How to predict the breakdown time / life of the equipment?

Question

Machine learning

Good day.
There is a small set of data on equipment failures, i.e. documented:
1. Date/time of failure
2. Type of failure
3. What equipment has failed (identification of instances by model and serial number);
A total of 130 breakdowns, some equipment appears more than once.
There is also data on about 160 pieces of equipment of the type: model, serial number, year of manufacture, counter of operations performed.
The hypothesis is that the probability of failure increases depending on the meter, and the operating time (obviously).
And here is further already zatyk.
There is little experience in building a model on the Azure platform, but then the classification problem was solved, and here, as I understand it, regression analysis is needed. What is the best way to compose data for a machine? What is the best algorithm to choose.
I want to try to predict for each equipment (model/serial number) its service life or the time of the next breakdown and/or the type of this breakdown based on the "case history" of the equipment, but I can't figure out how best to feed this data to the machine.
It would be nice if you tell me what to read on this topic, maybe someone had a successful experience in solving such a problem.

Answer 1

You need to read the theory of reliability.
As written above.
The breakdown probability can be approximately described by the following formula:
p(t) = a/(b*t) + c*exp(d*t) + e

Answer 2

There is such a section of Data Science, which is very optimistically called "survival analysis". The name is historical, since the chances of patients to survive for a certain period of time are initially analyzed. Or recover - depending on the level of optimism :-).
So, in the general case, this section just builds models aimed, among other things, at identifying the probability of equipment failure, the probability of a certain event occurring, attempts to predict the moment of equipment failure, etc. Well, from the most recent (literally yesterday) and the simplest description of this section of Data Science - you can read here: https://datascienceplus.com/survival-analysis-part-i/
(By the way, here it is explained, for one thing, and why the classical regression model does not work in this case, and how it differs from the survival model).
From the more fundamental - statsoft.ru/home/textbook/modules/stsurvan.html In
general, there is a lot of information on the topic. Key words - "survival analysis".

Answer 3

No, there is also a classification.
Multi-class classification, when there are several types of failure.
Look towards the classifier on neural networks with the output layer softmax activation and softmax regression. Read how the MNIST set is recognized using Feed-forward NN ( MLP ), there is the same task, only as a feature - pixels. https://www.tensorflow.org/get_started/mnist/beginners
3rd and 1st point is a feature. Only feature engineering is needed, because the lines
ufldl.stanford.edu/tutorial/supervised/SoftmaxRegr...