Options
Datamining for continuous process in manufacturing
Contributor II
Hello,
I'm a very beginner in Rapidminer and machine learning tools. I was wondering what is the best tools to apply for chemical industry continuous process. I mean we run big equiment on a 24hours basis to make different kind of products. We would like to improve the yield of one type of product that we produce may be 5 times a year. A campaign is usually 2 to 3 weeks depending on how difficult we achieve the right specs.
Some of process data are on a 1 hours basis, some are twice a day, some are every 2 hours. I've got may be 10-15 Xi and the quality of the product can be discibed by 4-5 Yi. some process data are numerical and some are attributes.
I'd like to see if there are some relationship between campaign when we get good/bad products. we've got a very low yield and from campaign to another, parameters are not the same.
really wondering where to start!? on this kind of process datas.
thanks
I'm a very beginner in Rapidminer and machine learning tools. I was wondering what is the best tools to apply for chemical industry continuous process. I mean we run big equiment on a 24hours basis to make different kind of products. We would like to improve the yield of one type of product that we produce may be 5 times a year. A campaign is usually 2 to 3 weeks depending on how difficult we achieve the right specs.
Some of process data are on a 1 hours basis, some are twice a day, some are every 2 hours. I've got may be 10-15 Xi and the quality of the product can be discibed by 4-5 Yi. some process data are numerical and some are attributes.
I'd like to see if there are some relationship between campaign when we get good/bad products. we've got a very low yield and from campaign to another, parameters are not the same.
really wondering where to start!? on this kind of process datas.
thanks
0
Answers
this is indeed a very intersting project. There are several things you can do. Funnily RapidMiner Germany has a similar research project in germany: http://fee-projekt.de/Publikationen.html (site is sadly in german).
There is a general idea of having a set of settings and a set of materials and a desired quality. You can learn a model which predicts the quality for a given parameter and material set. If you have this model, you can easily simulate what would happen if a change a parameter etc.
what you would need to built is a data set like this:
parameter1 parameter2 parameter3 material1 material2 QualityIn5Mins
The quality can be in 5 mins or per campaign. Afterwards this turn out to be a supervised regression task.
Best,
Martin
Dortmund, Germany
How do you treat Time serie datas? Is it ID for the date? And campaign number? Batch? Cluster? Or is it better to treat all datas as just one.
Difficult to explain
about time series: In the end you convert this again to one row (example) per time frame.
Lets assume you do it campaign wise (can also be in 5minutes steps or so). The you calculate features for the indivudual series. Things like:
Average
Maximum
Minimum
Slope
etc. But also stuff you know because you have domain knowledge e.g. Number of values above a threshold. What you then get is a table like you had before
campaignid avg_series1, maximum_series1.... avg_series5, maximum_series5 etc.
and you can do the same things.I am not sure about literature though. Possibly you can find things in the proceedings of our users conference.
~Martin
Dortmund, Germany
There's several posts on that site on RapidMiner for manufacturing processes so have a dig around. They also are the authors of one of the RM books.
Dortmund, Germany