Neural Network logging and tuning

[MeanMike]

I'm getting very close on my truck tune. It's a 2020 Silverado with an L84 5.3, 10 speed and an E90 Global A ecu. I've swapped out the DFM cam for an L8T cam, LS7 lifters and BTR beehive springs. I just installed a Magnuson 2650 with the Lingenfelter IAT Baro breakout box. Lingenfelter suggested running the factory 1 bar map to the supercharger inlet and ignore the pressure after the rotors. So, it never sees positive pressure or a PR above 1. This means it actually drops back a little at high rpm WOT as the inlet becomes restricted. PR ~.96 at higher rpm.

I've got the maf pretty close, but I seem to get variation in the upper flow ranges. Maybe my wideband is inconsistent, The sensor is old and seen a lot of miles in my CTSV. But the truck is running great as is.

My issue is with the Neural Network. I'm on about the 15th revision and I'm struggling to get the low and high PR ranges dialed in. I can't get a lot of data in my logs to fill that in and with all the filtering for the multiple cam angles, it's hard to get a lot on each histogram. What is the strategy for that? Do you set the cam to each angle in the tune and log each time you change it? Otherwise how do you get higher rpm data on the low cam angles and vice versa?

This is the closest I've gotten on the NN tune. I have individual histograms, but this is an overall so that I can watch it and fill in while driving. I have the trained NN file to look at, but it's not an uploadable file extension here. I can't seem to upload the log either, is there a file size limit?
NN datalog.jpg

[MeanMike]

I had to export a smaller range of the log to make something that would upload. But it's basically showing similar histogram results.

[smokeshow]

If you don't actually capture MAP, you won't be able to do the speed density portion. At least not with that raw data capture...speed density needs pressure.

[MeanMike]

I understand that. I know it’s part of the math. I had to fake in the Map torque model too. Lingenfelter didn’t say why they do this, but from reading here, I assume it’s something to do with the inaccurate throttle body airflow calculation by not also having the inlet pressure sensor. If anybody knows what pin to add that to on an E90, I’ll gladly do it.

I have a plan to use the 2.5 bar and tune that also. I actually have a base file ready with extended range for PR in the neural network trainer. I was trying to get this scenario dialed in during this week and swap to the 2.5 bar map this weekend and see what the issue really is if any.

[MeanMike]

I swapped it over to the 2.5 bar and loaded the corresponding tune. Other than the VE being very off, no ill driving effects for a quick lap around my neighborhood. So I guess I’ll start working on those VE tables tomorrow. But, same question, is it completely necessary to populate each cam angle table? Or is it best practice to retune the necessary spots in the VE tables if you make cam angle changes? If you fully tune each table, do you just “lock” the cam at specific angles in the tune and work one table at a time?

[CYN_CRVR]

I think it is important to understand they are not actually separate tables. There is one equation that uses the VE coefficients to determine the airmass calculation for all RPM, MAP/Baro, and cam angle values. HPTuners' VVE Table Editor is just a handy interface that tries to back-calculate the coefficients from the user-editable VE surface. This means that editing any one of the tables also affects all the other ones.

It's also worth noting that I've gotten strange results when using the VVE Editor for non-zero cam angle tables. I ended up freezing my cams at 0 to dial in my base table, then unfreezing the cams and manually adjusting the remaining cam-dependent coefficients. But I would only attempt this if you have a solid understanding of the VE equation.

With all this in mind, for most people I recommend just dialing in your base table (don't bother freezing the cam), and as you said, re-tune specific zones as necessary if you make cam angle changes later.

[MeanMike]

I didn't think about that, but it's definitely something to consider. I don't know squat about how a Neural Network works, but can it really be considered math or have an equation behind it that we could calculate outside of the model?

I spent time yesterday dialing in each cam angle starting at 0 and working through each higher angle successively. It's taking about a 30 minute drive per iteration to populate a histogram and I'm doing 2-3 iterations to get each table to +/- 5%. I'm manually changing any areas that I can't get in the histogram. I still get big negative error at the low PR regions that the motor can't hit.

First, I dialed in the 0 angle table (106 in my tune). Then added 5 degrees for the next, dialed it in and so on. I just logged and changed the +20 degree table (147) this morning. I went back and plugged into excel the original numbers from the 0 angle final result file and copied the 0 angle from the most recent result. There have been 6-8 edits of other tables in between. The values have changed by .001-.007. I might have tweaked the idle area of the 106 table somewhere along the way. So for mine, it doesn't seem that editing one table has a significant effect on others.

I hope to finish up the individual tables tonight. Then I'll put the cam timing tables back to stock and log the overall histogram and see what that gives me.

[MeanMike]

I went through a couple rounds of NN VE tuning. I tuned each cam position table separately. To do that I had to set each cam position table in the tune. Airflow/Camshaft timing/ Minimum, Low, Medium and High Baro and WOT. I have flex turned off so I didn't touch those tables. The cam timing change is half the Intake Valve position change is. So, 5 degrees in the table will move the IVP 10 degrees. I used smokeshows GMVE CL and PE math and filters from the same thread with my fuel trim cell value of 8 instead of 9. I also added filters for each cam position basically halfway between each table value. Example

PE filter for 149*
[2077.161]>143 and [2077.161]<154 and [2517.161.avg(200)]=0 and [2517.161.avg(-200)]=0 and (abs([50090.156.slope(250)])+abs([50090.156.slope(-250)]))<2 and [6310.avg(250)]=8

CL filter for 159*
[2077.161]>154 and [2077.161]<165 and [2517.161.avg(1500)]=0 and [2517.161.avg(-200)]=0 and (abs([50090.156.slope(1500)])+abs([50090.156.slope(-500)]))<2 and ([6310]=8 OR [6310]=14 OR [6310]=15)=0

This along with some careful steady state driving got me to +/- 3% error for all for all of the NN tables and the maf. The truck runs great with very little other changes.

