Losing G54 position ??

Hi all, and a happy new year .
I've had a rather peculiar problem with coordinate systems on two machines . Both are set to home with index so they should be accurate . But after a restart and rehome, G54 is not exactly where it was left .
Has anybody had this problem ?
Any ideas are welcome .

I think I did yesterday.

I use simple limit switches but I indicated my work piece. I just assumed I didn't notice the dial go an extra revolution but I have no confirmation this happened

I'd expect limit switches to have a bit of discrepancy . But I believe I've had this with index enabled on all axes .

When you say g54 is not the same do you mean the DRO offset is wrong or the physical position is actually wrong?

If it's steppers maybe lost steps or else maybe a loose coupling or mount of the encoder/motor etc. (you didn't mention how much different)

Chris

This is on a well built servo driven machine . And what happens is you set the center of the part today put it in as G54, do a bit of work on that and tomorrow after a restart and a rehome the part is off . Coupling are checked encoders are good . If the machine is not turned off while working the part, all is well .
And it's different every time . It's not an accurate number . It's not off by a complete turn of the screw .

So next I think I would set up a dial and see if the machine homes physically consistent. Either remove the final move part of the homing process or use g53 to remove any user coordinate system offsets.
My guess is that will not be consistent. If it is consistent then that would seem to indicate linuxcnc is not saving/loading the offset consistently (which would be odd)
You also didn't mention the rough error amount . are we talking inches, thousands, tenths?

The deviation varies a lot . From 0.5mm to a few mm . the screw is 10mm pitch but the operator has never seen a deviation of exactly 10mm . So the home switch is not tripped a thread early .
I'll try to run some more experiments this week .
Is there a possibility linuxcnc isn't saving properly ?

Most anything is possible - but that is very unlikely.
It sounds like mechanical to me - though a few mm is an awful lot.
Is this new behaivour or it's always been like that?
What type of encoders are you using?

Chris

It's been like that for quite some time . And I've had the chance to check everything mechanical . I've had my doubts about encoders but everything checks out fine .
The machine is a mori seiki with red hat fanuc drives . The original drives were replaced for a set of drives manufactured locally, that accept the weird fanuc signal . Output on these is not that difficult to use A B and C are standart . Only the commutation signals are a bit odd and are not differential .
The machine holds tolerances very well . I've personally checked motor couplings and they are solid . Bearings on the screw are good .

Cheers

P.S. I'm not entirely sure but I believe the problem gets worse if the computer is turned off incorrectly .

I ran a couple of experiments .
If power is turned off while linuxcnc is running it wipes the .var file clean . Nothing in it . And as lcnc is started it writes a file with all parameters = 0 .
Is this normal ?
machine runs 2.7.15 (updated today) on debian 9 with a mesa ethernet board .

P.S. What is the relationship between the .var and .var.bak files ? When is the bak file saved ?
Any details around the backup would be helpful .