Page 1 of 2 12 LastLast
Results 1 to 20 of 29

Thread: Neural Network Trainer - Sanity Check...

  1. #1

    Neural Network Trainer - Sanity Check...

    Someone please double-check my strategy here:

    The Neural Network Trainer accesses VE tables for several camshaft positions. My trainer file shows a table for five exhaust cam angles, five intake cam angles, and every combination of the two sets for a total of 25 tables. Since the Hemi only has one fixed ("exhaust") cam, there are only five tables used, which are the exhaust cam angles and only the corresponding intake angles.

    Here are the five tables I'm planning to use for training (PCM is an NGC4, 5.7 OS):

    Exhaust Intake
    89 139.5
    98.5 130
    108 120.5
    117.5 111
    127 101.5

    I believe I should ignore the other table combinations i.e., uncheck the box that says "use table for training"

    Am I doing it right??
    Last edited by rockystock; 03-08-2021 at 11:29 PM.

  2. #2
    Yes, those 5 tables look right for a FCA vehicle where the intake and exhaust cam angles are linked together. Make sure to uncheck "use table for training" for the tables you're not using before sending the file for training. This needs to be done every time before sending the file in for training.

  3. #3
    Quote Originally Posted by Victor@HPTuners View Post
    Yes, those 5 tables look right for a FCA vehicle where the intake and exhaust cam angles are linked together. Make sure to uncheck "use table for training" for the tables you're not using before sending the file for training. This needs to be done every time before sending the file in for training.
    Good to know. Thank you, sir.

    Just one more question at the moment: I'm starting to tune my network by addressing one table at a time. For example, I set the cam at full retard at all points (89? in this case), collect data, and then adjust only the table for 89? exhaust / 152? intake cam angle. Send back to the trainer, update the network, lather-rinse-repeat until the fuel trims look good at 89?. Then move on to the next table, and repeat the process for all five tables. Is this a good strategy, or is there a better approach??

    Thanks again!

  4. #4
    For best results, all tables that are marked for training should be filled in with accurate data. You can do a regular log using VCM Scanner and then filter out fuel trims by intake and exhaust cam angle. The last part of the NN FAQ explains how to set such filters:
    https://www.hptuners.com/products/ne...trainer/nnfaq/

  5. #5
    Potential Tuner
    Join Date
    Mar 2019
    Posts
    6
    How did you figure out which tables to use?

  6. #6
    Quote Originally Posted by hussainsrt View Post
    How did you figure out which tables to use?
    For the hemi camshaft, the most retarded exhaust angle (low number) corresponds to the highest intake angle (high number). I double-checked the numbers that appear in my trainer file by taking the total lobe separation (for the 392 cam that number is 241) and subtract the exhaust angle to find the correct intake angle.

    So my fully advanced angles are: 127 (exhaust) and [241-127=] 114 (intake)

    FYI - I'm running a stock 392 with an old NGC4, so the angles I'm using look more like the 5.7 cam. A fully advanced cam on a 392-based PCM is around 134 (exhaust).

    You could also log intake and exhaust cam positions to see their relationship.

    OK so I just figured out that my intake cam max and min angle is important to get right in the neural network training values. I made a math for the sum of the intake and exhaust center positions, and then threw it up on a log graph. My 5.7 PCM is showing 228.5 degrees for the total, so that is the number (not 241) that I need to use to set the correct training values. So the intake min angle is 228.5-127=101.5. Intake max angle is 228.5-89=139.5. I believe that if those numbers are not correct, the neural network will end up interpolating VE between the tables incorrectly...

    Hope that helps.
    Last edited by rockystock; 02-25-2021 at 10:45 PM.

  7. #7
    Quote Originally Posted by hussainsrt View Post
    How did you figure out which tables to use?
    The Neural Network FAQ explains this in the "Do I really need to modify all 25 VE tables for FCA vehicles?" section.

    https://www.hptuners.com/products/ne...trainer/nnfaq/

  8. #8
    Quote Originally Posted by rockystock View Post
    OK so I just figured out that my intake cam max and min angle is important to get right in the neural network training values. I made a math for the sum of the intake and exhaust center positions, and then threw it up on a log graph. My 5.7 PCM is showing 228.5 degrees for the total, so that is the number (not 241) that I need to use to set the correct training values. So the intake min angle is 228.5-127=101.5. Intake max angle is 228.5-89=139.5. I believe that if those numbers are not correct, the neural network will end up interpolating VE between the tables incorrectly...
    The neural network uses the camshaft angles defined in “Engine > Airflow > Variable Camshaft”. If those values are modified, the values for min/max intake and exhaust camshaft should also be changed in "Engine > Airflow > Neural Network".

