Do NOT average - that leaves no stability in the results as any one machine will effect all the others.
If it is IMPOSSIBLE to fingerprint individual machines and/or implement different profiles at proof/output stage THEN you could, conceivably, use the same profile for all.
I firmly believe...