Neural Network discussion

[tysonhughey]

So let me start this off with the obvious i am by no means a competent tuner. I am learning.
I've bean playing with the neural network on my charger that has the 5.7. car has no cats and i have an aem wideband that i cant even put the car into the enrichment enough range to use rite now because of it being winter and no traction. I'm working on the ide range and cruising range with the narrow bands. I'm seeing that I'm needing to pull by between -5% to -15% with in some areas in sport mode so higher rpm but low throttle showing -20% pulling in the trim. This seams excessive for a stock tune considering the exhaust should of leaned it out as it was. Is this normal? P.S. prior to removing my cats it was showing the same, so the changes made so it doesn't go pig rich worked in the tune.

also as a side question that i cant seam to find the answer to any ware. I have my wideband located in the gutted cat aka rear O2 bung. could this location potentially skew the wideband readings?
Log-0004-2C3CDXCT7JH187044.hpl
Log-0005-2C3CDXCT7JH187044.hpl

[LilSick]

not sure what year or which controller but you have the camshaft 30 degrees retarded at cruise

sport mode does not change the power output of the engine

[tysonhughey]

not sure what year or which controller but you have the camshaft 30 degrees retarded at cruise

sport mode does not change the power output of the engine

It is a 2018
Was doing 2 diffrent logs one in sport to identify areas at cruising at higher rpm

[Stoopalini]

What is your question about the Neural Network?

Post up the OEM calibrations as well as whatever was loaded when you took those logs.

[tysonhughey]

charger 007 cat Delete read.hpt02-10-2024 2018 chargeer octane red RT .010 all cam NN adjust.hpt

Here is the before working on the NN and making adjustments and after. also the stock read.04-20-2024 2018 charger octane red RT stock read.hptIt wont let me add the HTT files.
my question is I'm trying to make sure I'm going the rite way in understanding it? It just seams like the computer is reading rich and as a result is pulling fuel in the trims. the wide band is above 14.7 more than its below indicating slight lean that leads me to believe either (A) the stoich value is not 14.7 in the computer. its value is .0720 set that equals 13.888 afr but my understanding is it has other factors at play that make it equal to the 14.7. Am i correct in that understanding and if so and not if you have any insight on were its getting the error makeup from to get the almost 6% difference? (B) my wideband is having a skewed reading because of location and it should be showing richer than it is.

[Stoopalini]

What fuel are you running? I run 93 pump gas with up to 10% ethenol, thus I set my Stoich value for 0.0709 and it seems to work well. I also have NN enabled, and used the trainer to dial it in.

I also scaled out my neurons to 14 though. The process for doing that isn't too difficult, but can be tricky.

I see you zero'ed out the FA Enrich table for COT (34273), but the trigger thresholds are still set to 954*C. I would max out the fields for the COT and component protection triggers, so the triggers can never be activated.

I know it seems like zero'ing out the enrichment table should stop any additional fuel being added, but the trigger itself being activated can still cause weird things to occur (44636 as one example).

[tysonhughey]

02-10-2024 2018 chargeer octane red RT .011 thresh max out.hpt
I added the tune with those changes in it.
I run 90 only because that's the highest octane we get at the pump here in Alaska.
Looking back at the logs that i did it does look like it was advancing the spark a fair bit and thinking I need to revert back to pre NN changes and add the maxing out the threshold values to that and start NN adjustments over.
Im going to have to look into scaling out the neurons. What would be the benefit to that?

[Stoopalini]

I wouldn't revert the training you've done just yet. Try maxing out the thresholds 1st and see if you still show ~20% trims. If you do, then maybe it is the training.

Neurons are essentially the brain cells of the neural network. Sizing it to have enough, but not too many, can be tricky. A Hellcat has 11 neurons. I went with 14 because a previous tuner I worked with had good results using 14 on the GPEC2A controller, so I just went with it.

The right number of neurons improves the efficiency of the network and results in better accuracy in the output prediction. I was surprised when my R/T only had 6, so I scaled it out.

