Time-Series Regression Scenario #2: I have data from multiple sensors and want to predict remaining useful life (the amount of time a machine has before it needs repair or replacement).
My colleagues and I see this question with our customers in industrial automation who need to identify problems before they become dangerous or expensive. This time, you might want to use an LSTM network over machine learning regression. This approach reduces the need to identify features manually, which would be a significant task given multiple sensors.
Here's an example you can follow to see how to predict remaining useful life with an LSTM network.
Time-Series Regression Scenario #3: I have audio data I want to denoise.
Here you could use a CNN. The important thing about this method is to convert signals into images prior to passing them into the network. This means the signal becomes an image representation through Fourier transform or other time-frequency manipulation. Using images provides a way to see features you might not be able to visualize in the original signal. The network used can be a pretrained network designed for images, since a Fourier transform is essentially an image.
Here's an example demonstrating how to denoise speech using a CNN.
Another thing I should mention for anyone in Scenario 3 is that wavelets are also a semipopular method to extract information from time-series data and then use that as input for a CNN. There’s a nice write-up from researchers at UT Austin on how they converted brain signals to words and phrases using wavelets and deep learning.
Now, once again, you can do what you want here. It’s very possible that you could also use an LSTM network in Scenario 1, or a CNN in Scenario 2. These scenarios are just meant to give you a starting point.