Identify "steady-state" time series window

Question

I'm new with the time-series analysis.
I have several time-series (noisy of course) part of the same set of measurements (sampled simultaneously). The time series are the results of a stochastic process (dynamic system) where part of the time-series are inputs to the process and other time-series are the outputs.
I need to find a time window corresponding to a steady operation, i.e. where both inputs and outputs don't change over time (for a deterministic signal would say derivative over time almost zero).
Do you know any way to accomplish this task? The measurements are quite long, I need a rather computationally efficient method ideally without involving any visual inspection.

Thanks!

Answer 1

This could be defined as a classification problem over your time series.

One approach I can think of is to build an LSTM based classifier with a couple of dense layers for binary classification (whether the window is steady or not) and define your own loss function.

So in theory, you're training the LSTM network to sample batches of your time series and classify them as steady state or not.