In this section, a method to implement the previous detection criterion
is proposed (see Fig. 3 (click here)). The criterion uses the first Q
time lag (
) of the quantized correlation function 
.
To initialize the detector, knowledge of the receiver spectral shape is necessary.
Thus, at the beginning of the observation, the receiver is directly connected
to a noise generator so that the 
hypothesis is
forced. The quantized autocorrelation is estimated and the true
normalized autorrelation, 
, is deduced from
by applying the correlation correction function (Hagen & Farley 1973). Then, the
coefficients of the second order polynomial approximations, used to evaluate
the 
and
dependance on 
, are computed and stored.
The receiving system is then connected to the antenna, and observation starts.
At each clock cycle, the Q values of the quantized product for the Q first time lags
are computed by the correlator and stored into Q shift registers
of size N. For each time-lag, these shift registers represent a moving window
on the N last quantized products used to compute the final vector 
.
The component w0, which is issued from the null time-lag and is an estimate
of the input power
1 (see
note 
), is used to update and .
Then, the test function 
is computed and its value is
compared with the predefined detection level 
. If the test function
value is less than , no RFI is detected, and the
"first in" quantized products are sent to the final integration. If the
criterion value is greater, an RFI is detected and the final sum is suspended
until the test function value comes down below the detection level again.
The size of the implementation depends on the size Q of the vector
. In the next section, the influence of Q on the detector is demonstrated.
Figure 3: Implementation of the detector (dotted lines represent the
initialisation phase). The test function uses the Q first channels of
a P channels correlator
For a given RFI, the choice of the size Q strongly determines the detector
performance. From a spectral point of view, the detector carries out a
comparison between an estimated spectrum and a reference spectrum with
a spectral resolution inversely proportional to Q. By using values of
Q which are too small, the risk is the smoothing of relevant spectral
features of the RFI and therefore reduction in the quadratic error between
and the reference . In contrast, large values
of Q may reduce the detector performance because of a large induced
variance
.
In fact, the optimal value of Q must be chosen as a function of RFI
and the observational context.
Nevertheless, for multiple RFI detection or blind detection (no a priori information on RFI), the proposed detector can be modified to perform multiresolution criteria: the detector is sized for the largest value of Q (highest resolution) and criteria with intermediate resolution are obtained recursively.