Anomaly Detection now has multivariate and univariate

The Anomaly Detection service is a multivariate anomaly detection solution that generates a model by analyzing multiple timeseries signal data values that are highly correlated with each other.  In this release, rather than dropping signals whose values are weakly correlated with other signals, the model training API generates separate anomaly detection models for individual signals.

As part of this release, the model training API identifies and generates two types of models.  For a given data set, when signal values of multiple signals are inter-correlated, the API generates a multivariate anomaly detection model. For signals, whose values have weak (or no) correlations with other signals, the API generates separate univariate models for the individual signals.

The release highlights include:

  • Univariate models for non-correlated signals.
  • Ability to train only univariate models in completely uncorrelated datasets.
  • Ability to fallback to univariate models.
  • Parallel inference of multivariate and Univariate payloads to maintain consistent performance as before.
  • More efficient model packaging solution.

Component Updates


  • With this release, we have added support for a Univariate Kernel (One Class SVM) for detecting anomalies in univariate signals.  During model training, the service builds univariate anomaly detection models for uncorrelated signals and datasets.
  • When the requested FAP scores can't be attained, the Multivariate Kernel fallbacks to generating univariate models.


  • To deliver consistent performance, the backend runs inferencing on both multivariate and univariate signal payloads concurrently.
  • The inference API response and output now includes information about whether a particular signal has a univariate model. This information is in the details field.
  • The backend has been updated to use a more efficient model packaging solution.

User Interface

  • This release includes improved visualization graphs so that you can easily and quickly identify anomalies.  You now have the ability to select and view graphs for univariate, multivariate, or all signals.
  • When model training completes successfully, an information message about the number of models built is displayed.

Bug Fixes

  • When the detectAnomalies inference API request payload doesn't include timestamp attribute (values), the API call failed.This issue has been fixed.
  • The 'warning' attribute was missing in the model details file. This issue has been fixed.
  • When insufficient data points (rows) were provided during model training, the API was responding with an incorrect error message. This issue has been fixed.

You can find interesting information in the AI Blog.