1

What are some good time series classification methods where we have a single time series of some motor and a labelled training data indicating at what points did the motor fail, the rest are normal working. The time series classification model should be able to tell motor state normal or fail based given a set of time series features.

kjetil b halvorsen
  • 63,378
  • 26
  • 142
  • 467
user6460588
  • 103
  • 11
  • Could you provide an example dataset (possibly toy example)? – Nikolas Rieble Jan 05 '17 at 14:31
  • For example, I have a set of features like machine speed, temperature, noise etc based on which I want to determine if machine is in normal operation or has experienced a failure. The data of features is available in the form of a time series at different instants labelled at different instants. My interest is given the next set (or window ) of features like speed, temperature etc it should classify it as normal or failure – user6460588 Jan 05 '17 at 14:36
  • As you have only 2 classes involved, you have a lot of options that you can try. The accuracy of the results will mostly depend on how well the features actually represent the state of the motor. Did you have a look at regressions and SVM classifiers? For times series, HMMs work well too, especially if each feature can be well represented by Gaussian distributions (but it is a heavier process than SVM or regression) – Eskapp Jan 05 '17 at 15:13
  • But svm classifiers work well with independent examples not with time series. And also I want to mention to mention that the failure points in the time series are very less. Approx 4 points against 1000 normal ones. I don't think here two separate classes can be created. – user6460588 Jan 05 '17 at 15:35
  • You could try logistic regression, modelling the *probability* that state is failure – kjetil b halvorsen Oct 07 '18 at 02:02
  • Possible dup: https://stats.stackexchange.com/questions/50807/features-for-time-series-classification – kjetil b halvorsen Jul 25 '19 at 15:18
  • @user6460588 The issue with independence in SVM is that each observation is independent of the other. In your case, your observation is multivariate. You have some binary variable indicating failure or success, and then you have variables representing each value of each time series. In that regard, it is like an image classifier. The pixels of an image are related, but each image is independent. – Dave Jul 25 '19 at 15:19

0 Answers0