Global Journal of Pure and Applied Mathematics.
ISSN 0973-1768 Volume 14, Number 4 (2018), pp. 547-560
ยฉ Research India Publications
http://www.ripublication.com
Performance of IF Estimation for Variable-Slope
Linear FM Signals under Color Noise using
Time-Frequency Distributions
Safaa D. Al-Khafaji 1, Zahir M. Hussain2
1Faculty of Computer Science and Mathematics, 2MSc Student University of Kufa, Najaf, Iraq.
Abstract
This work presents a study on the performance of instantaneous frequency (IF)
estimation of linear FM (LFM) signal with different slopes under colored noise
(CN). Peak of a specific time-frequency distribution (TFD), the periodogram,
has been used for IF estimation of LFM signals under colored noise. Simulation
results of IF estimators of LFM signal for many slopes using TFD showed that
when the FM slope increases, frequency estimation will have more error for
lower signal-to-noise ratios (SNR).
Keywords: Instantaneous frequency (IF); Fourier transform (FT); colored noise
(CN); linear FM (LFM); time-frequency distribution (TFD).
1. INTRODUCTION
Many signals in practice are nonstationary, such as FM signal found in communication,
radar where there are many literatures studied it in order to present processing approach
such as Radar-Ambiguity transform, Wigner-Hough Transform, Nonlinear Least-
squares Method and Wigner-Ville distribution [1]. Those signals can be classified as
either mono-component or multicomponent signals [2]. Instantaneous frequency (IF) is
one of the essential parameter of non-stationary signal which extracts important
information about the changes of time-varying spectral in non-stationary signals. The
importance of instantaneous frequency which was a basic and fundamental tool for
describing multicomponent non-stationary signals is another reason of frequency
estimation, where the amplitude and phase can be computed using estimated frequency
[3][4]. In many applications, instantaneous frequency estimation of a nonstationary
signal is important. Instantaneous frequency of nonstationary signals can be described
by time-frequency distributions, such as the Wigner-Ville distribution (WVD) [5],
which produced an advanced concept of instantaneous frequency (IF) and produced
548 Safaa D. Al-Khafaji, Zahir M. Hussain
different time-frequency distributions-based estimation methods [6]. The definition of
IF has further been improved by Hussain and Boashash [7]. Estimation of instantaneous
frequency of FM signals using TFD is of greater accuracy than methods using FT and
correlation [8]. There are different types of noise encountered in different applications
which can fall under several categories. For example, in some applications noise can
affect the amplitude of a signal which would mean itโs additive in nature; noise can also
affect the signalโs phase or can be multiplicative [9]. In this paper, we consider colored
Gaussian noise; which will be added to LFM signal to test the performance of TFD in
extracting the actual frequencies for noisy LFM signal with different frequency slopes.
This paper is organized as follows: The analytic signal will be discussed in section 2.
In section 3 colored Gaussian noise will be presented. In section 4 IF estimation of FM
signal based on time-frequency distributions (TFD) is presented and lastly in section 5
results and discussion are presented.
2. THE ANALYTIC SIGNAL
We can produce the analytic signal in two ways [10]:
1. Direct method: producing the analytical signal by utilizing its definition. This way
is executed by utilizing Fourier transform (FT) of the signal, thereafter adjust the
negative frequency value to zero; the last step is to utilize inverse Fourier
transform. The disadvantage of this way is the existence unwanted rolls in the
signal.
2. The second method is producing the desired complex component of the signal by
utilizing the Hilbert transform (HT) filter, where the analytical signal ๐ง(๐ก) is
produced by adding the real part of signal ๐ฅ(๐ก) to this complex component
obtained from Hilbert transform as follows:
๐ง(๐ก) = ๐ฅ(๐ก) + ๐ป{๐ฅ(๐ก)} = ๐ฅ(๐ก) + ๐. ๏ฟฝฬ๏ฟฝ(๐ก) (1)
โด ๐(๐) = ๐(๐) + ๐[โ๐sgn(๐)]๐(๐) = ๐(๐)[1 + sgn(๐)] (2)
โด ๐(๐) = {0 ๐<02๐(๐) ๐โฅ0
(3)
The advantage of using the analytical signal is that we can sample at one-half the
Nyquist rate because analytical signal does not contain negative frequencies; this is a
useful property of analytical signal [11]. The analytical signal ๐ง(๐ก) associated with the
original signal ๐ฅ(๐ก) can be generated in MATLAB using the command hilbert(x) [12],
[13].
