Detect differences between cup anemometer bearings in a wind tunel

[VDuarte]

Hello. I'm doing a thesis in which I had to get audio files from anemometers in a wind tunnel, and, by analysing the .wav file, I was supposed to be able to detect if the bearings of the anemometer are faulty or not. However, from the data recieved, I haven't been able to detect any difference between a anemometer with good bearings (23 zip) and with bad bearings (24 zip). I should note that the wind tunnel is located in a industrial workshop, so there are lots of different uncatalogued noises, which makes this process even harder.

However, I'd like to ask: is my analysis wrong or is it just impossible to detect anything without knowing which noises to filter?

You'll probably have to change the string of the file location so that it opens the correct file (only 4m/s speed is available on the files I sent).

[CoastalMaineBird]

I don't have the S&V toolkit, so I can't tell exactly what you're doing (right or wrong).

I can suggest using the GABOR JTFA (Joint time-Freq Analysis).  method might yield some interesting insights into your data.

A rough analysis shows a clear difference between the good (23) and bad (24) files:

(NOTE: I did not take the time to properly scale the X and Y axes - it shows neither frequency in Hz or time in sec).

A finer analysis shows more detail:

GABOR shows the energy (intensity) vs. frequency (Y axis) vs. time (X axis).

You can see that on the good bearings, most of the sound energy is at the rotational speed, with some at DC.

On the bad bearings, the ratio between rotational speed and low frequency is much lower, and there's noise at intermediate frequencies as well.

I think there is a JTFA toolkit which offers a custom indicator graph, which handles the freq scaling for you.

[CoastalMaineBird]

Oops - the VI I used for this is attached here.

[VDuarte]

Steve,

Thank you for your answer. However I don't have access to the advanced signal processing toolkit, so I guess I can't make that analysis you've made. However, I think I can take the same conclusions from the data I get from the same points:

My problem is that, if I change the rotational speed (the pic above is at 4m/s in the wind tunnel, the pic below is at 5m/s), the graphs change drastically without any apparent relation (mostly to outside noises changing, and it's impossible for me to make all the tests in the same outer noise conditions):

And, if for example I take the 6m/s speed in the wind tunnel, the specters change again:

So it's being really hard (not to say impossible) to take any conclusions from all this data. Should I forget this, considering I can't isolate the noises correctly, and do other types of tests? Or am I missing something?

Thanks again though Steve for your analysis, and correct me if I'm stating something wrong please.

[CoastalMaineBird]

Oops - I didn't realize that Gabor was in the ADVANCED toolkit.

Still, we ought to be able to make something out of these results.

Take a look at this run, made with the same data but the frequency bins setting is down to 16:

What stands out to me is the RATIO between the energy at speed and the energy in the zeroth bin.

The good bearing has a high ratio and the bad bearing has a low ratio.

A Gabor spectrogram is just a sliding FFT, although there are probably mathematics to avoid redundant calculations:

 Imagine a 4000 sample signal.

Choose a block size N of say, 256.  This determines your frequency resolution.

Choose a step size T of say, 16.  This determines your time resolution.

Set an index I to 0.

repeat

    Extract a block of size N starting from index I

    Perform an FFT on that block.  That is the 0th "column" on your intensity graph.

    I = I + T

until done.

Not even sure you need to do the repetition.

Is there any hint from some other source of the actual rotational speed of the anemometer?

It looks like you might take that number, divide it by two and separate the FFT into energy BELOW that line, and energy ABOVE that line. 

If BELOW / ABOVE > 10 then the bearing is bad.

[CoastalMaineBird]

Another thought - use a pair of microphones in the wind tunnel - one (A) close to the anemometer, and another (B) farther away.  Subtract B from A in your analysis.  Both should pick up external noises equally and cancel them out, but the real noise will be stronger in A.

[CoastalMaineBird]

Here is a rough implementation of what I was describing, with NO TOOLKITS.

10:1 is not realistic, but there is a clear difference in the ratio.

If you could estimate the proper frequency, then it could be improved by selecting a few bins around that frequency, instead of my hi/lo scheme.

[VDuarte]

Regarding the usage of 2 microphones, even though I didn't use 2 at the same time (which indeed would allow me to subtract the waves) I've already done the test with the anemometer inside the wind tunnel, but with the cups blocked with duct tape. The problem is: I still can't see any noticeable difference between the stopped, the good and the bad bearings. Even if I can detect any difference at a certain speed, if I change the speed, those frequency spikes appear in the other bearing, which doesn't add up.

These onedrive links contain the data for all speed range for different conditions:

20-Bad Bearing (https://1drv.ms/f/s!AruXvturelaOgd8ytYaLSpX2NwxKYA)

21-Anemometer stopped inside the wind tunnel (https://1drv.ms/f/s!AruXvturelaOgd8zrSwlr4XHFOraVw)

23-Good Bearing (https://1drv.ms/f/s!AruXvturelaOgd80fapHTyTDhEzSgw)

And regarding all this data, I'm not being able to take that conclusion you've made for just the 4m/s data.

[CoastalMaineBird]

I guess I should have actually LISTENED to your audio data first.

You have people's voices loudest, then there's a radio / music player second, then there's the noise you're interested in. 

If you're seriously trying to diagnose a bearing issue, then you have to have sound from the BEARING.

It'll be hard to extract ANYTHING meaningful when a person is talking at 20 dB above the signal you want.

[VDuarte]

Exactly. That was the only data they were able to feed me, as there's no other way to acquire data in there without telling people from other departments to stop what they're doing, which, let's face it, it's pretty unfeasible in an industrial environment.

Guess I'll have to try to gather data from somewhere else for my thesis. Thank you anyway for all the support you gave me, it opened up new ways to me to analyse data (if I ever get valid data to analyse, that is... ahahahah).

Thanks again Steve!