Neural Network discussion

[LilSick]

i realize all that...

i use two channels to know if its in closed loop or not:

at a glance i use the fuel system status

but what i really use is relative pedal V cuz it does not lag at all

my trigger is set at 2.72v and relative pedal is always on the top row in my chart ^^^.

i really do not see the point in using a graph for his 1-2-3 blast when i can see everything clear as a bell on the chart

and what is clear is that his truck, like my car, is not very consistent in the lower gears!

mine is leaner in lower gears too

tom has seen my curve... its A LOT better than his rofl

smogged out pos controller or maybe a shitty tuner

[Stoopalini]

Don ... my vehicle does not have 370s. The OEM gear ratio on a Durango R/T is 3.09. The SRT has 3.70, but not an R/T.

The point of the log I posted (where you grabbed those screenshots) was just to test the WOT shift point RPM in 1-2 and 2-3. Yes, the EQ is still drifting off target 5-10% during WOT, and my assumption is that is due to forced induction (ie: lots more air coming in). In the meantime, I have commanded PE set to a safe range until I get around to licensing the PCM in the ANN Web Trainer again to tweak it further. It's actually a good example of what I was explaining earlier ... that the airflow model being off can impact WOT fueling ... even more with boost if the airflow model isn't tweaked for it.

My Stoich is 0.0709, and I am commanding 0.0209 in the upper RPMs when under boost. So 0.0918 final gives me a Lambda EQ target of 0.772 and it's tracking right around 0.830.

Stoops_Fueling.JPG

[LilSick]

EQ target of 0.772 and it's tracking right around 0.830.

you must be looking at a different log

i am seeing closer to .850 in that log





i was looking at a different sig

james's

he has the 3.70's

my bad

.830 huh? :

[LilSick]

i still say i can fix that ^^^ without paying for any NN trainer...

[Stoopalini]

Yeah, an average right around 0.830.

EQ_Error.JPG

Like I said, it ranges from a 5% - 10% variance from commanded when in PE Enrich mode and under boost (ie: above 1.0 PRatio). Which is why I think I can correct for it in the NN trainer; since that is a big piece of the airflow model.

I could fix it too without the ANN Trainer, but I'd prefer to fix the root cause in the airflow model rather than tweak the injector or PE data. I'd bet I could also adjust the cam position and have an effect on it too, but I'm happy with the VVT right now.

Not sure why you're so fixated on this though? The OP was asking about using ANN to dial things in, and I think the approach I outlined is a solid way to do it.

[LilSick]

I think the approach I outlined is a solid way to do it.

i pointed out (using pictures from your log) that the mixture can vary significantly from gear to gear

in other words - using a graph with a filter is not a solid way to "do it"

look at your log

your EQ is all over the place



i think the op and your help was on point until he moved from idle/cruise to wot

maybe i am wrong about ANN and WOT

i appreciate your efforts to help figure this out tho, i really do : )



i think you should make another log doing a dig, but staying in it for a couple more gears this time

[Stoopalini]

I dunno what you're seeing in the log man, but I see the EQ rising through the RPM range, topping out at 0.86, then dropping back down to 0.77 (commanded) at the shift and rising again through the next gear's RPM climb. I guess I just don't see that as being all over the place. It's pretty predictable. I also have an AEM X-Series wideband, not a Ballinger ... so there's that.

EQ.JPG

I'm pretty sure it's happening because the NN airflow model needs to be adjusted for upper RPM boost.

I know you don't agree the ANN airflow model has an impact on WOT fueling, and that's a point of disagreement. I don't see how it couldn't though, due to the NN providing the airflow input.

FAR PE targets set the fuel air ratio for enrichment. The PCM needs an airflow number to use in that calculation, otherwise it wouldn't know how much to engage the injectors to hit the PE FAR target ... where would the airflow input come from if not the airflow model? And in open loop, there's no feedback loop to tell the PCM if it's correct or not, so it blindly uses the PE amount even if it's wrong. I know you can tweak the injector data, so the PCM can use an incorrect airflow number and still hit desired EQ ... even if not matching commended. Or you can use PE table to do similar ... but in both those cases, the airflow would still be off and you'd just be compensating for it on the fuel side.

What I do know is I adjusted the NN in the upper RPMs and it absolutely had an impact on the WOT fueling. I was seeing much greater than 0.86 prior to doing the ANN adjustment. It always started at commanded, but was rising up to peak around 0.90 initially, and I was definitely not comfortable with that. So I adjusted the high RPM, high PRatio, rows of the virtual VE tables, and now it's topping out at 0.86 instead of 0.90.

