Events prediction with time series of continuous variables as features

Question

We have the feeling that behavior of a device in terms of continuous variables (fans speeds, temperatures, voltages, ...) has influence on rare events happening (components failures).

I now have to build a predictive model for that, to proof the influence.

Those continuous features are given as time series, and events are punctual.

I've made a model based on descriptive statistics of those variable (see this question) with decision tree, random forest, adaboost, and clustering but it doesn't works. I will still improve by balancing classes, but I'm convinced it is not the best approach.

I'm pretty sure that there are nicer algorithms for such predictions (this is quite common problem), but I don't find anything.

Do you have ideas?

Thanks a lot

PS: I'm working with Python and cython

Answer 1

First of all, you will not be able to prove anything with a model, you will have false positives/negatives. With a good model you may be able show what variables may be an indicator of component failure.

In problems like this feature generation can have the most important influence on the accuracy of the model. The time stamps can be used for aggregation. For example, you may aggregate metrics per device per hour. The metrics/features that you might create for input into the model might be average/max temperature or fan speed, rate of change in temperature or fan speed, number of seconds device was above some threshold temperature or fan speed, boolean indicator of voltage spike, etc. There could be any number of features you may create. You can then find which features are not strong predictors and remove these columns to reduce noise, if need be.

Answer 2

You can explore Joint modeling for Longitudinal and time to event data. Here survival model will be built for rare events and the linear models for longitudinal data.