I am trying to figure out the right fast fourier transform size.

In summary, the conversation is about using a Tascam recorder and Virtins Multi Instrument Software to record and analyze an environmental nuisance noise in the home. The noise source is suspected to be from machinery at a legal marijuana grow op, including heat pumps, exhaust fans, and other equipment. The issue is that when using different FFT settings, the results show varying levels of incoherent and tonal noise, including infrasonic frequencies. The question is raised on how to determine the most accurate FFT size for analyzing the noise. It is also mentioned that the noise may be caused by temperature drift and the presence of 60 Hz (or 50 Hz) frequency due to supply voltage.
  • #1
whariwharangi
12
0
I am using a Tascam recorder to record an environmental nuisance noise that is occurring in my home. I then use Virtins Multi Instrument Software, which includes an oscilloscope, band pass filter, and a spectrum analyser.

Noise source is probably machinery at a legal marijuana grow op. That would include heat pump, exhaust fans, smell filtration, and grow equipment. (I don't really know)

The problem is:

I use fft settings at 262144 and the spectrum analyser shows a lot of incoherent noise from 90 to 120 hz.

I then use fft settings at 4194304 which gives results showing strong noise frequency in 4 - 10 hz range while the incoherent stuff from 90 to 120 hz disappears. There are still strong tonal bands at 18 24 60 and 120 hz (among others)

I've been looking at data using the high fft setting in the belief that would be the more accurate. However, I'm driving myself bats looking for sources of infrasonic noise at 5, 7 9 hz etc and not finding noise at those frequencies outside. I'm not sure if its due to structure borne vibration caused by outdoors noise generating more noise inside.

It gets worse. If I use measurement microphones (Behringer ECM8000), instead of the built in recorder microphones, I get infrasonic noise frequencies and other tonal frequencies in the spectrum analysis for indoor stuff regardless of the fft setting. I only get the tonal bands from 18 hz and up outdoors. Where is the infrasonic noise coming from?

So the question is ... how do I determine what fft size to use to get accurate information about the noise I am analyzing?

I'd add slides but they are 2.25Mb and apparently too large for upload.
 
Last edited:
Mathematics news on Phys.org
  • #2
whariwharangi said:
I am using a Tascam recorder to record an environmental nuisance noise that is occurring in my home. I then use Virtins Multi Instrument Software, which includes an oscilloscope, band pass filter, and a spectrum analyser.

Noise source is probably machinery at a legal marijuana grow op. That would include heat pump, exhaust fans, smell filtration, and grow equipment. (I don't really know)

The problem is:

I use fft settings at 262144 and the spectrum analyser shows a lot of incoherent noise from 90 to 120 hz.

I then use fft settings at 4194304 which gives results showing strong noise frequency in 4 - 10 hz range while the incoherent stuff from 90 to 120 hz disappears. There are still strong tonal bands at 18 24 60 and 120 hz (among others)

I've been looking at data using the high fft setting in the belief that would be the more accurate. However, I'm driving myself bats looking for sources of infrasonic noise at 5, 7 9 hz etc and not finding noise at those frequencies outside. I'm not sure if its due to structure borne vibration caused by outdoors noise generating more noise inside.

It gets worse. If I use measurement microphones (Behringer ECM8000), instead of the built in recorder microphones, I get infrasonic noise frequencies and other tonal frequencies in the spectrum analysis for indoor stuff regardless of the fft setting. I only get the tonal bands from 18 hz and up outdoors. Where is the infrasonic noise coming from?

So the question is ... how do I determine what fft size to use to get accurate information about the noise I am analyzing?

I'd add slides but they are 2.25Mb and apparently too large for upload.

Hi whariwharangi! Welcome to MHB! ;)

Perhaps you can upload your slides to Google Drive (comes with gmail account), Dropbox, or Imgur?
They are all free of charge.

The only "setting" for an FFT that makes sense, is the number of sample points that always has to be a power of 2.
Indeed, your numbers are powers of 2.
In principle a higher setting should always give better results. Certainly nothing should "disappear".
That is, except if you're looking at artifacts caused by the Nyquist frequency.
If that is the case, there should be an identical spike symmetrically at the other end of the spectrum.
Is there?
Usually a low pass filter is applied first to get rid of Nyquist artifacts.

