Event Identification in Series I've just posted this question on Data Science SE which asks about machine learning methods to identify "events" in series (time-series or otherwise). I'm wondering if I should consider regressions too? 
The problem: I have series data, one variable is the continuous variable and the other is the measured variable. I want to identify any events that occur in the series. 
Example: 
The input data can be shown as a graph

The output should be used highlight any events in the series as such

My question here is, should I consider statistical methods (such as regressions) to solve this problem, or stick with machine learning? The method should be unsupervised, so I didn't consider regressions, but maybe I should?
For some example data, please see the question on Data Science.
Lastly, I don't think this post constitutes as cross-posting, as I am exploring a different route to the problem in the other question, but if you think am, please let me know. 
 A: This is a classical application of Statistical Process Control: you use the initial "in control" period to establish detection limits (e.g., by fitting a time series model to the data and calculating a one-step-ahead prediction interval), and whenever the series is outside these limits, you raise an alert.
This is actually supervised in a sense: given the data up to time $t$, you have an expectation for what should happen at time $t+1$, and if something unexpected happens, up goes the flag.
(See here for a motivation for short answers. Longer answers are always welcome.)
A: Time series data can have both auto-projective structure as well as latent deterministic structure i.e. the presence and waiting to be doscovered structue (pulses, level/ste shifts,seasonal pulses and/or local time trends ). The idea here is to identify both. 
Closely follow the work of Tsay here http://docplayer.net/12080848-Outliers-level-shifts-and-variance-changes-in-time-series.html and for a sometime useful software follow https://cran.r-project.org/web/packages/tsoutliers/tsoutliers.pdf which requires the user to pre-specify the form of the auto-projective structure (arima).
If the arima structure is not known then one needs to SIMULTANEOUSLY identify both as suggested here https://autobox.com/pdfs/ARIMA%20FLOW%20CHART.pdf in aniterative self-checking manner
