- Fri Jun 11, 2021 5:14 am
#51006
Hi.
Many commercial ECU's , f.e. Bosh Motronic have "AI" built in , wchich is basically simple pre-trained model able to detect change in input parameters and switch map to one fitting the trained model.
See https://eloquentarduino.github.io/2020/ ... -your-mcu/ to get some idea of example implementation.
Speeduino gets loads of multidimensional vector data, basic one is lambda response and saturation time vs. AFR setting computed out of MAP sensor , throttle position etc.
This obviously changes with air temp, humidity , pressure and so on, and normally Motronic detects that using similiar (though much older and simpler) "neural network" model, and them just switches map , interpolating basing on confidence rate.
This expands code size slightly, as one needs more maps for boundary/special conditions, but gives loads of cool features , like ability to detect and run with failed sensors (using failsafe maps) and thus bringing speeduino more close to "real" ECU's.
Best thing is though, if one uses decent uC , like teensy with real beef under the hood, it gives enough room for much more complex models than Motronic did allow, leaving it far behind. Like... really Far
Also one can use SVM to track and predict state of lambda , basing on actual engine state, allowing use of simple lambda instead of wideband and still getting "CSI zoom" effect in the resolution. SVM model can infer it from lambda response time, given there is testing fuel trim pattern to allow it coded in too.
Greetings and have fun
Many commercial ECU's , f.e. Bosh Motronic have "AI" built in , wchich is basically simple pre-trained model able to detect change in input parameters and switch map to one fitting the trained model.
See https://eloquentarduino.github.io/2020/ ... -your-mcu/ to get some idea of example implementation.
Speeduino gets loads of multidimensional vector data, basic one is lambda response and saturation time vs. AFR setting computed out of MAP sensor , throttle position etc.
This obviously changes with air temp, humidity , pressure and so on, and normally Motronic detects that using similiar (though much older and simpler) "neural network" model, and them just switches map , interpolating basing on confidence rate.
This expands code size slightly, as one needs more maps for boundary/special conditions, but gives loads of cool features , like ability to detect and run with failed sensors (using failsafe maps) and thus bringing speeduino more close to "real" ECU's.
Best thing is though, if one uses decent uC , like teensy with real beef under the hood, it gives enough room for much more complex models than Motronic did allow, leaving it far behind. Like... really Far
Also one can use SVM to track and predict state of lambda , basing on actual engine state, allowing use of simple lambda instead of wideband and still getting "CSI zoom" effect in the resolution. SVM model can infer it from lambda response time, given there is testing fuel trim pattern to allow it coded in too.
Greetings and have fun