Example for one of the faulty days

This has been done for everyday provided in the .csv file (over 35k rows)
The value of n and the window size for the rolling median can be calibrated for a more sensitive and a more corrective/smoother approach. And this can be ran multiple times to correct exactly for scenarios like the one in the picture.

After more smoothing

NOTE:

Only lower outliers have been considered. For our implementation, the outages present in the data were removed since even normal error in the logs or missing data could be interpreted as outages. We chose to remove that misunderstanding entirely.