Time series classification per sample (event detection)

Hi all,
Please help me with the following.

I want to train a model to detect events (posture change, can last up to several seconds) from acc data.
My input is multi-dimentional numeric data + binary labels (only for training, of course).
The model should then identify variable-duration episodes of posture transition in the given data.
It doesn’t have to be real-time → the model can look into the past and future samples.

I thought that RNN\LSTM can be a suitable option here, but can someone point to a good example or a comparable project?
I only see either future time-series prediction, or total time-series classification examples. Sometimes also anomaly detection (unsupervised, from data itself, without labels).

In my case I have long recordings likely containing many short events and mostly other activities.
Maybe it is not a good approach if I can’t find evidence that people use it?