Self learning EFI - Design Thoughts

[danyl]

Here's a wild idea I'm working on. Perhaps I can pitch it to someone.

It's a global design for a self learning EFI. The system would incorporate a Wide Band O2 sensor with an RPM meter and computer software, combined with an accelerometer such as the G-Tech, a programmable ECU and some statistics software.

I don't believe this is as far fetched as one might think. The hardware exists. The real hurdle is getting through the proprietary software from each manufacturer.

I'm looking as some graphs by Motec, Logitech, and Statistica. The Logitech software can plot real-time A/F against RPM. The G-tech can plot acceleration against RPM. Using this as a common denominator, the statistics package could be set up to compare various runs with alterations in injector timing and A/F changes.

The system would require "learning" runs on the same road. After a large enough sample of runs, the software could make reccomendations for changes in the injector timing. The changes could be instituted in small increments and the computer could then compare the gains/losses against the previous curves.

Statistica is very good at making predictions. A second "testing" phase could then occur as the reccommended changes are made to the ECU map. There may be many occasions where the changes were detrimental to efficency, but after a while the software could bracket-in the best solutions. This is very similar to how a Tank calculates a firing solution.

The best advantage of this system is that it could tailor itself to individual driving habits.

As the accelerometers get more advanced, G forces for inclines and curves could be taken into account, and the software could remap the ECU on-the-fly in situations were richer mixtures were needed for more power up hills and around curves, and then at other times, lean the mixture for high speed straight driving.

What do you think?

[tubut]

How about adding a GPS into the mix? This would add information about topology and route, which will make road segments and conditions more comparable (and fun to watch with a Google-maps overlay).

[Tifosi]

The idea seems sound, but the implimentation is likely to be way up there on the PITA scale.

[Greg Gordon]

I am not sure what you want it to "learn". Regardless of driving habits the required injector duration is determined by mixture requirements. All modern systems already adjust the mixture for max power when you want it, and max fuel economy under other conditions.

[Tifosi]

I believe what he's describing would be an aftermarket EFI system that didn't require any kind of mapping or setup by the installer other than connecting things.

Just a hook it up and it'll learn as it goes kind of thing.

[danyl]

EXACTLY!!!!!

The GPS would be nice, but the basic requirement would just be the on-the-fly feedback loop for autotuning.

Greg,

What I'm proposing is that just having the perfect static map for A/F at all RPMs is not enough. Sometimes different A/F ratios at different RPMs in different conditions MIGHT be plausible. That would be the real debate. People may have drastic driving habit differences that could be fine tuned with a "intelligent" mapping system.

In Innovate system has a five channel logging component that could record additional engine temp and throttle position. A marriage between these components and an "end result" monitor such as an accelerometer, could quantify the "seat of your pants" changes that a dyno/tuner would ultimately be after.

The system could try various tweaks during its learning mode and compare the results.

The counter argument of course, is that after the system went through all its trial and error changes, it would simply end up with a A/F based map that would have been available from the start on a dyno. But that's part of the debate. Would it really, or would we be suprised that it developed a slightly different reciepe for fuel mixture that gave more pronounced acceration and handling for each specific driver.

As far as the system being an overwhelming undertaking (PITA); I propose that the latest technology has now put this next generation of mapping within reach. If any reasonable shade-tree mechanic with a few thousand dollars could install a "smart" ECU and just drive around the same streets for a hundred miles or so, and gain the same results a professional tuner could...well wouldn't that be a great aftermarket product?

The hardware is there. Their software spreadsheets already have the necessary data. The missing link is the laptop software to combine the various inputs and then an ECU that allowed RAM changes of the map in real time instead of ROM flashing.

[Tifosi]

For myself, I'd rather see something that doesn't require a laptop.

There's nothing wrong with flash ROM, especially since it'll do what RAM won't: hold it's information once the power is cut.

That way one avoids having to own a laptop (many of us don't) all that 'only operates with XP and such-n-such Mhz processor and X amount of RAM' type stuff along with not needing any extra junk like cables and adapters to tote along or keep track of.

All built into one brainbox with not much more than a few dip switches to set up whether it learns an economy or performance type map, how many cylinders, the impendace of the injectors, auto or manual trans, the ignition type and whether you want it to learn or hold what it knows, along with an LED to indicate it's learning, another for 'ok, we're done' and a book that tells you the trouble shooting/failure code that the LED's would flash if there was a problem.

