I have dataset of following specification:

  • 512 samples taken at unevenly spaced intervals over the year
  • Each sample is an 8 second data from sensors with 4ms resolution
  • Samples are not labeled

For example, I have 5 samples taken on first day, then more then 10 samples take on 5th day and so on.

I want to cluster data to check if I can infer the mode of operation for the machine from single 8 second sample. Also, I want to measure the performance of the component over the year for predictive maintenance.

Currently I want to use self organizing maps for clustering purposes. I am new to this data science and am currently learning. The usual methods use evenly spaced samples. Also each samples in these cases is single input (Like stock value at the end of a day) instead of X second data taken at time Y.

My question is: How do I input such data into any model?


It depends how you want to cluster the data, but here are some options....


    You could, for example, completely ignore the timestamps and just seek to cluster the different modes of operation are based on the magnitude of the feature alone. Here, simply extract the values into a new feature list.

    Otherwise, you have to consider what is important about the samples. For example, does the fact that one snapshot containing 5 samples vs 8 samples distinguish it between modes? If so they, build your feature vector based off counting the number of samples in a day.

    In doing so, you will create single features for each day from various attributes in the day (e.g. magnitude, number of samples), enabling you to cluster them based off these features.


    Conversely, you could resample all the data into uniform intervals, but this would probably need lots of assumptions, and would therefore introduce noise.

  • $\begingroup$ Feature engineering sounds like a good idea. On thinking what is important about different modes. I think its their sequence. So different modes would have slightly different sequence of operation. So I would try reading on sequence clustering. But I am still not sure what to do about performance. Or how to input data for that. Also I am bit skeptical about resampling given the data set. $\endgroup$ – Muhammad Hamza Yousuf Apr 29 at 12:32
  • $\begingroup$ Sounds good. Performance would be looking at how identified modes change over time(?). This may become more clear later in the analysis. Please upvote/accept the answer if you found it helpful $\endgroup$ – WBM Apr 29 at 13:54

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.