DonnyFL Tatsu Data

[OldSpook]

Today when I was testing I pulled out the Donny FL Tatsu and ran it beside all the other moderators and the bare rifle.

I located the microphone 31 feet from the muzzle and 33 degrees (11 o'clock) left of the muzzle. This was done to reduce clipping. I still experienced some minor clipping with the bare rifle but we don't really care if the bare rifle appears to be making less noise than it really does, do we? Anyway it ain't much.

I should also mention this is on the Talon P which is a .20 caliber running 25 fpe.

I will further analyze this data when/if time presents itself.

Here is the audio shot for shot ordered bare, Tatsu, bare, tatsu ... bare, tatsu.

 

[subscriber]

Getting an accurate measure of how loud the bare muzzle is, sets the baseline to determine how effective the muffler under test is. Otherwise, how do we know how much the muffler did?

Now, clipping a few data points out of thousands would not seem to make much difference.

 

[WobblyHand]

Interesting data. Fortunately, there are two unclipped shots of the bare. This is good. I'll be able to measure them and see the difference between them. All of the Tatsu shots are un-clipped. What this does show is there is a bit of variation from shot to shot, for both bare and moderated shots. The nature of the beast, I suppose.

What's the standard deviation of the shot velocity and the mean velocity of this AG? If the standard deviation of a shot velocity is large enough, (and if the velocity is proportional to the noise) variations in velocity may swamp smaller things we are trying to measure. Dang, this is interesting as all get out, and not so easy to get great numbers. Guess this is what one would call AG variation of a shot.

But... if we could get say 20 good unclipped uncorked measurements, then we could get an estimate of the mean and standard deviation of the bare muzzle. If the standard deviation is sufficiently low, then we could use the mean value as the reference. Then moderated shots (at the same mic distance and angle) could be measured against that "valid reference".

Getting there. Probably need to slightly increase the microphone distance to reduce the incidence of clipping. Maybe a couple dB. Good work Best data yet.

 

[WobblyHand]

Now, clipping a few data points out of thousands would not seem to make much difference.

It never seems like it, at first glance. What if that impulse was the value 10 instead of 1, would that matter? Bet it might matter if it was 1000 or greater. Only way to know for sure is to take the data clipped and unclipped and do the actual numbers with statistically meaningful data sets. In radar we worked with statistics, there never was certainty, only probability. A single test was not that meaningful, it could have been spectacularly good, or bad. We did road tests, had a fleet of hundreds of vehicles running world wide tests. Each vehicel collected about 4TB of data a day. Actually this was required by our customers. We had to analyze that data. We had an exabyte disk farm. An exabyte is 1024 petabytes, or 1024*1024 TB. This was big data. I would also run hundreds, or thousands of simulation trials with random noise to get a feeling for the mean and standard deviation of the results. Our little band of brothers can't fund anything like that, but I hope you get the idea that this sort of work is statistical.

Likewise with these audio measurements in noise. We'd like to know with some level of probability (or certainty) that the sound is x dB lower +/- ydB for modelQ vs modelS.

I do know clipping ADC's is a sure way to generate crap in the spectrum, crap that never existed. It gives a distorted and untrue view of what is actually there. Been burned by this long ago, was a tough, tough lesson for me as a junior engineer. Got read the riot act by the chief systems engineer. That's why I've been such a stickler, ok PIA, in these threads. Seemingly little stuff matters a whole lot if you want instrumentation grade data. Once everything has been figured out, you can back off on some things safely, because you have a true initial picture and can actually see the quantitative effects of your changes and determine if they help or they are harmful.

 

[subscriber]

Clipping on a sound meter is like over exposing a photo. Some detail is gone. Boosting underexposure brings out detail, at the expense of noise. So, a slightly low a recording level seems better than a little too high.

 

[OldSpook]

Interesting data. Fortunately, there are two unclipped shots of the bare. This is good. I'll be able to measure them and see the difference between them. All of the Tatsu shots are un-clipped. What this does show is there is a bit of variation from shot to shot, for both bare and moderated shots. The nature of the beast, I suppose.

Why didn't I think of that?

