Multi-property Anomaly Alerts

[BradC]

Hi all,

I have set up a series of anomaly alerts on a demonstration pump rig. The demo rig has flow rate, pressure and temperature sensors, with anomaly alerts set up for each sensor. For the demonstration, the presenter can turn on the pump which then operates for 30 seconds before being automatically switched off. To train the anomaly detection I simply recreated a series of 30 second pump cycles. With the timing between each cycle being random.

The alerts were all trained with the following parameters:

Outbound Anomaly Rate	Min. Data Collection Time	Certainty
5 seconds	5 minutes	95.4  %

The goal of the demonstration is to have one of the valves in the rig be closed to represent an unwanted blockage in the pipe. This should then trigger one of the anomaly alerts.

After successfully training the anomaly alerts, I find that the pressure anomaly alert triggers as soon as the pump cycle starts. Which, from the expected use case is not anomalous. I have attached a picture of what I think the problem is.

From my understanding; the underlying model is doing a step-ahead prediction using a neural network for a series of sliding windows.  For windows 1 to 4 the current pressure states of the system are the same. The step ahead prediction for these windows are relatively easy to predict since it's relatively well-behaved. However, for window 5 the pressure state of the system is the same as windows 1 to 4 but the step ahead prediction is significantly different. One way of overcoming this is to include the current input state of the system as a variable for the step ahead prediction since this would help distinguish window 5 from windows 1 to 4.  

What is the best way of implementing multi-property anomaly detection in Analytics since it is not offered out of the box? 

Regards,

Brad

[cmorfin · Accepted solution]

Hi @BradC 

I do not know of a way to perform a multi-property anomaly detection ... I don't think it is possible at this stage.

What you can try is to increase the training time.

The anomaly detection works best with cyclic signals. If I understand correctly you are  stopping and starting it at random time, this makes the signal  non cyclic and is much more challenging for anomaly detection.
What you can try is to increase the training time, so ThingWatcher can see more of the behaviour and maybe work it out.
Also it is important to remember that this is based on machine learning/statistics, and due to the internal workings of ThingWatcher there is an expected delay between an anomaly occurring (or not) and ThingWatcher reporting it.
Also in your case, it might report an anomaly as soon as you start it, but if you let it run, does it not recognize the signal as being ok ?

The article https://www.ptc.com/en/support/article?n=CS248761 has some interesting answer to questions that goes with what I mentioned above.

Hope this helps

Christophe

[slangley]

Hi @BradC.

If the response provided by @cmorfin answered your question, please mark it as the Accepted Solution for the benefit of others with the same question.

Regards.

--Sharon