Anda di halaman 1dari 11

A New Time-Frequency Approach for Weak Chirp Signal Detection

Houng-Jyh Wang, Te-Chung Yang and C.-C. Jay Kuo Signal and Image Processing Institute and Department of Electrical Engineering{Systems University of Southern California, Los Angeles, California 90089-2564

ABSTRACT
Several methods have been proposed to detect chirp signals buried in white noise, including wavelet shrinkage and matching pursuits. However, most of these methods perform poorly when the signal-tonoise ratio (SNR) becomes very low (e.g. lower than 0 dB). In this work, we present a new time-frequency technique using the multi-window Fourier transform (MWFT) and local maximum indexing. The new method detects the location and frequency range of a chirp signal with SNR up to -6 dB.
Keywords:

imum index.

signal detection, time-frequency analysis, wavelet shrinkage, matching pursuit, local max-

1 INTRODUCTION

Chirp signal detection in a very noisy environment is important for many military and commercial applications. Several standard methods have been proposed for this task. One approach is based on the wavelet shrinkage technique 1], which denoises the given data in the wavelet domain. Another possibility is to represent the data by using a joint time-frequency representation such as the Gabor transform and matching pursuit. However, most of these methods perform poorly when the signal-tonoise ratio (SNR) becomes very low (e.g. lower than 0 dB). In this work, we examine the frequency energy distribution of the signal in the time-frequency plane, and introduce the multi-window Fourier transform (MWFT) and the local maximum index scheme to detect the weak chirp under a noisy environment. The multi-window Fourier transform is used to reduce the computation complexity, while a post-processing technique known as the local maximum index detection is adopted to enhance the output of the time-frequency analysis so that the signal can be detected more easily. The paper is organized as follows. We brie y describe the wavelet shrinkage and the matching pursuits algorithms, and point out their limitations in low SNR environment in Section 2. Then, we propose a new time-frequency analysis method to detect the signal in Section 3. The experimental results of detecting the weak chirp signal with SNR = -3, -6, -9 dB are presented in Section 4. Concluding remarks are given in Section 5

2 REVIEW OF DENOISING TECHNIQUES 2.1 Wavelet Shrinkage

Wavelet shrinkage is a well known de-noising technique which transforms noisy data from the temporal domain to the wavelet domain, and then applies a thresholding to the resulting coe cients, i.e. suppressing the coe cients smaller than a certain amplitude. For details, we refer to 1,2]. To test the performance of wavelet shrinkage, we generate two groups of sample data consisting of cosine and chirp waves as shown in Fig. 1 and 2, respectively. The amplitude of a discrete-time chirp

waveform is de ned as

The phase (n) can be expressed as

s(n) = A cos (n); 8n 2 0; 1; 2; ; N:


2

where fl and fu denote the normalized lower and upper discrete frequency bands, respectively, and ' is a constant phase shift factor. Then, the normalized instantaneous frequency is f (n) = f + (fu ? fl )n :
N l

fl)n g + '; 0 f < f 0:5; (n) = 2 ffln + (fu ? l u 2N

We choose fu = 0:1, fl = 0:0, ' = 0, A = 1 and N = 128 for the signal shown in Figure 2. The test signals are then corrupted by additive white noise with di erent noise power.
Cosine Signal:256 points 1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1 0

50

100

150

200

250

300

Figure 1: Test cosine signal consisting of 256 points.


Chirp Signal:128 points 1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1 0

20

40

60

80

100

120

140

Figure 2: Test chirp signal consisting of 128 points. The results of applying wavelet shrinkage to the test signals are shown in Figure 3(a)-(d). They were obtained by using the WaveLab Package developed at the Stanford University 3] with the nearly symmetric wavelet with 8 vanishing moments and the visual threshold method. We see that the wavelet shrinkage algorithm performs well for SNR = 10 dB case, but poorly for SNR=0 dB.

Sine Signal + Gaussian Noise : 10 dB 15 15

Chirp Signal+Gaussian Noise:10 dB

10 10

5 5 0 0 5

5 10

10 0

50

100 150 200 250 Result after Wavelet Shrinkage: Visu, Symmlet8

300

15 0

20

(a)

40 60 80 100 Result after Wavelet Shrinkage:Visu,Symmlet8

120

140

(b)

Sine Signal + Gaussian Noise : 0 dB 2 2

Chirp Signal + Gaussian Noise : 0 dB

1.5

1.5

1 1 0.5 0.5 0 0 0.5

0.5

1 0

50

100 150 200 250 Result after Wavelet Shrinkage: Visu, Symmlet8

300

1.5 0

20

(c)

40 60 80 100 Result after Wavelet Shrinkage: Visu, Symmlet8

120

140

(d)

Figure 3: The results of applying wavelet shrinkage to two test signals: (a) cosine signal with SNR = 10 dB, (b) chirp signal with SNR = 10 dB, (c) cosine signal with SNR = 0 dB, (d) chirp signal with SNR = 0 dB.