What's the standard deviation of the shot velocity and the mean velocity of this AG?

ES = 19, SD =3, MEAN 867

If the standard deviation of a shot velocity is large enough, (and if the velocity is proportional to the noise) variations in velocity may swamp smaller things we are trying to measure. Dang, this is interesting as all get out, and not so easy to get great numbers. Guess this is what one would call AG variation of a shot.

Oh now I remember why I didn't think of that.

But... if we could get say 20 good unclipped uncorked measurements, then we could get an estimate of the mean and standard deviation of the bare muzzle. If the standard deviation is sufficiently low, then we could use the mean value as the reference. Then moderated shots (at the same mic distance and angle) could be measured against that "valid reference".

I am sure you have time for that...

Getting there. Probably need to slightly increase the microphone distance to reduce the incidence of clipping. Maybe a couple dB. Good work Best data yet.

and I am sure you've got space for that...
Take the data you are getting and run with it or go get your own data.

 

[OldSpook]

It never seems like it, at first glance. What if that impulse was the value 10 instead of 1, would that matter? Bet it might matter if it was 1000 or greater. Only way to know for sure is to take the data clipped and unclipped and do the actual numbers with statistically meaningful data sets. In radar we worked with statistics, there never was certainty, only probability. A single test was not that meaningful, it could have been spectacularly good, or bad. We did road tests, had a fleet of hundreds of vehicles running world wide tests. Each vehicel collected about 4TB of data a day. Actually this was required by our customers. We had to analyze that data. We had an exabyte disk farm. An exabyte is 1024 petabytes, or 1024*1024 TB. This was big data. I would also run hundreds, or thousands of simulation trials with random noise to get a feeling for the mean and standard deviation of the results. Our little band of brothers can't fund anything like that, but I hope you get the idea that this sort of work is statistical.

Likewise with these audio measurements in noise. We'd like to know with some level of probability (or certainty) that the sound is x dB lower +/- ydB for modelQ vs modelS.

I do know clipping ADC's is a sure way to generate crap in the spectrum, crap that never existed. It gives a distorted and untrue view of what is actually there. Been burned by this long ago, was a tough, tough lesson for me as a junior engineer. Got read the riot act by the chief systems engineer. That's why I've been such a stickler, ok PIA, in these threads. Seemingly little stuff matters a whole lot if you want instrumentation grade data. Once everything has been figured out, you can back off on some things safely, because you have a true initial picture and can actually see the quantitative effects of your changes and determine if they help or they are harmful.

I probably would have sat down with you and explained that there are times when you can ignore clipping in a signal and times when you can not ignore it. I would not have needed the riot act to accomplish that. I would also have explained that you shouldn't take the appearance of an engineer as a measure of his competence. I look and sound like an old country boy, but I'm not like any old country boy you ever met. Take that to the bank in the future or don't engage.

We can do this all day long, in the end you will be required to SHOW me that disregarding 1 or two samples in a data set containing 4410 points is going to move the mean data point in the standard deviation enough to alter the answer we are seeking. We have sample data already which can quantify the clipping because I got the same moderators with no clipping, just like I said I would. So it can now be quantified. Maybe I need to do that, but I expect you are doing it right now. I look forward to your report.

I have other things to do which are more interesting than arguing minutia. The data we have is good enough BUT as promissed I have collected data which shows no clipping OTHER THAN the bare rifle. That means a smart guy like you will be able to properly rank (and quantify) the differences between the moderators. You will find that data in the other thread IF WE EVER GET AROUND TO TALKING ABOUT THE FOREST INSTEAD OF THAT DAMN MISSING TREE.

 

[WobblyHand]

Ok. No need to get testy. I spent a lot of time today looking at your data. Here it is, perhaps in an easier way to see, maybe not. Have some additional numbers as well, but not in a cute form. 32K samples, not quite centered on the impulses. Why? There's a tail which has energy which I want to capture somewhat.

