Prediction Coefficients

[gmorris]

Under the bus? Not sure how you threw me under the bus Just glad your liking it Josh - think your really going to like the gains that the ability to add timing will give you

I pretty much do injection timing different than anyone out the cc-rider... As Josh and others have seen - it even works rather well with stock cams...

I know it made a big improvement in my ability to run timing on my heads/big cam 418. I didn't realise you were also seeing improvement on stock cams. Damn...something else to do on the ZL1 lol.

[GHuggins]

You have to be a little more "gentle" with stock cams - other than this, yes it's nice

[Michael_D]

Why is that? Unless DynAir is doing unwanted things, which it isn't in the print screens above, there's no need to zero everything out.

I've been running with
- base 0.9 except cells 11, 17, 23, 29 where base is 0.95
- current 0.1 except the cells listed above where current is 0.05
- old 0.0
- corrected 0.1

Yesterday I changed corrected to 0.5 like it is in LS9 and noticed immediately that some of the problems at idle/during cold start with pure SD mode came partly back. This led me to think that the corrected gain may actually be some sort of weighting factor between MAF based part and GMVE based part.

The engine seemed to like this change on the general level, so I may experiment more by
- using more MAF based values (lower) at idle
- using more GMVE based values (higher) above the idle/low load

Edit. This is going in next based on the above assumption and the actual ave/min/max MAP vs speed vs TPS data in normal traffic:





Edit2. This is a comparison of the stock LS7 parameters shown in HPTuners and EFIlive, pretty much the same jsllc showed earlier, but with variable names (as interpreted by each company):



Btw, note how GM uses higher corrected gain in cells 2 & 3 where the very low speed cruising and tip-in + clutch off takes place.

Can you explain why one would want lower gain values in the higher MAP cells?

[barum]

You mean the corrected gain? The purpose was to rely more on the MAF and less on VE as there's much less dynamics on those areas.

However, having anything else than zeroes in this table (when running HCIE hardware) is guaranteed to give worse DynAir prediction with some occasional goofy results than with the pure row of zeroes.

[mowton]

Why is that? Unless DynAir is doing unwanted things, which it isn't in the print screens above, there's no need to zero everything out.

I've been running with
- base 0.9 except cells 11, 17, 23, 29 where base is 0.95
- current 0.1 except the cells listed above where current is 0.05
- old 0.0
- corrected 0.1

Yesterday I changed corrected to 0.5 like it is in LS9 and noticed immediately that some of the problems at idle/during cold start with pure SD mode came partly back. This led me to think that the corrected gain may actually be some sort of weighting factor between MAF based part and GMVE based part.

The engine seemed to like this change on the general level, so I may experiment more by
- using more MAF based values (lower) at idle
- using more GMVE based values (higher) above the idle/low load

Edit. This is going in next based on the above assumption and the actual ave/min/max MAP vs speed vs TPS data in normal traffic:





Edit2. This is a comparison of the stock LS7 parameters shown in HPTuners and EFIlive, pretty much the same jsllc showed earlier, but with variable names (as interpreted by each company):



Btw, note how GM uses higher corrected gain in cells 2 & 3 where the very low speed cruising and tip-in + clutch off takes place.

This is a great thread and a great chart barum. I like the way you presented the coefficients relative to the Zones. I love excel and the graphics we can generate. Quick question, did you set up the Zone boundary chart from the Editor table through "math" or long hand? I too have not had much luck zeroing the coefficients. I log Dynamic Air vs MAF to see when and where they deviate and work from there. Tuning both the MAF and VE (apparently becoming a lost art) seems to get me close. I think I had one car which was a Camaro L99 to LSA conversion where I had issues where I corrected with the coefficients. And if I remember right I merely pushed the High RPM Disable down to like 3000 rpm to cover the TIP-in/nonsteady state corrections in the NA regions.

I think also I needed to resond to a question on my previous post which was yes to "almost" zeroing" the Burst knock and no to the coefficients.

Thanks again for a great thread guys.

Ed M

Ed M

[Michael_D]

I'm not ready to throw in the towel and disable a table that has an attractive up side. Using the MAP you posted above provides a very good visual to help me understand what each corresponding cell references. What would be helpful, is to understand how GM came up with the values in the cells.

[barum]

Ed, the zone boundary chart was made manually. I haven't studied how, if possible at all, to do dynamic graphics through VBA. I use VBA for more tricky tasks which resemble ECM code execution etc. and it would be possible to calculate and populate a fixed layout like the zone boundary chart. However, if we are able to find parameter combinations through trial and error only, there's not much point in that exercise.

Michael, I take it you are referring to corrected gain? By all means, all efforts to tame these tables are welcome. What GM has done with Gen4 reads Matlab/Simulink all over it. VVE, Torque delivery calculation, DynAir filtering, you name it. This means the values are derived from Simulink models which are based on extensive test runs on stock engine. To be able to duplicate the work for a tuned engine is more or less impossible for anyone else than GM people who did the original work.