Here's a good read on the subject, if you are technically inclined: https://pmc.ncbi.nlm.nih.gov/articles/PMC8090920/


The trick is training the new neurons. If you just increase the neuron count, then export a file for training and run it through the web trainer .... The resulting trained file will have really messed up virtual VE tables because there are no preexisting weights or biases for the new neurons.

So what you have to do is export a file for training with the current setup (ie: before adding any new neurons), train it in the web trainer, then bring the trained file into VCM Editor. This should show you the current VE targets in the virtual VE tables. Copy all 25 virtual VE tables out, one at a time, and store the data in an excel file (I did one excel worksheet for each VE table).

Once you have all 25 copied and saved in an excel file, now increase the number of neurons in the ANN tab of the tune. Then export a new file for training, train it in the web trainer, and bring the result back to the NN Trainer in VCM Editor.

When you do that, the 25 VE tables will have some really strange values in them ... as expected due to the missing weights and biases. But you can fix that .... go back to your excel file and copy all of the stored VE values back to the VE tables in the NN Trainer in VCM Edit, overwriting all of the bad VE tables with the good data you saved previously. Now export the final results for training, and do the final step in the web trainer.

You can then apply the results from that final step to your tune and you will have weights and biases now populated correctly for all of the neurons you added, with targets using the same VE values you began the process with.

Once the neurons are added, and you have weights and biases populated with good data ... now you can just use the normal training process to dial in the VE with the benefit of more neurons.

[tysonhughey]

I wouldn't revert the training you've done just yet. Try maxing out the thresholds 1st and see if you still show ~20% trims. If you do, then maybe it is the training.

Neurons are essentially the brain cells of the neural network. Sizing it to have enough, but not too many, can be tricky. A Hellcat has 11 neurons. I went with 14 because a previous tuner I worked with had good results using 14 on the GPEC2A controller, so I just went with it.

The right number of neurons improves the efficiency of the network and results in better accuracy in the output prediction. I was surprised when my R/T only had 6, so I scaled it out.

Here's a good read on the subject, if you are technically inclined: https://pmc.ncbi.nlm.nih.gov/articles/PMC8090920/


The trick is training the new neurons. If you just increase the neuron count, then export a file for training and run it through the web trainer .... The resulting trained file will have really messed up virtual VE tables because there are no preexisting weights or biases for the new neurons.

So what you have to do is export a file for training with the current setup (ie: before adding any new neurons), train it in the web trainer, then bring the trained file into VCM Editor. This should show you the current VE targets in the virtual VE tables. Copy all 25 virtual VE tables out, one at a time, and store the data in an excel file (I did one excel worksheet for each VE table).

Once you have all 25 copied and saved in an excel file, now increase the number of neurons in the ANN tab of the tune. Then export a new file for training, train it in the web trainer, and bring the result back to the NN Trainer in VCM Editor.

When you do that, the 25 VE tables will have some really strange values in them ... as expected due to the missing weights and biases. But you can fix that .... go back to your excel file and copy all of the stored VE values back to the VE tables in the NN Trainer in VCM Edit, overwriting all of the bad VE tables with the good data you saved previously. Now export the final results for training, and do the final step in the web trainer.

You can then apply the results from that final step to your tune and you will have weights and biases now populated correctly for all of the neurons you added, with targets using the same VE values you began the process with.

Once the neurons are added, and you have weights and biases populated with good data ... now you can just use the normal training process to dial in the VE with the benefit of more neurons.

Thank you very much for the detailed information on that and the tips to getting it set up. I will have to play around with that so when it comes to the final revisions I can get it dialed it really good.

[tysonhughey]

I did another log session on the way to the gym and back. looks like it's getting a lot closer to my target of plus or minus 5. did another revision off that and it should be good until I can get some aggressive logging and WOT. So now my question is in the log graphs how you are setting yours up so you can get the information needed to know where to adjust considering it will be going into enrichment?