Due to that experiment, I'm inclined to believe the NN airflow model is having an impact on WOT calculations and I just need to make further adjustments to have actual EQ track the commanded EQ during WOT PE mode.

I'm not doing this on a track, so I cannot go WOT through 1, 2, 3, 4, etc ... I'm not comfortable doing 110+ MPH on my back roads here LOL!

[HaasExp]

look at your log

your EQ is all over the place

The actual EQ ratio looks pretty consistent to me compared to the 0.78 cmd EQ reference line.
Doesn't swing much either gear to gear

EQ.PNG

[tysonhughey]

thank you everyone for the input it is really appreciated.

My WOT goal is 12.5 but read that the hemi likes a bit richer than normal. is this true? WB when doing a small pull on the one clear part of the road at upper range of rpm was around 12.3. The problem I'm having rite now is no traction to get a good log I just want to make sure that i have all the information set up on my logger and an understanding of what i should be looking for and were to make the proper adjustments. When i did give enough throttle (I didn't have the logger in the car and i still have DEFCO enabled) it did seam like the enrich kicked in a bit later than I would think it should or just with un knowledgeable of the topic it seams. Is there a benefit to it kicking in earlier on a basically stock NA engine by changing the WOT Pedal thresh?

in what to modify correct me please if im wrong

FA Stoich .0720 = 13.88 AFR getting 14.7 on WB that =s 5.9% error

in power enrich aircharge and P-ratio set to desired AFR range with the 5.9% deducted to get the values for instance current ends in .0161+.0720=.0881 11.35 AFR add 5.9%that get me a 12.01

I would adjust the air charge and p-ratio tables to reflect the desired i think this will make the eq ratio graphs that I'm still trying to figure out how to setup corollate to the desired.

after some logs with WOT pulls getting good data in cells take that information to the Speed density VE tables bank 1 and make small adjustments (what is bank 2 for?)

with taking Stoopalini's method of adjusting the NN in the same areas as were the VE table or tables are being adjusted. Thought proses on this is if its not changing anything it wont hurt anything but if it does than it will give it the ability to adjust itself to changing environments heat / elevation. and will not be fighting the adjustments to the VE potentially.

[tysonhughey]

sorry about the delayed responses been hooked up with some solar classes for work.

[HaasExp]

FA Stoich .0720 = 13.88 AFR getting 14.7 on WB that =s 5.9% error

You really should just start using lambda on your wideband otherwise things can get very confusing if your fuel does contain some E.
While cruising in closed loop, the factory O2 sensors will always target lambda 1.00. this will be equivalent to the stoichiometric AFR of the fuel used. It will always target 1.00 lambda regardless of the stoich AFR.
As your wideband is displaying the pure gas scale (E0), it will multiply the lambda by the stioch AFR of gasoline, which will give you 1.00 x 14.7 = 14.7. This is why it displays 14.7 even though your stoich AFR is set to 13.88. Again, because its targeting lambda 1.00 in closed loop. So, there is no actual error. The wideband is reading correctly because its setup for pure gasoline and just converting the measured lambda accordingly.

Now an FA stoich of 0.0720 assumes you are running E15 or there abouts. If you are not actually running E15, then this should be changed to match the fuel you will usually be running. Otherwise, your closed loop trims will be off, and your OL error will also be off. This error will then just be baked into your VE tables to train the ANN with. Yes, it will still work but if you are going through all the effort to train the ANN then you may as well minimize all the errors from the data you are actually using to train it IMO.

[LilSick]

Now an FA stoich of 0.0720 assumes you are running E15 or there abouts.

that would make 100% sense except the factory uses numbers ranging from .0688 to .0720 and they have no idea who the customer is getting their fuel from : )

[Stoopalini]

When i did give enough throttle (I didn't have the logger in the car and i still have DEFCO enabled) it did seam like the enrich kicked in a bit later than I would think it should or just with un knowledgeable of the topic it seams. Is there a benefit to it kicking in earlier on a basically stock NA engine by changing the WOT Pedal thresh?

How do you know PE Enrich kicked in later than desired if you didn't have a logger connected?

FA Stoich .0720 = 13.88 AFR getting 14.7 on WB that =s 5.9% error

This is not the way to look at these numbers. HaasEXP explained already though. The Stoich parameter in the tune just tells the PCM what FAR is needed to reach a complete burn of the fuel you are using (ie: Stoich). These fields do not allow you to change the targeted AFR at all. Actually, the PCM doesn't use AFR anywhere. The 14.7 number is just something the gauge you have is using as stoich. As HaasEXP suggested, switching the gauge and your logging to Lamda instead will go a long way toward avoiding confusion.

