Can you expand a little more on how you expect the faults will affect the audio? If a fault will affect the sound of the machine in a predictable manner (e.g. if it will induce a high-intensity sound within a particular frequency range), you should be able to get away with a relatively simple pipeline: extract FFT or MFC (see link below for MFC), use the normalised/scaled coefficients as features, and apply a standard classification model (e.g. support vector machine, or logistic regression) to these. 

If you think the signatures of machine faults are somewhat more sophisticated, it might be worth looking into more sophisticated pipelines. Possibly the easiest thing to do would be to borrow from some relatively recent work in audio classification field using neural networks (see attached document). In particular if you have audio examples of failure and non-failure, this is probably a good approach.

If the latter approach interests you, I can provide links to implementations that would help in the language of your choosing. 

https://en.wikipedia.org/wiki/Mel-frequency_cepstrum

I enclose for you two useful papers that might help.

Prof. S. EL-Rabaie

Thank you for your suggestions!

@ Niail: Do you know of good implementation of convolutional DBN in R, Matlab or Python?

Thanks,

Sumeet

Hi Sumeet, 

Yes. There are a few that I know of.

I will stick with Python implementations right now. Keras is possibly the easiest that you can use (see link below). All options will have tutorials on-site. 

• Keras: https://github.com/fchollet/keras
• Theano: http://deeplearning.net/software/theano/
• Tensor flow: https://github.com/tensorflow/tensorflow
• Pybrain: http://pybrain.org/
• Torch: http://torch.ch/

Note that Keras is a wrapper for both theano/tensor flow so with this you can choose these two backends arbitrarily so long as these are installed on your machine. 

Let me know how you get on. 

Niall

Hi,

I have done some research in this area with my M.Sc. students. My research is divided into two main topics: Condition Monitoring and Music Information Retrieval (MIR).

Usually the phrase acoustic condition monitoring is used for the condition monitoring, but in reviews I received from some reviewers, they were considering ultrasound as for the acoustic signals only! The term audio would make them more satisfied if your are considering using signals within human hearing range. My reported work on the condition monitoring is mainly focused on the automobile engine fault diagnosis.

You may find a number of papers in my contributions to the research gate in my page if you were interested.

Regards

Peyman

Another approach could be to read more about the papers published from DCASE community

http://dcase.community/challenge2018/

http://www.cs.tut.fi/sgn/arg/dcase2017/index

http://www.cs.tut.fi/sgn/arg/dcase2016/index

Thanks Ankit!