I have an electronic component whose sensors record temperature, current and voltage values of various sub-elements. These readings are taken at regular intervals of time and I organized them as records of a dataset. In addition to these features, the dataset has a column with value 1 or 0 if at that moment the component is experiencing a malfunction or not. Here is an example of the dataset structure:

enter image description here

My goal is to build a model that is able to predict the occurrence of a malfunction with some time advance, so I think this is a multivariate time series prediction problem.

What approaches could I use to achieve this goal? I have read that it is possible to use Recurrent Neural Networks / LSTM but I do not have an adequate GPU.


1 Answer 1


I recently trained some multivariate LSTM regressions on a relatively slow, CPU-only laptop, and it worked. Training time didn't take too much, it's definitely feasible.

About your problem, I would definitely use some RNN architecture. LSTM are the most powerful RNN cells up to now, they have more memory and are able to learn longer sequences than GRUs, but they are slower than them. Choose based on you preference/circumstances.

The final layers must be Dense(), and the final output layer should have two nodes with softmax activation, to perform the binary classification. The most appropriate loss would then be some crossentropy measure.

  • $\begingroup$ Thank you! I'll set up an RNN then. If I also had other components (all structurally the same as this), each one would give me a different time series but all would refer to the same time period, how should I organize these datasets to train my network if I want it to recognize a malfunction occuring in any of them? $\endgroup$
    – Ingen 77
    Commented Aug 9, 2019 at 12:53

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.