How to train the anomaly detection module for long events (cycles of several hours)

[LR_9796586]

Based on these two previous questions related to the definition of a cycle and this one, regarding the time dependency on module. How should the following case be approached? 

Lets assume I have a process for which the cycles last 10 hours  each, according to the documentation at least five cycles should the used to train the model, on the other hand, according to this post all cycles should be in a time windows of 24 hours, in this case where the five cycles add up to 50 hours, how should the training be approached?

[nsampat]

@LR_9796586 ,

This appears to be a similar question to this post:

https://community.ptc.com/t5/ThingWorx-Developers/Can-the-Anomaly-Detection-be-trained-at-a-sampling-rate-and/m-p/734184#M52704

Please take a look at that thread, if you have any additional questions please let me know, otherwise if this answers your questions please mark this reply as "Solution Accepted"

Regards,

Neel

[LR_9796586]

Hi @nsampat the post you linked is not what I'm looking for. Instead of the "wating/dead" time during training process I'm referring to this statement in particular "The 5 cycles must be within the 23 hours and 59 minutes, otherwise its considered a new training window." from this post

If each cycle longs say, 8 hours as in the image below, which would be the approach to properly train the anomaly detection module? 

[nsampat · Accepted solution]

@LR_9796586 ,

The recommendation I provided, and the post you linked, still applies here.

The only way to simulate the training cycle is to submit all the data in a shortened period rather than actively training on live data.

The 5 cycles are a recommendation to have an assured trained model, but if your system is on an 8 hour cycle, 3 may be good enough but I cannot guarantee that it will be optimal.

Regards,

Neel

[LR_9796586]

Thanks @nsampat  

Is this limit a constrain of the algorithm used to detect the anomalies? is not clear from the documentation where this restriction comes from.  

[nsampat]

@LR_9796586 ,

I believe its the behind the scenes implementation of how our team designed the algorithm. I cannot really provide more information than that.

Regards,

Neel