I have data over a single, a machine includes different components, all the parts are interacting, the data are tracked for those parts, it tracks power consumption and many other relevant feature measurements. The longevity of the machine, from the seller, is specified to be around 2000-15000 hours. Due to this large range, it would be ideal to constrain the target for machine failure, with machine learning. Because we have tracked all the data, with the different components in the system. One specific component fails. Over a period of 9 months, the machine fails 5 times. Monthly average about 3-4 month. It is this, specific target we want to predict when the chances for failure are high.

Thus the specific instances for failure are limited. How do we deal with this, from a Machine Learning perspective, as opposed to giving a monthly average, it is not a safe bet? How do we train on limited instances for failure? This problem might not have one right answer, but many possible advice.


1 Answer 1


When you lack data you can use some theoretical knowledge on what should happen before the failure. For example for bearings a vibration in the frequency of the bearing should rise exponentially. The same goes for temperature, but there are too many variables that influence temperatures and it is not that easy to separate it. There are temperature sensors in some bearings. That should be more precise. Microphones can hear the problems like humans do. Ultrasound tools allow humans to hear the problems earlier.

You can read about some of these patterns in An introduction to predictive maintenance by R. Keith Mobley


Your Answer

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

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