I am working on a time series problem, trying to extract some valueable feautres at the latent space, my vibration data set is 100 times sequence, 112 samples, 1 feautre (since we take only one axis of accelerometer data)
I try to train using Lstm and CNN autoencoders separately, in both the cases in the training the loss is low, but in testing a new data set with the trained model. I get high loss, how to rectify it?
I am expecting to extract some feautres from the latent space