Log-0006-2C3CDXCT7JH187044.hpl
Log-0007-2C3CDXCT7JH187044.hpl

[LilSick]

So now my question is in the log graphs

how you are setting yours up so you can get the information needed to know where to adjust enrichment

open loop is the easy part

you already have that info on your chart

[Stoopalini]

For closed loop, the ANN is controlling things, and to get clean data, you should disable both COT/EOT as well as DFCO. You already have COT/EOT disabled due to removing the cats, but you still have DFCO turned on right now. Temporarily disable this so the injectors don't shut off during deceleration, causing actual EQ to go to 2.0 and throw off your logged data averages.

Once I have clean data, then for closed loop I:

• 1st look at EQ error (ie: What is being commanded compared to what is being read by the wideband). If this is off, then the PCM isn't able to achieve the commanded target. Your EQ error looks good in LOG7.
• Then I look at LTFT+STFT numbers. Yours looks to be reasonable, within 10%. There are some cells which are more than 10% variance, but this could be due to DFCO still being enabled. Within 5% is a good target to shoot for, but within 10% is absolutely reasonable and safe.
• Then I look at STFT and LTFT separately, to get an idea which one is driving the variance. For your log, it appears LTFT is making up the majority of the trim ... so I think some tweaks to the ANN virtual VE tables can certainly tighten this up for you.
• Finally, if I want to make adjustments .... I look at separate LTFT+STFT charts which filter the data by exhaust cam position. This is so you know which virtual VE tables to adjust in the ANN Trainer within VCM Edit, since the airflow model uses cam position as an input. I create 5 charts using the exhaust cam position values from the ANN trainer.

For open loop WOT, I look at EQ error with a filter against the FA Enrichment channel (you'll need to add this to your scanner channel list). Since COT/EOT and DFCO are disabled, if FA Enrichment is showing an adder greater than zero, then I assume it is due to PE Enrichment kicking in. Any EQ error showing up during PE Enrichment is where I look to make adjustments for WOT. There's some debate on the ANN air flow model having an impact here, and my personal findings from testing tell me that it does. So when I see EQ error during WOT, I will go look at my closed loop data for the same RPM range and see if there is a corresponding closed loop fuel trim variance there as well. If there is, I assume the airflow model needs adjustment and I tweak the ANN airflow virtual VE tables and test again. If there is not, then I adjust the PE Enrichment table values.

Of course, STFT should drop to zero during enrichment, but sometimes the GPEC2A PCM will carry trim values into the PE cycle. So this isn't a reliable way to identify PE. But do watch for it....

[LilSick]

For open loop WOT, I look at EQ error with a filter against the FA Enrichment channel (you'll need to add this to your scanner channel list).

why do you overcomplicate everything tom lol

EQ

that is all that matters.

first ya gotta know your target

not what the tune is commanding... but what you are looking for.

start there... not with graphs and filters

i realize you do not know what to adjust to get "there", but do you know where "there" is???

what EQ are you looking for at wot???

and do you have a baseline log???

[LilSick]

so you can get the information needed to know where to adjust..

PE

[Stoopalini]

why do you over complicate everything tom lol

I'm not over complicating it .... at least I don't think I am? Seems pretty straight forward to me. When I'm looking at data to adjust WOT fueling, I only want to see data relevant to when WOT PE is engaged ... hence I use the FA Enrich filter. I look at EQ error as an indicator of the airflow and fueling models accuracy.

I care what the tune is commanding because I want it to track in all situations, not just the one I happen to be looking at in the log at the time.

If I'm having to adjust PE Enrichment to read something like 0.70 EQ in order to hit an actual EQ of say 0.87, then something is probably off in the airflow model. Sure, you can make it work by doing it this way, but in my experience, it becomes unpredictable across a variety of scenarios, which isn't acceptable to me for a daily driver.

So I want the commanded and actual to match.