3. COLORED GAUSSIAN NOISE (CGN)
Most of the noise modeling in communications systems is based on white noise,
however, in addition to white noise; there is also colored noise which can also impact
Performance of IF Estimation for Variable-Slope Linear FM Signals under Colorโฆ 549
on communication systems due to limited bandwidth. Here we can generate color noise
via filtering white noise using FIR filer with a specific length N.
Finite impulse response (FIR) digital filters have impulse responses that โ(๐) contain
a finite number of non-zero samples. In a digital system, the output of a causal FIR
filter, ๐ฆ(๐) given a finite sequence ๐ฅ(๐) was input into the system can be represented
by the discrete-time convolution between โ(๐) and ๐ฅ(๐):
๐ฆ(๐) = โ(๐) โ ๐ฅ(๐) (4)
= โ โ(๐)๐ฅ(๐ โ ๐) = โ0๐ฅ(๐) + โ1๐ฅ(๐ โ 1) + โฏ + โ๐โ1๐ฅ[๐ โ (๐ โ 1)] (5)
๐โ1
๐=0
where โ๐ is utilized for โ(๐) for simplicity. Take the z-transform of both sides we
getting:
๐(๐ง) = โ0๐(๐ง) + โ1๐งโ1๐(๐ง) + โฏ + โ๐โ1๐งโ(๐โ1)๐(๐ง) (6)
Hence, the transfer function given by:
๐ป(๐ง) =๐(๐ง)
๐(๐ง)= โ0 + โ1๐งโ1 + โฏ + โ๐โ1๐งโ(๐โ1) (7)
From the difference equation or from the transfer function ๐ป(๐ง) we can implement a
causal FIR filter by utilizing delay elements and digital multipliers.
The Moving Average Filter (MA) is commonly used to smooth data to show general
trends and behavior of a sequence. Using a MA filter can improve our ability to make
a decision or interpret our data. In this work, a moving average filter will be used to
color noise simulation after filtering of Gaussian noise [12].
In MATLAB, simulated colored noise can be obtained by on Gaussian noise through
FIR filter [8]. In signal processing systems, the integrity and quality of systems can be
realized by understanding the statistical characteristics of the noise process associated
with the system. These noise processes are generated by electromagnetic or electronic
sources. Considering colored noise, the statistical model is filtered Gaussian [9].
Noise Model:
The main noise process is assumed to have the Gaussain probability density function
(pdf) with variance (power) ๐2 as follows [14]:
๐(๐ฅ) =1
๐โ2๐ ๐โ๐ฅ2/2๐2
(8)
where ๐ฅ is random variable and ๐ is standard deviation of the noise.
Modeled color Gaussian noise (CGN) in shown in the flowchart of Figure 1 [8].
550 Safaa D. Al-Khafaji, Zahir M. Hussain
Figure 1: Flowchart of modeling colored Gaussian noise (CGN)
4. INSTANTANEOUS FREQUENCY OF FM USING TIME-FREQUENCY
DISTRIBUTIONS (TFDS)
Non-stationary signals with time-varying frequency content (such as frequency
modulated (FM) signals and biomedical signals) cannot be analyzed effectively using
the Fourier transform (FT) as the FT cannot detect the time-varying features of the
signal. This is due to the time-averaging process (time-integration) used in the FT.
Time-frequency distributions (TFD) are double transforms from the time-domain into
the time-frequency domain representing the Fourier transform of the instantaneous
autocorrelation of the analytic signal. The simplest formula for time -frequency
distribution (TFD) is a windowed frequency distribution called the short-time Fourier
transform (STFT) is defined as [12]:
๐๐ง(๐ก, ๐) = โซ ๐ง(๐)โ(๐ โ ๐ก)๐โ๐2๐๐๐๐๐
โ
โโ
= FT๐โ๐
{๐ง(๐)โ(๐ โ ๐ก)} (9)
where ๐ง(๐) is the analytic signal, โ(๐ โ ๐ก) is a time window, ๐ is dimension of window,
Performance of IF Estimation for Variable-Slope Linear FM Signals under Colorโฆ 551
๐ก is time index and ๐ is frequency index. Solve the optimize problem in equation (22),
the instantaneous frequency (IF) can be estimate as mentioned in [2], [7] [15] [16]:
๐๐ ฬ (๐ก) = arg[max๐๐๐ง(๐ก, ๐)] , 0 โค ๐ โค๐๐
2 (10)
TFD is a revealer representation of the non-stationary signal because it shows the
distribution of signal energy over two-dimensional domain: the time -frequency
domain.
In this study, the periodogram (which is| ๐๐๐น๐|2) is use to estimate the IF of linear
frequency modulation (LFM) signal. Modelled and simulated IF estimation based on
TFD method is in the flowchart of Figure (2):
Figure 2: Flowchart for modelling and simulation IF estimation method based on
TFD for noisy FM signal.
