Control of a Cyclical System

[a_wright]

Hey all,

I want to design a control system for a batch manufacturing process. I do not have a good model for the system. I was going to try to use machine learning to create a difference equation. You can then convert a difference equation into a state space model. One of my textbooks describes a method to use using machine learning to model a system. However, I want the machine learning algorithm to be able to learn from previous iterations of this batch process, and I am not sure how to do that.

Does anyone have any ideas? Or any resources I could look at? The biggest problem right now, is that I don't even know where to start.

Thanks,
Abe

 

[Viktory]

There are risks to having a machine learning algorithm modifying the setpoints of a live control system.

Have a look into advanced process control. These systems are often used to modify setpoints in process facilities for optimization purposes. Many of them build simple models for process response based on past performance. It would be possible to do the same thing using machine learning (that is essentially what they are already doing). I suspect the difficult part would be figuring out how to teach your model properly so that a couple of strange results do not completely skew the model.

There are many resources available on machine learning. Simple google search brings up the below link, which has many resources at the bottom. I would recommend you start with a good understanding of control theory as almost all problems can be solved without complicating with machine learning. Once you are confident then bring in the machine learning, but be very clear what problem you are hoping to solve.

https://en.wikipedia.org/wiki/Machine_learning_control

 

[PNachtwey]

I want to design a control system for a batch manufacturing process.

It would really help to know what it is.

I do not have a good model for the system.

You need to know how many parameters are significant to your model. Then be able to write the differential equation.

I was going to try to use machine learning to create a difference equation.

First I would use system identification to find the values for the parameters of your model. If the system is SISO then a file of times, control outputs and feedback is required. A good initial guess is made for the values of your model parameters. The output of the model executed using the recorded control output to generate an estimated response. The estimated response use a sum of squared errors for each time period between the actual recored data and estimated data.
Usually the sum of squared errors will be high. There are algorithms that are very good at modifying the parameters in you model so the sum of squared errors is minimized. Then you have a valid model

You can then convert a difference equation into a state space model.

WHY?! Differential equations rule! The can be linear or non-linear. They are much more flexible than state space. In python one can use odeint() to integrate the non-linear equations.

One of my textbooks describes a method to use using machine learning to model a system.

Does your text book provide an example?

However, I want the machine learning algorithm to be able to learn from previous iterations of this batch process, and I am not sure how to do that.

Why!? Does each batch differ that much from the last?
It is possible use for 3 batch let minimization program minimize the values of the parameters for all three. Then one can use exponential averaging to update the valves of the parameters. This might be handy if conditions change a lot but slowly.

Does anyone have any ideas? Or any resources I could look at? The biggest problem right now, is that I don't even know where to start.

I have a YouTube channel called "Peter Ponders PID". It is much more in depth than most YT channels. Thee are 4 video on system identification. There are a couple where I use non-linear differential equations. State space is OK for learning but nothing in the real world is that perfect.

Peter Nachtwey
Delta Computer Systems

http://www.deltamotion.com

http://forum.deltamotion.com/

IFPS Hall of Fame Member

 

[IRstuff]

Machine learning works, until it doesn't; when it doesn't work, you'll have zero clues as to why. Unless you can train the algorithm with lots of failure examples, it will likely choke the first time something breaks.

TTFN (ta ta for now)
I can do absolutely anything. I'm an expert!

https://www.youtube.com/watch?v=BKorP55Aqvg

faq731-376 forum1529 Entire Forum list

http://www.eng-tips.com/forumlist.cfm