Forum Replies Created
-
Posts
-
A very good question!
Sometimes cutting occurs with immeasurably small wear. In that case, you would use the FIRST cut that the wear level was achieved on as your estimate.
For example, in case one, wear of 87.4 is achieved at cut 27; this value is maintained until cut 101. But if 87.4 were your target wear value, you would want to report 27.
Each training record contains one “wear” file that lists wear after each cut in 10^-3 mm, and a folder with approximately 300 individual data acquisition files (one for each cut). The data acquisition files are in .csv format, with seven columns, corresponding to:
Column 1: Force (N) in X dimension
Column 2: Force (N) in Y dimension
Column 3: Force (N) in Z dimension
Column 4: Vibration (g) in X dimension
Column 5: Vibration (g) in Y dimension
Column 6: Vibration (g) in Z dimension
Column 7: AE-RMS (V)The spindle speed of the cutter was 10400 RPM; feed rate was 1555 mm/min; Y depth of cut (radial) was 0.125 mm; Z depth of cut (axial) was 0.2 mm. Data was acquired at 50 KHz/channel.
Attribute (feature) extraction per se is up to the competitors. If you download the data on the main competition page:
https://www.phmsociety.org/competition/phm/10You will find that there are actually 1890 cut files available to work with!
What you will want to provide is the Nth percentile of a model of the distribution of maximum wear of all three flutes. N is up to you.
You will be provided (next week) the scoring function, so you can use that to select N. There will be a small penalty for over-prediction (i.e., being too conservative), and a larger penalty for under-prediction (i.e., being too liberal), so you will have to make that tradeoff.
The RUL estimate is implicit. You generate an estimated maximum wear for any flute, and when it crosses a fixed (and unknown to you) threshold, the tool is finished.
Leave one out (LOO) is a method of cross validation:
http://en.wikipedia.org/wiki/Cross-validation_%28statistics%29There are at least two ways to partition the data. You might train on two runs and evaluate on the third. Alternatively, you might train on two flutes within a run and evaluate on the third flute within a run.
But you are not obliged to use cross validation! If you have an approach you think will work better, use that…
You will want to predict the maximum expected wear after each cut for ANY of the three flutes in units of 10^-3 mm. There will be a small penalty for over-prediction (i.e., being too conservative), and a larger penalty for under-prediction (i.e., being too liberal).
No, 315 cuts is not failure. Each cut causes a particular amount of wear. When wear reaches a certain level, the cutter is worn out. That may occur in 200 cuts, or 600 cuts.
posted here:
http://www.phmsociety.org/node/442
( 50 KHz / channel )The files were compressed using the bZip2 algorithm. If your un-zipping software complains, make sure it is bZip2-compatible. 7-Zip is Windows open-source software that works well; Linux users can use the bunzip2 command; MAc users can use Stuffit.
