Feature Article: DOI. No. 10.1109/MAES.2020.

2997415

Real-Time Signal Generator for Noise Radar

Mehmet Ispir, Alper Yildirim, Tubitak Bilgem Iltaren

exponential, and Rayleigh distributions with higher accu-

INTRODUCTION
racy. The second method is called as the colored pseudo-ran-
Noise radar is a form of random signal radar that employs a dom noise (CPRN) generator. The amplitudes of the pseudo-
noise waveform as the transmitted signal and uses coherent random samples are generated with normal probability den-
processing of radar returns. It has been known for many sity functions (pdfs) and the power spectral densities with
years [1], [2] and has both civilian and military applica- uniform pdfs using the CPRN generator. Using our design,
tions. Because of the truly random transmitting signal, the radar engineer can easily generate noise signals with dif-
noise radars have many advantages compared with conven- ferent combinations of pdf and bandwidth and be able to
tional radars, including high interference immunity and analyze the effects of these noise waveforms on the target
high electromagnetic compatibility, very low probability detection performance. We present some preliminary results
of intercept, and good electronic counter countermeasure of the design and implementation of FPGA-based signal
capability [3]–[5]. These properties increase the noise generator for noise radar and discuss the different methods
radars’ importance in military applications. used to generate noise signals.
With the advancement of the technology, the steps of The article is organized as follows: In the second sec-
generating and processing signals in noise radars can be tion, we present background information on noise genera-
performed in digital environments [3], [6], [5], [7]. As a tion methods and algorithms on FPGA. Details of the
common prototyping platform for pseudo-random number FPGA implementation of noise generation methods are
generation, FPGAs can also be used for the generation of described next with the numerical results, and the final
samples for noise radar. There are many studies that present section presents a summary of the work and suggests
the results of FPGA-based arbitrary waveform generators topics for future research.
for modern noise radars [8], [9]. These studies focused on
generating pseudo-random sequences of different types
and give details regarding the correlation function. In addi- NOISE GENERATION METHODS
tion, FPGA-based noise radar applications are considered
in different number of studies. In [10], a software defined One of the most common implementations for pseudo-
noise radar based on an FPGA board is presented. random number generation is using linear feedback shift
In this article, our main motivation is to generate pseudo- register (LFSR). It consists of a shift-register, where the
random samples with different distributions and to create a bits at several positions (or taps) are combined with exclu-
testing environment for the target detection performance. sive-OR/NOR operations and feed into the least-signifi-
For this purpose, we developed a real-time digital signal cant bit of the register on each clock cycle. A maximal
generator on an FPGA by adapting and improving our previ- length LFSR can generate uniform random variables with
ous effective pseudo-random number generators for noise a periodicity of 2n
1, where n is the order of the polyno-
radar. Two different methods are considered for the noise mial. An example of x4 þ x3 þ 1 polynomial using the
generation in an FPGA environment. In the first method, LFSR is shown in Figure 1.
pseudo-random samples are generated by considering only
probability distributions. The pseudo-random noise (PRN)
generator can generate noise with uniform, Gaussian,

Authors’ current addresses: Mehmet Ispir, Alper Yil-

dirim, Tubitak Bilgem Iltaren, Ankara 06800, Turkey
(e-mail: mehmet.ispir@tubitak.gov.tr).
Manuscript received September 6, 2019; accepted May
15, 2020, and ready for publication May 22, 2020.
Review handled by Christoph Wasserzier. Figure 1.
0885-8985/20/$26.00 ß 2020 IEEE 4 Bit LFSR.

42 IEEE A&E SYSTEMS MAGAZINE SEPTEMBER 2020

Authorized licensed use limited to: Carleton University. Downloaded on September 13,2020 at 05:28:36 UTC from IEEE Xplore. Restrictions apply.
 Credit: Image licensed by Ingram Publishing

