2
$\begingroup$

I am working on an AI project to predict the life time of an industrial tool. the data I have represents the consecutive Power values of the spindle during each use of the tool to produce a new piece. The target (restNum) is a percentage value representing how much I could use this tool before replacing it. I calculated it manually according to the available information in the database. I tried to extract new features from this Power data and considered it as time series data for each use to produce a new piece and used the The R package tsfeatures to generate these features for each piece production. When I analysed the correlation between each feature and the target (restNum) using Orange Tool, I noticed that there is always low correlation between them and the target. enter image description here

When I also draw a scatter of this data, the low correlation is also clear, so that for any value of a specific feature is mapped to all possible values of the target. For example you can see the scatter of two features (x axis) and the target (y axis).(It is the same with the other features) enter image description here enter image description here

How is it possible to improve the processing of this data to prepare it for generating a regression model of the target. The model has not been specified yet and I want to compare between different models. But according to my tests the model could not learn from this data and the accuracy is not more than 70% in case of using ltsm neural network for example. Any ideas or hints to choose it?

Improve this question
$\endgroup$

1 Answer 1

Reset to default
0
$\begingroup$

This is a super general question; pre-processing data for regression or classification is not a universal process and is very data-dependent.

Sometimes the data does not contain strong signal - do you have a strong reason to expect more than 70% accuracy? Do you have a good reason to expect correlation between your predictors and the outcome? Perhaps there is an unobserved confounder -- for example, the life time could be largely due to unobserved differences in manufacturing quality combined with random variation. Or perhaps your predictors and outcome are completely unrelated.

More information is needed to approach a decent answer for your questions.

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.