Neural network classification improvments

hi all,

I have trained a neural network using fitcnet, the data is a simple classification of hole numbering. the data is as per the simplified attached picture. but essentially the center of the hole xy point, the angle and distance relative to the center of the whole blade, the whole blade major axis and finally the hole number a value from 1:319.

the problem i have found is two-fold.

1. i think it may be 'over-fitted' it works well on the synthetic data but performs worse on real data
2. 2. it rountinly will identify multilple holes as the same number e.g there would be 4 hole number 3's now in the real data sometimes a single hole can split into a pair of much smaller holes they stay rougly within bounds of the original hole.

The data into the training was a single row of info for one hole listing the above info including its class effectivly treating each time i then ask it to predict the classification of a new hole being a 'clean-sheet' with no knowledge of already predicted holes.

Does anyone have any suggestions for possible improvments?

many thanks tim

[Rik]

Have a read here and here. It will greatly improve your chances of getting an answer.

It looks like your question is not actually about Matlab.

For example, it is unclear to me what you want with the numbers. Are you simply counting holes, or do the numbers have actual meaning. In the latter case duplicates are to be expected, in the former case you shouldn't be including the counts in the training, since they're an emerging property.

I suspect your question is much better suited to another forum, since it looks like you need to revise the design of you method. Once you have a method you wish to implement, you can ask a question to that effect here.

[Tim]

Hi thanks for coming back to me. On your 2 points 1. The cooling holes numbers are very specific. They allow identification of each in relation to the design intent. The project is looking at the way these cooling holes deform with time. The problem is I have tens of thousands. I’ve tried a few different methods (the Hungarian for 1) but each have failed to be robust enough for my use case. Hence why I’ve been playing around with a neural network and I guess the question is a little vague as I know the result I need and have yet to find a robust method

[Rik]

If these are holes you're tracking over time, would it be safe to assume they will not move much? That would mean you can crop the image to a small patch surrounding the original/previous location of the hole and then detect which hole is the closest to the center. That way you can keep the identification, but keep using the fact it is an emerging property.

But with the rest of your question it is still not clear to me how this is a question I can help with. I would strongly suggest playing around with a chatbot first and then post a more refined question on a more specialized forum.