If you fudge the injectors data and PE data to hit the actual EQ target on a nice 80*F sunny day, but the commanded is far off from the actual ... then one morning the temp is down in the 20s, with a different baro due to a storm blowing in, and now all of a sudden the actual EQ isn't tracking the same anymore because the airflow model has issues.

I believe this is what happened with my vehicle when I had it professionally tuned. It ran great for a day, but the shop was 3 hours away, located at sea level. The wife drove it home, weather changed (and we're at a different elevation), and it started having stalling and lean spike issues. When I started over and applied the approach I outlined above to the tune, all of those issues were resolved and it runs very predictable/stable in all scenarios now. On a 105* summer day or a 25* cold winter morning, the startup, idle, and part throttle cruise are all stable now.

Once you have commanded and actual matching, it's pretty easy to adjust the desired EQ by just making changes to the PE tables.

I'm not sure why you wouldn't work to get commanded and actual EQ in alignment on a daily driver?

[LilSick]

I believe this is what happened with my vehicle when I had it professionally tuned. It ran great for a day, but the shop was 3 hours away, located at sea level. The wife drove it home, weather changed (and we're at a different elevation), and it started having stalling and lean spike issues. When I started over and applied the approach I outlined above to the tune, all of those issues were resolved and it runs very predictable/stable in all scenarios now. On a 105* summer day or a 25* cold winter morning, the startup, idle, and part throttle cruise are all stable now.

Once you have commanded and actual matching, it's pretty easy to adjust the desired EQ by just making changes to the PE tables.

why did you have it dyno tuned by the ocean if you live on a mountain?

why are you going into PE or WOT in 25 degree weather?

the controller and the narrow bands keep the mixture at whatever you have calibrated as stoich at idle and at cruise...

he is asking about aggressive logging (whatever that is) and WOT.

first lets see what his target is

.87 is lean for any gen III imo

they seem to like fuel : )

[HaasExp]

I'm not over complicating it .... at least I don't think I am? Seems pretty straight forward to me. When I'm looking at data to adjust WOT fueling, I only want to see data relevant to when WOT PE is engaged ... hence I use the FA Enrich filter. I look at EQ error as an indicator of the airflow and fueling models accuracy.

You are not overcomplicating anything at all mate. Filtering is the way to go so you eliminate all the "noise" that will otherwise skew your data. I also use FA>0 as an OL filter. I also use stoich FA/(stoich FA + FA enrich) to recalculate the commanded EQ as the actual commanded EQ PID refresh rate is quite slow regardless of what you set the polling too.

[LilSick]

I also use 1/ (FA+the stoich FA) to recalculate the commanded EQ as the actual commanded EQ PID refresh rate is quite slow regardless of what you set the polling too. [Yeep] agreed. but this part:

You are not overcomplicating anything at all mate. Filtering is the way to go so you eliminate all the "noise" that will otherwise skew your data.

we are talking about WOT fueling

measured in lambda

hopefully with an accurate sensor/controller

here are toms first 3 gears

a screen shot snap shot of 5600 rpm in each gear

its all over the place!

[LilSick]

36 mph at 5600 in low gear...

you should be ashamed of yourself lol

i know that truck has 370's in it

that thing would need 40's on it to go 36 mph at 5600 in low gear lol

4.7 seconds zero to sixty does not sound exaggerated tho

but both those things cannot be true

perplexed...

[HaasExp]

we are talking about WOT fueling

measured in lambda

Yes and that's what the filter is doing. Making sure the graphs to determine the EQ error are only populated by WOT and OL data. Filter by FA enrichment (I missed the enrich in my previous post so edited it). If FA enrichment is > 0, its a pretty surefire indicator that PE is enabled, and you are in OL mode. IF you AND that with a relative pedal voltage > than your PE enrichment trigger, then you can also filter out COT enrichment events.
The EQ ratio is measured by the wideband in lambda. The target EQ or EQ cmd is also in lambda. You can see it for yourself in the screenshots you posted.