Shifting data when reading from binary file using fread

I am trying to read experimental data from a binary file using fread. While the resulting data is read almost correct, the data is slightly shifting in a seemingly unpredictable manner. An example of this is given below (note that the direction of shifting is different for all signals):

While the result should look more like:

I am using the following code:

fid = fopen('sound_s3');
data = fread(fid,[1024, inf],'int32'); % reading in the 1024 signals of ~50k samples each
fclose(fid);

data = data/2^31; % converting bit to pressure

% plotting the examples
figure
plot(data(1,1:20000)');
xlabel('sample');ylabel('pressure')

figure
data = data - smoothdata(data,2,'movmean',1000);
plot(data(1,1:20000)');
xlabel('sample');ylabel('pressure')

Does anyone know what could cause this issue? I have tried basically all 'fread' options, and I am sure that the data is stored in 32 bit precision.

Note that I am sure that the signal is measured and should look like the second image.

Thanks in advance!

[Walter Roberson]

Did you try using big endian?

[dpb]

Walter's probably got it but just in case there's something really funky going on; give us the vendor datasheet that describes the data format used by the data acq card/system.

[Lars Janssen]

It isn't big endian, I had tried that already.

I am not sure what excactly the vendor datasheet should contain but I found the following about the data acq system:

Decimator Normalization Factor: 8.003663365181634

Fixed Point Fraction Length: 31

Sampling Frequency: 46875

I am not exactly sure what a Dec. Norm. Factor is and will look into it now. Might it be the cause of my issues?

[Walter Roberson]

Comparing that information to the plot, what I suspect is that you need to read the values with 'int32=>double', then divide by 2^31, and multiply by 8.003663365181634 . However it would not surprise me if instead what you need to do is read as '*uint32', and then

neg = data >= 2^31;
data(neg) = data(neg) - 2^31;
scaled = double(data) ./ 2^31 * 8.003663365181634;
scaled(neg) = -scaled(neg);

[dpb]

Me, neither, by the exact name w/o any more context.

What I see is Fixed Point Fraction that makes me think isn't just an integer but a float in an integer represenation.

Who's the vendor/what's the device? Somewhere there's got be sample code to read the data; oftentimes there will be a Matlab version, too.

[Lars Janssen]

@Walter Roberson, in my case I care mostly about relative values. So a scaling factor does not solve my problem. Both reading as 'int32=>double' and '*uint32' did not change the values.

@dpb, the device I am using is:

https://www.sorama.eu/cam1k

Specs:

https://static1.squarespace.com/static/5aec19464eddecf8ba36f73e/t/5c879949b208fcd94f3836a2/1552390476125/Sorama+CAM1K+Specifications+2018.pdf

However, they have a built in analysis tool. For my specific goal I can not use this tool, but I assume this is why it is not easy to find any sample code.

[Lars Janssen]

They also want you to buy a Credit package as a requirement for analyzing measurements using their tool. Which might be another reason why this information is not easily accesible.

[Walter Roberson]

The architecture is a bit messy; and Yes, they don't want people doing their own analysis.

https://pure.tue.nl/ws/files/52023982/20170222_Vliet_2016_bb844317.pdf

Alternatives: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5795332/

[dpb]

Didn't see a thing that described the format...that's simply rude.

At this point unless you can find some other Rosetta stone, would seem next step would be to take a file and decode it with their tool and reverse engineer the file structure with the few hints that are given....or, find a non-proprietary solution.

[Walter Roberson]

I got lost in the question of bit depth. One part says that the daq are single bit and then a second later it was talking about 32 bits per sample.

[Lars Janssen]

Thanks for the help anyways! I will try to look if I can further reverse engineer the files. If I find anything I will let you know.