If I want to train a convoluted NN on time series but I cannot decide where to split the data.
I see that other people use jumping window over the input. so the feed say 20 sec of observation as 1 sequence into a CNN.
I cannot do that as splitting the observation by fixed size will most definitely break important patterns - ie first part of a pattern will go into the end of the current seq and the rest into the beginning of the next seq.
I can however find a sensible solution by preprocessing data and finding places of much smaller significance and make seq cutting in the middle. but then the seq length will vary greatly.
Can I still use CNN ? or this idea is silly?
How else people extract features from time series using CNNs ?