When I've come accross this type of problem in the past I've found the following often helps:

For analysis we use Profitrace, using this tool we can see the waveform of the signal on each slave which helps to identify when a network is close to a problem but just working. The tool does allow you to see the waveform for each slave from a single node, but when there is an issue I've found it better to check each node locally.

Common issues are network lengths (I note you've checked this)

Too many terminators, (some connectors and instruments have internal terminators that do not isolate the rest of the network and in some cases the terminator can only be checked by dismantling the equipment).

Not enough terminators. You are usign FOCs, so a terminator is required at the connection to the FOC, and the terminating equipment.

From memory, you have to pay attention to capacitive correction and minimum cable lengths according to the speed of the bus. but any fault would be apparant on the waveform shown on your analyser.

A damaged Fibre-Optic cable. On a previous job we found a FOC was repeatedly crushed by a disgruntled employee, the fault was highly intermittant , but was traced usign the v-out measurement on the FOC interface.

Add a redundant ring. As you are already using FOC, redundancy may be an option for you.

Noise interferance seem the most likely. Equipotential bonding is always cited as the solution here but I know in some steelworks installation it can also be the cause of the problem when dirty and clean earths and the local earths are all noisy. Problems I have found before are: