Why do our devices fail?

Question

I would like to figure out why some devices of my company fail. Therefore, I'm able to use a list in which around 300 devices are listed together with about 70 parameters while only half of it is numerical, the others are mostly ok/not ok or quite a bunch of comments. The latter is hard to use for an analysis, I guess? However, the list contains mainly devices which failed within their first year (which means our warranty kicked in ^^).

I'm thinking about how to tackle this task in means of methods. At the moment I'm eyeballing on the data via scatter plots / correlations. I'm aware of various methods but I wonder which makes really sense instead of applying methods I can hardly draw insights from. My next move will be a PCA and/or a EFA, also already wonder, which makes more sense? Does a PCA really reveal information about the most influence (here)? Or instead a SEM? I've never worked with the latter one but it seems like what I'm looking for? However, it seems I have to know the so-called "latent" variable upfront resp. what it could be but this is what I am looking for? Could you evaluate those methods and/or provide some other useful approaches?

Looking at a biplot of your data where the points are coloured by whether they pass or failed is not guaranteed to be insightful, but it is a good start in an exploratory data analysis. — Galen, Mar 15 '22 at 20:50
Thank you both, this is helping! So far, I lack information about devices without failures.. have to check whether I can get some. — Ben, Mar 16 '22 at 06:29
btw, in case such data are not available, do you think I can create artificial devices by creating random parameters from intervals as they should be? — Ben, Mar 16 '22 at 07:56
@Ben, last comment: I don't think this is a good idea, as there is no guarantee whatsoever that this creates "realistic" data. You'd need to have a fairly reliable model for how parameters for non-failing devices should look like, not just intervals. — Christian Hennig, Mar 22 '22 at 08:46
If you don't have data about devices that did not fail to make a comparison between devices that failed and did not fail (yet), then you might still learn something if you have data about the type of failure. For instance you mentioned that you know the time untill failure. Patterns that you find there might help you get some clues about the causes of failure. — Sextus Empiricus, Mar 22 '22 at 09:03

Escherichia · Answer 1 · 2022-11-16T20:27:42.090

1

I'm not an expert, but here's my guess:

It may be useful to do feature extraction from the variables that consist of unstructured text data. Useful features that come to mind include the length of the comments and their sentiment.

It may be useful to come up with several models for your data that can be used to classify a device as failing in the first year or not. I would consider using a random forest as one of these models. Then, looking at these models, you can see which variables were important.

I think that essentially what you are trying to do is causal inference. You want to know the cause of the failures. This is a good resource on causal inference.

Once you have an idea of what may be causing the failures, then you could do a randomized, controlled experiment for verification.

edited Nov 16 '22 at 20:27

answered Mar 15 '22 at 20:35

Escherichia

124

Thank you! I think I'll follow this. Could you also please point out whether it makes sense to apply a PCA before applying a RF? Is a PCA useful in general, here? – Ben Mar 22 '22 at 06:29
1

I'm not an expert, but this answer seems to suggest that using only the top components from a PCA as features in a Random Forest may be beneficial. – Escherichia Mar 22 '22 at 20:26

Why do our devices fail?

1 Answers1