Differences Between Model Types

The anomaly detection system offers three different types of model that can be used depending on how much data you have about your process.

You choose between these model types when creating a new anomaly detector.

If your process is reliable and repeatable, you should be able to use either the Predictive or Generative models. If your process is less predictable, more inconsistent or you only have limited sensor coverage, the simpler Support Vector Machine models might be the best fit.

The predictive model is the most powerful, but also the one that needs the most complete picture of your process.

This type of model builds a unique design of parallel neural network to try and predict what each asset property should be, based on the overall context of your asset.

However, this type of model may not work very well when you have randomness or changes caused by events or inputs outside what is 'visible' via the inputs to the AI model. If you aren't able to train an accurate version of this AI, we suggest then trying the other two models.

This produces a value predicting model.

The generative model uses an autoencoder neural network to predict the overall balance of your asset properties based on the state of the system.

This type of AI model is less likely to indicate specific problems (ie. 'Tension Too Low'), but will still give indications of which parts of your asset are out-of-balance.

Because it focuses on the overall balance of inputs and outputs, it still provides useful indication of if a system is running normally.

This produces a value predicting model.

Support Vector Machines provide a simple 'yes/no' signal showing if the asset is performing in an unusual way.

Unlike the other options, this type of AI isn't able to suggest anything about what the problem may be - just that the system is not working as expected.

The other anomaly-detection models try to predict specific values that your process should have. Some applications - thanks to external or environmental factors or limited sensor coverage - aren't that repeatable. Support Vector Machines work well when the target values are more vague or lie within a range rather than consistently being a specific number.

This produces a outlier detection model.