I have a high dimensional sensor data and many variables are highly correlated. Is it a good idea to get principal components without removing correlated features to reduce dimensions?

My final objective is predictive maintenance (classification). So, will dimension reduction remove the important information used to predict machine failure?

  • 3
    $\begingroup$ It depends. For instance, I can imagine there is a sensor sitting somewhere that has a flatline readout--no noise at all--but sometimes when the machine is about to fail in a particular way, it gives a strong response even when none of the other sensors notice. That's why the sensor is there! If you were to do PCA on a dataset without an instance of that particular failure, your sensor would be the first variable dropped, because it has zero variability. This thought experiment suggests you need to tell us more about the nature of the machine, the sensors, and the data you will be using. $\endgroup$ – whuber Dec 4 '19 at 19:46
  • $\begingroup$ @whuber Thank you for your insights. The sensor data are collected every 10 mins time intervals but they have many missing values. So, I aggregated the daily data and computed mean, median, min, max and SD of all sensor variables which resulted into many highly correlated variables. I am new at this so please pardon my mistakes. Thank you. $\endgroup$ – Arch Desai Dec 5 '19 at 19:05

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.