2.2 Matching Pursuits

Matching pursuits decompose a signal into a linear expansion of waveforms selected from a redundant dictionary of functions 4]. The matching pursuits algorithm provides a general procedure to determine an adaptive signal representation by computing the signal distribution in the time-frequency plane. To give an example, we use the matching pursuits routine in the WaveLab package 3] with the Haar basis to test a 1024-point chirp signal within a 2048-point data sequence with additive white noise. The Image Atom Phase Plane of applying matching pursuits with SNR= 40, 3 and -3 dB are shown in Figs. 4, 5, 6, respectively. The result degrades for SNR = 3 dB, and becomes very poor for SNR = -3 dB.

3 A NEW TIME-FREQUENCY DETECTION APPROACH

In this section, we propose a new method to detect the chirp signal location and its frequency range by performing the multi-window Fourier transform on the signal and then using a local maximum index scheme to display the signal location and frequency range in the time-frequency plane. For higher SNR values, only the application of the local maximum index scheme to the result of short-time Fourier transform (STFT) works properly. By comparing Fourier and wavelet bases, we choose the Fourier transform as our basic transform,

Phase plane: SNR=40db,Chirp Length:1024 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0

Frequency

0.1

0.2

0.3

0.4

0.5 Time

0.6

0.7

0.8

0.9

Figure 4: Denoising with matching pursuits with SNR = 40 dB case, where the black region denotes the chirp signal location and its signal frequency range (the upper side corresponding to a higher frequency value).
Phase plane: SNR=3db,Chirp Length:1024 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0

Frequency

0.1

0.2

0.3

0.4

0.5 Time

0.6

0.7

0.8

0.9

Figure 5: Denoising with matching pursuits with SNR = 3 dB. since the wavelet bases are not well suitable for a narrow-banded signal. To make a balance between the performance and speed, we propose a technique called the multi-window Fourier transform (MWFT) which can enhance the adaptivity of the short-time (or windowed) Fourier transform while maintaining low computational complexity. This transform is similar to the Gabor transform in spirit, but easier to apply in implementation. For a given data sequence m(n) with a total length of NT , the N-point STFT:

MN (ns ; k) = FFTN ns ; m(n)] =

ns +N ?1 i=ns

m(i)e?ji 2Nk

where ns is the starting point of the STFT window. By moving the starting point ns of STFT, we can focus on di erent portion of the data sequence. Usually we select ns = n N 2 , n = 0; 1; 2; so that each adjacent window of STFT overlaps. The time-frequency plane of STFT can be obtained by aligning every MN (k), where the y-axis is the frequency-axis with index 0 k N ? 1 and the T x-axis is the time-axis 0 < 2N N . For multi-window Fourier transform, we allow the choice of di erent window sizes, for example, N = 32; 64; 128 with di erent overlappings. Through temporal and frequency interpolation, we can interplate every time-frequency plane so that they have the same resolution. By adding them together, we get the time-frequency plane of the multi-window Fourier

Phase plane: SNR=3db,Chirp Length:1024 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0

Frequency

0.1

0.2

0.3

0.4

0.5 Time

0.6

0.7

0.8

0.9

Figure 6: Denoising with matching pursuits with SNR = -3 dB. transform representation, ( ; ! )1. The basic idea of local maximum index scheme is to determine the maximum frequency energy distribution location in each window. For a given time index = 1; 2; : : :; tT , we can compute the largest energy value within a window
f `L ( ) max ( ; ! ); !=1 L

which can be normalized via

and determine the frequency index which gives the largest energy value
f `idx ( ) = INDEX! =1 L ( ; ! ):

`L ( ) `img ( ) = max( `L ( )) 100;

The plot of `idx ( ) as a function of is called the local maximum index gure. We show the timefrequency analysis plane and local maximum index gure of applying the 2-level MWFT2 and local maximum index scheme to a 1024-point chirp signal buried in white noise with SNR = 3 dB in Figs. 8 and 9, respectively. The one-level MWFT results with window sizes equal to 64 and 128 points for SNR = 3 dB are shown in Figure 7 for comparison. In the upper gure with window size equal to 64, the vertical axis denotes the frequency and the value 0 is 0.0 (radius), 32 is and 64 is 2 . In the lower ugre with window size equal to 128, the value 0 is 0.0, 64 is and 128 is 2 . For both window sizes, the results both look very good. Based on Figs. 8 and 9, we can see that the signal has been detected and its location (x-axis) and frequency range (y-axis) is clearly displayed. For this case, this chirp signal starts from the 512th point with signal length equal to 1024 points. By assuming that the frequency of a narrow-band chirp signal to be detected is a smooth function of time while noise has a at wide-band spectrum, the signal part can be highlighted in the time-frequency plane as shown in Figure 8. Otherwise, due to the characteristics of random noise, no special frequency will be enhanced so that we are not able to see a smooth curve in the local maximum index gure.
1 2

Where is the time index and ! is the frequency index. = 1, 2, 3, : : : , tT and ! = 1, 2, 3, : : : , f . For this example, the level number, L, is 2, so the rst window size, Ns , is 64 and the second one is 128.