As for infrasonic noise, that could be cause by temperature drift.
When electronic equipments heats up, it's not uncommon to see low frequency noise.
The first measurements on a new day could show less low frequencies.

Furthermore, there should be a narrow peak at 60 Hz (or 50 Hz depending on where you live) due to the supply voltage. This frequency permeates everything. And there may be secondary peaks at multiples of that frequency, say at 120 Hz.

Anyway, if you hear environmental noise, consider that only frequencies of 20 Hz up to 20000 Hz are audible.
It seems to me that it makes sense to check first between those frequencies.
 
  • #3
I like Serena said:
Hi whariwharangi! Welcome to MHB! ;)

Perhaps you can upload your slides to Google Drive (comes with gmail account), Dropbox, or Imgur?
They are all free of charge.

The only "setting" for an FFT that makes sense, is the number of sample points that always has to be a power of 2.
Indeed, your numbers are powers of 2.
In principle a higher setting should always give better results. Certainly nothing should "disappear".
That is, except if you're looking at artifacts caused by the Nyquist frequency.
If that is the case, there should be an identical spike symmetrically at the other end of the spectrum.
Is there?
Usually a low pass filter is applied first to get rid of Nyquist artifacts.

As for infrasonic noise, that could be cause by temperature drift.
When electronic equipments heats up, it's not uncommon to see low frequency noise.
The first measurements on a new day could show less low frequencies.

Furthermore, there should be a narrow peak at 60 Hz (or 50 Hz depending on where you live) due to the supply voltage. This frequency permeates everything. And there may be secondary peaks at multiples of that frequency, say at 120 Hz.

Anyway, if you hear environmental noise, consider that only frequencies of 20 Hz up to 20000 Hz are audible.
It seems to me that it makes sense to check first between those frequencies.

I am not hearing the noise; I feel it as if there is vibration. If I do hear anything ... rumbling in the walls ... I go outside and its silence (unless I use earphones with the recording equipment)

I actually did identify part of the problem was from a blower on a furnace. Basically it turned his air system into a pipe organ blowing at 18hz. You couldn't hear it standing behind his house but the recording showed clearly when it was on and off.

Here is Scotland document on Low Frequency Noise as in introduction:
http://www.gov.scot/resource/doc/158512/0042973.pdf

Thanks for the information about google. I will try to upload.
 
  • #5
Slides are here:

https://drive.google.com/folderview?id=0B0yOczVuy47tcTRRTHBUT0daUzA&usp=sharing

https://lh4.googleusercontent.com/oawQXOZ0YyeDr0qGQLGB8IT5uOX_Ho_maNowqmxlFjEUJQhgZKNI79zf9c94_0VtwrVAxzhJyx5oa24=w1549-h991-rw

https://lh4.googleusercontent.com/TsYgAqsAD2Mrvd6W9ZED612wpDfVR4AaLEv4TJi75sZzl6UKOwBtTfqSrt2W2_0ZXN9snD3qNBtOZc4=w1549-h991-rw

https://lh3.googleusercontent.com/_LC4UrpBEtnVGeF5yFDTjXfjcLoHhr-r6TFAUsutYCNodgbf1hEZSkQMi8Jc8d9DkEtcDnV7jkVocXI=w1549-h991-rw
 
Last edited by a moderator:
  • #6
whariwharangi said:

Ah okay. The frequencies between 100-120 Hz do not disappear.
It's just the vertical scale that effectively gets automatically adjusted.
Due to the higher resolution, the peaks at 60 Hz, 20 Hz, and 120 Hz become more pronounced.
These are really very narrow band peaks, and the vertical scale gets rescaled to fit the 60 Hz peak.
Although it's slightly off, I presume that the 60 Hz peak is the frequency of the supply voltage in your neighborhood, and the 120 Hz peak might be its first harmonic.

Can you change the vertical scale, so we can see other frequencies again?

Btw, frequencies near 0 Hz would tend to drown in the "pink" or 1/f noise.
Since the general shape looks indeed like an 1/f curve, I'm assuming that we're just seeing pink noise.
 
  • #8
I like Serena said:
Ah okay. The frequencies between 100-120 Hz do not disappear.
It's just the vertical scale that effectively gets automatically adjusted.
Due to the higher resolution, the peaks at 60 Hz, 20 Hz, and 120 Hz become more pronounced.
These are really very narrow band peaks, and the vertical scale gets rescaled to fit the 60 Hz peak.
Although it's slightly off, I presume that the 60 Hz peak is the frequency of the supply voltage in your neighborhood, and the 120 Hz peak might be its first harmonic.

