Chapter 7 __________________________________________________ Host Computer Commands
VAISALA______________________________________________________________________ 287
value has 12-bits of precision and a dynamic range of approximately
190 dB. The large dynamic range is necessary to cover the full range of
data. In summary:
The resulting voltage span is actually ± 4 × V
MAX
.The extra factor of four
is built into the format so that transient excursions above the full scale
input voltage can still be encoded properly. These may arise for time series
data that have been processed by an IIR clutter filter.
An improved "High-SNR" packed floating format is also available that
offers nearly the same dynamic range, but provides a 6 dB improvement in
SNR, that is, a commensurate improvement in sub-clutter visibility of
-78 dB versus -72 dB.
0916-114
The High-SNR packed format is similar to the legacy packed format
except that it uses one extra mantissa bit and one fewer exponent bit. The
dynamic range lost in the exponent is recovered through a formatting trick
known as "soft underflow", that is, the mantissa is allowed to become
unnormalized when the exponent is zero.
To decode this format when the exponent is non-zero, first create a 13-bit
signed integer in which bits zero through ten are copied from the Mantissa
field, and bits eleven and twelve are either 01 or 10 depending on whether
S is 0 or 1. Then, multiply this by 2**(exponent-25), where the exponent
field is interpreted as an unsigned 4-bit integer.
To decode the High-SNR format when the exponent is zero simply
interpret the mantissa as a 12-bit signed integer and multiply by 2**-24.
A complete analysis of the noise properties of the floating point codes
would be fairly tricky. For the High-SNR format, the 12-bit mantissa with
hidden normalization bit will vary from 2048 to 4095. The SNR will
therefore vary from 66 dB to 72 dB and we can assign a mean value of
69 dB. Another 9 dB of useful range is contained within the code as
follows:
- In a floating point encoding format, the notion of fixed additive
quantization noise is not really correct. For a signal having a given
power, the additive noise within each instantaneous sample will scale
down according to the magnitude of that sample. The ensemble of
To Next Page
Loading...