Since each next sample in LFSR is twice the current The Box–Muller algorithm can be represented by the
value with the addition of 0 or 1, a strong correlation is following two equations:
present between the successive numbers. This can be pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
seen from Figure 2, where two successive samples x0 ¼
2 ln ðu0 Þ cos ð2pu1 Þ (2)
ðx; x þ 1Þ from the sequence generated by the LFSR is pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
plotted. x1 ¼
2 ln ðu0 Þ sin ð2pu1 Þ (3)
This disadvantage of the LFSR structure can be elimi-
nated with use of leap-forward or multiple parallel LFSR where u0 and u1 are random variables with uniform distri-
structures [11]. Another method is use of combined LFSR bution (13) in the range of [0,1], i.e., a ¼ 0, b ¼ 1, x0 and
structures like Tausworthe generators [12]. To compare x1 are random variables with Gaussian (Normal) distribu-
the improvement in terms of correlation of successive tion (4) (zero mean (m) and unit variance (s 2 ), Nð0; 1Þ).
samples, Figure 3 presents correlation of samples of Taus- The Box–Muller algorithm as described generates two
worthe generator. independent Gaussian random numbers, so if only a
Using both leap-forward LFSR and Tausworthe gener- Gaussian is required only one of these is needed. As can
ators, uniformly distributed pseudo-random samples be seen in the equations, Box–Muller method can be used
whose pdfs given in (13) can be generated to produce random variables with Gaussian distribution

(4) by using uniformly distributed random numbers
b
a ; for a  x  b
1
fðxÞ ¼ (1) 1 ðx
m22 Þ
0; for x < a or x > b: fðxÞ ¼ e 2s
1 < x < 1: (4)
2ps 2
In (13), a and b are the extreme values, which can be
The structure of the Box–Muller algorithm provides
taken by the random number.
basis for generating expontial, Rayleigh, and Ricean dis-
Next distribution that finds common usage in differ-
tributed samples. Pdf of Rayleigh distributed x random
ent areas is the Gaussian distribution. There are differ-
variable is
ent methods for generating normally distributed
samples such as Ziggurat [12], inversion [13], Wal- x
x2
fðxÞ ¼ pffiffiffiffiffi e 2s2 0  x < 1: (5)
lace [14], and Box–Muller [15]. Noise samples with s2
Gaussian, exponential, and Rayleigh/Ricean distribu-
tions are generated using the Box–Muller algorithm, By definition, Rayleigh random variable is equal to the
which is widely used in the literature [16], [15], [17]. square root of the sum of squares of two independent Normal

Figure 2. Figure 3.
Correlation of successive samples which are generated with 32-bit Correlation of successive samples which are generated through
LFSR structure. 32-bit Tausworthe generator.

SEPTEMBER 2020 IEEE A&E SYSTEMS MAGAZINE 43

Authorized licensed use limited to: Carleton University. Downloaded on September 13,2020 at 05:28:36 UTC from IEEE Xplore. Restrictions apply.
 Real-Time Signal Generator for Noise Radar

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi i.e., x0  Nð0; s Þ and

2
distributed random qvariables,
x1  Nð0; s 2 Þ, r ¼ ðx20 þ x21 Þ where r  RayleighðsÞ.
Noise samples with Rayleigh distribution are generated by
using (6), which can be obtained by using the Box–Muller
equations for Normal variables
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
r¼
2 ln u0 : (6)
Figure 4.
The pdf of a Ricean distributed random variable is as Usage of IFFT for colored noise generation.
follows:
method of generating narrowband colored noise is to pass
x
ðx2 þm2 Þ xm an independently generated white noise through a band-
fðxÞ ¼ pffiffiffiffiffi e 2s2 I0 2 0  x < 1: (7)
s2 s pass linear time invariant (LTI) filter. The CPRN genera-
tor provides extensive control in noise signal design. As
By definition, Ricean random variable is equal to the opposed to the traditional technique, where the transfer
square root of the sum of squares of two independent Normal function of the LTI filter determines the PSD of the sig-
distributed random variables, i.e., x0  Nðm cos ðuÞ; s 2 Þ nal at the output, it is possible to generate colored noise
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
and x1  Nðm sin ðuÞ; s 2 Þ, r ¼ ðx20 þ x21 Þ, where with a given continuous PSD and with a given first-order
marginal pdf.
r  Riceðjmj; sÞ and u are any real number.
Usage of inverse fast Fourier Transform (IFFT) for
By definition, if u is uniformly distributed random
generating noise samples ([19]) can be implemented by
number,
ln ðuÞ= is a random number with exponential
supplying the frequency, phase, and amplitude values to
distribution. The pdf of an exponentially distributed x ran-
an IFFT algorithm to obtain the signal that is composed of
dom variable is as follows:
parallel cosine signals with different amplitude and phase

values. Figure 4 shows the idea of using IFFT for generat-
e
x ; if x  0
fðxÞ ¼ (8)
0; if x < 0: ing noise samples.

