I am working with failure prediction application of an artificial neural network. My data set consist of 29 days (data utilization trace over a 29-day period, CPU usage, machine failure, ..etc).

Output: time to failure (TTF)(e.g. TTF at day 30).

So we use most of the days for training, then we predict the next day. All feature (E.g CPU usage) is available, expect the failure information.

If I need to calculate time to failure TTF in these 29 days, I will have some null values.

For example: assume machine has a different failure in 29 days. failure at day 3, 10, and day 20. Now the TTF after day 20 is not available, because we do not know when was the next failure after day 29. AND I need to use most of 29 days of data to train ANN. I believe it is not a good idea to use only 20 days for training.

I replace the missing value by the mean strategy but I got a bad result.

  • $\begingroup$ What you describe seems to fall into the area of Machine Learning called „survival analysis“ (I.e. almost like regular supervised learning up to the fact that the target variable and not the features themselves are missing from time to time). There are multiple ways how to approach such a problem, the most prominent probably being Cox Hazard models or turning a classifier into a survival model: benkuhn.net/survival-trees. $\endgroup$ – Fabian Werner Aug 23 at 8:36
  • $\begingroup$ @FabianWerner, yes I believe it makes sense, and ANN might not good for this data. Since I have enough data, I will choose the model that fit the data distribution $\endgroup$ – jou Aug 24 at 17:55

Your Answer

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

Browse other questions tagged or ask your own question.