5. SIMULATION RESULTS
In MATLAB, we design three FIR filters with lengths W=01, W=01, W=01 for
simulating CGN as in Figure (1). In this simulation, SNR ranges from โ20 to 80 ๐๐ต.
For the experiments performed in this paper we used a LFM noisy signal simulated with
CGN added to it, this signal has the following form:
๐ฅ(๐ก) = ๐ด โ sin(โ (๐ก)) + ๐(๐ก)
552 Safaa D. Al-Khafaji, Zahir M. Hussain
where ๐ด is the amplitude, which for these experiments was set to ๐ด = 1 volt and โ is
the angle function for the LFM signal which can be used to calculate the instantaneous
frequency of our signal:
๐ค๐(๐ก) =๐โ
๐๐ก, and ๐ค๐(๐ก) = 2๐ ๐๐(๐ก) radians/sec
This means that the instantaneous frequency in Hertz can be given by:
๐๐(๐ก) =1
2๐
๐โ
๐๐ก
We set the angle function to:
โ = (๐0 โ ๐ก + ๐ผ โ๐ก2
2) โ (2๐)
Where ๐ผ = [0.01 10 50] is the slope values vector, and ๐0 = 28 Hz, and ๐(๐ก) is CGN
process.
This results in the final LFM signal given in Equation (11):
๐ฆ(๐ก) = ๐ด sin( (๐0 โ ๐ก + ๐ผ โ๐ก2
2) โ (2๐))
+ ๐(๐ก) (11)
The simulation of this signal was performed with a time length of ๐ฟ = 10s, a sampling
interval ๐๐ = 0.01s, sampling frequency ๐๐ = 10 ๐ป๐ง, and a total number of samples of
๐ = ๐ฟ/๐๐ .
Monte Carlo simulations were performed with ๐ = 30 realizations. We used a Hilbert
transformation (HT) to obtain the analytic signal ๐ง(๐ก) associated with the original
signal ๐ฅ(๐ก) before estimation, to avoid aliasing by using Equation (1). We estimated the
IF by taking the peak (max) of the TFD. Figure (9) shows the contour plot of the TFD
for the noisy LFM signal with the theoretical computed IF of a noiseless FM signal
(given by the dotted line).
Finally, we draw a mid-time IF estimation for each SNR in Figures (10), (12), (14) with
slopes [0.01, 10, 50], respectively, and calculated the relative mean-square error (MSE)
of the IF estimation at mid-time of the LFM signal for each SNR for different values of
filter length (W) in Figures (11), (13), (15) with different slopes.
Figure (3) shows colored Gaussian noise (CGN) when window of filter W=30 and a=1,
by passing Gaussian noise through this FIR filter.
One may compute autocorrelation function by using ๐๐ โ xcorr function, where ๐๐ is
sampling period. It is obvious that the auto-correlation function of Gaussian operation is
nearly a weighted delta function in the lag domain, which means that their samples (in
the time domain) are weakly-related to one another. Figure (4) shows the autocorrelation
for colored Gaussian noise (CGN).
Performance of IF Estimation for Variable-Slope Linear FM Signals under Colorโฆ 553
Figure 3: Colored Gaussian noise (CGN) when ๐ = 30.
Figure 4: Autocorrelation of colored Gaussian noise (CGN) when ๐ = 30.
554 Safaa D. Al-Khafaji, Zahir M. Hussain
Figure 5: (A) Time versus LFM signal, (B) Time versus AWGN, and (C) Time
versus noisy signal (y)
Figure 6: (A) IF versus LFM signal, (B) IF versus AWGN, and (C) IF versus noisy
signal (y)
A
B
C
A
B
C
Performance of IF Estimation for Variable-Slope Linear FM Signals under Colorโฆ 555
Figure 7: 3D Spectrogram plot of noiseless LFM signal.
Figure 8: TFD slice at time index n=10.
556 Safaa D. Al-Khafaji, Zahir M. Hussain
Figure 9: Contour plot of the TFD of noisy LFM signal.
Figure 10: Mid-time IF estimation of LFM signal versus SNRs with different W
when E=0.01.
Performance of IF Estimation for Variable-Slope Linear FM Signals under Colorโฆ 557
Figure 11: MSE of IF estimation at mid-time of LFM signal versus SNRs with
different W and E=0.01.
Figure 12: Mid-time IF estimation of LFM signal versus SNRs with different W
when E=10.
