BMW M3/M4 s55 Engine RPM Redline Projections via Technical Sound Analysis

[hyperboost]

I bet BMW will make it 7500rpm

[bradleyland]

Originally Posted by swamp2

Completely agree.

Although the low pass filtering, fourier techniques and stereo vs. single channel stuff are all completely irrelevant, they just don't affect the answer. My error was recording through speakers rather than internally. Even trying to repeat my process showed considerable error. After switching to recording with Soundflower the shift burble was much more recognizeable visually in the traces. I also got much more recognizable peaks in the spectral results. I agree the observed frequency is right around 330 Hz perhaps with a worst case range of about 325-335 Hz. Thus with a Doppler correction from 90-97 mph this provides an rpm range of about 7250-7550 (with rounding to 50 rpm).

This is how science works - basic peer review. Thanks for keeping me on my toes. I will revise my OP in a way to not hide my shortcoming and I'll give you the credit due.

Thanks!

Yeah, after I went through all those extra steps, I started questioning my result, because I was only getting around 330 Hz, and I was worried that by trying to distill the result, I had actually corrupted the data, haha Even without the extra pokes and prods, the spectrum plot is remarkably useful.

[MCaro]

Originally Posted by bradleyland

I had similar questions, so I did some exploration of my own.

Why did you choose a hanning window?

I think he probably chose Hanning window because of the dynamic range of the audio data. Although I looked at rectangular as well, and the plot looked pretty good after I applied a lowpass filter.

Did you try a spectrogram to get better time/frequency resolution?

Spectrographs of mechanical sounds are really difficult to interpret. I've attached a screenshot of a short section if you're interested in looking at it. This is the region just before the upshift.

Have you tried running the audio through an adaptive FIR filter to remove white noise?

I'm not really sure what an adaptive FIR filter is. From what I gather, it would required input coefficients that would allow me to narrow the frequency range. I just used a lowpass filter as you can see in my full write-up.

==============

My analysis shows a result that's closer to 7400 RPM. If you download my files, you can open the .aup files with Audacity (it's free). Running spectrum analysis on the audio just before the upshift always gives me values closer to 330 Hz, not 350-360 Hz. Not sure why exactly.

Full write-up: http://upload.bradlanders.com/bimmer...ound-analysis/

Copy/paste of my Markdown source (links won't work, so you'll have to click here for the full version):



EDIT: Removed the spectrogram attachment, because it makes the thread so wide. Just download the Audacity files and have a look for yourself.

Great response, I'm going to work through it when I am back on Sunday, I would like to compare possible solutions. Luckily, it should not take too long to examine it using python or MATLAB. Do you mind PMing me the audio file?

[bradleyland]

Originally Posted by MCaro

Great response, I'm going to work through it when I am back on Sunday, I would like to compare possible solutions. Luckily, it should not take too long to examine it using python or MATLAB. Do you mind PMing me the audio file?

I don't think you can receive PMs yet because you only have two posts. All my source files are available in my full write-up though.

http://upload.bradlanders.com/bimmer...ound-analysis/

[MCaro]

Originally Posted by bradleyland

I don't think you can receive PMs yet because you only have two posts. All my source files are available in my full write-up though.

http://upload.bradlanders.com/bimmer...ound-analysis/

Great, thanks!

[swamp2]

Probably an obvious step but for anyone playing around with this try to establish a baseline predictive capability using some dyno videos - no Doppler and a known redline.

[turbolag]

So...

[ASAP]

still here, still waiting to hear official numbers! Not worried about my chances at all!

[swamp2]

We may be losing sight of some important details in our (myself included) obsession with a high redline. I sure hope that the upper ranges of the engines rpm range are both useable and required and most importantly - exciting to use.

I'm digging the bet ^ just above!...

[Mangler]

I took a 1/3 octave snapshot of the highest rev point at the 3:06 point in the video. There are two peaks at 300 and 500hz but all the energy is around 1200hz and they are not harmonics. How are you deriving the 370hz for the calculations? When I am back to a PC ill post a screenshot. Don't know why we can't just upload pics from mobile. So stupid.

T

[swamp2]

Originally Posted by Mangler

I took a 1/3 octave snapshot of the highest rev point at the 3:06 point in the video. There are two peaks at 300 and 500hz but all the energy is around 1200hz and they are not harmonics. How are you deriving the 370hz for the calculations? When I am back to a PC ill post a screenshot. Don't know why we can't just upload pics from mobile. So stupid.

Automobile noise is very broadband on nature. Some sources are tonal but those. simply mix with the broadband ones. Sources include, engine, exhaust, wind and tire. It isn't necessarily that the combustion frequency always has the most power. As I posted earler the fundamental combustion frequency is often the second (lowest) peak in the spectral content. That being said I'm not sure what the first peak is in those cases.

