I'm very new to the world of signal processing and looking for some guidance.

I have 5 kHz accelerometer data from a car and looking to process this to remove some of the high frequency noise and also determine whether some of the 'spikes' in the data are in fact real events (Some are 6G, so I believe they may not be 'real').

I've read lots of conflicting approaches but one that I keep seeing is the use of a low pass Butterworth filter. I've tried this using python and found that the reduction of noise goes hand in hand with the reduction of magnitude of 'spikes', so I'm now concerned about the removal of data that could be 'real'.

enter image description here

I then went to look at the frequency domain and found most of my high power events are at lower frequencies, giving me 4 distict events (I think);

enter image description here

  • How do I determine which approach to use in order to remove any erroneous data points?
  • Is there a better method/approach of doing this that I'm yet to come across?
  • $\begingroup$ well, a spike very much has very high-frequency content. That's literally what makes it spikey; so, a filter that removes high frequencies (i.e., a low pass filter) is designed to reduce the amplitude of spikes. It's the whole idea here! You'll need to come up with a signal model that describes what your "real" signal looks like, which properties it has, so that you can (and we can help you) derive a sensible processing chain for that :) $\endgroup$ Sep 24 at 10:47
  • $\begingroup$ Hi Marcus, thank you for the reply. I'll need to understand better what is meant by signal model before I can come up with one. After some reading, I'm still left a bit confused. Can you elaborate on the term a bit? $\endgroup$ Sep 24 at 13:19
  • $\begingroup$ there's really not that much beyond plain English there: a signal model is just a model, i.e. a description on how your signal "comes to be". So, where do these spikes come from, which shape do they have, what about the spikes do you care about, where does your noise come from or which shape does it have... such things. $\endgroup$ Sep 24 at 13:21
  • $\begingroup$ "Events" don't translate into spikes in the frequency domain. A spike in the time domain gets spread across all frequencies. Rather, continuous sine wave tones in the time domain get gathered up into spikes in the frequency domain. Those spikes you see in the FFT are frequencies at which your accelerometer data has tones -- I suspect they're harmonics of the engine's rotational frequency, but they could be any other repetitive signal prevalent in the car. $\endgroup$
    – TimWescott
    Sep 24 at 15:02
  • $\begingroup$ If you feel any "clunks" when you're driving, those are events the accelerometer will detect as spikes. You don't say how old the car is or the driving surface, but 6G is probably mild for a floor-mounted accelerometer registering a rock kicked up by a wheel. If it's vertical, it could be a pot-hole. $\endgroup$
    – TimWescott
    Sep 24 at 15:05

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.