The stoich value doesn't matter for closed loop at all, as there is feedback from the narrow band O2 sensors when running in closed loop. This allows the PCM to adjust the fueling in real time, with intent to hit 1.0 lambda. In open loop, the Stoich is used along with any active enrichment (if activated). Just know that open loop is not the same as PE enrich mode though. Open loop can happen without PE enrich activated.

EQ error is not something to look at when operating in closed loop. Use the fuel trims instead, assuming you are hitting 1.0 lambda. You want to know how much the PCM is needing to adjust to hit 1.0 and fuel trims will tell you that (ST plus LT trims). The only time EQ error may matter in closed loop is if the trims cannot add or remove enough fuel to hit 1.0. If that's happening, then there's bigger problems at play though.

in power enrich aircharge and P-ratio set to desired AFR range with the 5.9% deducted to get the values for instance current ends in .0161+.0720=.0881 11.35 AFR add 5.9%that get me a 12.01

Not really. Just to reiterate, AFR isn't really something you should be using here. The GPEC2A controller does not use AFR anywhere. That said, the PE tables are only used when the PCM is in open loop and PE Enrichment is being commanded. The % change to make in the cells within these tables needs to be informed through your open loop, PE Enrich mode, WOT log data. I am assuming your actual lambda is matching your targeted lambda though. If these are not matching, then you need to make other adjustments elsewhere 1st. But as described earlier in this thread, some folks will adjust the PE tables anyway and not worry about actual matching targeted (or commanded). For your non-boosted application, I wouldn't expect this to be too far off though, and you've already experienced the NN Trainer ... so why not tweak it to align the actual to commanded?

If actual is matching commanded through the RPM range, then your airflow model is setup well and making small adjustments to the WOT fueling becomes really easy. You just dial in the PE table. And you can target whatever value you want. The trick is then determining what lambda is best for your motor. That can be figured out on a dyno or with time slips at the track.

For my boosted application, things need quite a bit of tweaking in the boost area since the operating system wasn't setup for boost (ie: PRatio being higher than 1.0). But this doesn't apply to you ... so apologies for all the noise in this thread related to that, a lot of it probably isn't relatable to your scenario ... but if you understand it well enough, it will provide you some good insights into what is going on.

I would adjust the air charge and p-ratio tables to reflect the desired i think this will make the eq ratio graphs that I'm still trying to figure out how to setup corollate to the desired.

The aircharge and PRatio PE tables are only for when the PCM is in open loop and PE enrichment is being commanded. The FA Enrich channel in VCM scanner will show you any and all enrichment happening, not just PE. For example, it will show you COT. Logging this is a good way to know what is going on though.

after some logs with WOT pulls getting good data in cells take that information to the Speed density VE tables bank 1 and make small adjustments (what is bank 2 for?)

The Speed Density VE tables have extremely limited bearing on your setup. The Bank 1 and Bank 2 VE tables in the Speed Density tab are not used when the Neural Network is turned on. The NN provides the airflow mode instead. If you tune NN off though, the PCM will then use these VE tables instead. The NN is much better though, as the airflow model with NN enabled can account for many more variables than speed density can, and it's a prediction model, so it is faster in responding. Cam position is one example of this. The NN can adjust VE (ie: airflow model) for the position of the cam which Speed Density cannot.

That said ... There is some thought that the Speed Density VE tables can sometimes be referenced when NN is enabled, like during transients as an example. For that reason, I will adjusted mine to suit my setup. If it is referenced during transients, then I'm covered and won't see small anomalies in the data. If it is not referenced, then nothing is lost/gained by tweaking them.

Just know that when NN is enabled, the Speed Density VE tables are not really used by the PCM for closed loop nor open loop operation.

The Bank 1 and Bank 2 refer to the cylinder banks. Driver's side and Passenger's side ...

with taking Stoopalini's method of adjusting the NN in the same areas as were the VE table or tables are being adjusted. Thought process on this is if its not changing anything it wont hurt anything but if it does than it will give it the ability to adjust itself to changing environments heat / elevation. and will not be fighting the adjustments to the VE potentially.

Not sure what you mean by this. The Virtual VE tables accessed through VCM Editor's "VE Neural Network Trainer" are where you make changes to the VE. Not the Speed Density VE tables (unless you turn off the Neural Network that is).

You can ignore the Speed Density VE tables for now. The stock setup of those should be fine if they are being referenced during transients.

