Signal translation with Seq2Seq model

Question

I'm currently doing some research on signal processing and I got a dataset which includes the signal in itself and its "translation".

So I want to use a Many-to-Many RNN to translate the first into the second.

After spending a week reading about the different option I have, I ended up learning about RNN and Seq2Seq models. I believe this is the right solution for the problem (correct me if I'm wrong).

Now, as the input and the output are of the same length, I don't need to add padding and thus I tried a simple LSTM layer and TimeDistributed Dense layer (Keras):

model = Sequential([
    LSTM(256, return_sequences=True, input_shape=SHAPE, dropout=0.2),
    TimeDistributed(Dense(units=1, activation="softmax"))
])

model.compile(optimizer='adam', loss='categorical_crossentropy')

But the model seems to learn nothing from the sequence and when I plot the "prediction", it nothing but values between 0 and 1.

As you can see, I'm a beginner and the code I wrote might not make sense to you but I need guidance on few questions:

• Does the model make sense for the problem I'm trying to solve ?
• Am I'm using the right loss/activation functions ?
• And finally, please correct/teach me

Answer 1

I am skeptical that machine learning is the right tool for this problem. I would look for a more direct solution, perhaps using peak detection or changepoint detection, or some other form of classical method for time-series analysis.

If you do use machine learning, cross-entropy loss is not the right loss for your problem, and you will definitely need to change that.