[mowton]

Michael, I take it you are referring to corrected gain? By all means, all efforts to tame these tables are welcome. What GM has done with Gen4 reads Matlab/Simulink all over it. VVE, Torque delivery calculation, DynAir filtering, you name it. This means the values are derived from Simulink models which are based on extensive test runs on stock engine. To be able to duplicate the work for a tuned engine is more or less impossible for anyone else than GM people who did the original work.

By looking at it the way you presented it (which again I think is great), I wonder if we can handle it much like the Torque Model Airmass A-D where I log the Delivered Eng Torque (in this case the MAF vs Dynamic Air differences) vs. RPM/MAP and based on the RPM (in this case assume Zone 17, 23 and 29 for Wide open throttle stabs) adjust the coefficient to bring the Delivered Torque in line with an existing Dyno curve, advertised engine dyno curve or "reality" (in this case reduce the transient induced filter differences between the MAF and Dynamic Airflow). This instead of zeroing it all and losing the positive effects of correction which GM so graciously provided us in the drivabiliuty area's?

Ed M

[Michael_D]

Ed, the zone boundary chart was made manually. I haven't studied how, if possible at all, to do dynamic graphics through VBA. I use VBA for more tricky tasks which resemble ECM code execution etc. and it would be possible to calculate and populate a fixed layout like the zone boundary chart. However, if we are able to find parameter combinations through trial and error only, there's not much point in that exercise.

Michael, I take it you are referring to corrected gain? By all means, all efforts to tame these tables are welcome. What GM has done with Gen4 reads Matlab/Simulink all over it. VVE, Torque delivery calculation, DynAir filtering, you name it. This means the values are derived from Simulink models which are based on extensive test runs on stock engine. To be able to duplicate the work for a tuned engine is more or less impossible for anyone else than GM people who did the original work.

Thanks for the explanation. Yes I was referring to the corrected gain.... Your explanation is way, way over my lil pee brain though.... I was hoping for a simple way to use the GM modifiers to apply to re calibrated MAF / VE tables.

I flashed your transient settings and went for a couple drives. Not knowing exactly what to look at in the logs to see if it improved or went the other direction, I just observed AFR tip in changes under acceleration. I do notice less of a delay for AFR to catch up to commanded when PE is triggered. So I believe your settings may have helped. Nothing funky noticed, so that's always a good sign...

[JBZ]

so just to make sure im understanding this correctly. the way the correct gains are weighted is higher number more geared towards VE and lower number weighted more towards MAF...is that right or am I backwards? so in Barums screenshot in Cell 8 you are relying more on VE and 3,9,10,11,23,28,29 more weighted on MAF and all the rest of the cells are completely middle of the road not giving MAF or VE any weighting advantage. is that correct?

[JMsquared]

I have another noob question?
How do the Zone boundaries affect the tune, and what process is needed to change them?
Also, I'm seeing Engine load never exceeds 100% I,m forced induction, so Im thinking I should see more than 100% . If that's true where do I change this?

Thanks

[Giambals]

Bump from the dead.

I've been trying to understand the dynamic airflow coefficients (on a dual vvt engine) and stumbled across this thread. My coefficients are a little different, but the same idea applies.


Capture.PNG

I dug up the GM patent 7,010,413 B2, which describes this airflow model https://www.google.com/patents/US701...p4AZcQ6AEIJzAA and I have some thoughts and questions about this model

The dynamic airflow prediciton is calculated as:

According to the present invention, the predicted CAFP is calculated as follows: CAF P ⁡ ( k + 1 ) = a 1 ⁢ CAF E ⁡ ( k ) + a 2 ⁢ MAF ⁡ ( k ) + a 3 ⁢ MAF ⁡ ( k - 1 ) + b 1 ⁢ MAP ⁡ ( k ) +   2 ⁢ MAP ⁡ ( k - 1 ) + b 3 ⁢ MAP ⁡ ( k - 2 ) + c 1 ⁢ TPS ⁡ ( k ) + c 2 ⁢ TPS ⁡ ( k - 1 ) + c 3 ⁢ TPS ⁡ ( k - 2 ) + d 1 ⁢ UMAF ⁡ ( k ) + d 2 ⁢ UMAP ⁡ ( k )
where k is the current time event.

First off, I think a major point that is being missed is that in order for the model to remain stable the following must be true:

To ensure steady-state accuracy, the predictor coefficients are constrained according to the following equations:
a 1 +a 2 +a 3=1
b 1 +b 2 +b 3=0
c 1 +c 2 +c 3=0

This means that the sum of the coefficients for CAF_e(k) (Current time step cylinder air flow estimate), MAF(k)(MAF current time step) , and MAF(k-1) (MAF last time step) must = 1.

The sum of coefficients for MAP(k) (MAP current time step), MAP(k-1) (MAP last time step), and MAP(k-2) (MAP two time steps ago) must = 0

The sum of coefficients for TPS(k) (TPS current time step), TPS(k-1) (TPS last time step), and TPS(k-2) (TPS two time steps ago) must = 0

I believe a lot of the limited success of increasing/decreasing these coefficients is due to adhering to this rule.

