# Analysis of time series data where its structure changes

I'm relatively new to time series analysis and am working on characterization of time series data of a VM’s resource utilization (is its usage of CPU variable, how variable is it, magnitude and duration of bursts). Furthermore, I will not necessarily know the characteristics of the time series data (periodicity, seasonality, trends, etc.) and will need to perform any sort of exploratory analysis programatically due to the number of VMs.

For example, I have the following time series where the nature of the time series changes drastically roughly midway. What is the best way to detect such a change in the time series given the scenario? I’ve looked at change point detection and have experimented with ruptures. The granularity of the change points detected is dependent on the method and parameters to the cost functions.

Also, would it be better to analyze the individual segments of the time series for periodicity, seasonality, etc.?

• If you wanted an online method to detect only very noticeable changes then you could use the Shewhart chart: make an estimate of the mean $\mu$ and standard deviation $\sigma$ of the "normal" behaviour, which I'm assuming is the type of observations that are present at the start of your chart. Then you can signal a change to abnormal behaviour whenever you see an observation $|X_t - \mu| > 3\sigma$. – Alex May 18 at 8:09
• Thanks, @Alex, I'll look at that. – Jan Kho May 19 at 19:04