I am dealing with IOT data from a mechanical machine. On the input I have ~100 features that are measured every minute. On the output, I have labels of zeros and ones, where zero indicates the absence of the event and 1 indicate the presence of an event. The event represent the failure for the machine in place. Therefore, the goal is to predict at every time step the remaining "minutes" for a failure to occur. I would like to know how to tackle this problem, and if possible for some material to read.

Is there a way to know which features in the past leads to a failure in the future if I'm using an LSTM?

  • $\begingroup$ Instead of remaining minutes to failure, which is extremely difficult to actually predict accurately, would a probability of survival to at least an arbitrary time t be okay instead? This sounds very close to a survival analysis problem, where you are interested in estimating the survival function via. a parametric accelerated time model or perhaps a Cox PH model with ad hoc estimators. $\endgroup$ – aranglol Jan 21 at 6:44

Try boosting and random forests. You would need a long enough time series. A good reference (with examples in R) is

James, G., Witten, D., Hastie, T., & Tibshirani, R. (2017). An Introduction to Statistical Learning: with Applications in R. Springer.

Usefulness of particular features can be assessed via variable importance plots.

  • $\begingroup$ This is basically a generic answer for any supervised learning problem (and to be honest, dangerous advice in general especially with time series that may exhibit trends in which gradient boosting and random forests are poor at due to being tree based learners). Magic multipliers are commonly used to fix this problem but this is by no means widespread or based on sound theory. $\endgroup$ – aranglol Jan 21 at 6:43
  • $\begingroup$ @aranglol Trends can be captured in some of the features. A careful choice of features is important, of course... My response is not generic since I, personally, had success with applying random forests to a similar problem. What highly specialized answer did you expect when the problem formulation occupies only 8 lines?... Give a peer the benefit of a doubt before penalizing his/her response. $\endgroup$ – stans Jan 21 at 7:04
  • $\begingroup$ In your answer, the OP clearly states uniqueness in their dataset with respect to the time to failure aspect. The problem is clearly more unique then just a generic time series problem. Furthermore, if you think features may benefit the problem at capturing trends then explain this in your answer and elaborate, or explain in your answer how you were successful in the past with a similar problem. Linking two webpages with no explanation isn't helpful for the OP or anyone else when no context or advice is given. $\endgroup$ – aranglol Jan 21 at 14:24

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.