Hello, I am looking at an unusual problem for a client: There are ~20 CO sensors in the factory, each one generates a number every 15 minutes.
Each sensor puts out a time series with 96 steps every day. I need to come up with a generative model for these (for automatic anomaly detection). Each sensor is identical, so generative process for the data can be assumed to be the same.
As a first pass, I averaged out each component of the series, to come up with a mean time series. Autocorrelation over the elements of this mean series shows clear first order dynamics, and an ARMA model fits quite well.
However, I am worried that this is the best way to model this kind of data, since each sensor is deployed in its own environment and the standard deviation of each component of the mean series is significant. So, a model trained on the mean might miss outlier events in individual series.
I wonder if there is any guidance on how to approach this kind of problem?
Thanks for any pointers!
[link][comment]