  9. #9
    Potential Tuner
    Join Date
    Mar 2019
    Posts
    6
    Quote Originally Posted by rockystock View Post
    For the hemi camshaft, the most retarded exhaust angle (low number) corresponds to the highest intake angle (high number). I double-checked the numbers that appear in my trainer file by taking the total lobe separation (for the 392 cam that number is 241) and subtract the exhaust angle to find the correct intake angle.

    So my fully advanced angles are: 127 (exhaust) and [241-127=] 114 (intake)

    FYI - I'm running a stock 392 with an old NGC4, so the angles I'm using look more like the 5.7 cam. A fully advanced cam on a 392-based PCM is around 134 (exhaust).

    You could also log intake and exhaust cam positions to see their relationship.

    OK so I just figured out that my intake cam max and min angle is important to get right in the neural network training values. I made a math for the sum of the intake and exhaust center positions, and then threw it up on a log graph. My 5.7 PCM is showing 228.5 degrees for the total, so that is the number (not 241) that I need to use to set the correct training values. So the intake min angle is 228.5-127=101.5. Intake max angle is 228.5-89=139.5. I believe that if those numbers are not correct, the neural network will end up interpolating VE between the tables incorrectly...

    Hope that helps.




    Quote Originally Posted by Victor@HPTuners View Post
    The Neural Network FAQ explains this in the "Do I really need to modify all 25 VE tables for FCA vehicles?" section.

    https://www.hptuners.com/products/ne...trainer/nnfaq/
    thanks!

  10. #10
    Quote Originally Posted by Victor@HPTuners View Post
    The neural network uses the camshaft angles defined in ?Engine > Airflow > Variable Camshaft?. If those values are modified, the values for min/max intake and exhaust camshaft should also be changed in "Engine > Airflow > Neural Network".
    I've definitely learned that the cam angles defined in the Training Values are extremely important. My intake angles were a few degrees off from what they should have been. I think the ANN was then interpolating between the tables, but basically venturing into the other calculated tables that should remain unused. So the incorrect values were not so bad with stock ANN parameters because those intermediate tables were not so far off; but I've noticed the Neural Network Trainer generates some pretty wild "unused" tables to get close to the tables I'm intending to use. Hope that made sense

  11. #11
    OK so I think I've worked through the earlier issues. Thanks, Victor, for your help. I've now imported the stock 392 VE tables into my 5.7-derived system, and with the correct cam training values the entire VE operating range is running much closer to where it should.

    I'm now trying to get a set of trained tables that more closely match what I request. I've tried using the log data and doing a "Paste Special -> Multiply by %" and then manually smoothing out the tables (somewhat) before exporting. So far, the resulting tables change by the requested amount in some places, but often only change by a percent or two. For example, my idle region needs to drop by around 10% in the fully advanced table (using LTFT+STFT, the fuel trim shows -10% in the idle region). The Trainer result only gives me around 2%. I'd guess that if I try to force a larger change, it may skew the table in other areas, or it may skew the other tables...

    So the question is - are there any tips or tricks to get a more accurate set of results? Here are some options I may try but hoping a tip might save some time...

    1. Start with a "Paste Special -> Add" OR
    2. Double the fuel trim error and do a "Paste Special -> Add" or "Multiply by %"
    3. Make sure the tables are reallllly smoothed out before exporting


    Thanks again for your help!
    Last edited by rockystock; 02-27-2021 at 06:46 PM.

  12. #12
    Quote Originally Posted by rockystock View Post
    OK so I think I've worked through the earlier issues. Thanks, Victor, for your help. I've now imported the stock 392 VE tables into my 5.7-derived system, and with the correct cam training values the entire VE operating range is running much closer to where it should.

    I'm now trying to get a set of trained tables that more closely match what I request. I've tried using the log data and doing a "Paste Special -> Multiply by %" and then manually smoothing out the tables (somewhat) before exporting. So far, the resulting tables change by the requested amount in some places, but often only change by a percent or two. For example, my idle region needs to drop by around 10% in the fully advanced table (using LTFT+STFT, the fuel trim shows -10% in the idle region). The Trainer result only gives me around 2%. I'd guess that if I try to force a larger change, it may skew the table in other areas, or it may skew the other tables...

    So the question is - are there any tips or tricks to get a more accurate set of results? Here are some options I may try but hoping a tip might save some time...

    1. Start with a "Paste Special -> Add" OR
    2. Double the fuel trim error and do a "Paste Special -> Add" or "Multiply by %"
    3. Make sure the tables are reallllly smoothed out before exporting


    Thanks again for your help!
    I'm glad my answers helped you out!

    As for tips, I think the main one is that the neural network trainer always tries to minimize average error across all tables. That means that some areas might get results closer to what you want, while others will be farther. If there is a particular area where the neural network trainer is struggling (like your idle region), you can try to set the VE lower/higher than what you want in order to "force" the neural network to train that area better. For example let's say you want the VE in a particular cell to be 70%, but the trained neural network outputs 75%. Try setting the VE of that cell to 65% and do another training. In most cases that should fix the issue.

