How to use Anomaly Alerts

[drieder]

Hello everyone,

I have successfully created an anomaly alert for a temperature property with default settings (Outbound Anomaly Rate = 5secs, Min. Collection Time = 1min) exept of 90% certainty. During the training the temperature was quite constant at 27 degrees.

My goal is to detect whenever there is an high temperature raise and I hoped to achieve this with the anomaly detection. So I created a subscribtion to the event "Anomaly Monitor Status" and the property "temperature". Unfortunately the service in the subscribtion never triggers when I raise the temperature to 35 degrees. 

Also the AlertHistoryStream doesn't show new alerts. This is how the history looks like currently:

The top most alert was created after anomaly training finished. Since then not a single alert was created eventhough the temperature exceeded the normal temperature by far.

Any Ideas what could be the cause? Did I subscribe correctly to the alert?

Any help is appreciated very much.

Best Regards,

Dominik

[cmorfin]

Hi Dominik

I would think we can set aside the subscription for now. If you do not see any new anomaly alert triggered then the subscription won't trigger either.

However I am wondering if in your case the anomaly alert is the best. If the values are fairly constant, maybe a above/below or in range alert might be more suited and more accurate.

The anomaly alert is based on statistics, therefore not 100% accurate. It needs a few datapoint to react to an anomaly. If there is only one value outside the usual range it is very likely it won't be flagged as anomaly if subsequent values are back in the range.
Similarly when an anomaly has been detected, it will need some more time to detect that it has gone.

If you are not already using it, the anomaly mashup can be useful to see the trained data and current values in a graphic way.

However as indicated above, if your temperature is suppose to be fairly constant in a range, other type of alert might be more appropriate.

Hope this helps.

Kind regards

Christophe

[drieder]

Hello Christophe, thank you for your reply. I thought the anomaly detection would make sense since the average temperature changes throughout the day/night and through out the year, what makes it hard to define a border at which my "alarm" will trigger.

I hoped to detect "unnatural" temperature behavior this way. Do you know how I can get the Anomaly Alert to be reseted? This event never happens...

Best Regards,

Dominik

[cmorfin]

Hi  Dominik

The anomaly detection uses machine learning and is best suited for cyclic signals.
If you have signals that may have several normal states (for example summer temperature is ok between 25 and 30, while winter temperature is ok between 10 and 15), then you will have to retrain the model  to suit the new normal state or have several alert setup. If you only train during one normal state, the anomaly detection will likely report all as anomaly during the other normal state.

To reset the alert, you can try the 

DisableAlertsForProperty and  EnableAlertsForProperty  services on the Thing.

Kind regards

Christophe