Out of 5 bare shots I could capture (the first one was too close to the start to treat the same way) the square root of the sum of the squares is nan, 5.848, 6.202, 5.804, 5.156, where the 2nd and 5th ones are known to be unclipped. In dB that
is -inf, 15.34 ,15.85, 15.27, 14.25

For the 5 Tatsu shots, which are all good, they are: [2.362, 2.829, 2.794, 2.303, 2.44 ] in dB they are [7.47, 9.03, 8.92, 7.25, 7.75] You can thank me for suppressing gratuitous numbers after the decimal point.

The clipped waveforms are different spectrally, check out the green waveform for the clipped bare shot, although it may not matter if all we are after is a single number, post integration. I think sound tone and distribution matters. Here's everything in picture form. You want the real processed data, I can attach my code (158 LOC, written today). It's in python/scipy/numpy which is MatLab like in using N-d arrays and matricies. I've used it a long time. Pretty much gave up MatLab, even during my working life, darn toolboxes blew my budget, or purchasing would take so long to make a buy, that I had to get the job done and do it in python anyways.

Cheers. Have fun. I did today. Geeked out on a little signal processing. Good to do that every once in a while. Dusts off the cobwebs.

 

[OldSpook]

Think that tail that you're talking about is on the beginning of the shot? The gentleman over on GTA who has a laser trigger setup shows that we are collecting noise before the pellet exits the barrel. That's why I'm trimming the start as close to the "pop" as I think prudent.

@karl_h 's input maybe very valuable. I'm trying to find specifications on the microphone on this Motorola but I'm betting that these devices are sampling into 16 bits.

I'll see if I can get a look at the raw data.

 

[WobblyHand]

Think that tail that you're talking about is on the beginning of the shot? The gentleman over on GTA who has a laser trigger setup shows that we are collecting noise before the pellet exits the barrel. That's why I'm trimming the start as close to the "pop" as I think prudent.

@karl_h 's input maybe very valuable. I'm trying to find specifications on the microphone on this Motorola but I'm betting that these devices are sampling into 16 bits.

I'll see if I can get a look at the raw data.

The little bubble on the beginning might be the valve or trigger, or something early in the shot cycle. If you trim the beginning how are you going to window it? As you well know decent window functions are near zero at it's start and end. The signal is multiplied point by point by the window. I keep the peak at least somewhere near the maximum of the window. Yes, it captures the tiny start energy, and I also look at the long tail of energy from 500ms out to around 750ms, at least on this graph. The tail certainly has more energy than the the tiny bubble at the front. If you start too early the window function is low valued. Here's a Hamming window with the dashed black line, illustrating what I mean.

Hey you can do it any way you want, but if you suppress a lot of the main signal, don't think you are doing yourself a favor. Could be wrong though. If so, show me how to do it better. Always willing to learn a new thing or two. Cheers.

As for the captured data quantization, it can be determined by the smallest delta value you can find in the data. Shouldn't be too hard to figure out. If it is 2**-15, then its a 16 bit converter. 2**-11, it's a 12 bit converter, etc. Doubt it is 24 bits...

 

[OldSpook]

The little bubble on the beginning might be the valve or trigger, or something early in the shot cycle. If you trim the beginning how are you going to window it? As you well know decent window functions are near zero at it's start and end. The signal is multiplied point by point by the window. I keep the peak at least somewhere near the maximum of the window. Yes, it captures the tiny start energy, and I also look at the long tail of energy from 500ms out to around 750ms, at least on this graph. The tail certainly has more energy than the the tiny bubble at the front. If you start too early the window function is low valued. Here's a Hamming window with the dashed black line, illustrating what I mean.

Hey you can do it any way you want, but if you suppress a lot of the main signal, don't think you are doing yourself a favor. Could be wrong though. If so, show me how to do it better. Always willing to learn a new thing or two. Cheers.
View attachment 431476
As for the captured data quantization, it can be determined by the smallest delta value you can find in the data. Shouldn't be too hard to figure out. If it is 2**-15, then its a 16 bit converter. 2**-11, it's a 12 bit converter, etc. Doubt it is 24 bits...

Agreed most likely not 24 bits.

I explained my reasoning in the first post in this thread. I am not trying to understand any signal that I cannot change. The only part of the signal that matters to me is the part that happens during and after uncorking until decompression reaches perhaps a couple of atmospheres. By that time the pellet is meters down range and the noise that is being received is purely echo and or harmonics in the gun. There is nothing a moderator can do to change that part of the signal.

