I always seem to forget how to convert the calibration
dump information into code for doing a calibration, so
I'm writing this mostly for myself.
Boards may have one of 4 calibrations statuses, depending
on how well the calibration code is trusted. These are:
STATUS_NONE, the default for no information; STATUS_SOME,
meaning that a dump has been converted to initial code,
but not tested; STATUS_DONE means that the output of a
STATUS_SOME dump has been checked, and is correct;
STATUS_GUESS is a marker that code has been converted
from a previous version of the code, but not checked.
The NI E series boards have several internal voltages that
can be measured, and also several calibration DACs that function
similar to adjustable resistors on old data acquisition boards.
The information we need is which DACs affect which measurable
voltages; then we can write calibration code that adjusts those
DACs until the voltages are within spec.
Usually, there are DACs (or multiple DACs) that are added to
an analog input signal: 1) before the variable gain amplifier
("pre-gain"), 2) after the variable gain amplifier ("post-gain"),
3) between the board's stable voltage reference and the
reference input to the ADC ("gain offset"), and 4) before a
unipolar-to-bipolar adjuster ("unipolar offset"), or other
equivalent circuit.
In addition there are DACs that adjust the output voltages and/or
reference voltage inputs to a D/A converter. These are pretty
intuitive once analog input is understood, and is dependent on
correct analog input calibration.
The measurable quantities are 0 volts and an internal voltage
reference near 5 volts, and can be measured at any gain. The
interesting combinations are:
ai, bipolar zero offset, low gain
ai, bipolar zero offset, high gain
ai, bipolar voltage reference, low gain
ai, unipolar zero offset, low gain
The unipolar zero offset may not be available on some boards.
In a STATUS_NONE dump, for each measurable quantity and each
calibration DAC, the DAC is varied throughout its entire range
and the quantity measured. The data is linearly fit, and if
the slope is statistically non-zero, a line is printed:
caldac[0] gain=1.26(11)e-7 V/bit S_min=235.659 dof=254
The information given is caldac index, slope (gain) and slope
error (in parenthesis, modifying the last two digits of the
slope), and two statistical parameters S_min and degrees
of freedom. S_min and dof will be roughly similar for a
good fit. If S_min is more than a factor of 4 greater than
dof, this is probably not a good fit. Typically this means
that the DAC doesn't affect the measureable strictly linearly,
or there is systematic noise. The latter seems to common in
E series boards, so I'm not too worried about the following
dump where there are S_min/dof ratios above 4.
Here's an example dump, generated by a STATUS_NONE dump for
a pci-mio-16xe-10, with the analog output section removed:
Warning: device not fully calibrated due to insufficient information
Please send this output to <ds@schleef.org>
Id: comedi_calibrate.c,v 1.21 2001/10/10 22:07:53 ds Exp
Driver name: ni_pcimio
Device name: pci-mio-16xe-10
Comedi version: 0.7.61
ai, bipolar zero offset, low gain
offset -6.795(14)e-3, target 0
caldac[0] gain=1.26(11)e-7 V/bit S_min=235.659 dof=254
caldac[2] gain=3.96840(14)e-4 V/bit S_min=1390.18 dof=254
caldac[3] gain=4.348(11)e-6 V/bit S_min=258.75 dof=254
caldac[8] gain=5.4659(69)e-7 V/bit S_min=386.361 dof=254
ai, bipolar zero offset, high gain
offset -2.4224(55)e-4, target 0
caldac[0] gain=3.61(45)e-9 V/bit S_min=247.26 dof=254
caldac[2] gain=3.96644(48)e-6 V/bit S_min=351.927 dof=254
caldac[3] gain=4.063(46)e-8 V/bit S_min=272.024 dof=254
caldac[8] gain=5.46305(30)e-7 V/bit S_min=314.035 dof=254
ai, bipolar voltage reference, low gain
offset 4.992959(13), target 5
caldac[0] gain=-4.4928(11)e-5 V/bit S_min=1111.4 dof=254
caldac[1] gain=-2.792(11)e-6 V/bit S_min=248.971 dof=254
caldac[2] gain=3.96488(14)e-4 V/bit S_min=1059.18 dof=254
