====AI Anomaly Detection====

The **AI Anomaly Detector** addon creates //anomaly detectors// for assets.

These came about after conversations with our customers. When an issue was found on-site that they didn't have an alarm for, they mentioned how nice it would be if they could have an alarm that detected //all of the things they forgot to make an alarm for//.

While intended as a joke, it got us thinking. And after some experimenting, we built one using common anomaly-detection techniques. While they were effective, we discovered the following...

  * Most of these algorithms simply provided a 'Yes' or 'No' response, and couldn't explain **why** they were triggering.

  * They would often be overly sensitive to channels of data that weren't very important.

Our solution was to use **Generative AI** to create an //AI model// of the asset(s). This helps to not only explain //why// an alert has been raised, but also lets users tune the alerts to ignore the less-relevant properties of your assets.

===How It Works===

Firstly, it uses your historical data to create a machine-learning model of how your device behaves.

With the right example data, this model has ability to predict what the condition of the asset //should// look like in different situations - for example, on particular products, at different speeds or in different environmental conditions.

It then //compares that prediction with the live values from the asset//, and highlights where the prediction and your actual values are different.

===Example===

For a very basic example, let's say we built a model of a **colour sensor** that makes up part of our cupcake baking machine.

It learns from our history, looking at the colour of our various products, such as chocolate, blueberry, apple and vanilla cupcakes.

When we next produce vanilla cupcakes, the model suggests that our final colour should be a light gold - but when the real-world value shows as a deep brown, we immediately raise the alarm.

Our anomaly detection does this sort of comparison, but can also do it on complex assets with many different values that need monitoring.

===Building a Model===

To build an AI anomaly detector, you first need to build the **model**.

1) Search for the asset you want to detect anomalies on and open the dashboard, \\
2) Press the **AI Anomaly Detector** button, \\
3) If there isn't a model for that asset already, you'll be able to press **Create Model** to build a new model. \\

Building the model can take a while - normally from 30-90s, but it depends on the size of the model and the hardware ARDI is installed on.

For more details, see [[creating|creating anomaly detector models]] to build models, or [[twinengine:tuning|tuning AI anomaly detectors]] to improve their accuracy.

===Using a Model===

Once building is complete, you can [[view the model state|view the model state]].

You can also get this information via an [[API|API call]], or take the model you've created and run it [[live against your process in real-time|live against your process in real-time]].

The system also supports the simpler [[svm-detectors|SVM-based anomaly detectors]].