Adaptive algorithms

[robert joseph dingli]

Continuing our familiarisation

Matt wrote
> 
> My search of optimization in the literature has been a little weak, so I 
> was unaware of anything like that done at Ohio or Melbourne.  I'd be 
> curious to get exact numbers for references on the subject.  Our system 

I'll pass on the SAE ref # when I find where I put my copy of the FISITA 94
proceedings.  The title was "Adaptive Control of Air-Fuel Ratio for
Optimizing the Efficiency of a Lean Burn Natural Gas Engine." 
Proc of FISITA 94 Vol 3

Other work into self tuning optimization and general automotive
adaptive control systems includes research at

UMIST (assoc with Lucas) (P Wellstead, P Scotson, etc)
Ohio State University (G Rizzoni, W Ribbens, F Connolly, etc)
Ford / Ricardo (A Beaumont, A Noble, A Scarisbrick)
UC Berkeley, MIT, Princeton  (J K Hedrick, D Cho, J Moskwa, C Nesbit, etc)

plus a car load of others

I'm in the process of writing a reference list.  Mail me later for exact
details.

> holds fuel injected per cycle constant during an eight revolution dither 
> cycle.  The MAP is dithered about 1% high for four revs and back for four.  
> This, while the spark goes through a two rev advanced two rev retarded 
> repeating cycle. From this we estimate TORQUE fluctuation.  If there 
> is a torque fluctuation a small correction is made in the direction of the 
> improvement.  In this way we work toward the most efficient air fuel ratio 
> and spark timing.  (Remember that if fuel flow is held constant the max 
> torque equivalence ratio is also the most efficient.)  Corrections are 
> extermely small because cycle to cycle variation is orders of magnitude 
> larger than the fluctuation due to our dithering.  On average it works 
> quite well.  The concept is similar to a lock-in amplifier, if you are 
> familiar with those.  Future work will involve getting fast response 
> without wild wander from the noise in our signal.  Also, we would like to 
> try the system on an engine with single point fuel addition (TBI or 
> carburetted).

Now that you mention it I remember reading one of your papers (SAE 892142).
The system is a basic hill climbing technique that is unfortunately very
susceptible to noise and multiple extrema.  The work I'm implementing is
different in that it forms a 2 or 3 dimensional quadratic model and 
statistically fits available data to estimate the position of the peak.
The system can then 'jump' straight to the estimated optimum and re-tune 
from there.  The system is very immune to noise and tunes in within about 3
or 4 iterations.  New data is incorporated recursively and is inherently 
filtered in the algorithm.  

The current work is aimed at calulating engine output (torque or efficiency)
in real time without using torque transducers.  I'm experimenting with
various crank speed based algorithms.

The fact the the work has been applied to NG engines is primarily because
of funding issues and secondly because of the fact that such engines will
happily run at very lean AFR's with efficiency peaking further away from
stoic than for petrol.  More recent work is into optimising a lean burn
petrol engine which uses a hyrogen assisted ignition system allowing AFR's
as lean as lambda = 5.

> RECENT REFERENCES:
> US Patent 5168853
> SAE 940546, 940547
> ASME ICE-Vol. 22, p73-83, October 1994
> (the system is running much better now than in any of those papers)
> 
> Thanks for your interest.
> 
Thanks for the references.

Robert 


-- 
----------------------------------------------------------------------
             Robert Dingli           r.dingli at ee.mu.oz.au

Power and Control Systems                 Thermodynamics Research Lab
Electrical Engineering                    Mechanical Engineering
   (+613) 344 7966                           (+613) 344 6728
  University of Melbourne, Parkville, 3052, Victoria, AUSTRALIA
----------------------------------------------------------------------