caldac[3] gain=4.318(11)e-6 V/bit S_min=437.441 dof=254
caldac[8] gain=5.4810(70)e-7 V/bit S_min=404.213 dof=254
ai, unipolar zero offset, low gain
offset nan, target 0
caldac[2] gain=3.96773(39)e-4 V/bit S_min=158.236 dof=107
[The explanation gets a little fuzzy here]
The resulting function for calibration will look something like:
void cal_ni_pci_mio_16xe_10(void)
{
postgain_cal(ni_zero_offset_low, ni_zero_offset_high, XXX);
cal1(ni_zero_offset_high, XXX);
cal1(ni_reference_low, XXX);
cal1(ni_unip_offset_low, XXX);
}
You get to fill in the XXX's. The post-gain calibration DAC will
be the one for which the ratio of caldac slopes for the low and
high gain measurables is similar to the ratio of input ranges for
low and high gain. This ratio is typically 100 or 200, and really
should be printed by the program. Thus, for this dump, we choose
caldac[2], since the ratio is very nearly 100. We don't choose
caldac[0] or caldac[3], because the gains are smaller, and the
ratio isn't exactly 100 or 200.
Next is the pre-gain calibration. Adding a voltage before the
amplifier will affect every input range selection equally, so the
pre-gain cadac slope will be nearly equal for both bipolar zero
offset at low and high gain. In this example, it would be caldac[8].
Next is the voltage reference calibration. The caldac controlling
the voltage reference adjustment is proportional to the offset,
so the correct caldac will typically be the one that has a large
slope for the bipolar voltage reference measurement, but a small
slope (by a factor of 2e4, here) for the zero offset measurements.
It could be any of caldac[0], caldac[1], or caldac[3], or possibly
all of them. We'll choose the caldac with the largest slope for
rough calibration, then use the one with the smallest slope for
fine calibration, namely caldac[0] and caldac[1].
This is one way that STATUS_SOME is useful, because you can calibrate
the zero offset, then get a much better idea which other channels
are likely to be for the voltage reference.
Note that we haven't done anything with caldac[3]. It clearly
does something useful, but until we attempt a coarse calibration,
it's not certain what it does. It turns out to be a fine postgain
adjustment.
In this example, there doesn't appear to be a caldac that affects
unipolar zero offset, so it will not be used in the final function:
void cal_ni_pci_mio_16xe_10(void)
{
postgain_cal(ni_zero_offset_low, ni_zero_offset_high, 2);
cal1(ni_zero_offset_high, 8);
cal1(ni_reference_low, 0);
cal1(ni_reference_low, 1);
}
There are a number of functions that are useful for optimizing a
given caldac, each optimized for different cases. The inconsistently
named postgain_cal() and cal1() measure the observable(s) at a
number of points throughout the entire caldac range, and then do a
linear fit to determine the optimum value for caldac. These functions
are good if the caldac dependence is strictly linear. They are also
useful if the target value for the observable is at the endpoint of
the measurable range, as when measuring unipolar zero offset, since
the functions automatically compensate for bad input values.
The function cal_fine() is useful for fine-tuning of the results of
cal1(), especially if the dependence is close, but not quite linear.
The goodness of the linear fit is quantified by the S_min value in the
log -- an S_min value that is approximately the same (within a factor
of 2 or 3) as dof (degrees of freedom) indicates a good fit. An S_min
value that is about 10 times dof indicates that fine tuning is probably
necessary. An S_min value that is many orders of magnitude larger than
dof indicates that linear fitting should not be used.
The functions cal_binary() and cal_postgain_binary() are used when
the caldac dependence is highly non-linear. It does a binary search
in the range of the caldac to find a decent value. Once a decent
value is found, cal_fine() should be used, since the caldac dependence
should be relatively linear in a small range around that value.
