I am working on a time series data for which I intend to impliment machine learning model for detecting change point in time series data.

This data is recorded fom machinary and we have to predict process is completed. Now we know when process is completed(actual time at which process is getting completed) and we have to link it to the data that we have recorded from sensors. Our goal is to build model trained on this currently available historical data which will be able to predict the time at which process is getting completed for the future data.

One of the sensor data look like this:

enter image description here

Now my question is I know that process gets completed at 11:50, so I have created target variable which is encoded as 0 from the beginning and then 1 form 11:50 onwards. This data I fed to Neural networks to predict the variable. Is this a correct approach? Is there any sophisticate method for encoding the target variable than just labelling it as 1 after certain timestamp as it is time series data and my goal is to predict time at which process will get completed for the next sesnor data?


1 Answer 1


If you are trying to detect (as soon as possible ) when it finished then I would use Intervention Detection procedures http://docplayer.net/12080848-Outliers-level-shifts-and-variance-changes-in-time-series.html to form a hybrid model of memory (arima) and latent deterministic structure (pulses/level-step shifts, seasonal pulses and local time trends).

I would take the values up to 11:50 and incremenatlly add future values one at a a time AND then determine when the statistical model identification process suggested a permanent level shift based collectively on the new observatioms .

This would require you to specify either implicitely or explicitely the minimum # of future values (L) before a statistically significcant level shift was declared/ proclaimed/found at a specific level of confidence (Z).

I have recently incorporated a third "given" and that is the minimum magnitude of the level shift into my research and practice. In this way the user can minimize false level shift conclusions as not being "LARGE ENOUGH OR SUBSTANTIAL ENOUGH " to warrant attention.

If you can share your data ( or some proxy ) in a csv file I will try and help further.

  • $\begingroup$ Thanks a lot for your answer. Also the paper looks something I can build the work here. But apart from ARIMA if I had to use any machine learning algorithm, for which if I have to create input-output combination, what will be the best approach for encoding? What does it mean to incrementally add feature as you have mentioned in the answer? $\endgroup$
    – user247481
    Jun 4, 2019 at 8:05
  • $\begingroup$ Suppose I have a time series and I want to detect a level shift of at least L units using no less than K values in that group with at least Z% confidence. AFAIK there is no machine learning to be done here (just statistical-learning) as what you are trying to uncover is latent deterministic structure ( e.g. a level shift) while accounting for (conditioning for) memory effects and machine-learning does not do this.. $\endgroup$
    – IrishStat
    Jun 4, 2019 at 8:27
  • $\begingroup$ thank you for the clarification. $\endgroup$
    – user247481
    Jun 4, 2019 at 8:51

Your Answer

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