Vibration and sound analysis of large data sets

[knapkerd]

I am recording large amounts of data during testing that will be analyzed afterwards.  The data is recorded at 40kHz (6 signals) for about an hour, which amounts to ~5GB TDMS files.  I then need to perform several sets of analysis included in the Sound and Vibration toolkit, such as 1/n octave, colormap and order tracking.  From what I understand, the nature of these types of analyses must be performed at the full 40kHz and must be conducted on the entire data set in order to get the various resonance frequencies and octaves from the entire test.  I have looked at other posts about large data sets, but they seem to deal with problems that allow for breaking up of the data set into smaller pieces or taking every 10th or 100th data point.  Since I cannot do either of these with the type of analysis I am performing (am I right about this) I was wondering if anyone knew of a good way to handle all of this data?

I am attaching a quick example I worked up of how I want to analyze the data.  I am also attaching a picture because the Signal and Vibration Toolkit is needed.

The computer this is being run on is a 1.2 GHz Dell D430 laptop with 1GB of RAM running Windows XP.

Any suggestions are welcome as I dont really know where to start.  Thanks.

--Robert

Message Edited by knapkerd on 06-11-2008 09:46 AM

[preston_johnson]

Thanks for the posting,

Storing data into TDMS files is indeed the recommended method to store the large data sets.  Judging from your picture, you have a good start at post processing the data. 

There are some good tutorials in the Sound and Vibration Measurement Suite Help Files.  Here is the overall outline for S&V applications:

http://zone.ni.com/reference/en-XX/help/372416A-01/svtconcepts/sv_concepts/

Here is the general entry point into our Sound and Vibration Measurement Suite Documentation:

http://zone.ni.com/reference/en-XX/help/372416A-01/

Do you have any specific questions regarding your analysis results?

[knapkerd]

My question is more in regards to the large amount of data I am reading in.

First, I am wondering what is the best way to read in large TDMS files?  I am currently trying to read in channels separately from a nearly 8GB file (6 channels), but I am geting memory errors.  I am thinking that maybe I should read in the signal in several sections.

The second question is with regards to analyzing such large data sets with the VIs shown in my first post.  Will there be any issues with analyzing channels that are each over 1 GB?  Should I instead analyze parts of the signals separately and combine the analysis (if this is possible)?

Also, with regards to the S&V outline, are each of the steps (such as limit testing) required to move forward, or are they suggested based upon what type of analysis I am doing?  I am very new to sound and vibration signals and analysis.

Unfortunately, I have not acquired any actual data because the testing is months away, so I am simulating signals.  So, I do not have questions about the actual results.  Also, I have not gotten past the TDMS read memory issue, which means I have not tested analysis the large data sets with the Sound and Vibration VIs.  The picture in my first post is more for a representation of the things I want to analyze.

Thanks for you help.

--Robert

[preston_johnson]

I would read in the data files in sequential blocks, reading all 6 channels from each block.  There are inputs to the TDMS read files that allow for start index and duration. 

You can trend the results of the analysis of the block by block analysis to give you a total data file result.  For example, you might track the unbalance of gearbox frequency magnitude over time or speed.  In this way, you produce a result (the trend chart) over the entire data set, yet are analyzing it block by block. 

The steps in the analysis diagram are listed as optional.  If you are not doing limit testing, the you can ignore this step. 

Let us know how your simulated data sets work. 

[knapkerd]

Would this trending work with the colormap, order tracking (RPM vs. order) and octave analysis?  If so, how would I combine the block results into one graph?  Sorry, I am new to these sound and vibration measurments and dont yet fully understand how they work (although those links you posted helped).

Also, is there a way to determine the size of the block I am reading in?  I am assuming that I could take the size of one element and then the size of the block would be    (block size)=(element size)*(# of signals)*(# of elements read).  What is the size of one element in a signal?

Thanks for your help.

--Robert

[knapkerd]

I forgot to mention, the analysis is being conducted on data captured using displacement transducers, accelerometers and microphones.  Not sure that this is important, but I thought I would mention it.

--Robert

[preston_johnson]

The Block size is the lenght of time you set the data acquisition driver to capture data in each loop.   For example, you might capture data at 15000 samples per second for 1 second and write this 1 second of data to the open file. 

To trend and combine all the blocks into a loop, you need to obtain analysis results from each block such as the order magnitude and phase, then plot them on a trend chart, or buffer them in memory for later plotting.

[knapkerd]

So for octave or FFT analysis, a waterfall graph would be a type of trend graph (is this correct?).  Is there a way to "trend" a colormap or an order vs. RPM graph?  It seems like sequential analyses of these graphs could not be plotted together (unless some sort of averaging works).

Also, is there a big performance hit if I grab more than just one block of data?  If I am writing 8000 samples per loop to a file, which ends up with approximately 144million samples and I wanted to read in something more like 100kSamples and analyze that way, would it be slower?  The only reason I ask is that I have been timing the reads from the TDMS file for different #of elements read and increasing the size of the read to something on the order of 100kSamples seems more efficient.  But, I have not yet put this data through the SVT analysis VIs.

Thanks for you help.

--Robert