You have some unpredictable 1/100 or 1/1000 defect that occurs long after production and sale.
Just how do you go about isolating the cause, and testing a solution? Make 5 changes, and put through a production batch of 1000 units, and then do accelerated testing? If 5 fail from one batch, and 2 from the rest, is there even enough statistical power to confirm that you've come across a solution? And you just burned through 5000 units.
Sounds like fun trying to solve this kind of problem.