Can you change the vertical scale, so we can see other frequencies again?

Btw, frequencies near 0 Hz would tend to drown in the "pink" or 1/f noise.
Since the general shape looks indeed like an 1/f curve, I'm assuming that we're just seeing pink noise.

Yes 60 Hz is supply voltage frequency. I get 120 Hz if I record noise from transformers. The recording data shown here was taken out of doors about 30 meters from any building on a school field using a battery powered recorder. Most of the audible sound from the building is drowned out by a 1/5 hp heat pump. (heat pump is propped up on a 2x4 to give some idea of standard of installation being used)

I'm finding emf from wiring makes no difference house power on or power off for indoor recordings even with phantom powered microphones.

It does affect recordings made when I connect geophones to the recorder with strong induced 60 Hz present that is not present when house power is off. Geophones are showing there is vibration present and at frequencies mirroring noise recorded outside and inside but I haven't figured out if the vibration is powerful enough to be felt or if I am just 'feeling' the noise.

I am uploading link to data using y scale x 5 and full screen.
 
Last edited:
  • #9
What we see is:
\begin{array}{c|c}
\text{Resolution }\Delta f & 100-120 \text{ Hz amplitude} \\
\hline
0.34\text{ Hz} & 0.3 \\
0.17 & 0.15 \\
0.01 & 0.01
\end{array}

This is expected.
If we look at half the band width, the corresponding amplitude is half as high.
That is, except at the points where we have a spike, in which case the amplitude remains the same, and the spike becomes more pronounced.

At the highest resolution, the signal between 100 and 120 Hz apparently drowns in white noise, although we can still see a couple of isolated spikes.
That white noise might be caused by interference with the sampling frequency.
What is the sampling frequency?
If we do an FFT with a frequency resolution that is close to or lower than the sampling frequency, we can expect some artifacts.
 
  • #10
I like Serena said:
What we see is:
\begin{array}{c|c}
\text{Resolution }\Delta f & 100-120 \text{ Hz amplitude} \\
\hline
0.34\text{ Hz} & 0.3 \\
0.17 & 0.15 \\
0.01 & 0.01
\end{array}

This is expected.
If we look at half the band width, the corresponding amplitude is half as high.
That is, except at the points where we have a spike, in which case the amplitude remains the same, and the spike becomes more pronounced.

At the highest resolution, the signal between 100 and 120 Hz apparently drowns in white noise, although we can still see a couple of isolated spikes.
That white noise might be caused by interference with the sampling frequency.
What is the sampling frequency?
If we do an FFT with a frequency resolution that is close to or lower than the sampling frequency, we can expect some artifacts.

The sampling frequency of the recorder is 44.1k, format wav 16 bit.

Apparently I can increase the sample rate to 96k.

The sample rate on the Spectrum Analyser/Oscilloscope is determined by wav file parameters; no matter what settings I make, it goes to the wav file numbers when I open the file.

The recordings are approximately 1 minute long (often estimated in the dark ... and therefore usually a bit longer). I have to be careful of wind and intrusions of noise from traffic.

When I am trying to make direction finding calculations some frequencies vary a lot in amplitude or are incoherent (microphone axis crossing time differences changes during the duration), possibly due to change in load, reflection, and reverberation.

I have connected a microphone directly to the computer to observe real time analysis and note there is a lot of 'wander' in some of the frequencies.

The observed information is the sound from heat pump seems to vary a lot, particularly on starting, suggesting the electronics are using variable speed resolution to get maximum efficiency. (Which I thought was causing the 100 - 120 hz information at the lower fft sizes)
 
Last edited:
  • #11
whariwharangi said:
The sampling frequency of the recorder is 44.1k, format wav 16 bit.

Apparently I can increase the sample rate to 96k.

The sample rate on the Spectrum Analyser/Oscilloscope is 48k. I can set it at various levels to 200k ... though the higher sample rates would probably make the computer crash.

The recordings are approximately 1 minute long (often estimated in the dark ... and therefore usually a bit longer). I have to be careful of wind and intrusions of noise from traffic.

