Frequency Domain Analysis of an Accelerometer Signal

I have a signal streaming in from an accelerometer (as shown in the photo below). As you can see, around the 80,000th sample, an event occurs.

I am trying to employ DSP techniques to be able to recognize when this event occurs. I am looking at rolling standard deviation, rolling mean, and setting a threshold to trigger the event notification.

However, I am at these forums to see what your opinions are regarding frequency domain analysis. I am relatively novice at frequency domain engineering, and I was wondering if this event can be recognized by diving into the frequency domain?

Could I do a Short-Time Fourier Transform? FFT?

What are some recommendations on how to recognize when this event occurs?

Hi,

My opinion:
* FFT is too much effort. And after the 80.000th sample there is a increased amplitude, not an increased or otherwise modified frequency.

My ideas:
* use a high pass filter to get rid of DC. Then perform a continous RMS calculation.
* high pass filter + rectifer + averaging.

***
But this are just assumptions, because we don´t have a clue about expected timing, sampling rate, resolution, noise...

Klaus

The FFT analysis is useful when it is made in a sample set of values with repetitive patterns. With the above compressed graph it is not possible to distinct if besides the variation in the amplitude of the signal, if there is also any other change on it. If your goal is to actually use this method, try magifying and spline plott specific regions of the curve to see if there is any more information not visible on that scale.

Are you restricted to real-time methods or is this analysis done after a test is run? Specifically, can you use all samples to determine if an event occurs at sample N, or are you restricted to only samples before N?

gary_feesher said:

I have a signal streaming in from an accelerometer (as shown in the photo below). As you can see, around the 80,000th sample, an event occurs.

Click to expand...

Let me try: I assume that the data is basically a time series. The sample number is basically a time domain uniformly carried out.

There is a change in acceleration around the stated point; amplitude increases over time.

Acceleration is directly related to force; rapid vibrations means the force is periodic. Change in amplitude can be also change in frequency or frequency may stay the same.

You should try a FFT and see what you get; it is just a tool to help you understand better the underlying physical process.

But what is a short time Fourier transform?

Short term FT is any use of the DFT (the FFT is a DFT) using a set of time intervals that are less than the entire length of the signal. The goal is to estimate the spectrum and how it changes over time. This is different than time analysis which only measures how the signal changes over time and spectral analysis which measures how the signal changes over frequency. The results are not unique across all parameters of the transform.

AM radio is a classic example. It can be viewed as a frequency that has an amplitude that changes over time. Using algebra, it can also be seen as multiple frequencies that don't change in amplitude over time.

The STFT also competes with wavelet transforms in this area. Likewise there are other heuristics to give results that are more meaningful in a given application.

Frequency Domain Trigger is implemented in Real Time Spectrum Analyzer.

See Figure 1-7 at page-7.
**broken link removed**