Noise

To make this method work properly, the signal component of interest may not vary greatly during a conversion. However, another criterion for a successful enhancement of the resolution is that the input signal has to vary slightly when sampled. This may look like a contradiction, but in this case, variation means just a few Least Significant Bytes (LSB). The variation may be seen as the noise component of the signal. When oversampling a signal, there may be noise present to satisfy this demand of small variations in the signal. The quantization error of the ADC is at least 0.5 LSB. Therefore, the noise amplitude has to exceed 0.5 LSB to toggle the LSB. Noise amplitude of 1-2 LSB is even better because this will ensure that several samples do not end up getting the same value.

Criteria for noise when using the decimation technique:

Normally, there will be some noise present during a conversion. The noise can be thermal noise, noise from the CPU core, switching of I/O-ports, variations in the power supply, and others. This noise will in most cases be enough to make this method work. In special cases though, it might be necessary to add some artificial noise to the input signal. This method is referred to as dithering.

Figure 1 (a) shows the problem of measuring a signal with a voltage value that is between two quantization steps. Averaging four samples will not help, since the same low value will be the result. Figure 1 (b) shows that by adding some artificial noise to the input signal, the LSB of the conversion result will toggle. Adding four of these samples halves the quantization steps, producing results that give better representations of the input value, as shown in Figure 1 (c). The ADCs ‘virtual resolution’ has increased from 10 to 11 bits. This method is referred to as decimation, and will be explained further in section Averaging.

Another reason to use this method is to increase the signal-to-noise ratio. Enhancing the Effective Number of Bits, ENOB, will spread the noise over an increased binary number. The noise influence on each binary digit will decrease. Doubling the sampling frequency will lower the in-band noise by 3 dB, and increase the resolution of the measurement by 0.5 bits.