Non-Parametric Tire Data Fitting, Presentation in NYC

[ColumbiaMatt]

Hi FSAE Forums,

I did a final project on fitting semi-parametric and non-parametric models to TTC data.

I'll be giving a presentation about it at the New York Open Statistical programming meet up on August 17th at ebay nyc.
Registration Link: http://www.meetup.com/nyhackr/events/232858104/

I wanted to share my paper on the forums because I don't think anyone has done this before, and I think its an interesting way to tackle the problem from looking at the raw data.

https://www.dropbox.com/s/0hr8ru9bgsduf07/Sheridan_Matthew_Tire_Data_Report.pdf?dl=0

I was hoping to hear people's thoughts on it. A few things

1. I removed information identifying the tires analyzed and the conditions at which they were analyzed. I don't think I include anything in the paper that breaks any TTC rules for sharing among the public.

2. I didn't have time to go in depth with selecting the best model and evaluating their effectiveness. This was more of a "look, this can be done!".
If anyone wants to go farther with it, I would love to help.

3. I have some background on vehicle dynamics in the beginning of paper. If anyone has any nitpicky comments about how I explain certain concepts, please make them!
I need to make sure I am presenting all the nuances as accurately as possible at my presentation!

Best,

Matthew Sheridan

[BillCobb]

Clearly you've chosen to include data for your fitting dataset from both + and - slip input velocity. This manifests itself as an apparent hysteresis in the raw curves which can confound almost all function fitting functions. Several possibilities as a way around this quandry:

1) Select positive slip velocity only OR else negative slip velocity subsets of your data. Increasing slip (magnitude) is generally more appealing to simulationists because that's when your car is heading to the wall or a turn apex instead of running away from it. Doing both and weighting the resulting fit coefficients into a single group of blended values can be done, too.

2) OR, remove the hysteresis using cross correlation techniques (candling) which will also give you a rough estimate of the tire's relaxation distance. The RXY function in Matlab will do this nicely for you in one swell foop. The resulting peak cross-correlation occurs at some index of the sample (time) rate. Time, speed, reciprocal, you ought to know the rest of the story. With the hysteresis removed you now have a much cleaner multivariable cloud to investigate. Note: I've posted these methods on the TTC Forum a while back, including fitting an Fy(alpha,Fz,gamma) function using the Matlab Spline toolbox to produce a Simulink callable function that mimics a tire test machine. With this, you can reinstall the first order relaxation effect and see how it effects the look of your output data. This works for Mz and Fx processes, too. In reality, though, relaxation is higher order that just plain 1 (espcially Mz which I also showed on the TTC Forum.

I'd much rather see slip angle used with regard to the tire lateral and longitudinal VELOCITY vectors instead of steer angle vs travel angle because of more advanced analysis that would reference the axle sideslip angles (for which the tire is a good part of). If the tire is stationary, is there a slip angle ???

Your own choice of raw data shows the typical flaw in Pacejka's 'magic' form: There are often linear regions in the Fy-slip data produced by parts of the tire that engage the roadway in a designated loading order (breakers, sidewall, belt edges). Tires with a LOT of load reserve (like race tires or special construction recipes can reveal this). Your fitting function's weighting vector needs to be told which region it's to value and what's to be sidelined. Passenger car tires (round and black ones), usually don't have these special regions because they are chosen by size and load capacity to have less load reserve (Meaning they might run at 80% to 95% of their T&RA load just sitting there in a parking lot. Not so for race tires.