How does Stream Analytics Anomaly Detection work?

Question

I am new to Azure and am testing out their in-built Anomaly Detection function (AnomalyDetection_ChangePoint).

I have connected live streaming data from a temperature sensor in an office with readings that hover around 28C when the aircon is on and 30C when the aircon is off.

May I know if the model will read the normal range to be 28-30C and only detect temperatures outside this range as anomalies? Or can it take 28C to be the normal and detect 30C to be an anomaly?

Greatly appreciate any help to better understand how this function works. Thank you!

Answer 1

Hello @116dreamer ,

Azure Stream Analytics offers anomaly detection in several ways: spike and dip detection and changepoint detection.

This is based on a mathematical model so it involves a statistical confidence level.

Instead of comparing incoming values to (fixed) values, this detection compares values with previous values over a certain time span.

If the new value difference is detected with a certain confidence, it is seen as an anomaly.

So you basically fill in:

• the value to check
• the number of messages expected
• the duration
• the confidence

I suggest running a training set through Azure Stream Analytics and testing it before you take it into production.

I also suggest using this next to a more traditional (fixed) value comparison.

Check out my blog post too.