[bradleyland]

Originally Posted by Mangler

I took a 1/3 octave snapshot of the highest rev point at the 3:06 point in the video. There are two peaks at 300 and 500hz but all the energy is around 1200hz and they are not harmonics. How are you deriving the 370hz for the calculations? When I am back to a PC ill post a screenshot. Don't know why we can't just upload pics from mobile. So stupid.

T

A few questions:

1) What tool are you using to perform the FFT?
2) What windowing function and size are you using?
3) What do you mean by a 1/3 octave snapshot? Do you mean a bandpass filter; if so, what specific frequency range?

Below are answers to my own questions

1) Swamp and I are both using Audacity's "Plot Spectrum" feature, so it'd be worth running it through another tool. The math should be the same, regardless of the tool though.

2) I've tried Hanning (preferred), Hamming, and Rectangular at 512 window size. All produced highest amplitude peaks at 320 Hz - 330 Hz

3) I applied a lowpass filter, but it's completely unnecessary for the FFT analysis because of the windowing function. It just eliminates noise, where "noise" is defined as anything that falls outside what could reasonably be assumed to be the primary cylinder firing impulse.

In general, I don't see much amplitude at all in the 1200hz range when using rectangular/hamming/hanning windows at a window size of 512. How are you extracting audio from the video? I used a tool called AudioHijack to capture system audio internally, but a better method would be to download the video data, then extract the audio stream.

[MCaro]

Originally Posted by bradleyland

A few questions:

1) What tool are you using to perform the FFT?
2) What windowing function and size are you using?
3) What do you mean by a 1/3 octave snapshot? Do you mean a bandpass filter; if so, what specific frequency range?

Below are answers to my own questions

1) Swamp and I are both using Audacity's "Plot Spectrum" feature, so it'd be worth running it through another tool. The math should be the same, regardless of the tool though.

2) I've tried Hanning (preferred), Hamming, and Rectangular at 512 window size. All produced highest amplitude peaks at 320 Hz - 330 Hz

3) I applied a lowpass filter, but it's completely unnecessary for the FFT analysis because of the windowing function. It just eliminates noise, where "noise" is defined as anything that falls outside what could reasonably be assumed to be the primary cylinder firing impulse.

In general, I don't see much amplitude at all in the 1200hz range when using rectangular/hamming/hanning windows at a window size of 512. How are you extracting audio from the video? I used a tool called AudioHijack to capture system audio internally, but a better method would be to download the video data, then extract the audio stream.

Just for another perspective, I used MATLAB to compute the PSD of the signal. Please see the attached images. Notice there is a peak adjacent to the one I have selected

If you're interested in the code, it is as follows:

clear; clc; close all;
% Read the file in in
[clip, fs] = wavread('m3clip.wav');
% Shorten the clip and only select one channel
clip = clip(4*fs:fs*(4.3),1);
% Have a listen to make sure it is okay
soundsc(clip , fs)
% Display the length of the clip
disp(sprintf('The length of the clip is %1.2f s',length(clip)/fs))
% Apply a hamming window
clip = hamming(length(clip)).*clip;
% Create a linearly space independent time vector
t=linspace(0,length(clip)/fs,length(clip)); 
% Find the coefficients for a lowpass butterworth filter
[B,A] = butter(5,600/(fs/2),'low');
% Filter the signal
clip = filter(B,A,clip);
% Find the coefficients for a highpass butterworth filter
[B,A] = butter(5,200/(fs/2),'high');
% Filter the signal
clip = filter(B,A,clip);
% Pad the signal with zeroes to increase the FFT's efficiency
nfft = 2^nextpow2(length(clip));
Pxx = abs(fft(clip,nfft)).^2/length(clip)/fs;

% Create a single-sided spectrum
Hpsd = dspdata.msspectrum(Pxx(1:length(Pxx)/2),'Fs',fs);  
plot(Hpsd); 
title('Power Spectral Density for the F80 M3 Sound Analysis Thread')

Let me know if I made any mistakes in my script.

[Mangler]

Here is the RTA snapshot that I took with an RTA tool on the iPad. Not nearly the level of analysis you guys are capable of. This was taken at 3:04 in the video clip where the engine noise is the loudest passing by. Not scientific at all just interesting that there is a peak at the frequencies that you center on and then the bulk of the energy is much higher. How are you selecting the frequency that you are?