  13. #13
    Quote Originally Posted by Victor@HPTuners View Post
    The neural network uses the camshaft angles defined in “Engine > Airflow > Variable Camshaft”. If those values are modified, the values for min/max intake and exhaust camshaft should also be changed in "Engine > Airflow > Neural Network".
    Means if intake and exhaust tables are changed, I have to change Min and Max also under “Engine > Airflow > Variable Camshaft” and then match it in "Engine > Airflow > Neural Network" ??

  14. #14
    Quote Originally Posted by RedShark_JK View Post
    Means if intake and exhaust tables are changed, I have to change Min and Max also under “Engine > Airflow > Variable Camshaft” and then match it in "Engine > Airflow > Neural Network" ??
    My point with that post was that if you make changes to the camshaft angles in "Engine > Airflow > Variable Camshaft", you should also write the new min-max ranges in "Engine > Airflow > Neural Network". It's up to you to decide if you want to modify the tables in "Engine > Airflow > Variable Camshaft".

  15. #15
    Quote Originally Posted by Victor@HPTuners View Post
    My point with that post was that if you make changes to the camshaft angles in "Engine > Airflow > Variable Camshaft", you should also write the new min-max ranges in "Engine > Airflow > Neural Network". It's up to you to decide if you want to modify the tables in "Engine > Airflow > Variable Camshaft".
    What I mean is, there is also Min and Max parameters same page with variable camshafts other than Min and Max in Neural Network

    Both Min & Max should be changed or only in NN as it's one used for training ?

  16. #16
    Quote Originally Posted by RedShark_JK View Post
    What I mean is, there is also Min and Max parameters same page with variable camshafts other than Min and Max in Neural Network

    Both Min & Max should be changed or only in NN as it's one used for training ?
    I think you might be referring to the parameters in the "Variable Camshaft" page. Those parameters are used by the PCM to determine which spark tables to use for the final spark calculation.

    I believe the most important parameters to set for the Neural Network are the "Training Values" under the "Neural Network" page/tab. I've found it is very important to set the minimum and maximum cam angles to the most extreme angles your setup can attain, especially if you let your cam use all of its range in the "variable camshaft" tables. I'd recommend using a set of stock tables under "Variable Camshaft" so the cam will definitely hit max advance when cold and at idle, and will probably be at max phase at part-throttle cruising. Next, log your intake and exhaust camshaft center position channels. Make a note of what your max advance and max phase angles are (e.g., for a ~2010 5.7, your max advance will be approx. 127 exhaust / 101.5 intake; max phase will be 89 exhaust / 139 intake). Cam lobe separation for 5.7 is 228.5, so make sure the angles you use in your "training values" always add up to equal 228.5, or whatever the correct angle is for your specific cam. For example, the Exhaust Cam Max + Intake Cam Min should add up to 228.5 for the ~2010 5.7.

    As I've been tuning my setup, I've found that collecting five good sets of data (one set for each valid cam angle combination in the trainer), and then modifying each of the 5 tables in one trainer file, and finally exporting the changes for training has yielded the best results so far. This approach requires setting the cam position in "variable camshaft" to one value in all of the Normal, Max TIP, WOT and Max Airflow exhaust tables (and the corresponding angle in the intake tables) so that the cam stays fixed in that one position over the whole VE table as you drive around. Then repeat for all five positions. I think that once you have the neural network trained well enough, you should be able to set your "Variable Camshaft" tables to whatever you want, and the neural network should be able to properly estimate airflow everywhere.

    One word of caution if you try tuning this way - watch your spark timing especially when you tune at the more advanced angles. I had to drop my part throttle base table by about 4 degrees above 300 mg aircharge.

  17. #17
    Tuner in Training 14HemiRam's Avatar
    Join Date
    May 2020
    Location
    Oklahoma City, OK
    Posts
    18
    Curious Ive been trying out the Neural Network Trainer and I am having a hard time with it. Is there somewhere I can get some good info on how to use the trainer? I haven't found much on this subject. Thanks

  18. #18
    If you upload your tune file, I might be able to help you.

  19. #19
    Tuner in Training 14HemiRam's Avatar
    Join Date
    May 2020
    Location
    Oklahoma City, OK
    Posts
    18
    That would be great I'll send it over. Appreciate it.

  20. #20
    Tuner in Training 14HemiRam's Avatar
    Join Date
    May 2020
    Location
    Oklahoma City, OK
    Posts
    18
    BTW which tune would you like the one before the Neural Network tune, basically the one I used to upload, or the tune after the neural network touched it?