Dealing with gaps in time series with TWA

[plpt]

My use case consist of machines sending data periodically (when machines are on). I want to predict failures.

I created a dataset with computed features for each on cycles and  then create a model on TWA.

But features are complexes to compute inside TW, so putting this model to production will be unconvinient.

I have 2 questions to make a better/easier use of TW for this use case (and for my personnal understanding of TW) :

• Is there a way that Thingworx can handle breaks in time series to create a model?
• Can Anomaly detection understand this kind of pattern and don't mind of time gaps between the next cycle (see joined image of a normal and failure cycle).

[nsampat]

@plpt ,

Thank you for posting to the PTC Community.

Regarding gaps in information and using Anomaly Detection with Time Series, have you review this article Using Time Series in ThingWorx Analytics 

There is information about sampling, using goal history for predictions, look back, and implementation.

Regarding the use of Anomaly Detection, we also have this hub article, Using Anomaly Detection in ThingWorx

Please let me know if the documentation provides you with the information you are requestion.

Regards,

Neel

[plpt]

Thank you for your response @nsampat, I did see these articles.

From what I understood time series model can not contain time intervals without data.

But since it's common in the industry maybe someone can give some tips and feedback from their experience.

Best regards,

Philippe

[nsampat]

@plpt ,

What you are describing can be handled by the lookback feature in Analytics Manager.

Please take a look at this article: How is the lookback window populated when using time series model in Analytics Manager?

Please note, this works if a value is missing for the interval being retrieve, it is then substituted with the previous value reported.

Regards,

Neel

[plpt]

Thank you for this precision, it does work for scoring but my question was about time blanks for model creation.

Best regards,

Philippe

[nsampat · Accepted solution]

@plpt ,

For blanks in model creation, you would have to revisit your data collection process.

Unfortunately the only suggestion I have for that is ETL on the data input end.

Regards,

Neel