient of expansion matched to the rest of the device, and this is then covered by an opaque layer to block out ambient IR. As a result of this compound construction and all its failure modes, optos have a very high failure rate compared to other components. You seem to be aware of this though as you mention considering vibration, heat etc.
The manufacturers sometimes overlook one of these factors. There were some cases of optocouplers a few years back (which used white plastic rather than black bodies, I think) which didn't seem to do the last function. They responded to ambient light / IR too. You'd think it was a basic property of the device to NOT respond to this, but perhaps they moved manufacture to a new factory and didn't emphasise the importance of the resin type in the training. Or perhaps they were fake components. If nothing else explains the problem, it may be worth checking the supply chain to determine the provenance of these (expensive) devices.
The manufacturer's QA department should be able to comment on the batch numbers written on the opto - a photo would probably be best if possible, as this may give them other clues if they ARE fake like "all our optos are in a different colour body".
When I had similar trouble a few years ago I asked the manufacturer's QA department for failure rate information. It took a little while but they produced it, which is when we realised how awful cheap optos were, and switched to high quality ones (like the type you're using!). One of the failure modes mentioned in the report was, IIRC 20 years later, something about the resin going opaque as it aged! You should be able to confirm that if you have an original circuit around, you could compare its opto output with that of a new opto.
Does the circuit work OK if you drop a new opto into this position?
When all is said and done though, it sounds like your customers are getting a bit overexcited about a single failure. As the application is not unimportant (I assume this means it is for the defence industry), it is a good idea to kick off an investigation to learn from the event and improve the circuit if possible; but statistically a single failure in one year conveys too little information to extrapolate.