T

[bradleyland]

Originally Posted by Mangler

Here is the RTA snapshot that I took with an RTA tool on the iPad. Not nearly the level of analysis you guys are capable of. This was taken at 3:04 in the video clip where the engine noise is the loudest passing by. Not scientific at all just interesting that there is a peak at the frequencies that you center on and then the bulk of the energy is much higher. How are you selecting the frequency that you are?

T

Yeah, this looks like the kind of app that performs analysis through the mic, which doesn't work so well for this application. When you do Fast Fourier Transforms, you're breaking a signal down in to its constituent components. Every time you reproduce the sound, you lose fidelity. This is especially true when you're playing the sound over a speaker, then re-recording it. You're suffering acoustic loss at the speaker and the mic. Add to this the fact that the mic in the iPad isn't a reference grade component, and your result is going to include all kinds of errors.

How are you selecting the frequency that you are?

The RTA app you're using takes an audio sample, performs Fast Fourier Transforms, then creates a histogram of the amplitude and frequencies derived from the source material.

The tools we're using allow us to import the actual source file (audio extracted digitally from the video), so there is no reproduction error, and we can operate on the full range of frequencies present in the source audio. Granted, there's some loss in the compression that occurs when the video is rendered, but from there, it's digital reproduction, so the operations are lossless.

Download Audacity on your computer, then get the source files from the end of my write-up:

http://upload.bradlanders.com/bimmer...ound-analysis/

[bradleyland]

Originally Posted by MCaro

Great, thanks!

The download link at the end of my write-up was broken, but I fixed it now, so you can download the full Audacity source files if you'd like to explore a bit.

[swamp2]

You CANNOT average into 1/3rd octave or any other similar averaging techniques with this method. You need full spectral data. If you don't understand this, you probably should not be blindly throwing software at this problem...

[MCaro]

Originally Posted by bradleyland

Yeah, this looks like the kind of app that performs analysis through the mic, which doesn't work so well for this application. When you do Fast Fourier Transforms, you're breaking a signal down in to its constituent components. Every time you reproduce the sound, you lose fidelity. This is especially true when you're playing the sound over a speaker, then re-recording it. You're suffering acoustic loss at the speaker and the mic. Add to this the fact that the mic in the iPad isn't a reference grade component, and your result is going to include all kinds of errors.

How are you selecting the frequency that you are?

The RTA app you're using takes an audio sample, performs Fast Fourier Transforms, then creates a histogram of the amplitude and frequencies derived from the source material.

The tools we're using allow us to import the actual source file (audio extracted digitally from the video), so there is no reproduction error, and we can operate on the full range of frequencies present in the source audio. Granted, there's some loss in the compression that occurs when the video is rendered, but from there, it's digital reproduction, so the operations are lossless.

Download Audacity on your computer, then get the source files from the end of my write-up:

http://upload.bradlanders.com/bimmer...ound-analysis/

Another viewpoint: Aside from what has already been stated, consider the size of the speaker on your iPad and its frequency response. Then also consider what part of the frequency spectrum the mic is tuned for. That should provide you with some intuition as to why you're seeing peaks at higher frequencies.

Originally Posted by bradleyland

The download link at the end of my write-up was broken, but I fixed it now, so you can download the full Audacity source files if you'd like to explore a bit.

Thanks, I grabbed the audio file from your write up to create the PSD in MATLAB (see above).

Originally Posted by swamp2

You CANNOT average into 1/3rd octave or any other similar averaging techniques with this method. You need full spectral data. If you don't understand this, you probably should not be blindly throwing software at this problem...

I agree. Using software without a fundamental understanding of the material (in this case, signal processing) is a bit like trying to run before learning to walk -- don't do it.

[Kayani_1]

This is positive news and great job to the guys who figured it out. So 8K red line from a turbo 6 now we are talking. All it needs is a proper hydraulic steering on par with last gen 911 and keep the price hike to minimum and we might have a great car at hand.

[BMW269]

@swamp2


You did a sound analysis quite easily, could do it to figure out if the engine is I6 or V6 (90°)?

Also, the "spies" cannot stand anywhere around the N-Ring, so there are places where the can cannot be filmed/photgraphed. The drivers apparently where told to shortshift/not rev to max at places where spy photographs are standing.

[NISFAN]

Can someone use the sound trace to measure drop in frequency after a gear change? That will give much more important data than RPM MAX.

[bradleyland]

Originally Posted by NISFAN

Can someone use the sound trace to measure drop in frequency after a gear change? That will give much more important data than RPM MAX.

It's possible, but you need a good clean upshift where the driver is WOT (or close) before and after the shift. Basically, a hard acceleration run at an easily estimated speed.

Watch this video and post any timestamps of regions where you find that:

If you do the footwork, I'm happy to do the analysis.