====Live Anomaly Detection====

Running the [[start|AI]] via the web interface or [[API]] is useful, but it's **slow**. Every request has to load, run and unload the AI on the server, which takes a considerable amount of time.

If you want to have live alerts or feedback from your anomaly detector, you can instead download your model and run it via our //modelhost// system.

This keeps the AI in memory and re-runs the model whenever there's changes in the state of your assets, giving you up-to-the-second anomaly detection.

It also lets you use the information in a variety of places, making it available via OPC-UA, MQTT, Prometheus, REDIS, write it to databases etc.

===Downloading your Model===

When signed in as an Administrator, open the dashboard for the asset and click **AI Anomaly Detector**.

Click **Live Model**.

Then click **Download Model Files**.

This will give you a ZIP file containing the Python tools required to run your AI model live. This can be run on your ARDI server, or on any other system that has access to the ARDI server (ie, you might have a dedicated AI system).

===Running your Model===

You run your model by launching **model.py** in Python.

On linux, this can be done with the following command...

<code>
python3 model.py
</code>

or in Windows by running **run.bat**. If you don't have Python in your PATH, you might need to edit the run.bat file to give it the path to your installation of Python.

===Viewing Outputs===

You can view the outputs of your model using an OPC-UA client, such as **UAExpert** on Windows. The default port for the model is 5555 - when running on the local machine, the full address will be **opc-tcp:/ /localhost:5555**.

You'll have the following values...

^Path^Usage^
|<asset_name> Anomaly Detector/Constants/* Tolerance|Difference thresholds for each property|
|<asset_name> Anomaly Detector/Inputs/*|The real-world, live values for each property|
|<asset_name> Anomaly Detector/Outputs/Model *|The output/modelled values for each property|
|<asset_name> Anomaly Detector/Outputs/* Error|0 if the signal is normal, otherwise 1|
|<asset_name> Anomaly Detector/Outputs/Anomaly|0 if the asset is normal, otherwise 1|

For instance, if your asset was named 'Motor 1' and had the property 'Temperature', your OPC-UA output would include...

**Motor 1 Anomaly Detector/Constants/Temperature Tolerance** \\
**Motor 1 Anomaly Detector/Inputs/Temperature** \\
**Motor 1 Anomaly Detector/Outputs/Model Temperature** \\
**Motor 1 Anomaly Detector/Outputs/Temperature Error** \\
**Motor 1 Anomaly Detector/Outputs/Anomaly** \\

If the //model temperature// is ever more than //temperature tolerance// away from the real-world //temperature//, both the //temperature error// and //anomaly// values will be '1'.

===Merging Multiple Models===

Although you will download the models individually, it's possible to combine each of your anomaly detector AIs into a single, common script. This will make them all available from the one OPC-UA server, making it much simpler to record data and feed it back into ARDI or your control system(s).

Your **model.py** file will contain several functions. 

**AnyTrue** and **ExecuteModel** are always exactly the same, so you only need to include them once in your shared script.

The **AnomalyXXX** function (for example, asset #12 will have a function called //Anomaly12//) can be copied into a common script. Remember to include the decoration (the line beginning with '@') as well.

This combined script can then be run as a service, giving you a constantly-running library of live anomaly detector AIs that can communicate directly with industrial systems.