Statistical Data Analysis - Vehicle Drive Cycles

[moesterl]

Hi Everyone,

I'm looking to get a little shove in the right direction. I have a massive amount of refuse truck data which includes such parameters as speed, gps location, fuel use, dump weight, idle time, etc.

What I would like to do is determine how similar two drive cycles are in order to see statistically significant fuel economy benefits between two different vehicles.

The data I have was collected while the trucks were in use for over a year of operation and was not a "controlled" experiment. As of right now I simply calculate a bunch of drive-cycle parameters for two different trucks and see how similar they are. However, I'd like to use actual statistics to prove that the datasets can in fact be fairly compared.

Please let me know if you have any ideas or statistical concepts I can look into.

Check out the attached excel table to see results.

Thanks,
Matt

 

[magoo2]

What you've presented is summary data. I'm guessing you have the actual data by day and this is what you ought to use.

Let's select one result - mpg. Run the descriptive stats on that for both trucks. You'd have 34 values for one; 53 for the other.

I would then be tempted to run a 2-sample t test. You're trying to find out is the mean value the same or not for the 2 trucks.

I would repeat this for tons per gallon, avg driving speed, etc.

I normally use MiniTab for my statistical analysis. Other pprograms are also available.

Hope this helps.

 

[IRstuff]

You make calculations on the "Driving Fuel Economy," but it's unclear whether that's simply a gross calculation, or something involved with more instantaneous fuel usage measurements.

I don't see how you could do that analysis unless you have the actual fuel pump rate data or equivalent. Just having got a Prius, it's interesting to see how much your instanteous mileage fluctuates as a function of acceleration, road conditions, etc.

Just from my short experience with the Prius, it's clear to me that unless you have a detailed and accurate model of the driving of the truck w.r.t. gas consumption, anything other than a gross comparison would be hard to validate.

TTFN

FAQ731-376
Chinese prisoner wins Nobel Peace Prize

 

[moesterl]

Magoo - thanks, I think the t-test for each drive-cycle parameter would give me a good idea if the datasets can be compared.

IRstuff - The raw data is recorded directly from the CAN bus at 1hz. I simply added up the fuel rate L/s at each second to come up with a total fuel usage. Agreed, it's not a controlled experiment which makes comparisons difficult.

Right now, all of my statistics are derived from summing each channel second by second, a gross calculation. Alternatively, I could perform statistics by day, using mpg, avg mph, drive%, etc for each day.... however, then each day get's equal weight ignoring the difference between a day the truck traveled 20 miles and a day it traveled 3 miles. Make sense?

 

[IRstuff]

I guess that I would at least try to separate the different components of the route and look at the consumption as a function of those components.

TTFN

FAQ731-376
Chinese prisoner wins Nobel Peace Prize

 

[GregLocock]

drivers are worth anything up to 10% variation in mpg in trucks.

I don't know whether you can accuont for that or if it is just a noise factor.



Cheers

Greg Locock


New here? Try reading these, they might help FAQ731-376

http://eng-tips.com/market.cfm?

 

[FeX32]

You can come up with your own index first. Maybe something that is not exactly frequency based but along the those lines. Could use for example 1/time where time is the idle time, or distance time ect. The other axes will be the fuel consumption or other relative measure you want to compare.
After deciding on an index for your data (or parts of the data) you can use a statistical correlation measure like

http://en.wikipedia.org/wiki/Cross-correlation

to measure similarities and differences in the data.
If the data is unique and your index you come up with is unique ( of course if you decide to do it like this) then be sure to publish the results (if your company is alright with this)




Fe

 

[GregLocock]

What you probably need to do is called PCA - principal components analysis, in which the relative contribtion of each factor to the result are established.

http://en.wikipedia.org/wiki/Principal_component_analysis

Cheers

Greg Locock


New here? Try reading these, they might help FAQ731-376

http://eng-tips.com/market.cfm?

 

[moesterl]

Guys - Thanks so much, I'm going to research your suggestions and give them a try with my data.

I'll let you know how it pans out.

-Matt