[tysonhughey]

Ok understood Lambda. thank you once again for everyone with their help and patience it is very appreciated.
I went to the gas station that i usually get my fuel from and asked what the ethanol content is in their fuel he told me 0%. If that is correct based on what Haasxp said i would of already baked an error using NN by missing step 1 of setting Stioch by changing it to FA .068 for non ethanol fuel. I'm still unsure on if it is non ethanal so i have a tester on the way because none are local. This way i can be for sure i know my targets.
I am hitting a Lambda of 1 at idle and at cruise and it is very stable not bouncing around like it was at the start with most i see being within the +/- 5% in my STFT and LTFT. Sadly i will be back to adjusting for those ranges if the test results in non ethanol. Potentially will not have to make any changes based on what Stoopalini in that it does not use those values in close loop mode but in the case wouldn't hurt to look over those areas.
I'm probably looking to stay in a richer range working up to .84 lambda at max range. FA would be deciphered on what Stioch is set to.

does this look correct for my EQ error graph setup?
Screenshot 2025-02-15 102913.png

I also added FA enrich to my gauges monitored and changed the AFR values to read in Lambda
hopefully by the time i get back from my vacation i will be able to get a log in so i have something to play with while I'm at work for 2 weeks after.

the goal is to have the understanding with adjusting minimalist changes in the factory tune so that when i go to boosted situations and adding factors of scaling the tables for the different map Bar and injector scaling its not so overwhelming and i can slowly begin adding in boost and make adjustments accordingly. I still don't know how I'm going to accurately or even safely adjust the spark tables without a dyno but that's a bridge i will cross when i get there.

[LilSick]

you can listen to those guys if you want but i am telling you...

you are wasting your time.

because of lag the graphs requires too many hits to work for one or two wot pulls

ditch yer gauges too and just use the chart for this

tuners been using charts for wot fueling since the mid 1980's : )

the NHRA let us monitor the flow rate with electronics... we just could not 'control' it electrically

and no, that graph is not going to work. it is lacking break points and more

[Stoopalini]

I guess it depends on the goal. If the goal is to just get WOT fueling into a safe place and be done with it, then maybe a waste of time ? but if the goal is to learn how it?s actually working, and how all the controls function, then it?s time well spent IMO.

I?m so glad we have way more modern tools than they did 40+ years ago The advancement of the neural network is definitely a game changer. I imagine it?ll get more and more advanced over the next 10 years too.

[LilSick]

a NN really is a game changer...

just not for WOT.

prove me wrong - i love to learn!

and my ego can take it : )

show us your graphs from each pass next time you go to the track

[Stoopalini]

I've already tested it and don't really need to prove it more, but here is some views to show it. And Don, you know very well I have no plans, nor the desire, to take the wife's SUV to the track LOL!

Here is before the adjustment to the ANN. Top of 1st gear, EQ maxed out at 0.90 with a VE of 90.5. The VE being reported here aligns to the virtual VE tables in the NN trainer.This was OEM virtual VE values, ie: prior to me doing any training of it.

EQ_Error_Before_ANN.JPG


This next one is after the ANN adjustments. I increased the virtual VE values at the upper RPM, upper PRatio cells, for the cam intersection I had set for WOT. Still a 1-2 shift, EQ maxes out at 0.86 now instead of 0.90, with VE now at 101.5 instead of 90.5, and you can plainly see the obvious improvement of the EQ curve leaning out through the RPM increase.

EQ_Error_After_ANN.JPG


I think you re-enabled NN on yours right; After we had that long discussion on how it works? If so, you can easily test this for yourself next time you go to the track.

Oh ... and also notice the relationship between VE and EQ in the graph ... prior to training the NN, the blue VE line in the graph drops more as the RPMs increase, and the red EQ rises proportionally. After I scaled the virtual VE tables out to 1.5 PRatio, increased the values, and trained it .... the blue VE line now dips less and the red EQ line rises less.

Yes, still work to do, but it clearly shows me the airflow model in the NN definitely has an impact to WOT fueling. The reported airmass further supports this as well.

[LilSick]

you cannot compare up at the peak like that

torque intervention is already in effect

post the logs

or pics at 5600 like i did

second pic has 7 degree less timing in it at that point

speaking of which, what is going on with the timing nearing the top of 2nd gear bro?

[Stoopalini]

The point was to show the WOT EQ impact from NN training. It clearly shows that. I chose the top of 1st because that was the highest WOT EQ error on both logs. 5600 rpm would show the same thing, just a lesser EQ error …. Just look at the charts.

According to the chart, KR kicked in at top of 2nd. Pulled timing.