Filtering data from impact hammer test
Filtering data from impact hammer test
(OP)
Dear all,
I have acquired some data by using instrumented hammer and 8 accelerometers.
The data has been acquired on a railway track on a timber sleeper.
Now, I want to analyse the data in the frequency domain.
Usually I am interested in the lower frequency range up to 1000Hz.
When I calculated the corresponding FRFs for every accelerometer some of them are very noisy (channel 6,7, and 8). Channel 6,7 and 8 are the furthest from the impact location.
The rest of the channels (1,2,3,4,5) have more cleaner FRF curves with visible peaks.
I also tried to filter out the data, but I think I got lost in this part since I don't understand how the filtering is applied.
Basically I need few tips how to start getting more cleaner FRFs. So, my questions are following:
1. How to decide in general which digital filter should be applied? Is there a rule for this?
2. Is it recommended to always filter out the data? Can other info from the signal be lost?
3. What is the best filter I should apply in order to get more clean FRFs for my application?
Many thanks in advance!
I have acquired some data by using instrumented hammer and 8 accelerometers.
The data has been acquired on a railway track on a timber sleeper.
Now, I want to analyse the data in the frequency domain.
Usually I am interested in the lower frequency range up to 1000Hz.
When I calculated the corresponding FRFs for every accelerometer some of them are very noisy (channel 6,7, and 8). Channel 6,7 and 8 are the furthest from the impact location.
The rest of the channels (1,2,3,4,5) have more cleaner FRF curves with visible peaks.
I also tried to filter out the data, but I think I got lost in this part since I don't understand how the filtering is applied.
Basically I need few tips how to start getting more cleaner FRFs. So, my questions are following:
1. How to decide in general which digital filter should be applied? Is there a rule for this?
2. Is it recommended to always filter out the data? Can other info from the signal be lost?
3. What is the best filter I should apply in order to get more clean FRFs for my application?
Many thanks in advance!
Emina B.





RE: Filtering data from impact hammer test
Anyway, filtering hammer data is tricky. I'm hoping you still have the raw data, as your only hope is time domain filtering.
The usual problem with the force channel is too much noise after the impact has decayed. I like to use pretrigger and a rectangular window.
Filtering the accelerometer channels in the time domain is a bit trickier as they usually ring for a lot longer (remember to apply the same pretrigger to them). I've had some success cleaning up a channel with exponential windowing, but if the data ain't there, it ain't there. What do your coherence plots look like?
Incidentally if the data is anything like clean then I prefer not to use windows, at least as a check, so I can see what artefacts the windows have introduced.
Once you have FFTed the transfer functions and averaged them (how many averages did you use?) then there is not much you can do.
This forum has an FAQ with some links to signal processing on line resources.
Cheers
Greg Locock
New here? Try reading these, they might help FAQ731-376: Eng-Tips.com Forum Policies http://eng-tips.com/market.cfm?
RE: Filtering data from impact hammer test
If you are in the position of performing more tests, try accelerometers with higher sensitivities and try multiple impact locations. This would help you check reciprocity and help the curve fitter.
Good luck,
Sze Kwan (Jason) Cheah
RE: Filtering data from impact hammer test
Fig 1. Force and Acceleration for all impacts and all channels
Fig 2. Force and acceleration for impact at sleeper right end side
Fig 3. Force and accleration for impact at sleeper right end side-zoomed in
Fig 4. Accelerance functions for channel 1 for impact at sleepers end on the right hand side; top of railhead on the right side, top of railhead on the left side; and sleepers end on the left hand side . Channel 1 is located on the end of the sleeper on the right end side.
Fig 5. Accelerances for all channels when the impact is done on sleepers end at right hand side.
Fig 6. Accelerances for all channels when the impact is done on sleepers end on the left hand side.
Fig 7. Accelerances for all channels when the impact is done on topf of railhead on the left hand side.
Fig 8. Accelerances for all channels when the impact is done on topf of railhead on the right hand side.
I have used +/-2g and +/-5g range MEMS DC accelerometers. Sensitivity for 5g acc was within 394 to 411 mV/g, whereas the sensitivity of 5g acc was within 998 to 1016 mV/g.
The impact hammer is PCB 086D50 with 5,5kg and soft tip.
