1
\$\begingroup\$

This switching artifacts occur at first data points.

enter image description here

Why does it lift the entire noise floor? Is it because all the frequencies in the white noise got amplitude? or only lower frequencies which is akin to DC effect in time domain baseline shift? Do you call this noise floor shift also baseline shift?

enter image description here

In the following, I skpped the first 1000 samples in Matlab and FFT it which flattens the entire noise floor.

enter image description here

\$\endgroup\$
3
  • \$\begingroup\$ What is the FFT of the transient, with the remaining sample (after it) zeroed? What does this say about the plots shown? \$\endgroup\$ Commented May 8 at 5:51
  • \$\begingroup\$ Its not zeroed. It measuring white noise with all 3 inputs shorted (in+, in-, ground) \$\endgroup\$
    – Jtl
    Commented May 8 at 7:16
  • \$\begingroup\$ Are you familiar with the concept of superposition? What is the definition of linearity, and is FFT a linear transformation? \$\endgroup\$ Commented May 8 at 17:27

1 Answer 1

1
\$\begingroup\$

FFT has no concept of transient signals; everything in the time-domain data record is assumed to be repeating over and over. The last sample in the data record is followed by the first sample in the data record. That's why the lowest FFT bucket is not really "DC offset", but rather a bucket that contains the lowest band of frequencies in the signal.

A common way of dealing with this problem of artifacts near the beginning and end of the time-domain record, is to apply a "windowing" function which reduces the amplitude of the samples near the beginning and end. There are a lot of different windowing functions with different trade-offs, because using any windowing function at all introduces some distortion in the form of "frequency smearing" across adjacent frequency buckets.

In your particular case, an even better solution would be to discard the initial part of the time-domain data record.

Note: if you're using a SAR ADC, those tend to have a lot of charge injection, especially when measuring high-impedance sources. So you may have to actually acquire the data at speed, and then discard the result until the measurement is stable.

\$\endgroup\$
2
  • \$\begingroup\$ So you mean to say bec of the transient. The lowest band of frequencies get higher in amplitude? But why does it not transfer to the top of the band with the noise floor in same lower position. If you can point some illustrations pls do so. Kindly elaborate your first paragraph. Tnx. \$\endgroup\$
    – Jtl
    Commented May 8 at 7:19
  • \$\begingroup\$ No, the frequencies do not change. The shape of the repeating transient and its long irregular shape are equivalent to a whole bunch of sinusoids of many amplitudes and phases, so including that transient in the data puts all of those components into a whole bunch of the frequency buckets, spilling out over much of the band. The noise floor is not actually increased: the transient’s FFT is so wide that there’s no clear view of the noise floor. \$\endgroup\$
    – MarkU
    Commented May 8 at 19:43

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.