I read your article about denoising autoencoder for x^2 wave form .. It’s really helpful and thank you .. I tried to improve this idea to sin wave but not only the same wave that copied again and again … I created 1000 different signal and 1000 different noise .. I need your help to know how to make perfect autoencoders .. any help or advice will be helpful.. thanks
If you’re training on a variety of sine based wave forms that are noisy (e.g. sin(x), sin(2x), 4sin(x), …), you need a lot more than 1000 samples. If you find poor performance, the size of your dataset is likely the issue. My suggestion would be to increase the dataset by a thousand fold, if possible, then train again. Neural networks really need a lot of data to work properly! Why it worked in my case is because I trained with an equal base form all the time, and it seems that you are not doing that.
still finding difficulty of building the autoencoder.. I am using different noise variance for each 1000 pure signal of total 20,000 .. I don’t know is the variety of the noise signals is the problem or there are other problems..
What is the layer structure of your autoencoder? What are the hyperparameters?