1
$\begingroup$

I have three variables measured at a sensor: Temperature (T), Humidity (H), and Methane Concentration (PPM). There are physical reasons why changes in T and H will influence PPM. I am interested in removing changes caused to PPM by changes in T and H. What I would like to see is an expected value of PPM with effects of delta T and H removed. Below is a plot over several days of measured values of T, H, and PPM. Additionally, this is one of many sensors. I need a way of generating a model for each individual sensor as this correlation is specific to component tolerances.

Plot of T, H, PPM over Time

I'm looking for direction on where to start with this. What algorithm would you use? What's the simplest solution to try and get an expected PPM reading that minimizes the effects of delta T and H?

$\endgroup$

1 Answer 1

2
$\begingroup$

Here are 2 interesting algorithms :

  • Multivariate LSTM. LSTM cells are great to find patterns having cycles with 50 to 300 timesteps. Be aware that it is quite sensitive to noise, so apply smoothing algorithms to see if predictions are better.
  • Random Forest. Even if it hasn't a long memory like the LSTM cells, Random Forest is great to find correlations between signals. In some cases, Random Forest have even better results than LSTMs.

In addition to that, you seem to have very precise data and it might be not necessary for prediction tasks. In fact, having too precise data could reduce prediction accuracy, because algorithms have to memorize more data and hence can produce more prediction errors.

Consequently, I recommend reducing the sampling rate as low as possible, without altering the overall data quality. Such a simplification should be "humanly understandable" (= not too precise and not too simplified). You can do this by replacing the values of 10 records with 1 using their mean value. Otherwise, you could get poor predictions and too many calculations.

$\endgroup$
1
  • 1
    $\begingroup$ Thank you! I will be investigating both these options further and also consider doing some smoothing. $\endgroup$
    – coomie
    Aug 3 at 23:08

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.