Device failure prediction?

Question

Machine learning

Good afternoon, connoisseurs

Such a situation: I have dump files with device state parameters at different jobs. Each dump file can contain a different amount of data: for a couple of hours, for a day, or for several days. At some jobs, the device failed, that is, I know in which dump file there is data about the failure of the device.

I want to make a predictive model that will report if the probability of failure in the next job is very high (or better, if it tells how long the device will fail).

Can you please tell me where to start such a project? Where can I read about the implementation of such projects?

Answer 1

Today such problems are really successfully solved by machine learning methods.
What you are talking about - a "dump file" - is essentially a time series with the recorded readings of your device (if I understand correctly). Plus, you have information about when the device failed during your measurements. All this together constitutes a "training sample". Your task is to find such patterns in the behavior of the system that anticipate the onset of a breakdown signal.
I don’t know what you specifically do, but I have met with such tasks in detecting aircraft equipment failures, in detecting failures in the operation of nuclear power plant equipment, in detecting the initial stage of a disease, in the task of detecting intrusions in computer networks. Of course, without the competent support of a specialist in the subject area, the problem cannot be solved (or very difficult to solve). His task is to explain, as a first approximation, what parameters can influence the appearance of an alarm, your initial task is to check his "testimony", most likely agree with him, and maybe correct this list (it happens).
After that, your task is to build a classifier. A classifier is an algorithm, defined either functionally, or logically, or programmatically, which will take your futuredump files, and at the output to give a forecast (make a diagnosis) about the approaching event, and in fact - to attribute this data to one of the classes, in the simplest case - to the "normal" or "pre-emergency state" classes.
In machine learning, this approach is called supervised learning. in the beginning you train (build) a classification algorithm, and then you use it for prediction. There are a lot of methods for constructing classifiers, from a trivial kNN to methods based on various neural networks, from time series analysis methods to systems based on fuzzy calculations. And it is impossible to tell you in absentia which of this set of algorithms is more appropriate to apply specifically for your task. Here you will either need to deal with them yourself (I indicated the general direction),
You can read about it anywhere - from popular science articles like
https://habr.com/ru/post/460467/
https://towardsdatascience.com/how-to-implement-ma...
https://www .kdnuggets.com/2017/04/datarpm-predicti...
to more serious work (it makes no sense to bring them here if you reach the level when you need them - I think you will find them for your subject area. If not, write with what I can help).
In any case, the task you set is not easy, but interesting. You will not get anything intelligible quickly and from a swoop, but if you find the strength and interest in yourself to go deeper, the results can be deep and meaningful.
Good luck!.

Answer 2

In general, this is one of the disciplines of any physics.
Go to a website such as VSU and download the curriculum.
Purely practically duplicate systems and arrange elections, for a system with 3 components, consider 2 matching and discard different
https://en.wikipedia.org/wiki/%D0%A2%D0%B5%D0%BE%D...

Answer 3

Q: How often do devices fail?
If device failure is a fairly frequent occurrence, then you can try to solve the classification problem, and, in the future, return the "failure probability".
If failure is rare, then read about "autoencoder for anomaly detection".