When I am trying to make direction finding calculations some frequencies vary a lot in amplitude or are incoherent (microphone axis crossing time differences changes during the duration), possibly due to change in load, reflection, and reverberation.

I have connected a microphone directly to the computer to observe real time analysis and note there is a lot of 'wander' in some of the frequencies.

The observed information is the sound from heat pump seems to vary a lot, particularly on starting, suggesting the electronics are using variable speed resolution to get maximum efficiency. (Which I thought was causing the 100 - 120 hz information at the lower fft sizes)

Your highest resolution FFT has 4194304 samples (or frequencies) over a period of about 74 seconds.
That corresponds to 4194304/74 = 57 kHz... which is uncomfortably close to 44.1 Hz...

So indeed, it makes sense to sample at a higher frequency, or measure for a longer period of time, or use a lower FFT number.

Btw, is a low pass filter included in your setup?
To get rid of Nyquist artifacts, the signal, before sampling, when it is still analogue, should pass through a low pass filter of 44.1/2 = 22 kHz (or whatever half the sampling frequency is).
Nyquist artifacts can show up as false low frequency spikes that are really high frequency spikes.
 
  • #12
I like Serena said:
Your highest resolution FFT has 4194304 samples/frequencies over a period of about 74 seconds.
That corresponds to 4194304/74 = 57 kHz... which is uncomfortably close to 44.1 Hz...

So indeed, it makes sense to sample at a higher frequency, or measure for a longer period of time, or use a lower FFT number.

Btw, is a low pass filter included in your setup?
To get rid of Nyquist artifacts, the signal, before sampling, when it is still analogue, should pass through a low pass filter of 44.1/2 = 22 kHz (or whatever half the sampling frequency is).
Nyquist artifacts can show up as false low frequency spikes that are really high frequency spikes.

I gather from this that the 262144 fft size is more appropriate than the 4194304 fft size. Yikes.

Yes there is low pass filter included in the software.

I don't have an analog filter that would be placed between microphone and recorder.

I have been using the software band pass filter to isolate frequencies for direction finding.
 
Last edited:
  • #13
whariwharangi said:
I gather from this that the 262144 fft size is more appropriate than the 4194304 fft size. Yikes.

Yes there is low pass filter included in the software.

I don't have an analog filter that would be placed between microphone and recorder.

I have been using the software band pass filter to isolate frequencies for direction finding.

Software can only do it if it controls the analogue-to-digital conversion.
I'm afraid that once the signal has been sampled (digitized), it's too late.
Ideally the recorder applies a low pass filter before sampling - maybe it does?
 
  • #14
I like Serena said:
Software can only do it if it controls the analogue-to-digital conversion.
I'm afraid that once the signal has been sampled (digitized), it's too late.
Ideally the recorder applies a low pass filter before sampling - maybe it does?

The manual for the recorder makes no statement about the frequency range for which it will record, only that it is designed to record music and in particular bass and drums.

There is a high pass filter with three settings ... which would block the noise I am trying to record. But there is no mention of a low pass filter.

Otherwise ... there should be a cut off frequency for the microphones.

The measurement microphones provide a data sheet showing flat response from 20 hz to 20 khz. They also state that there is flat response to 15 Hz ... and if the spectrum analysis is right its showing noise in the mHz range (albeit you have to take the data with a grain of salt) There is no mention of high frequency cut off though.

The problem has been to explain why I am getting low frequency signals indoors that I am not recording outdoors. It might be the noise is attenuated and distorted by walls leaving only LFN. It could be the walls are making noise in response to the structure borne vibration. Now it appears there may be issues with the process of fft sizing not being the best fit.

I'm not sure what to do now. I've got about a thousand recordings (which includes noise mapping and learning about the best way to get direction). The issue of fade out seems to be with indoors recordings and not so much with outdoor recordings. That being said ... the data provided on this thread is an outdoor recording.

I have proceeded with a letter to the grow operator along with some data that I have gathered showing there is noise in my home generated by his equipment. It seems likely that I will need to go to civil court to get the noise to stop ... so I do need to get ducks in a row. I guess there is reasonable certainty even with the fft issue ... and I guess there is no way to get absolute certainty without grow operator cooperation.
 
Last edited:
  • #15
whariwharangi said:
The manual for the recorder makes no statement about the frequency range for which it will record, only that it is designed to record music and in particular bass and drums.