Up to this point, pseudo-random samples are gener-

ated by only considering the probability distribution. In FPGA IMPLEMENTATION
addition, there are other methods that focus on generating
samples with specified probability distribution and power Our real-time signal generator is created by combining
spectral density (PSD). In [18], a method is proposed to PRN and CPRN generators. The main block diagram is
generate realizations from a wide-sense stationary random given in Figure 5. The PRN generator is able to generate
process with a given continuous PSD and with a given samples with uniform, Gaussian, expontial, and Rayleigh
first-order marginal pdf. This method is based on the addi- distributions. The CPRN generator block generates col-
tion of cosine signals whose amplitude, phase, and fre- ored pseudo-random samples with the IFFT method. Out-
quency parameters are random variables. puts of these two generators are switched through a
The method uses the following equation: multiplexer and routed to the pulse generator block, which
PRN generates continuous and pulsed signals. Also the
1 X M pulse repetition frequency of the generated signals can be
X½n ¼ pffiffiffiffiffiffiffiffiffiffi Ai cos ð2pFi n þ Fi Þ (9) adjusted as constant, staggered, and random values.
M=2 i¼1
Finally, controller block provides peripheral component
where Ai , Fi , Fi are random variables with pdfs pA ðaÞ, interconnect express (PCIe) based control signals from
pF ðfÞ, pF ðfÞ and represent amplitude, frequency, and host PC for controlling and selection of distribution
phase, respectively. By changing these pdfs, the power
spectrum density and statistical distribution of the noise
samples can be controlled independently and samples
with desired characteristics can be obtained much more
easily. The derivation of the pdfs, i.e., pA ðaÞ, pF ðfÞ,
pF ðfÞ required to generate random samples with Gauss-
ian, Gamma, and K distributions are given in [18].
We investigated the method in [18] in terms of
usability of real-time generation of noise samples on
FPGA [19]. In [20], we compared the colored noise gen- Figure 5.
erator method with the traditional method. The traditional FPGA block diagram.

44 IEEE A&E SYSTEMS MAGAZINE SEPTEMBER 2020

Authorized licensed use limited to: Carleton University. Downloaded on September 13,2020 at 05:28:36 UTC from IEEE Xplore. Restrictions apply.
 Ispir and Yildirim

method, an approximation of an fðxÞ function is imple-

mented in three steps x and fðxÞ are in the ½a; b interval.
In the first step, x is transformed into y, which belongs to
½a0 ; b0 . Next, fðyÞ is calculated using an approximation to
the function as a second step. The approximation of fðyÞ
function in ½a;0 b0  interval is obtained by equally dividing
the ½a0 ; b0  interval and calculating the cn coefficients for
the polynomial equation that are given in the following
equation [15]:
Figure 6.
PRN generator block diagram. fðyÞ ¼ ððcn  y þ cn
1 Þ  y þ . . .Þ þ c0 (10)

parameters and can transfer digital noise samples to the cn coefficients are calculated using min–max approxima-
host PC. The details regarding the main blocks, PRN, and tion that is based on minimizing the maximum difference
CPRN generators are presented below. of the error value between the real and the approximated
function values. Finally, by implementing function spe-
cific interval reconstruction as explained as above, fðxÞ
function in ½a; b is obtained. Approximations are imple-
PRN GENERATOR mented to one unit in the last place (ulp) to provide true
The PRN generator is designed by improving our previous Gaussian PDF model.
work based on the Box–Muller algorithm ([17]) and capa- For the implementation of ln() function, the following
ble of generating uniform, normal, Rayleigh, and expo- equations are used:
nential distributions. The main block diagram of the
design is given in Figure 6. x ¼ M2f (11)
Uniformly distributed samples are generated by two 32-
bits Tausworthe RNG blocks. Each block uses the Taus- ln ðxÞ ¼ ln ðMÞ þ f  ln ð2Þ: (12)
worthe noise generation method, which provides better ran-
domness with lower resource utilization [15]. The python The main purpose in these equations is to obtain a
code for a 32 bit Tausworthe generator is given in Figure 7. range reduced lnðÞ value for piecewise polynomial
The use of two Taus RNG blocks is necessary for model- approximation. In these equations, x and M take values in
ing Normally distributed samples with Box–Muller method. [0,1) and [1,2), respectively. f takes integer values in
Uniform samples with 48-bit can model the normally distrib- [1,48] and determines the number of necessary shifts for
uted samples with 8.15s [17]. In the ideal case, this value transforming x to M. In (12), lnðxÞ can be calculated by
must be 1. The distribution quality of the generated samples approximating lnðMÞ with polynomial approximation.
depends on this value and increases with it. We used a second-order polynomial approximation for
To implement, sqrtðÞ, lnðÞ, and sin/cos functions calculating this value.
piecewise polynomial approximation is used [15]. In this For the implementation of sqrtðÞ function, the follow-
ing equations are used:

pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
pffiffiffiffiffi
M  2f=2 ; for f even
M  2f ¼ pffiffiffiffiffiffiffiffiffi (13)
2 M  2ðf
1Þ=2 ; for f odd:

Calculation of sqrtðÞ in the Box–Muller algorithm,

first range reduction of the p
2 ffiffiffiffiffiln ðxÞpvalue,
ffiffiffiffiffiffiffiffiffi [0,66.15], is
realized using (13). Next, M or 2 M is calculated
by using a first-order polynomial
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiapproximation. After the
range reconstruction,
2 lnðxÞ is obtained in the range
[0,8.15].
Due to symmetry of the sine and cosine functions,
only the first quarter period of the cosine function with a
first-order polynomial equation is approximated and used
for other quarters of sine and cosine functions. For sin/cos
functions, the look-up table methods can also be used,
Figure 7. which could be faster, whereas due to minimum area con-
Sample python function for generating Tausworthe samples. siderations polynomial approximation is chosen.

SEPTEMBER 2020 IEEE A&E SYSTEMS MAGAZINE 45

Authorized licensed use limited to: Carleton University. Downloaded on September 13,2020 at 05:28:36 UTC from IEEE Xplore. Restrictions apply.
 Real-Time Signal Generator for Noise Radar

Table 1.

Resource Utilization for Noise

Generator Block

Resource Type Used

Slice Register 46732 (%5.39)

Figure 8. Slice LUTs 26465 (%6.1)

CPRN generator block diagram. Block RAMs 20 (%1.36)
DSP48E 27 (%7.55)
After realizing the required functions of the Box–
Muller algorithm, Gaussian distributed samples are
obtained. By using x2 and sqrt blocks in the Figure 6, For the IFFT block we use our previous FFT/IFFT
square root of the sum of squares of Gaussian distributed implementation in overlap-add filter design [21]. This
samples are taken and Rayleigh distributed samples are design is implemented with radix-4 architecture and utilize
obtained if v ¼ 0, else Ricean distributed samples are the FPGA resources given in Table 1. Due to low FPGA
created. resource usage and ease of implementation, 256 point
radix-4 is found suitable compared to radix-2 and split-
radix architectures [14].
CPRN GENERATOR Resource utilization of the single instance of noise
generator design on a Xilinx Ultrascale+ RFSoC FPGA is
The block diagram for CPRN generator is given in
given in Table 1. With a single instance, 250 million noise
Figure 8.
samples/s can be generated. To increase the throughput,
This block can generate noise samples with Gaussian
multiple instances of the noise generator can be used by
pdf and uniform PSD. Amplitude Gen block is designed to
duplicating generator block with a simple generate
generate Rayleigh distributed samples. The same Box–
statement.
Muller structure for Rayleigh distributed samples in the
PRN block is used for this block. Phase Gen block is used
for generating phase samples that is independent of the
characteristics of the amplitude and the frequency sam- RESULTS AND DISCUSSIONS
ples. These samples must have uniform distributions and In order to show the effectiveness of the design, first PRN
are generated by using Tausworthe RNG blocks. generator is used and digital noise samples are generated
Amplitude and phase samples are used for obtain- with different distributions by adjusting PRN parameters
ing the real and imaginary inputs to the IFFT. If we such as mean and variance and controlling switching mul-
represent these values as IFFTreal , IFFTimag , respec- tiplexer through PCIe-based control registers. Generated
tively, they can be calculated as in (14), where (N) is samples are fed into both 12 bit DAC and DMA interface.
the length of IFFT, A is the amplitude, and f is phase
with sin/cos functions

IFFTreal ¼ N=2  A  cos ðfÞ

(14)
IFFTimag ¼ N=2  A  sin ðfÞ:

The frequency generator block generates uniformly

distributed samples for defined uniform frequency inter-
val. To generate noise samples within the defined fre-
quency range, the FFT bins in IFFT block must be fed
according to the following equation:

Fi  Fs Fi
IFFTbin ¼ ¼ (15)
Fs =N N

where IFFTbin is the corresponding frequency bin num-

ber, Fs is the sampling frequency, and Fi is the generated Figure 9.
frequency values. 50 000 digital samples generated with PRN Block.