The 2g acc were overranged when the impact location was in their vicinity. The impacts were done at both sleepers end and on top of the rail head at both rail seats. So in total I had 4 groups of impacts and within every group 3 hammer impacts are done.
I will be able to do more testing, but first I want to have a plan how to post-process the data and the future one I get.
Thank you,
Emina
RE: Filtering data from impact hammer test
As a casual observer, my guess is the soft tip provides excitation at only the low frequencies (<300Hz?). That may explain why the response higher frequencies are 'noisier' compared to the lower frequencies for all channels. This is effectively a physical low pass filter. A way to confirm this is to plot the autopower spectrum of the hammer. If the above is true, try switching to a harder tip.
Good luck,
Sze Kwan (Jason) Cheah
RE: Filtering data from impact hammer test
Walt
RE: Filtering data from impact hammer test
I'd always include the phase plots and coherence
Also try checking the reciprocity, and try the effect of different force levels.
Cheers
Greg Locock
New here? Try reading these, they might help FAQ731-376: Eng-Tips.com Forum Policies http://eng-tips.com/market.cfm?
RE: Filtering data from impact hammer test
The accelerations are plotted to 1200 Hz, and OP said a "soft" hammer tip was used. How "good" is the data above 100 Hz, without seeing Coherence plots?
Walt
RE: Filtering data from impact hammer test
This is the autopower spectrum of the force for impact on top of the sleeper
I have used a soft tip, brown color. Also my interest now is for frequencies under 1000Hz.
The following figures show FRF, pahse and coherence for two accelerometers and impact on on top of the sleeper. Also, I have plotted coherence^2 for these.
I am not so sure can I get anything from this data. I was expecting to be able to obtain first five mode shapes and corresponding natural frequencies. I haven't used triggering throughout the testing. The data acquisition software was SignalExpress from NI.
I don't quite understand what you mean with reciprocity (of what?).
I haven't applied any filters on the force or acceleration data. I applied detrend (by subtracting the mean) to acceleration data to remove the gravity effect. The FRfs I have calculated as cross spectrum
between the force and acceleration, divided by the autospectrum of the force. I used for this cpsd function in Matlab with Pwelch spectra estimate; rectangular window with 50% overlapping;window length = NFFT/4.
Do you think I can proceed further with data analysis, e.g. filtering out the frequencies of interest?
Is there anything else I could do with the data in time domain?
Thanks,
Emina
Emina B.
RE: Filtering data from impact hammer test
Walt
RE: Filtering data from impact hammer test
Cheers
Greg Locock
New here? Try reading these, they might help FAQ731-376: Eng-Tips.com Forum Policies http://eng-tips.com/market.cfm?
RE: Filtering data from impact hammer test
The auto-power is traditionally shown in the log-y axis similar to transfer function. Maybe that would show it trending down sharply after 300Hz. Do you get something similar to Blue line of Figure 1 of https://www.uml.edu/docs/oct98_tcm18-189815.pdf? Please compute it as described below.
Please see: https://www.uml.edu/docs/aug01_tcm18-189833.pdf
Each impact of the hammer produces vibration that decays within it's own time block. You have 12 hits so there are 12 averages of those blocks. The data in between the time blocks are mostly noise and thus discarded.
Reciprocity is very useful. Here's a good write-up on it: https://www.uml.edu/docs/Dec08_tcm18-189884.pdf.
From your description, am I right to guess you're acquiring the time signal with NI/Signal Express and post processing the data in MATLAB with your own code? If so, I agree with Strong's recommendation on first verifying your code with known structure to get known results.
A suggestion for next step is to extract the right time block with pre-triger and block length containing enough data samples such that the signal decays completely within the block. Rectangle window can thus be used. Each corresponding impact hammer signal block should have the same time stamp. Compute the transfer function and coherence for each accelerometer channel. Avoid digital filters.
I'm not familiar with Welch's PSD. I usually just use the fft function straight up to compute the cross & auto-powers.
p/s: I'm a fan of Pete's concise style of describing modal concepts. Here's more:
https://www.uml.edu/Research/SDASL/Education/Modal-Space.aspx
https://www.uml.edu/docs/Young_Engineer_IMAC21_2003_assembled_MACL_tcm18-189941.pdf
Good luck,
Sze Kwan (Jason) Cheah