IPSW 2019 work
So far, this is just preliminary presentation stuff.
To check out the data structure,
jupyter notebook
and select the "Exploring [...].ipynb" notebook. Then you'll have to change the filename=....h5 line.
plot.py ../path/to/h5File.h5 --draw_boxes
O'Shea convolutional modulation recognition
O'Shea autoencoder signal identification
Unsupervised structured signal identification
Modulation identification using higher order cumulants
Clustering on autoencoder lower dimensional representation
Perhaps an autoencoder could work. Would need to train on a bunch of raw data. After training, we would pass new data x throught he autoencoder model and obtain f(x). If the mean-squared error distance between x and f(x) is high, then x is probably anomalous. We could then add x to the training data so that the model f could learn from the anomalies. Problem is the input data x must always be the same shape. The pro is that this could be implemented using a GPU via the PyTorch package. Paper: https://arxiv.org/pdf/1807.08316.pdf
Some good references on coding autoencoders using the PyTorch package: https://medium.com/@vaibhaw.vipul/building-autoencoder-in-pytorch-34052d1d280c https://github.com/L1aoXingyu/pytorch-beginner/tree/master/08-AutoEncoder
PyTorch is a great Machine Learning python library. It allows for easy backpropagation which is needed for training. Futhermore, it enables GPU usage for faster computations. Documentation: https://pytorch.org