If I can't change it then I'll spend my time studying it after I understand what's going on in the first 10 milliseconds.

 

[WobblyHand]

Agreed most likely not 24 bits.

I explained my reasoning in the first post in this thread. I am not trying to understand any signal that I cannot change. The only part of the signal that matters to me is the part that happens during and after uncorking until decompression reaches perhaps a couple of atmospheres. By that time the pellet is meters down range and the noise that is being received is purely echo and or harmonics in the gun. There is nothing a moderator can do to change that part of the signal.

If I can't change it then I'll spend my time studying it after I understand what's going on in the first 10 milliseconds.

Whether we like it or not, it comes with the acoustic signature. It's what we hear. And yes, it's different depending on the environment.

Here's the first set of bare shots with the time domain plotted in dBs. Is each pointy spike an echo? Could very well be. Why such a long tail? Is that a continuum of echos or something else? Interesting. First echo is @ 46ms delay, (~16m), next is roughly double that, could be a reflection. There's a couple of them. Then there's something distinct at 201ms delay (~69m). Using 344m/s for the speed of sound, for these calculations.

The moderator can't change the echos pursue, but it sure can reduce the signature of the main event, so all the echos are quieter, since the stimulus is quieter.

 

[JungleShooter]

You guys ROCK!!

Thanks for taking the jump headfirst into this particular rabbit hole — it will benefit many of us in the long run.


I really enjoy seeing what's possible in our worldwide AG community when we all contribute what we have been given.

Matthias

 

[OldSpook]

Whether we like it or not, it comes with the acoustic signature. It's what we hear. And yes, it's different depending on the environment.

Yes it does, but there is nothing we can do after the decompression to alter that acoustic signature. As you mention below the only time we can influence that is before decompression.

Here's the first set of bare shots with the time domain plotted in dBs. Is each pointy spike an echo? Could very well be. Why such a long tail? Is that a continuum of echos or something else? Interesting. First echo is @ 46ms delay, (~16m), next is roughly double that, could be a reflection. There's a couple of them. Then there's something distinct at 201ms delay (~69m). Using 344m/s for the speed of sound, for these calculations.

Temperature this AM was about 40 degrees. I am at sea level + 20 feet. The signature which shows up very clearly in the shots for the Tatsu at ~202ms after the uncorking are of the impact of the pellet on a cinder block 40 yards from the shooting position. It won't be in every shot. I wasn't trying to hit it. It was just yhere. The geometry is explained in the first post in this thread and includes an explanation of the timing of that spike. I didn't try to compensate much for temperature but used 1075 fps for the velocity of sound this AM (1050+temp(f)). I remember thinking when I picked the number it might be a bit low.

The uncorking event takes about 7.5 ms on most air rifles. That is the time from the pellet exiting the muzzle to the pellet getting down range a couple of meters.


Stan's Posts here are very helpful in analyzing the uncorking and decompression events.

The moderator can't change the echos pursue, but it sure can reduce the signature of the main event, so all the echos are quieter, since the stimulus is quieter.

That is exactly correct. Almost ALL of that signal after about 7 or 8 ms is echo and harmonics in the physical structure of the gun. You can't alter it unless you alter it's source.

So lets think about what we are going to hear. First there will be the "tick" of the trigger, then springs and things decompessing and hammers smacking valves, and finally air venting down a tube pushing a plug. These things all generate harmonics in the gun. Finally (maybe 12 ms, 10 on a gun with a FAST lock time) we hear the uncorking even before there is much if any air exiting the moderator. All of these things are apparent in Stan';s posts. Then the pellet exits the moderator and a blast of air follows it in less than a millisecond. Uncorking is over and decompression begins. You can bound decompression by locating the point where the aggregate RMS peaks and begins to decline again which appears to be about 2 ms after the pellet exits the muzzle of the moderator.

So that is what we have to work with, everything we do we have to do in that 5 to 7 ms period between muzzle exit and 2 meters down range. After that we are hearing the echos and harmonics in the gun.