I've prepared a patch as suggested by Hans-Georg Wunder and Jeff McBride.
In addition I've removed the ability to completely leave out the integral
action by setting Ti to zero. The velocity form of the PID algorithm _needs_
the integral action.
>> Hello List,
>>
>> I think there's a small bug in the moving-average filter in
>> xmlauto.cxx
>>
>> I noticed that the output from it was always out a bit and
>> checking with a calculator showed that it seemed to be dividing
>> by the number of samples + 1 instead of just the number of
>> samples.
>>
>> subtracting 1 from 'samples' in line 702 seems to fix the problem
>> and as 'samples' doesn't seem to be used elsewhere I think it's
>> safe. Possibly implies that the number of samples may be one
>> less than specified but I'm not familiar enough with c++ to spot
>> it.
Roy Ovesen:
You are right. I would suggest resizing input[] to (samples + 1) instead.
Change lines 654 and 661 to:
input.resize(samples + 1, 0.0);
That way we average over the number of samples as configured.
I've added some features to the PID controller:
Ability to set desired sampling interval in seconds. Use <Ts> under <config>
to set the desired sampling interval of the PID controller.
Example:
<config>
<Ts>0.1</Ts> <!-- desired sampling interval -->
<Kp>-0.05</Kp> <!-- proportional gain -->
<beta>1.0</beta> <!-- input value weighing factor -->
...
...
</config>
Ts defaults to 0.0, so if you don't set it it samples at the highest possible
frequency.
Add an offset to the input variables (input and reference).
Example:
<reference>
<prop>/controls/flight/elevator</prop>
<scale>-1.5</scale>
<offset>1.0</offset>
</reference>
Note that <scale> has higher precedence than <offset>, regardless of the order
that they appear in the config file.
I've added some digital filters to the autopilot. They are all low-pass
filters that filter away high frequency signals/noise. There are 4 different
filters:
1. Exponential - The algorithm is essentially the same as the one used in the
fgGetLowPass() function.
2. Double exponential - Two exponential filters in series. This filter has a
"steeper" frequency response curve. It filters "better" than the single
exponential.
3. Moving average - Averages a number of inputs.
4. Noise spike - limits the amount that the output value can change from one
sample to the next.
Filters 1 and 2 are characterised by it's filter-time in seconds. For filter 3
you have to set the number of input samples to average over. For filter 4 you
set the maximum allowed rate of change as [1/s]. Since the sampling interval
(dt) isn't constant we have to calculate the maximum allowed change for every
update.
Example of a double exponential filter with filter time 0.1 seconds, that is
1/0.1 = 10 Hz.
<filter>
<name>pressure-rate-filter</name>
<debug>true</debug>
<type>double-exponential</type>
<input>/autopilot/internal/pressure-rate</input>
<output>/autopilot/internal/filtered-pressure-rate</output>
<filter-time>0.1</filter-time>
</filter>
This would go in the autopilot configuration file.
I've also removed the filtering of the "pressure-rate" helper value, use the
new filters if you want to filter it! ;-)
This patch adds the ability to do a simple scaling of input without having to
add hardcoded helpers. Example:
<reference>
<prop>/autopilot/settings/vertical-speed-fpm</prop>
<scale>0.01667</scale>
</reference>
Add FGPredictor class to xmlauto. Add support for horizontal navigation based
on flight track as opposed to heading. Add crosstrack-error support to nav.
Simplify error adjust calculation for horizontal nav (better interception).
Fixed potential divide by zero that was producing nan issues in the xmlauto
code.
I've done some changes to xmlauto.cxx.
Only calculate the derivate filtering if derivate time Td is greater than
zero. This means that one can set Td=0.0 in the xml file to completely remove
the derivate action. (Setting Td to zero in the current version would lead to
a division by zero and crash.)
Setting the integrator time Ti to zero doesn't make sense, right! I've
modified so that setting Ti to zero results in the integral action being
completely removed.
seem to be fully deterministic in P-only mode. This old simple controller
does what I expect, so it's good for calulating stage #1's of multi-stage
controllers.