46 IEEE A&E SYSTEMS MAGAZINE SEPTEMBER 2020

Authorized licensed use limited to: Carleton University. Downloaded on September 13,2020 at 05:28:36 UTC from IEEE Xplore. Restrictions apply.
 Ispir and Yildirim

Figure 10. Figure 12.

Uniformly [uniformð0; 1Þ] distributed samples. Exponentially [exponentialð1Þ] distributed samples.

Figure 11. Figure 13.

Normally [normalð0; 1Þ] distributed samples. Rayleigh [Rayleighð1Þ] distributed samples.

After generating the noise samples, digital values are case; generated noise samples are given with the compari-
transferred to host PC via PCIe interface. One million son of pdf and PSDs. Also, spectrum plots of the samples
snapshots from continuously generated samples for corre- are included. As seen from the results, colored noise sam-
sponding distributions are given in Figure 9. ples can be generated with arbitrary uniform frequency
The main reason behind using digital samples instead bands.
of analog signal through DAC is to prevent the inherent As stated in Section II, the CPRN generator pro-
negative effect of DAC internals, i.e., the anti-aliasing fil- vides extensive control in noise signal design. The noise
ters, on distributions. The nonideal responses of the antial- samples generated with the CPRN generator are not
ising filters (Chebychev, Gaussian and Legendre family of subject to filtering, unlike the traditional method where
filters [22]) change the smoothness and sharpness of the the colored noise samples are generated by filtering the
DAC response and hence the correlation between the digi- white noise samples, and therefore they are not affected
tal samples. by filter characteristics. The continuous and the pulsed
To test pdf distributions of the samples generated, his- waveforms can easily be generated in various combina-
togram plots of digital samples are obtained and corre- tions by using noise samples with different bandwidths
sponding pdf curves are fitted to the each of the histogram and center frequencies.
results. The results are given in Figures 10–13 for uni-
form, Gaussian, exponential, and Rayleigh distributions,
respectively. As seen from the figures, histograms of the
samples obtained coincides with the probability density
CONCLUSION
curve of the corresponding distributions. This article presents the design of real-time pseudo-
Next, the CPRN generator is used and colored noise random number generators for different probability
samples with Gaussian pdf and uniform PSD are obtained. distributions. Also, normally distributed samples with
Three different frequency bands, such as [0.1,0.2], uniform power spectral distribution have been gener-
[0.2,0.3] and [0.3,0.4] are considered for testing the sam- ated. This design can be used for testing noise radar
ples generated and results are plotted in Figure 14. In each detection performance for different noise distributions.

SEPTEMBER 2020 IEEE A&E SYSTEMS MAGAZINE 47

Authorized licensed use limited to: Carleton University. Downloaded on September 13,2020 at 05:28:36 UTC from IEEE Xplore. Restrictions apply.
 Real-Time Signal Generator for Noise Radar

Figure 14.
Simulation results for CPRN generator based samples. (a) Colored noise samples with PSD Uð0:1; 0:2Þ. (b) Colored noise samples with PSD
Uð0:2; 0:3Þ. (c) Colored noise samples with PSD Uð0:3; 0:4Þ.

