If you use appropriate derating guidelines for electrolytic capacitors (keeping the ripple current in check, ensuring they stay well below their maximum temperature rating) they can last a very long time and still provide adequate performance. I've seen capacitors come out of 15 year-old equipment and still 'work' to some degree.
Fans have a limited lifespan depending on how fast you operate them. They tend to gradually fail over time.
If you have a primarily surface-mount design, consider conformal coating to help limit exposure to the elements and mitigate oxidation, etc. Products in the field for years which have some air exposure can very easily get several millimeters of dust, which makes a very effective thermal blanket.
Make sure that, if you're using RoHS parts, that you're not using parts prone to tin whiskering. (This is another area where conformal coating can be helpful).
PLC manufacturers would like you to believe that their software is more reliable and more thoroughly tested. My impression is that the core OS components of PLCs are usually quite well designed and tested, but the drivers for external hardware (motion systems and the like) are often libraries hacked together by applications engineers and then passed around the company. The hardware in PLCs is often antiquated-- a lot of them are running old, hot Geode processors.
When you buy a PLC from Allen-Bradley, B&R, Siemens, or any of the other big players, you're mostly paying for support when things go wrong. Their hardware is made with the same manufacturing processes as Arduinos, and there's nothing magical about the real-time operating systems running on PLCs that makes them bug-free. But, I think that support is often worth paying for. If it's a machine that costs the company $1M every day it's not operating, I'd be damn sure that when something went wrong, there was a team of professionals who could help fix it, not just me and Google. For the specific case of light curtains or other safety interlocks, I would want to make sure that the manufacturer had a hefty insurance policy in place, rather than a statement that tries to disclaim all merchantability for any particular purpose.
Even so, if I were designing (for example) a bit of simple pneumatic actuation for some fixture, and I was willing to shoulder the support burden of fixing the machine when it broke (or if I wasn't able to get the resources allocated to pay for the PLC), and safety wasn't an issue, I'd happily use an Arduino.
I'd probably prototype the system with an Arduino and then rewrite the code in pure C once it was working, so that my code was the only code on the microcontroller.
Best Answer
I have some background in high reliability electronics for space applications. The way you define "reliability" here is the key.
If you're thinking just in terms of the the random failure rate, a properly derated and protected flight-qualified MOSFET-based assembly can easily beat any equivalent relay.
But that's not the only thing to consider when chosing between those two technologies, of course. That would be too easy. :)
Sorry for raising more questions that answers, but your question can't be easily answered without having a broader view of the design problem.