There is a high pass filter with three settings ... which would block the noise I am trying to record. But there is no mention of a low pass filter.

Otherwise ... there should be a cut off frequency for the microphones.

The measurement microphones provide a data sheet showing flat response from 20 hz to 20 khz. They also state that there is flat response to 15 Hz ... and if the spectrum analysis is right its showing noise in the mHz range (albeit you have to take the data with a grain of salt) There is no mention of high frequency cut off though.

The problem has been to explain why I am getting low frequency signals indoors that I am not recording outdoors. It might be the noise is attenuated and distorted by walls leaving only LFN. It could be the walls are making noise in response to the structure borne vibration. Now it appears there may be issues with the process of fft sizing not being the best fit.

I'm not sure what to do now. I've got about a thousand recordings (which includes noise mapping and learning about the best way to get direction). The issue of fade out seems to be with indoors recordings and not so much with outdoor recordings. That being said ... the data provided on this thread is an outdoor recording.

It should be possible to verify.
If we inspect the high frequency spectrum, and if it's a "mirror" of the low frequency spectrum, then I think we're looking at Nyquist artifacts.
If we don't, then presumably someone has taken measures to apply a low pass filter.

Btw, I've also noticed that your signal had significant amplitudes in the first couple of seconds that did not show up atterwards. That sounds like some kind of pollution of the signal...
 
  • #16
I like Serena said:
It should be possible to verify.
If we inspect the high frequency spectrum, and if it's a "mirror" of the low frequency spectrum, then I think we're looking at Niquist artifacts.
If we don't, then presumably someone has taken measures to apply a low pass filter.

Btw, I've also noticed that your signal had significant amplitudes in the first couple of seconds that did not show up atterwards. That sounds like some kind of pollution of the signal...

I make a statement at the start of the recording. It helps keep track of where the recording was taken, orientation of microphones etc. Thats probably the 'pollution' you are observing.

There isn't usually much above 200 Hz. If I use the lower fft settings the harmonics show up but even that isn't much above 1000 Hz.

There is sometimes a small spike at about 21 kHz ... not sure what is causing that. Perhaps someone has an animal chaser.
 
Last edited:
  • #17
Here is data for recording made with 150 seconds. fft size 131072 and 4194304. I also put in 1khz for the 131072 fft size to show the harmonics. Nothing significant shows above 1 kHz.
https://drive.google.com/open?id=0B0yOczVuy47tQWpoM3pLSENHNU0

It doesn't look like length of time of recording is making any difference. This one is indoors with windows open so some reverberation might be present.
 

FAQ: I am trying to figure out the right fast fourier transform size.

1. What is the Fast Fourier Transform (FFT)?

The Fast Fourier Transform (FFT) is an algorithm used to rapidly compute the discrete Fourier transform (DFT) of a sequence or signal. It takes a time-domain signal and converts it into its frequency components, allowing for analysis and manipulation of the signal in the frequency domain.

2. How do I determine the right FFT size for my signal?

The right FFT size for a signal depends on several factors, such as the sampling rate, desired frequency resolution, and length of the signal. Generally, the FFT size should be a power of 2 and should be large enough to capture the frequency content of the signal without excessive oversampling.

3. What happens if I choose the wrong FFT size?

Choosing the wrong FFT size can result in inaccurate frequency analysis of the signal. If the FFT size is too small, important frequency components may be missed, leading to a loss of information. If the FFT size is too large, it can result in oversampling and longer processing times without any significant improvement in frequency resolution.

4. Can I use an FFT size that is not a power of 2?

While it is recommended to use an FFT size that is a power of 2, it is possible to use other sizes with some FFT algorithms. However, this may result in slower processing times and may not provide the same level of accuracy as a power of 2 FFT size.

5. Are there any other factors to consider when choosing an FFT size?

In addition to the factors mentioned earlier, it is also important to consider the windowing function used on the signal. Some windowing functions may require a larger FFT size to accurately capture the frequency content of the signal. It is also important to consider the trade-off between frequency resolution and processing time when choosing the FFT size.

Similar threads

Replies
2
Views
1K
Replies
13
Views
2K
Replies
4
Views
2K
Replies
1
Views
2K
Replies
19
Views
4K
Replies
1
Views
6K
Back
Top