0
$\begingroup$

Context:

I have thousands of IoT device time series. Multiple metrics are tracked per time period. In order to perform predictive maintenance, I need to detect anomalies.

Problem:

There are multiple categories of devices or devices from the same category with different firmware versions. I wonder if I should train one model per:

  • time-series (find anomalies per device)
  • group of time-series (figure out if one device behaves differently from its cohort)
  • one single big model to rule them all (is there a strange cohort/category of devices)

Or maybe even a combination of all of the above?

I am particularly interested to get this question answered for models like https://matrixprofile.org/ or an autoencoder based approach.

Improve this question
$\endgroup$
3
  • $\begingroup$ What kind of anomalies are you looking for / do you have? Is it on device characteristics, like firmware issues, network issues etc - or on the machine/setting that the devices observe? The latter is typically more the case for predictive maintenance $\endgroup$
    – Jon Nordby
    Commented Sep 8, 2020 at 7:11
  • $\begingroup$ If it is the latter (machine/environment) that is important, then it the characteristics of the machines/env that primarily should be considered when choosing approach. And hopefully device characteristics like firmware version can be ignored, or controlled for $\endgroup$
    – Jon Nordby
    Commented Sep 8, 2020 at 7:14
  • $\begingroup$ Probably more with regards to the quality of network connectivity. $\endgroup$
    – Georg Heiler
    Commented Sep 8, 2020 at 7:45

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.