0
$\begingroup$

I have converted a time series to a symbolic sequence and want to use machine learning to detect anomalies. In this specific case, every time series represents the same process. So a similar, but not exactly the same pattern is repeated again and again as long as the system is in its non-anomalous state. So something like

abacab abacab ababcab abacabc

Sequences in an anomalous state are supposed to deviate stronger. The strength of deviation should be expressed in the output of the anomaly detector.

Im struggling to find a comparable problem in literature. With what type of method is this typically solved? Can someone point me to a paper?

Improve this question
$\endgroup$
5
  • 3
    $\begingroup$ What's your reason to convert TS to symbolic sequences which is usually more fit to dfa? Imagine your normal cases for the same process are all stationary and anomalies are all nonstationary, then it's apparently hard to judge while in TS at least if your TS is an ARMA process you can use unit root test to easily tell. $\endgroup$
    – cinch
    Commented Dec 28, 2022 at 20:03
  • $\begingroup$ Hi. I got the idea for conversion from papers on symbolic aggregate approximation and I thought this representation is a good starting point for anomaly detection. $\endgroup$
    – sinpalabras
    Commented Dec 28, 2022 at 20:40
  • 2
    $\begingroup$ You would be better off describing your problem without presupposing a particular set of procedures or encoding would be part of the solution. Such a presupposition unnecessarily limits the scope of responses and, in the worst case, can eliminate superior solutions altogether. $\endgroup$
    – whuber ♦
    Commented Dec 28, 2022 at 20:42
  • $\begingroup$ I agree with the others that this sounds like it's better handled as a time series sequence, where one can use models like Hidden Markov Model with constraints, or a neural net like LSTM/GRU to model the sequences. See also xyproblem.info $\endgroup$
    – Jon Nordby
    Commented Dec 31, 2022 at 23:16
  • $\begingroup$ Sorry for the delayed reply. What if the symbolic representation is based on apriori knowledge and is therefore a way of knowledge enhanced feature engineering/dimensionality reduction. It is done to see if this is beneficial to anomaly detection compared to using the time series. $\endgroup$
    – sinpalabras
    Commented Feb 14, 2023 at 19:49

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.