F/A feedback oscillations

[Dale Ulan]

...
> (constant load) where you can assume that the system is linear but 
> any transients (accel, decel) will be hard to control and probably 
> require seperate control.

Usually, the feedback system is 'zeroed' during transients, because it
would cause inaccurate learning.
> 
> I believe that the best approach would be to have a traditional 2D 
> load-speed map (like open loop control) and use something like an 
> adaptive least mean squares algorithm to tweak the points in the map.
> The adaption algorithm would use a simple dynamic model of the engine 
> (such as the one you included) as a reference model for the adaption 
> process.

In fact, normally, the speed-load map (or airflow curving map)
is left alone in ROM, and a small array (from 16 to 64 values) are
modified by the feedback process.

If you don't mind reading a paper a few times, SAE paper 860594,
Development of a High-Speed High-Precision Learning Control
System for Engine Control, by Naoki Tomisawa of JECS, and
Hiromasa Kubo and Shoji Furuhashi of Nissan is worth reading.
> 
> This adaption process can tollerate quite noisy sensors, with 
> the noise only slowing the convergence to the optimal control 
> strategy.

Actually, the trick is to filter out the engine pulsations without
degrading the performance of the steady-state control system. If
you can get your steady-state control algorithms to work properly
during all driving conditions, then you can add feedback to make
the system better.

Because of the differences in time lags, feedback control is not
appropriate for engines in the traditional sense. Engines like
feed-forward control (which is basically what present systems do), with
a small amount of feedback. The trick is to use proper models for
doing feed-forward control, rather than spending your life in a dyno
lab.

Sensor filtering is not trivial, either. Proper filters have no lag-time, 
which is not the case with most electronic filters. This is not easy to
do, so in modern engine control theory, an observer (or forward model)
is built to mathematically simulate the noisy or filtered variable.
This is where Extended Kalman Filters and fun stuff like that come in...

An example of a neat observer is in a paper (1994), Transient
Air Flow Rate Estimation in a Natural Gas Engine Using a Nonlinear
Observer by Robert Weeks and John Moskwa.

Papers 910258, 930856, and  920289 are also good. The first two are
by Elbert Hendricks and others, which are what I call essential reading.
The third paper covers operating characteristics of Zirconia sensors.

-Dale