2010 First International Conference on Pervasive Computing, Signal Processing and Applications

Frequency Estimation Based on Modulation FFT and MUSIC Algorithm

Tong Li, Yinhui Tang

Department of Information Engineering
Academy of Armored Force Engineering
Beijing, China
[email protected]

Abstract—Compared with traditional Fast Fourier Transform

(FFT), modulation FFT preserves its characteristics of low II. MODULATION FFT AND MUSIC ALGORITHM
complexity when used to estimate the frequencies of real-
valued signals. But there are spectral leakage and picket-fence A. Modulation FFT Algorithm
effect, frequency resolution is still difficult to achieve high Modulation FFT[2] shift spectrum of the N-point real
accuracy. Multiple Signal Classification (MUSIC) algorithm sequence firstly, so that the spectrum is no longer
can achieve high frequency estimation accuracy, but it requires symmetrical. Then compute this new signal’s FFT, and
searching the spectral peak in the full range frequency domain, reconstruct the N spectral lines to indicate the original real
and this takes a long time. To meet the needs of high accuracy sequence’s positive frequency components.
and real-time performance when estimating the hopping
Assuming a real-valued signal whose sampling rate is
frequencies, an algorithm combining modulation FFT and
fs
MUSIC algorithm is proposed, which can achieve high , and real part of its Fourier transform between − f s / 2
frequency resolution and reduce the search range of MUSIC f s / 2 is shown in Fig. 1.
algorithm’s spectral peak. Simulation results show that this and
algorithm can achieve high frequency resolution, and the real-
time performance can meet the demand of frequency
measurement.

Keywords-component; frequency estimation; modulation FFT;

MUSIC algorithm; frequency resolution

I. INTRODUCTION
While handling a N-point real-valued sequence with
traditional FFT, because of the symmetry of the
spectrum[1],only N/2 spectral lines can be used to estimate
the frequency. In reference [2], real-valued signal is loaded
into complex-valued signal by time domain modulation. Figure 1. Spectrogram before and after shifting frequency
Then with relations between the modulated signal and the
original real signal, spectrogram whose resolution is twice as Let X cR ( f ) denotes real part of its DTFT. With
high as the traditional FFT. But there are sampling time’s
constraint and influence of spectral leakage and picket-fence periodicity and symmetry of the real sequence’s spectrum,
effect, frequency resolution is still very limited. Multiple we know
Signal Classification (MUSIC) algorithm [3] which is based
on model of harmonic signals can estimate frequency to X cR ( f ) = X cR ( fs − f ) f s / 2 < f ≤ f s
infinite precision. But this algorithm has to search spectral , . (1)
peak in the full range frequency domain, which need a large
amount of calculation and the real-time performance is Now the interval between two spectral lines or the frequency
affected. resolution is Δf = f s / N = 1/ T .
To meet the needs of real-time performance and high To take full advantage of the real sequence’s value to
precision, combined with characteristics of modulation FFT improve the frequency resolution, it’s modulated to a
and MUSIC algorithm, a modulation FFT-MUSIC algorithm complex sequence. This is equivalent to the frequency
is proposed, which search the spectral peak in a specific shifting in frequency domain. Here we consider shifting the
range defined by modulation FFT algorithm. This algorithm
spectrum 0.25Δf to the left firstly, and then compute its
can achieve high frequency resolution, and greatly reduce the
computation of MUSIC algorithm FFT. Now the expression in time domain we get is

y (n) = x(n) ⋅ e− j⋅1/4⋅2π ⋅n / N (2)

978-0-7695-4180-8/10 $26.00 © 2010 IEEE 526

525
DOI 10.1109/PCSPA.2010.132
10.1109/PCSPA.2010.125
and its spectrum is shown as the solid dots in Fig. 1. With ⎡ x (n) ⎤
Fig. 1 we know that after frequency shifting, the N spectral x ( n) = ⎢ # ⎥ = A ( w)S( n) + W (n)
⎢ ⎥ (4)
lines are no longer symmetry, and all of them can be used to ⎢⎣ x(n + M − 1) ⎥⎦
estimate the signal frequency effectively.
⎤ ⎡ A1e 1 1 ⎤ ⎡ w( n)
j ( ω n +ϕ )
⎡ 1 1 " 1 ⎤
Like what Fig. 2 shows ,the following N/2 lines are ⎢ jω1 jω2 ⎥⎢ ⎥ ⎢ ⎥
inserted to the first N/2 lines at one point intervals reversely. e e " e jωr ⎥ ⎢ A2e
j (ω2 n ) +ϕ 2 )
⎥ + ⎢ w( n + 1) ⎥
=⎢
The new spectral lines and the original real sequence’s ⎢ # % ⎥⎢ # ⎥ ⎢ # ⎥
⎢ j ( M −1)ω j ( M −1)ω ⎥⎢ ⎥ ⎢ ⎥
DTFT coefficients’ relationship is ⎣⎢ e
1
e 2
" e j ( M −1)ωr ⎦⎥ ⎣⎢α n e r n
j (ω n +ϕ )
⎦⎥ ⎣ w( n + M − 1) ⎦
YR′ (k ′) = X cR ((0.5k ′ + 0.25)Δf ) . (3) T
Define a( wi ) = ⎡⎣ 1 e jωi " e j ( M −1)ωi ⎤⎦ and

