This was built to check the ultra-low frequency noise in voltage regulators and references that formed part of the readout electronics of a Micro Bolometer array system. These , like most IR detectors have DC-coupled video signal chains. As such they are very susceptible to thermal drift and 1/f noise. Most semiconductor data sheets only publish noise spectra down to a few Hz. Below is an excerpt from the data sheet of the Texas Instruments TL1431 voltage reference. No data is shown below 10Hz which is precisely the region where we need maximum stability.
A spectrum analyser was designed and built to check the noise density of the TL1431 and other references. The hardware is shown below. Chopper stabilised op-amps were used to reduce 1/f noise. When collecting data, the analyser ran on Lithium cells in order to remove the possibility of PSU noise. The DC component of the signal was subtracted using a switched capacitor prior to amplification and digitisation. The 18-bit data was transmitted via USB to a host computer, at up to 20ksps, where Python scripts were used for both Noise Floor analysis and Line Magnitude analysis.
With inputs grounded the intrinsic noise of the analyser was measured. The results are shown below. The left-hand panel uses Hanning apodization and gives the noise density of wide band noise. It will not give correctly scaled results for narrow-band line sources. For this, the Flat-Top apodization function needs to be used instead, giving the result shown in the right-hand panel. No narrow-band sources were visible.
The image below shows the noise with the inputs floating. The 50Hz mains frequency and its first harmonic are clearly visible.
The TL1431 was then connected to the analyser giving the result below.
The noise waveform (shown in the lower panel) is dominated by low frequency noise. Its spectrum (upper panel) shows the 1/f noise below 10Hz, something that was not apparent from the manufacturer’s data sheet. This 1/f noise will be directly superimposed upon the sensor’s video signal. If special measures are not taken, for example using blind reference-pixel signal subtraction, then the SNR of the detector will be degraded.