0
$\begingroup$

I'm interested in data analysis and machine learning and have taken some online courses about "R" and "statistical data analysis" so far. So I would consider myself maybe as an advanced beginner.

Now I'm trying to apply some basic machine learning concepts on a practical example that somehow differs from the examples in the courses. I would like to develop a failure detection routine for a solar heating system. More precisely: I would like to detect air in the solar loop by analysing temperatures, pressure and flow values.

My problem: So far I mostly worked with datasets where all the observations were independent. Now I have a time series of measurements of - lets say - three temperature values, one pressure value, one flow value, weather data and the information, whether the solar pump is currently running or not.

All this data is of course somehow interrelated. The logging rate is around 60 seconds. I have data of a solar system running under normal condition, and data for states ranging from "a lot of air" down to "almost no air" in the loop. (Unfortunately I have no exact quantification of the amount of air).

Now I don't know how to choose a suitable ML method for this case. All the concepts I learned about so far don't seem to fit here. Does anyone have some advice or can give me a few buzzwords for methods that could work here? At the moment I feel stuck because I don't know what to look for.

Thank you in advance!

$\endgroup$
0
$\begingroup$

Anomaly detection usually works by assuming some "normal" (in the sense of "correct", "non-anomalous") distribution of the measured values. This assumed distribution is usually "normal" (now meaning "Gaussian"), but it need not necessarily be so.

Thus, deciding whether a tuple of measurements values indicates an anomaly is basically answering the question:

How probable is it to obtain values as extreme or more extreme than the measured ones?

If the probability is below some pre-defined threshold (say, 0.01), you'll classify the values as anomalous.

In your case, with multiple measurement values, you are dealing with a multi-dimensional probability distribution, e.g. a multi-variate normal distribution, but the above basic idea still applies.

| cite | improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.