A( w) = [a( w1 )a( w2 )…a( wr )] .Now covariance matrix

of the observational data vector
x (n) is

R  E {x (n)x * (n)} = APA H + σ w2 I . (5)

Characteristic decomposition of R [5] is

⎡ Σ′ 0 ⎤ ⎡ U sH ⎤
R = UΣU H = [ U s , U n ] ⎢ ⎥⎢ H ⎥ (6)
⎣ 0 σ w I M − r ⎦ ⎣⎢ U n ⎦⎥
2

Figure 2. Spectrogram before and after rearrangement

We can get the imaginary part by inserting imaginary where Σ′ contains r large eigenvalues, U s is the signal
part of the following N/2 lines to the first N/2 lines at one subspace, U n is the noise subspace.
point interval reversely.
From Fig. 2, N independent N independent lines in the Pseudospectra can be constructed as
range of 0 to f s / 2 is achieved. As a result, the interval
1
between adjacent two lines is changed to 0.5Δf . From P( w) = . (7)
a H ( w)U n U nH a( w)
equation (3), after computing FFT of the modulated signal
shown as equation (2), we can get spectrum estimation of the Then according to peak value of the pseudospectra, we can
original real sequence using the relationship that the k th estimate exact frequency of each harmonic.
spectral line ( k = 0,1" N ) is corresponding to frequency MUSIC algorithm can achieve high frequency
(0.5k + 0.25)Δf . resolution, and it is close to Cramer-Rao lower. Reference
[5] shows that with a fixed time window width, resolution
Assume that IF signal is sampled at 10MHz, and and estimation accuracy of MUSIC algorithm decrease with
N=256.Thus the estimated frequency is f s / N ≈ 39.1kHz the decrease of SNR, and in particular noise, they increase
with the increase of the time window width.
with traditional FFT, and f s / 2 N ≈ 19.5kHz with
Computation of this algorithm is mainly reflected in the
modulation FFT. Frequency resolution is doubled. In characteristic decomposition and the spectral peak searching.
traditional VHF frequency hopping radio, frequency hopping
interval is 25 kHz. But because of spectral leakage, picket- Calculated amount is O ( M
3
) . From equation (7), U n U nH
fence effect and environmental influence, it’s necessary to remains the same when frequency changes, so it just needs to
find an algorithm with high frequency resolution to replace 4 5
compute once. Set M = 30 , r = 3 ,then M = 8.1× 10
or improve this modulation FFT algorithm.
complex multiplications are calculated every time equation
B. MUSIC algorithm (7) is computed. If step value of the frequency domain search
For a process formed by r complex sine waves is set to 2π fs / 10 4 , 10 4 frequencies’ pseudospectra need
9
r to be calculated when searching for peaks, that is 8.1× 10
x(n) = ∑ Ai e j (ωi n +ϕi )
+ w(n) complex multiplications, which means a large amount of
i =1 calculation.

where w( n) is the additive white noise. The sequence

consists of M sampling values [4] is

526
527
 III. FREQUENCY ESTIMATION COMBINING MODULATION Generte x(n)
FFT WITH MUSIC ALGORITHM
Traditional FFT can detect integer harmonics, it has low Compute y ( n ) = x ( n )e
− jnπ / 2 N

complexity, computation and can be easily implemented in

hardware. But for non-integer harmonic, because of spectral Compute FFT of y(n)
leakage and picket-fence effect, the frequency resolution is Rearrange the spectral
limited. Modulation FFT overcomes the shortcoming of lines and get Y ′( k ′)
traditional FFT that half of the spectrum is redundant, and it
Determine the lines which
can double the frequency resolution. MUSIC algorithm has is larger than mean of Y ′( k ′)
high frequency resolution, but requires the full range
frequency domain when searching for spectral peak, and that Compute the search domain
will result in a large computation. Based on this, modulation A = [0.5k ′Δf , 0.5( k ′ + 1) Δf ]
FFT and MUSIC algorithm are combined. Firstly modulation
Use MUSIC algorithm to
FFT algorithm is used to estimate frequency roughly and search the peak in
then MUSIC algorithm is used to estimate accurately. In this determined domain
way, we can achieve very high frequency accuracy and
calculation is greatly reduced. Determine the estimated
frequency
Given a N-point real sequence x(n) which is sampled at
f s , the specific steps of modulation FFT-MUSIC algorithm Figure 3. Diagram designing process

