Machine learning model to evaluate oscillator sound?

Question

Machine learning

Hey guys,
I'm thinking about the next problem and somehow I'm not sure about the solutions. I'm wondering how others would have developed this system and how correct my approach is.
There is a company that monitors their power generators to determine when the generators need repair based on the sounds they make. The generators are constantly monitored by microphones and the audio data is stored for processing. I need to develop a machine learning model that will determine if a generator needs repair. Three main questions:

Supervised and unsupervised learning? (supervised vs unsupervised)
I think the supervisory approach is better, because we will be loading labeled data into the model - as labeled data, I understand the sounds of generators, two sets (needs maintenance / does not need).

What architecture is ML?
Recurrent/Recursive Neural Networks (RNNs). They are used for sequential tasks such as speech processing, etc.

What preprocessing steps will your pipeline include and why? Pre-processing pipeline
I thought that I need to create lists of data, load them, but no more ideas come to mind...

Answer 1

An interesting task. In different productions - generally solvable. (For me - somewhat personal - back in the late 80s of the last century, I wrote a system in which the presence of cavities-shells, i.e. marriage, in large cast products was determined by sound). Today, acoustic diagnostics is one of the not very fashionable, but really very useful applications of machine learning.
The question about supervised vs unsupervised is formulated not quite correctly. If you have already accumulated a couple of thousand samples of records, and for each of them you know whether it corresponds to the normal or pre-emergency state of the equipment - then there are no questions, we apply the learning method. However, the question remains - which one of them, but here it is necessary to experiment-analyze and investigate. But if such a sound libraryyou don't have labeled samples - then we are thinking about using clustering methods (i.e. unsupervised methods)
About whether it is worth using neural networks here is a debatable question. Keep in mind that modern classification methods (or clustering with subsequent classification) using the same trees on problems like yours often show better accuracy than neural network-based methods.
But in any case, recurrent networks in this case are a very dubious solution. After all, you do not have the task of predicting, identifying anomalies in the time series, and not even the task of dynamic recognition of patterns (sounds) in the signal. Your task is to simply classify multidimensional data represented as a series. And this, as a rule, does not imply a deep correlation dependence within the data. Moreover, (as far as I can remember) your task may be more efficiently solved by moving from the time to the frequency representation of the audio signal, and then in general the data is nothing but a multidimensional static set of feature values ​​(in this case, frequencies).
Good luck with your research. There will be specific questions - ask. How can we help.