The next piece of the puzzle is the differentiation of small and large MAF/MAP transients, which are handled by the UMAF and UMAP variables.

The component UMAF is governed by the following equations:
UMAF(k)=MAF(k)−MAF(k−1)−MAFDEL
if MAF(k)>MAF(k−1)+MAFDEL, otherwise
UMAF(k)=0
where MAFDEL is a predetermined constant (gain limit) that differentiates between small and large transient behavior in MAF. If there is small transient behavior in MAF, then UMAF is set to zero. The component UMAP is governed by the following equations:
UMAP(k)=MAP(k)−MAP(k−1)−MAPDEL
if MAP(k)>MAP(k−1)+MAPDEL, otherwise
UMAP(k)=0
where MAPDEL is a predetermined constant (gain limit) that differentiates between small and large transient behavior in MAP. If there is small transient behavior in MAP, then UMAP is set to zero. Thus, the components UMAF and UMAP enable accurate calibration of the predictor coefficients during small or large transient behavior.

The ideal behind this looks like there is a threshold on the rate of change of MAP and MAF, which when exceeded, are captured by additional airflow estimate from UMAF and UMAP and their coefficient. The patent says that these coefficients are not constrained by their sum, and the model will remain stable. Hopefully HPTuners can add MAFDEL and MAPDEL variables so this piece of the model can be completed.

The last piece of the algorithm that I am not totally understanding is how to handle the 'Corrected' and 'Cam Transient' coefficients, since they're not described in any patents I could find.

I came up with a spreadsheet to look up the dynamic prediction coefficients and speed-density coefficients based on the PR and RPM boundaries, so I have a decent way to check the model. I have had good sucess with the speed density calculation, but am hoping for some input on understanding the last piece of the dynamic air model!

[GHuggins]

What do the prediction coefficients look like in 3d graph format - if that's what you meant by using your spreadsheet to look at it?

[rao]

Very interesting. Hopefully someone who is much better at math than I am will sort all of this out. Until them everything stays zeroed out

[Giambals]

My spreadsheet is similar to Barum's above. The coefficients do not map out well, since the coefficients aren't related in that way.

What I have been doing is taking vehicle logs and looking up coefficients for each datapoint to calculate predicted airflow and comparing the ECM's prediction. Without the Cam Transient, MAPDEL, MAFDEL, and CAFERR components, I havn't been able to match the estimate well.

Hopefully this sparks some more discussion on the topic, because I think there is a lot of work to be done in this area!

[Giambals]

Very interesting. Hopefully someone who is much better at math than I am will sort all of this out. Until them everything stays zeroed out

To disable the dynamic air model, I'd be curious to check if zeroing out MAP and TPS coefficients, while making all the MAF coefficients sum to 1 would work better. Just a thought.

[barum]

To disable the dynamic air model, I'd be curious to check if zeroing out MAP and TPS coefficients, while making all the MAF coefficients sum to 1 would work better. Just a thought.

This is already in use in some of the OEM SW versions. I've tested that as well, but it does not disable the whole dynamic air model, only the parts related to MAP and TPS. I did try several different versions, even the "average of the three most resent ones", but did not get the response I was looking for.

[Giambals]

This is already in use in some of the OEM SW versions. I've tested that as well, but it does not disable the whole dynamic air model, only the parts related to MAP and TPS. I did try several different versions, even the "average of the three most resent ones", but did not get the response I was looking for.

I wonder if that behavior is because we can't see/edit the MAPDEL and MAFDEL gain limit variables, so UMAP and UMAF are still coming into play?

Once the transient air model is eliminated (not that it is necessarily a good thing to do...), air calculation should just be some blend of steady state VE and MAF airflow, I believe.

Any idea how "Cam Transient" is used? Im assuming it is just something like:
(current intake cam position - last timestep cam position) * cam transient coefficient?

[sssnake]

To disable the dynamic air model, I'd be curious to check if zeroing out MAP and TPS coefficients, while making all the MAF coefficients sum to 1 would work better. Just a thought

I tried this for months (and my current tune still includes this methodology). However, I still end up with the maximum MAP value feeding the equation as 105KPa even though my "MAP Estimated Max" value is set to 235KPa. As a result I have to set my "Dynamic Airflow High RPM Disable" to 1750 RPMs or I get a flat spot in the dynamic air output. As there is lag in the MAF sensor during airflow changes (when compared to the MAP data), I am seeing issue in the transient fueling. Enough issue that unless the car is fully warmed up I can kill the engine with throttle transients (peg the throttle let off the gas and it will die before settling in to idle - no it is not a throttle follower issue). Does anyone know if the HPT staff is looking into ways to alter the maximum value for the MAP sensor that this algorithm uses?

[pannetron]

Another bump to this old but fascinating thread. It boggles me how different the OEM 2012 ZL-1 settings for predictive coefficients are compared to same time frame CTS-V. It appears that the ZL-1 calibration folks spent a LOT more time tuning predictive airflow.