used to estimate the frequencies are as follows: resolution of traditional FFT is 1/N=0.00390, frequency
1) Like Equation (2) shows, y(n) is achieved through resolution of modulation FFT is 0.5/N=0.00195. That is,
modulating x(n) in time domain to shift its spectrum frequency resolution has been doubled. During simulation,
0.25Δf to the left. traditional FFT takes 0.0228 seconds, and modulation FFT
2) Calculate y(n)’s N-point FFT to get N spectral lines takes 0.0316 seconds, with an increase of 0.0088 seconds. It
is shown that compared with traditional FFT, computation
Y (k ) . of the improved FFT increases small.
3) By Insert the following N/2 lines to the first N/2 lines at
one point interval reversely after calculating their conjugates,
we get Y ′( k ′) , in which the kth spectral line is
corresponding to frequency (0.5k + 0.25)Δf .
4) Calculate mean of all the spectral lines, and determine
which is greater than this mean.
5) For each line determined by step 4), find their search
field of MUSIC algorithm
A = [0.5k ′Δf , 0.5(k ′ + 1)Δf ] .
6) Search the spectral peak using MUSIC algorithm in the
Figure 4. Spectrum comparison between traditional and modulation FFT
search field A, and determine the estimated frequencies.
The programming flow diagram is shown in Fig. 3. The spectral lines obtained by traditional MUSIC
algorithm’s search in full range frequency domain and
IV. PERFORMANCE SIMULATION modulation FFT-MUSIC algorithm’s search in limited
frequency domain are shown in Fig. 5. Traditional MUSIC
Consider a stochastic process by algorithm takes 9.9983 seconds, and modulation FFT-
MUSIC takes 0.7396 seconds. That is, consumed time is
x(n) = cos(2π f1n) + 2 cos(2π f 2 n) + cos(2π f3n) + w(n) (8) reduced twice.
The estimated frequencies using different algorithms are
shown in Table I. The estimation accuracy of modulation
where w(n) is white noise whose variance is 1, FFT-MUSIC is defined by the frequency search step, which
f1 = 0.0534 ， f 2 = 0.2023 ， f 3 = 0.4156 (all are
is set to 1× 10 / 2π = 1.59 × 10 here. The results show
−4 −5

that compared with traditional FFT algorithm, modulation

normalized frequencies), M is 30, and number of samples is FFT-MUSIC algorithm’s estimation accuracy is improved
256.
Simulation is done using MATLAB as what Fig. (3)
shows. The spectral lines using traditional FFT and
modulation FFT is shown in Fig. 4, where frequency

527
528
 MUSIC algorithm are combined in this paper. To estimate
frequency accurately, after pre-estimating with modulation
FFT, the interesting range frequency domain are determined,
which MUSIC algorithm is used to search for spectral peak
in. Compared with traditional FFT-MUSIC algorithm,
because frequency resolution of modulation FFT algorithm is
doubled, modulation FFT-MUSIC algorithm’s search range
is reduced twice, and computation is also reduced about
twice. It can be used to estimate frequencies of FSK signals,
and especially when estimating FH signals’ parameters, this
algorithm can be used because of its good frequency
resolution and real-time performance. Simulation results
Figure 5. Comparison between traditional MUSIC and modulation FFT- show that this improved algorithm is effective, has a high
MUSIC
frequency resolution, and its run time has been reduced
obviously compared with traditional MUSIC algorithm.
TABLE I. ESTIMATED FREQUENCY COMPARISON BETWEEN EACH
ALGORITHM
REFERENCES
f1 f2 f3 Expectation [1] Hu Guangshu, Digital Signal Processing. Beijing:Tsinghua University
- of Deviation Press, 2003,pp.125-127.
Theoretical
value
0.0534 0.2023 0.4156 - [2] Yang Zhijian, Ding Kang, “A Novel Algorithm Based on Modulation
FFT for Improving Frequency Resolution,” Journal of Vibration and
FFT 0.054688 0.20313 0.41406 1.9267 ×10 −4
Shock, vol.27, pp. 319-321, 2008.
Modulation
0.053711 0.20215 0.41504 1.330 ×10−4 [3] Schmidt R O, “Multiple emitter location and signal parameter
FFT
Modulation estimation,” IEEE Trans. On Antennas and Propagation,vol. AP-34:
0.053253 0.20232 0.41551 7.2333 × 10 −5
pp.276-280,March 1986.
FFT-MUSIC
[4] Petre Stoica, Randolph Moses, Spectral Analysis of Signals. Beijing:
V. Conclusion Electronics Industry Press，2002.
[5] Jiang Yi, Research on Harmonic Frequency Estimation Based on
MUSIC algorithm has high frequency resolution, Non-parameter Methods. Cheng Du: University of Electronic Science
estimation accuracy and stability. It has been widely used in and Technology of China, 2008.
radar, sonar signal processing. But MUSIC algorithm needs [6] Liu Xuejun, Liu Chang, Wang Jingzhi, Xing Jisheng, “Interharmonic
to search spectral peak in the full range frequency domain, Parameter Estimation Based on FFT and MUSIC,” Power System
which bring on a large amount of computation, and real-time Protection and Control, vol.37,pp.37-40,2009.
performance is severely affected. Modulation FFT and

528
529