0
$\begingroup$

I have a machine that moves with one axis in the same direction (basic position A to end position B). While driving, the torque is measured and recorded every 10 milliseconds. This looks something like this:

Time [t] | Torque [Nm]
0 | 2.5
10 | 4.0
20 | 4.7
30 | 5.6
40 | 6.0
50 | 5.5
60 | 5.1
70 | 3.2
80 | 2.1
90 | 1.5
100 | 0.2

The torque is therefore almost identical for a new machine. On a "very old" machine, the situation is different: the torque is constantly increased (signs of axis wear), or there are some outliers in the record (e.g., torque greatly increased at 50 milliseconds compared to a new machine).

What I need now is a prediction for wear or failure of the axis!

What I have:

Very much data from an axis movement coming from a new machine (the data should therefore be considered "GOOD").

What I do not have:

Data of an axis movement of an old machine (defective axis, increased torque, axis wear, etc.)

My basic idea is:

I want to use Machine Learning to train the data to the "GOOD" state. With the help of machine learning, a prediction (probability of a defect or wear) or a classification (GOOD or BAD) should then be carried out.

My approach:

Regression or classification (supervised learning)

My question is:

Does this approach make sense, or would you choose a completely different approach? Is it even possible what I intend to do?

Thanks in advance

$\endgroup$
4
  • $\begingroup$ Hi @Toberto! I wrote a research paper about Predictive Maintenance, that will definitely help you. Especially the part about feature engineering. It's in review, so when I'll be able to share it I go back to you. :) $\endgroup$
    – wind
    Jan 18, 2019 at 9:33
  • $\begingroup$ Hi @wind, that sounds very interesting - I look forward to hearing from you. Could you already tell me in advance if my project is possible and the approach is the right one? $\endgroup$
    – Toberto
    Jan 18, 2019 at 9:38
  • $\begingroup$ It depends on several aspects, like amount of your data, contamination of the data, frequency of errors, etc. You rather need to have some examples of the error state. Here is my presentation about approach to prototyping of such system and some challenges you can meet in your project. youtu.be/wZaTrkIKtE0 I reccommend to start from it. All in all- your direction, in general, make sense, but the success depends on more aspects than you wrote about. $\endgroup$
    – wind
    Jan 18, 2019 at 9:50
  • $\begingroup$ So basically you have a good class upon which you want to train a model, and then make predictions on new data to detect abnormalities? One-class SVM would be one option. $\endgroup$ Jan 18, 2019 at 10:14

0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.