A laptop just doesn't seem like a gotta-have for such a system, unless you wanted to do the programming yourself, in which case there would be no need for the self learning to begin with other than to correct for slight mis-programing errors made by the human.

[turbolarespider]

You might want to search some SAE papers- I've read some suggestions for algorithms to self calibrate the spark.

Fuel is pretty easy- tune the open loop controller so that the closed loop controller has to do very little. It's how we do real production work.

Air/fuel is easy. Spark is the hard one.

Even harder is VCT.

If I could suggest- find one of the aftermarket systems that will run a "script" file- one where you have an input to try some numbers in the calibration automatically. With that, you can have two programs running at the same time- one running the car, the other optimizing the calibration.

Not easy, that's for sure. We take HOURS of data to map our engines- if there was such a thing as autotune, I'm sure we'd be doing it to save time and money.

Eric

[danyl]

Yes, and it's those hours of data I think this system could gain from. To make reasonable statistical predictions, one needs a VERY large sample size. This is far more than can be done in a few tuning sessions. From my readings on the various aftermarket ECUs, and even those with Auto-tune functions, none of them seem to data long much more than a few minutes of real world driving.

I'm thinking the next evolution should (or may even will) be a smart ECU that logs hundreds of hours of A/F and acceleration data and then tweaks the map accordingly.

I think you will admit, that at present most tuners are still using a good precentage of "art" in thier tweaks.

The closed loop systems make adjustments to target the A/F ratio only. None are measuring power at the rear wheel. If this was truly enough, then we'd be seeing the NASCAR and other racing entities letting thier ECUs do all the work. They don't however! They go back in and make thier own changes based upon experience.

The average Joe can't afford a pit crew or hours with an experienced tuner. But what if for a few thousand dollars, he could take his modified ametuer machine and let a statistical based system, with some form of rear wheel power monitor do the work for him. It could make small and subtle changes based upon its estimates and then actually measure the change. Those changes that were beneficial would be kept and those that were poor would be discarded.

The system could have options for choosing fuel economy vs. power/acceleration.

Tifosi is correct, the system could eventually have it's own CPU and input/output lights/buttons, but that of course would add to the cost.

[danyl]

I spoke with the guys at Innovate today. They have already designed a product very similar to what I am talking about. I ordered one. It arrives in a few days. They are also discussing merging it with one of the programmable ECU manufacturers. What they had not considered yet was the "trial" portion of my concept.

But this may support the merit of my idea.

[slyalfa]

I think one of the missed things. why you make it rich is to cool the mix so you can add more advance.
with out knock. I think max power is hard to do with out major risk to the motor. you need to log EGT as well, melted parts are also not good. I think some simple firmware upgrade could help out on the MPG part.
but there is still a risk of melted parts. but much less under light load. just moving the timming for a fixed TPS and RPM to get the smallest PW for the same mix would boost MPG. but the best is knowen to be where damage starts aka too hot and micro dets. almost any of this would need ION feedback and EGT to work at all with out breaking the motor.

BTW I am a hardware guy for the VEMS ECU.

[bcal]

I the MS world we have megalogviewer which records and logs data and can give suggestions for a revised VE table.
I have used it to great effect with both NB and WB sensors though the WB is much better provided you get your AFR table correct to start with.

Only problem is you need to carry a laptop in the car but there is someone working on a plug in memory stick.

MS also has autotune using Megatune but I've never had much success with that.

[180OUT]

I think you could probably use neural net software w/a genetic algorithm to "learn" the operational range of the engine and then predict an optimum fuel delivery/advance curve. Probably, this wouldn't be a "drive and forget" installation, however. You'd still need a laptop or PDA. Still, it'd be cool to just drive around for a few days and then have the EFI/ECU program itself with a perfect performance map. That's quite a step away from replacing emulsion tubes in Webers.

[bcal]

Quote:

Originally Posted by 180OUT I think you could probably use neural net software w/a genetic algorithm to "learn" the operational range of the engine and then predict an optimum fuel delivery/advance curve.

[180OUT]

[quote]

Quote:

Originally Posted by bcal By genetic I gather you mean generic