2
$\begingroup$

I am working on predictive maintenance and get temperature data from assets. In few months or few days asset remains down and we do not get temperature value. In this scenario i cannot fill data with missing value techniques. Also cannot give some number because even 0 and -1 are valid values for temperature. How to deal with such data?

I am thinking of putting very big value for such columns which is not possible as temperature. Please suggest.

$\endgroup$
  • 1
    $\begingroup$ Why can't you use traditional imputation techniques like filling with mean or forward or backfill? Also; what are you intending to predict? Is this going to be a time series forecasting project or something? $\endgroup$ – Dan Scally Aug 19 '19 at 7:06
  • $\begingroup$ I am trying to predict the failure of asset 2 weeks in advance. These are the values when asset was not running at all, asset was down. $\endgroup$ – Neeraj Sharma Aug 19 '19 at 7:12
1
$\begingroup$

From your question it seems that you are easily able to identify the periods assets were down for maintenance and simply wonder how to best code that in the data?

If that is the case I would simply add a new column of information marking assets as either 1 - active or 0 - down for each period. If need be you could even code it into active/down/maintenance, etc.

This would be additional valuable data for your prediction model (maybe assets are more or less likely to go down after a maintenance period). Additional it would allow you to simply mark the sensor data as missing NA because you will be able to differentiate between a broken sensor and an asset on maintenance from your other data.

As @Tasty213 suggested any good classification algorithm might have been able to do this as well but it seems you are simply able to add this data yourself.

| improve this answer | |
$\endgroup$
0
$\begingroup$

I would advise that you first of all have a good look at the data you currently have available then see what it looks like with the various standard data imputation methods.

Secondly is temperature the only feature? If it is, you will almost certainly need more features in order to get a good model.

Finally there are some algorithms for which having legitimate temperature values shouldn't throw it off. You could perform classification into 'no maintenance' and 'maintenance', using a KNN classifier. All the -1 temperature values which occur only when the units are disabled would be clumped to together easily identifying the special case.

| improve this answer | |
$\endgroup$
0
$\begingroup$

Welcome to the site! I would start answering you by presenting you with another question: Is this a problem?

Most data scientist that do what you do collect their data via sensor readings and are usually swimming in data, tons of it is available. So, is the two-week gap in your data really all that significant? Something tells me that might still be OK to model even if you were to remove this time period completely.

So, have you verified that you actually have a problem on your hands?

| improve this answer | |
$\endgroup$
  • $\begingroup$ yes my data comes from sensor. So i have two ways to get missing values. One sensors goes bad. Secondly sensors not active as the whole unit is shut down. $\endgroup$ – Neeraj Sharma Aug 20 '19 at 7:25
  • $\begingroup$ @NeerajSharma I was making the point that you may not need to do this. Exactly why do you need to fill in the data? If you're doing this by sensor then you most likely have a TON of data and enough for modeling. If I was working on a sensor-based project and I was stuck because of a small time gap then I would revisit my entire approach to the project. $\endgroup$ – I_Play_With_Data Aug 20 '19 at 11:49
  • $\begingroup$ ok will try to not fill and see ... thank you $\endgroup$ – Neeraj Sharma Aug 22 '19 at 5:41

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.