automated feature generation with matlab?

I am searching for a good method to generate automated features from my dataset, which I can use to put in a classifier. My dataset contains signals (for example acceleration), so it is numeric.

I already tried autoencoder and sparsefilter, but they seem only to work properly if there are features available already.

My aim is to generate features from the dataset. Is there another algorithm I can try? Maybe a better neural network (with more hidden layers)? And if I can use a better neural network, how can I get the features out of it?

Every suggestion is very much appreciated!

Answer 1

[Ramanuja Jagannathan]

Can you provide the problem you are trying to classify, so that a better solution could be provided. Nevertheless, if the objective is to do classification on the 1 dimensional data, you could have a look at the classification learner app. https://www.mathworks.com/help/stats/classification-learner-app.html

You could also have a look at the Predictive maintenance toolbox for analyzing and labeling machine data. https://www.mathworks.com/help/predmaint/index.html

Also, have a look at our examples on neural network problems, one of which might be helpful for your problem. https://www.mathworks.com/help/nnet/examples.html

[Artem Bykanov]

Thanks for your response! I am trying to classify whether a fault exist or not. And if it exist where. In the end I have 6 classes to classify for.

For this problem I have time dependent data which are generated from acceleration sensors. Now there is the possibility of generating features with functions, which take for example the max value of the data. But my aim is to generate other automated (not physical) features, which I can give to the SVM to classify.

concluded: For the SVM I need features, which I want to generate directly from my dataset. So the input has to be the acceleration signals and teh output should be features.

Thanks in advance for helping me with this problem!

Answer 2

[Artem Bykanov]

Thanks for your response! I am trying to classify whether a fault exist or not. And if it exist where. In the end I have 6 classes to classify for.

For this problem I have time dependent data which are generated from acceleration sensors. Now there is the possibility of generating features with functions, which take for example the max value of the data. But my aim is to generate other automated (not physical) features, which I can give to the SVM to classify.

concluded: For the SVM I need features, which I want to generate directly from my dataset. So the input has to be the acceleration signals and teh output should be features.

Thanks in advance for helping me with this problem!

Answer 3

[Bernhard Suhm]

This sounds like a good use case for applying our new AutoML capabilities. Wavelet scattering generates features automatically from sensor (and image) data, and then automated model selection with fitcauto finds an optimized model. The process is describes on this page.