Running it with the LSA map sensor after the rotors as my MAP signal gave no big issues. I'm still playing with driver demand tables and some trans tuning, but it would be considered good enough by most people.

[Pavetim]

How did you get it to run MAF only?

[MeanMike]

You can’t do that with an e90 ecm. You have to tune both the maf and nn at the same time.

[Pavetim]

That's what's confusing me how you do both at the same time. So you can't just crank up the rpm to 8000 on your min HL ? Per that description that should turn off NN and hence run MAF only? I got the credits and think I have my individual cam handles setup on my scanner now I want to see if my NN/VE and my MAF are good. I will need for sure when I do my intake and cam anyways

[MeanMike]

I'm not expert, but seems like raising Min Range HL is just going to lower the min rpm value in your NN table. But the actual NN/VE can interpolate beyond those bounds.

The biggest thing to understand is that NN/VE and maf should roughly match at steady state. Once you can wrap your head around that, then you can start looking at filtering out non steady state (transient) data in your histograms. Then you can get into how you actually calculate your errors. I recommend you read this thread a few times. https://forum.hptuners.com/showthrea...ight=smokeshow

You can start by logging these parameters and see how closely they match.
Screenshot 2025-01-20 062708.jpg

If you haven't changed the cam don't bother trying to edit the NN. The NN is pretty good and it seems like it can learn quite a bit.

It was frustrating for me that I couldn't even look at the NN tables without using credits and then I only got 30 days. So, message me your email address and I'll email you a trained base file that you can open to look through.
Screenshot 2025-01-20 063648.jpg

[Pavetim]

I already paid for the trainer and I want to practice so I know how to do it all when I do my other mods. and thtas the conundrum here so I log those 3 and the MAF or NN veers off from one another then How do I know what one is wrong cause NN and MAF use different scaling graphs so can't just directly compare them unless you know a way to directly correlate them. Unless I just use the graph vs time but that's a pain in the ass to look at. NN uses pressure ratio vs RPM and MAF uses Hz or can I. Build a graph like NN but use MAF airflow as the data points? And why can't you just disable MAF i mean is.t that the point of having these multiple systems in case maf fails we have VE? Everybody keeps saying maf is always enanled. It if you fail it in the tune and unplug it then it cannot be enabled right? I will read that post ypu linked up but that's gen 4 and makes it more complicated with all that math, I mean so much easier to just disable and unplug math takes 30 seconds but all that math is a pain if you don't and I really don't understand it all yet

[MeanMike]

Smokeshows thread explains how you know how to correct VE/NN vs Maf using Dynamic Airflow and AF error. Those airflow numbers are the same units g/s, but they are derived in two different ways. Maf works best in steady state, but doesn't handle transitions well. VE is a better source for airflow prediction when in transition. Dynamic airlfow is a blend of both VE and Maf using either when they work best. I can't remember where I read it, but I believe Smokeshow is on the patent for that strategy.

If you unplug the maf, you'll lose your barometric pressure (thus pressure ratio) and IAT. Both are used for the VE calculation.

If you want to tune one of these correctly, you're going to have to have a decent understanding of the math. The math is the same for Gen 4 and 5. If you don't want to learn it, you'll be better off paying someone else.

[Pavetim]

OH ok that makes sense all this time if somebody just said that. I didn't realize this engine had the MAF and IAT and MAP in one sensor

[Pavetim]

Don't suppose anybody wants to make a little $$ and set this up for me i am a little lost. I can send my channel, graphs, configs and everything and you setup my math's, graphs, filters, etc. I can tell my MAF is off now I changed my intake just at first glance. I would pay somebody $100. Conditions are good in seattle now for tuning but really want to get this done before I head to montana in 2 weeks. Think I got spark and knock where I want it just want to dial in my MAF and NN/VE.

[4wheelinls1]

Don't suppose anybody wants to make a little $$ and set this up for me i am a little lost. I can send my channel, graphs, configs and everything and you setup my math's, graphs, filters, etc. I can tell my MAF is off now I changed my intake just at first glance. I would pay somebody $100. Conditions are good in seattle now for tuning but really want to get this done before I head to montana in 2 weeks. Think I got spark and knock where I want it just want to dial in my MAF and NN/VE.

You can copy the AFR percentage error formula in the scanner from AFR error, just make the PIDS VE airflow and MAF airflow. Build a Graph with the same axis as the NN trainer and log the math parameter you just created; I call mine airflow error. Be steady and get a reasonable hit count. You can also replicate the graph multiple times for reach cam position in the trainer and filter each by cam angle +/- 4 degrees or similar. Copy past by percentage the error into the appropriate NN can angle and smooth logically based on the acquired data for each angle.

[1scls1z]

Don't suppose anybody wants to make a little $$ and set this up for me i am a little lost. I can send my channel, graphs, configs and everything and you setup my math's, graphs, filters, etc. I can tell my MAF is off now I changed my intake just at first glance. I would pay somebody $100. Conditions are good in seattle now for tuning but really want to get this done before I head to montana in 2 weeks. Think I got spark and knock where I want it just want to dial in my MAF and NN/VE.

When starting out tuning it doesn't hurt to pay a reputable remote tuner for a base tune and channel config. They will likely even help you setup the custom maths needed to put together the correct formulas for histograms. Or you can spend weeks, months, or years trying to figure the shit out on your own.