2
$\begingroup$

I am working on a live sensor data set and looking for abnormal patterns (leading to a machine fault condition) from the available data set.

I am learning and new to the world of data science, but comfortable with Python. I have few questions that I am looking to get suggestions.

  1. What kind of algorithms would be best suited for this case?
  2. What are the basic steps for doing a predictive analysis in python?

Please correct me if my questions are not correctly framed.

$\endgroup$
7
  • $\begingroup$ Depends on your dataset's Complexity.. Welcome to the Site.. $\endgroup$ – Aditya Mar 22 '18 at 18:40
  • $\begingroup$ I am dealing with time series data, symptoms of faulty conditions depends on multiple parameters. $\endgroup$ – Vikas Gaikwad Mar 22 '18 at 18:43
  • 2
    $\begingroup$ What will and what won't work, their is no assurance to that point in ML....It's just we say that it should work in this case... Doing an EDA is the first step, checking which components get replaced frequently,are there parts which stops working on same day, time of the week etc.. doing feature engineering (adding new features) and them applying a DL model which suits my understanding is what I would do.. $\endgroup$ – Aditya Mar 22 '18 at 19:02
  • $\begingroup$ This is a really broad question - you are asking for information about a methodology that incorporates most of statistics... $\endgroup$ – Spacedman May 27 '18 at 7:10
  • $\begingroup$ what kind of sensor do you have? $\endgroup$ – Francesco Pegoraro Sep 24 '18 at 14:23
0
$\begingroup$

You are talking about anomaly detection, and there are many approaches. If you can create a training set, one-class SVM is a place to start, but even simple control charts can be useful, particularly with live streaming data.

$\endgroup$
0
$\begingroup$

If you have a dataset with labeled anomalies, then you can use a binary classification approach. If the anomalies are not labeled, then you have to look into outlier/novelty detection. The scikit-learn documenation has a good overview.

$\endgroup$

Not the answer you're looking for? Browse other questions tagged or ask your own question.