558 Safaa D. Al-Khafaji, Zahir M. Hussain
Figure 13: MSE of IF estimation at mid-time of LFM signal versus SNRs with
different W and E=10.
Figure 14: Mid-time IF estimation of LFM signal versus SNRs with different W
when E=50.
Performance of IF Estimation for Variable-Slope Linear FM Signals under Colorโฆ 559
Figure 15: MSE of IF estimation at mid-time of LFM signal versus SNRs with
different W and E=50.
CONCLUSIONS
Time-frequency analysis has been used to estimate the IF law for linear frequency-
modulated (LFM) signal with a Gaussian model for colored noise. From the results we
noted that the problem of noise is less serious if we use TFD since TFD is 2D not 1D;
hence TFD spreads noise on a plane not on a line, making it less dense. Effect of LFM
slope on IF estimation using spectrogram: for different slopes, TFD can estimate
frequency correctly; this is because TFD spreads noise. However, when slope E
increases, frequency estimation may not reach steady-state fast. In figures 11 and 12 it
goes beyond SNR=30dB for large W. More error at lower of index of middle time.
REFERENCES
[1] Zhang, X., Liu L., Cai J., and Yang Y., โA Pre-Estimation Algorithm for LFM
Signal Based on Simplified Fractional Fourier Transformโ, Journal of Information and Computational Science, pp. 645-652, 2011.
[2] El-Jaroudi A. et al., Instantaneous Frequency Estimation and Localization. A
book chapter in Time-Frequency Signal Analysis and Processing: A
Comprehensive Reference, B. Boashash Editor, Elsevier 2016.
[3] Liao, Y., Phase and Frequency Estimation: High-Accuracy and Low-Complexity Techniques, M.Sc. Thesis, Worcester Polytechnic Institute, United
States, 2011.
[4] Emresoy, M. K., and El-Jaroudi, A., โIterative Instantaneous Frequency
560 Safaa D. Al-Khafaji, Zahir M. Hussain
Estimation and Adaptive Matched Spectrogramโ, Elsevier, Signal Processing,
vol. 64, no. 2, pp. 157โ165, 1998.
[5] Cohen, L., โTime-Frequency Distributionsโ, Proceedings of the IEEE, vol. 77,
no. 7, pp. 941-981, 1989.
[6] Boashash, B., Jones, G., and O'Shea, P., โInstantaneous Frequency of Signals:
Concepts, Estimation Techniques and Applicationsโ, Proceedings of SPIE, โAdvanced Algorithms and Architectures for Signal Processing IVโ, vol. 1152,
pp. 382-400, 1989.
[7] Hussain, Z. M., and Boashash, B., โAdaptive Instantaneous Frequency
Estimation of Multicomponent FM Signals Using Quadratic TimeโFrequency
Distributionsโ, IEEE Transactions on Signal Processing, vol. 50, no. 8, pp.
1866-1876, 2002.
[8] Al-Khafaji, S. D., Hussain, Z. M., and Katrina Nerille, โFrequency Estimation
of Fm Signals Under Non-Gaussian and Colored Noiseโ, International Journal
of Applied Engineering Research, vol. 12, no. 22, pp. 12342-12356, 2017.
[9] V. P. Tuzlukov, โSignal Processing Noiseโ, 2002.
[10] Boashash, B., Ed., โTime-Frequency signal analysis: methods and
applicationsโ, Longman Cheshire, Melbourne, Australia, 1992.
[11] Spataru, A., โTheorie De La Transmission De /โInformation -1: Signaux Et
Bruits, translation from Editura Technica, โTheory of Transmission /
Information -1: Signal and Noiseโ, Bucarest, Romania, 1970.
[12] Hussain, Z. M., Sadik, A. Z., and OโShea, P., Digital Signal Processing,
Springer, Germany, 2011.
[13] Oppenheim, A. V., and Schafer, R. W., Discrete-Time Signal Processing, NJ:
Prentice-Hall, Englewood Cliffs, New Jersey, United States, 1989.
[14] Hogg, R. V., and Craig, A. T., Introduction to Mathematical Statistics,
Macmillan Publishing Co., Inc., New York, 1978..
[15] Hussain, Z. M., and Boashash, B., โAdaptive Instantaneous Frequency
Estimation of Multicomponent FM Signalsโ, Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP 2000), vol. II, pp. 657โ660, Istanbul, 2000.
[16] Hussain, Z. M., and Boashash, B. โDesign of Time-Frequency Distributions for
Amplitude and IF Estimation of Multicomponent Signalsโ, invited paper,
International Symposium on Signal Processing and Its Applications
(ISSPA'2001), vol. 1, pp. 339-342, 2001.