In the picture, there are data and a target in the time series. The data is padded until it reaches the max length. The target is marked by humans. It pins down the starting point and stopping point of the pulse.

The first idea is to encode the time series to be a text-like DNA sequence and then apply machine translation to it. My inspiration comes from English and German translation (aka word replacement).

Update 27 May 2021:
I found a similar experiment in the kaggle


Suppose I use seq2seq neural network train on this dataset. Later on, I test with the same data but shift the time to the right-hand side 1000 points. Will it be able to detect new starting points and new stopping points?

If not

Can a neural network detect the starting point and stopping point?



1 Answer 1


As a general principle for: process the sequence by splitting into smaller, overlapping analysis windows. The window length should be slightly longer than your event of interest. In such a window you can compute features that characterize the event, such a the difference between min and max inside the window. With such features the events shown here should be rather easy to separate out. You can used the labeled data to train a binary classifier on each window, like Logistic Regression or Random Forest.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.