Siemens SINUMERIK 840D sl Function Manual page 424

Compensations (K3)
Neural quadrant error compensation
Requirements for neural QEC
An essential requirement for implementing QEC with neural network is that the errors
occurring on the workpiece at quadrant transitions are detected by the measuring system.
This is only possible either with a direct measuring system, with an indirect measuring
system with clear reactions of the load on the motor (i.e. rigid mechanics, little backlash) or
with suitable compensation. With indirect measuring systems, any backlash that might occur
must be compensated by backlash compensation.
Learning/working phases
QEC with neural network involves the following two phases:
● Learning phase
● Working phase
The learning phase can be executed for several (up to 4) axes at the same time. For further
information about neural network learning, see Section "Learning the neural network".
The learning and working phases and the resulting neural QEC are purely axial. There is no
mutual influence between the axes.
Saving characteristic values
On completion of the learning phase, the calculated compensation data (characteristic
values in user memory) including the network parameters (QEC system variables) must be
saved in a file selected by the operator. These files are named "AXn_QEC.INI" per default.
Loading characteristic values
These saved and learned compensation data can be loaded back directly to the user
memory in the same way as part programs.
When the part program containing the tables is loaded, the compensation values are
transferred to the NC user memory. The characteristic values become effective only after
compensation has been enabled.
Characteristic values cannot be written when the compensation function is active (machine
data
MD32500 $MA_FRICT_COMP_ENABLE (friction compensation active) must be set to 0
and must be active).
424
A certain type of response is impressed upon the neural network during the learning
phase. The relation between the input and output signals is learnt. The result is the learnt
compensation characteristic that is stored in the non-volatile user memory. The learning
operation
is activated or deactivated from the NC part program with special high-level language
commands.
During the working phase, additional speed setpoint pulses are injected in accordance
with the learnt characteristic. The stored characteristic does not change during this
phase.
Function Manual, 03/2009, 6FC5397-1BP10-4BA0
Extended Functions