What's new
What's new

Machine Learning for Wire EDM

Motorsports-X

Hot Rolled
Joined
Nov 22, 2014
Location
Texas
As most of us in here know, the wire EDM process is insanely variable and complex. When you start to think about all the processes and parameters that are involved -wire type, speed, tension, spark gap, voltage/amperage, flushing, offsets, sensitivities, etc) - its almost amazing that it works at all. As a machinist my first job is to cut steel; but as an engineer and naturally curious person I like to understand the underlying physics of how things work the way they do (such as why HSM allows you to run such hi feed rates)

I thought a few weeks ago that it would be nice for "us" to have a calculator/database that allows us to input cut parameters, machine, and material specs etc. Over the course of years and thousands of inputs, you would be able to query your particular process and get a good statistically sound set of parameters that give you a fast, precise cut.. (with outliers ignored)

Then I got to thinking, that would be a lot to ask thousands of people to input all those parameters on a regular basis, so why not teach the machine to drive the same way tesla is teaching its cars to steer?

I found this paper that goes into depth on machine learning and how its applicable to wire edm.

http://www.iaeme.com/MasterAdmin/UploadFolder/20320130406019\20320130406019.pdf



WHat would your input be on this? would you have a problem networking and sending your cut data back to a grid to be processed?
 
Hi Motorsports-X:
I believe the paper to be rather naive with regard to the real-world state of the art and the power of data mining for this sort of application.
The proposal appears to be to try to refine a set of inputs everyone can use by taking the collective experience of many machines taking many cuts and attempting to apply data mining techniques to them in order to be able to predict what your machine must do to make your best possible cut when you want it to.
I'm unconvinced that this sort of data mining is the most appropriate way to approach the problem which, after all is to optimize the cut I want to make today, on my machine in its current state of repair on my job.

My understanding of data mining is that it seeks to predict that a Starbucks customer will be more likely to order a donut with his coffee if Starbucks presents the buying opportunity a certain way based on an analysis of coffee and donut buying behaviour of gazillions of customers and finding patterns in that aggregate that can be exploited.
What's crucial is that no individual customer's behaviour matters; it's the aggregate that informs how Starbucks presents donuts and coffee.
If a particular customer is resistant to the allure of a Big Data driven coffee and donuts campaign, nobody dies.
It's power is that it can point out unexpected and counterintuitive correlations to try, based on sifting a big mass of even the most trivial details of real world experience for those new correlations.

The EDM guy's situation is totally different in that he MUST be able to cut his job every time, and the wrong settings kill it dead.
Taking an oversized glorified average is not helpful because it cannot anticipate all of the variables and they matter for every individual who attempts to do the job.

A simple example:
Joe maintains his machine meticulously.
Fred is a slob.
Joe's data will not work with Fred's working style and vice versa, simply because of the important consequences of the different maintenance standards.
So in order for the group data to be effectively pooled and used, Fred's machine has to let the world know that Fred is an asshole who doesn't look after his toy, and it has to complain to the database in detail, not just in generalities, otherwise Fred's data pollutes the pool.

Contrast that with machine learning of the kind used by some modern sinkers.
They monitor the cut continuously and are constantly tweaking different combinations of the variables they can control; seeing which combinations work better and which do not.
As they gain experience, they consolidate those successful patterns into the initial settings for the next job and continue to try to improve the performance from job to job and also from beginning of job to end of job.

But it's all done with the data from ONE machine and it does not say to itself..."Joe and Fred and Charlie did it this way and it worked out good, so I'm gonna do it this way too".

This kind of "Neural network" "Fuzzy logic" thinking appears to have been all the rage in the late 1990's but I haven't heard anything recently; so either everybody's doing it now, making it old news, or it didn't work and everybody's keeping quiet.

I could certainly be wrong, but I can't see this new "data mining everything" fad really providing much useful new insight into how to set up a job.

Cheers

Marcus
Implant Mechanix • Design & Innovation > HOME
www.vancouverwireedm.com
 
Group/Machine Learning, As used in the tesla Automobile, does happen to use the "joe, fred, and charlie" data.. but the data is still collected by an intelligent and accurate platform. It Then combines all the vehicles senors to form a conclusive opinion, and then applies that opinion across multiple units to check the responses. If 40 cars go around a corner at 35, with no issue, and then one does it at 36, it will try that corner at 36 on more vehicles to get an accurate measure of response. if it appears to be working, then the new value is 36. The vehicle will not let the car exceed certain parameters like g load, acceleration, distance from centerline.. etc.

sure you get some dirty data and outliers once in a while...BUT.. if that outlier is on the "better side" those are actually what helps push the boundaries. When you collect from 20,000 points you end up with a very good baseline of what will work. The "learning" portion is pushing up from there. and 20,000 inputs push faster than your single machine can. Freds dirty machine data would never affect the pool because the cut is always slower.. and when fred sees that his machine never runs the industry average, he will get out the scrub pad.

By contrast, the single machine solution, as well as it does work.. will never learn as quickly. if you cut D2 everyday, and all of a sudden need to cut O1, then your machine is going to have to spend time learning that. Data mining will pull O1 settings from the other 5000 machines like yours that already cut O1.



In wire EDM we have many different machines, controls, etc.. so It would not be a single platform for a single machine. Unless somehow the companies agreed on standard variables to use.
 








 
Back
Top