# Normalizing data before or after extracting time domain features

I have 100 time series (with 200 instances each) datasets each corresponding to a particular activity. I want to perform supervised classification for the activity. I want to use time domain (time-invariant) features for each time series to perform the classification.

Example dataset

| Time | Column A | Column B |
|------|----------|----------|
| 1    | 19.45    | 0.32     |
| 2    | 22.5     | 0.89     |
| ...  | ...      | ..       |
| 200  | 33.11    | 1.23     |


100 such files.

After extracting time-invariant features from each time series I get 100 s rows as below.

| Column A minimum | Column A mean | Column B minimum | Column B mean | Label |
|------------------|---------------|------------------|---------------|-------|
| ...              | ...           | ...              | ...           | Push  |
| ...              | ...           | ...              | ...           | Pull  |


I have 100 rows (1 row corresponding to each dataset originally) in this new dataset.

Now I perform classification on this generated dataset. My question is, if I want to normalize/standardize the dataset, which would be a better way? Scaling before the extraction of time-invariant features or after generating the time domain features?