Our future work will be the design of a noise radar [8] S. Lukin, O. Zemlyaniy, and K. Lukin, “FPGA based ran-
simulator by adding a receiver correlator (matched dom waveform generators for noise radars,” in Proc. 16th
filter) structure to the noise generator design with a Int. Radar Symp., Jun. 2015, pp. 777–782.
radar target generator. [9] K. A. Lukin, O. V. Zemlyaniy, D. N. Tatyanko, S. Lukin,
and V. Pascazio, “Noise radar design based on FPGA
technology: On-board digital waveform generation and
REFERENCES real-time correlation processing,” in Proc. 18th Int. Radar
Symp., Jun. 2017, pp. 1–7.
[1] L. Guosui, G. Hong, and S. Weimin, “Development of
[10] K. Lukin, S. Lukin, J. Moreira, and R. Spielbauer,
random signal radars,” IEEE Trans. Aerosp. Electron.
“Software defined noise radar on the basis of FPGA based
Syst., vol. 35, no. 3, pp. 770–777, Jul. 1999.
SPOS board,” in Proc. 3rd Int. Asia-Pacific Conf.
[2] K. Kulpa, Signal Processing in Noise Waveform Radar,
Synthetic Aperture Radar, Sep. 2011, pp. 1–2.
ser. Radar. Norwood, MA, USA: Artech House, 2013.
[11] P. P. Chu and R. E. Jones, “Design techniques of FPGA
[Online]. Available: https://books.google.com.tr/books?
based random number generator,” in Proc. Mil. Aerosp.
id=agIgAgAAQBAJ
Appl. Programmable Devices Technol. Conf., 1999, vol. 1,
[3] G. Galati and G. Pavan, “Noise radar technology as an
no. 999, pp. 28–30.
interference prevention method,” J. Elect. Comput. Eng.,
[12] G. Zhang, P. H. W. Leong, D.-U. Lee, J. D. Villase-
vol. 2013, pp. 146 986:1–146 986:6, 2013.
nor, R. C. C. Cheung, and W. Luk, “Ziggurat-based
[4] T. Thayaparan and C. Wernik, “Noise radar technology
hardware Gaussian random number generator,” in
basics,” Defence Res. Develop. Canada Ottawa, ON,
Proc. Int. Conf. Field Programmable Logic Appl.,
Canada, Tech. Rep. 2006-266, 2006.
Aug. 2005, pp. 275–280.
[5] K. Lukin, P. Vyplavin, O. Zemlyaniy, V. Palamarchuk, and S.
[13] D. Lee, R. C. C. Cheung, J. D. Villasenor, and W. Luk,
Lukin, “High resolution noise radar without fast ADC,” Int. J.
“Inversion-based hardware Gaussian random number gen-
Electron. Telecommun., vol. 58, no. 2, pp. 135–140, 2012.
erator: A case study of function evaluation via hierarchical
[6] X. Xu and R. M. Narayanan, “Impact of different correla-
segmentation,” in Proc. IEEE Int. Conf. Field Program-
tion receiving techniques on the imaging performance of
mable Technol., Dec. 2006, pp. 33–40.
UWB random noise radar,” in Proc. IEEE Int. Geosci.
[14] D.-U. Lee, W. Luk, J. D. Villasenor, G. Zhang, and
Remote Sens. Symp. (IEEE Cat. No.03CH37477), Jul. 2003,
P. H. W. Leong, “A hardware Gaussian noise generator
vol. 7, pp. 4525–4527.
using the Wallace method,” IEEE Trans. Very Large Scale
[7] J. Sachs et al., “Ultrawideband pseudonoise sensors,”
Integr. Syst., vol. 13, no. 8, pp. 911–920, Aug. 2005.
Appl. Radio Electron., vol. 12, no. 1, pp. 79–88, 2013.

48 IEEE A&E SYSTEMS MAGAZINE SEPTEMBER 2020

Authorized licensed use limited to: Carleton University. Downloaded on September 13,2020 at 05:28:36 UTC from IEEE Xplore. Restrictions apply.
 Ispir and Yildirim

[15] D. . Lee, J. D. Villasenor, W. Luk, and P. H. W. Leong, [19] M. Ispir, O. Ozdil, and A. Yildirim, “Coloured noise gen-
“A hardware Gaussian noise generator using the Box– eration with IFFT,” in Proc. IET Intell. Signal Process.
Muller method and its error analysis,” IEEE Trans. Conf., Dec. 2013, pp. 1–4.
Comput., vol. 55, no. 6, pp. 659–671, Jun. 2006. [20] A. Yildirim, M. Ispir, and O. Ozdil, “Method for generat-
[16] G. Box et al., “A note on the generation of random normal ing noise radar signals,” in Proc. Int. Conf. Radar Syst.,
deviates,” Ann. Math. Statist., vol. 29, no. 2, pp. 610–611, Oct. 2017, pp. 1–5.
1958. [21] O. Ozdil, M. Ispir, E. Onat, and A. Yildirim,
[17] O. Ozdil and A. Yildirim, “Hardware noise generator,” in “Implementation of FPGA-based FFT convolution,” in
Proc. IEEE 19th Signal Process. Commun. Appl. Conf., Proc. IET Int. Conf. Radar Syst., Oct. 2012, pp. 1–4.
Apr. 2011, pp. 538–541. [22] K. Gentile, “The effect of DAC resolution on spurious
[18] S. Kay, “Representation and generation of non-Gaussian performance,” A Technical Tutorial on Digital Signal
wide-sense stationary random processes with arbitrary Synthesis, Section 4, Analog Devices Inc., Norwood, MA,
PSDs and a class of PDFs,” IEEE Trans. Signal Process., USA, 1999.
vol. 58, no. 7, pp. 3448–3458, Jul. 2010.

SEPTEMBER 2020 IEEE A&E SYSTEMS MAGAZINE 49

Authorized licensed use limited to: Carleton University. Downloaded on September 13,2020 at 05:28:36 UTC from IEEE Xplore. Restrictions apply.