I have reviewed a paper that uses image processing and computer vision techniques to achieve similar things as you are trying. Basically they transform the vibration signal into an image by cutting "chunks" of it and putting each one in a row and then using LBP or SIFT features and a classifier to obtain if the engine is fault and what type of fault it is. It worked reasonably well, so you might want to explore this kind of techniques.

The title of the article is:

Fault diagnosis of induction motors utilizing local binary pattern-based texture analysis

I have reviewed a paper that uses image processing and computer vision techniques to achieve similar things as you are trying. Basically they transform the vibration signal into an image by cutting "chunks" of it and putting each one in a row and then using LBP or SIFT features and a classifier to obtain if the engine is fault and what type of fault it is. It worked reasonably well, so you might want to explore this kind of techniques.

The title of the article is:

Fault diagnosis of induction motors utilizing local binary pattern-based texture analysis

you may try to use wavelet transform to identify the spurious peaks that existed in the vibration signal. It will be good indicator to detect the faults in the noisy signal.

Miuel, I read the paper, it seems interesting, but they did not mention how they evaluate healthy or faulty condition, with which indicator, I mean before training the system to detect the type of fault we should find out a way to detect fault and healthy state?

and do you know how to use LBPs in Matlab ?

I have no idea about them

Prakash, there are a lot o papers on wavelet and it is not easy to find something new in wavelet because a lot of researchers have one a lot on ot, Do ou have experience in signal analysis too? I have some questions

yes. I am also working on signal processing and wavelet transform area only. Do you have any specific question in signal processing? We can share our thoughts..

I was not suggesting that you use LBP specifically, but the computer vision techniques to do signal processing. If you are more comfortable with SIFT or SURF, they might work as well.

You might find LBP articles and codes (Matlab, C) in the webpage of the university of Oulu:

http://www.ee.oulu.fi/~miguelbl/LBP-Software/index.html

And some specific code optimized for some platforms in my webpage as well:

http://www.ee.oulu.fi/~miguelbl/LBP-Software/index.html

RE:How to detect fault of gear by vibration signal?

I don't know the answer of the question but I suggest you to ask [email protected]

Best wishes,

Dejan Momcilovic

Hello,

I propose for you to characterize the default as an irregularity in the signal, thereafter, you use an algebraic approach to detect all irregularities presented in your signal. The algebraic approach showed good performances in terms of reducing noise as well as highlighting irregularities

for more informations, you can check my publication " Algebraic Approach for R-peak Detection in the ElectroCardioGram (ECG) Signal"

Good luck

Nehla

We can calculate the frequency of the meshing gears. You analyze your time signal through FFT and you have the frequencies spectrum, in which you can identify the meshing-gears frequency. Define it on a button of your computer and each time you push that button you will have the new height and the history of changing the height of that frequency. Your problem is solved.

If you know the meshing frequencies and the potential defect frequencies of your components (bearings) you can use band pas filters to extract the interesting frequency ranges and then you can calculate a trend value of that frequency range. These trend values will show you any changes which occur in those frequency ranges.

If you analyse in time domain, you can develop a cross-correlation of time history, acquired regularly

In my research on this matter I have found the answers in Mobius Institute.

It seems that most people try to detect the defaults of gears and bearings by using several methods in the theoretical domain. If you come to reality and you have to solve the problem through real measurments, then the wavelet and the similar is useless. In such a case the following things matter: the machine with the faults, the datalogger with its software, and you. So, try to get a nice datalogger and its software and be trained in measuring and detecting the faults of machines. In my ResearchGate section you may find the paper Maintenance of Mechanical Drivers and Fault Diagnosis of Machine Elements, which will give you help.Good luck to you.