Total length:2048,Chirp start at:512,Chirp length:1024 10 20 30 40 50 60 10 20 30 40 50 Image Result After Applying FFT:Window Size 64 60

20 40 60 80 100 120 5 10 15 20 25 Image Result After Applying FFT:Window Size 128 30

Figure 7: The time-frequency plane of applying one-level MWFT transform and local maximum index scheme to a 1024-point chirp signal with SNR = 3 dB with window sizes 64 and 128.
Total length:2048,Chirp start at:512,Chirp length:1024,SNR=3dB 10

20

30

40

50

60 10 20 30 40 50 Image Result after applying MULW & Local Index Max 60

Figure 8: The time-frequency plane of applying two-level MWFT (Ns = 64, L = 2, window sizes: 64 and 128) and local maximum index to a 1024-point chirp signal corrupted by white Gaussian noise with SNR = 3 dB.

4 EXPERIMENTAL RESULT

In this section, we will show the results of applying MWFT and the local maximum Index scheme to detect the chirp signal with white Gaussian noise of SNR value = ?3, ?6 and ?9 dB. We rst examnine the generation of the chirp signal in Section 4.1, and then discuss the chirp signal and general narrow band signal detection in Sections 4.2 and 4.3, respectively.

4.1 WEAK CHIRP SIGNAL GENERATION

Ps ; SNR = 10 log10 P n where Ps means the signal power , and Pn is noise power. In the implementation, we keep the power of chirp signal to 1:0 and change the noise power to meet the required SNR value.

By using the similar formula mentioned in Section 2.1, we generate the chirp signal by choosing fu = 0:2, fl = 0:0, ' = 0, A = 1 and N = 1024. Then, we generate the 2048 points of white noise by a Gaussian random value generator and mix them with the chirp signal starting from 512th point to 1535th point ( 1024 points chirp signal in total ). The SNR is de ned as below:

Total length:2048,Chirp start at:512,Chirp length:1024 70

60

50

40

30

20

10

0 0

10

20 30 40 50 Result After Applying Local Maximum Index Method

60

70

4.2 CHIRP SIGNAL DETECTION

Figure 9: The local maximum index gure of applying MWFT (Ns = 64; L = 2) and local maximum index to a 1024-point chirp signal corrupted by white Gaussian noise with SNR = 3 dB.

We show the time-frequency plane result in Figure 10(a) and the corresponding local maximum index in Figure 10(b) after the application of the 2-level MWFT with Ns = 64 and the local maximum index scheme to the chirp signal of SNR = -3 dB case. Based on these these two gures, one can locate the signal position which matches our simulation data: around from 512th to 1535th points with 1024 s points in total. Because Ns = 64 and L = 2, the time domain (x-axis) scaling at N 2 = 32. That is, each point in the x-axis denotes 32 points in the time domain. The smaller Ns is, the more precision we have in the time domain. However, smaller Ns also means the less precision in the frequency domain. In this work, we choose Ns = 64 to make a balance between the time and frequency domain precision. The time-frequency analysis and the local maximum index results for SNR = -6 dB are shown in Figs. 10(c) and 10(d), respectively. Our algorithm still performs reasonably well, except for a few points missing in the local maximum index gure. However, by adding more levels of MWFT, the performance improves. We show the results of the time-frequency analysis and local maximum index gure of applying the 3-level MWFT to the data sequence to be tested with SNR = -6 dB in Figure 11(a) and Figure 11(b), respectively. The results are better than those obtained by using 2-level MWFT. Even though in theory, the frequency energy will be enhanced at the location of signal in time-frequency plane, the maximum frequency energy may sometimes be determined by noise in some particular regions. This e ect will be more frequent for weaker SNR cases, because noise may be coherent to form one particular frequency band and match the STFT window size in a certain region so that it is judged as one narrow band signal in the time-frequency domain. Another problem associated with the weak SNR cases is that the noise power is often larger than the signal power and the local maximum index may report an incorrect position as a signal which is actually a noise in single window STFT. However, by performing MWFT, we can reduce the probability of misdetection and improve the detection results by increasing the level number. In practice, a larger level number means more computation and we have to make a balance between computation complexity and precision. Generally speaking, L = 3 or 4 is a good choice. For the experiments in this section, L = 2 is su cient for SNR up to -6 dB. Finally, we examine the case with SNR=-9 dB. The results with Ns = 64 and L = 2 are shown in Figure 11(c). It is not easy to nd out the signal location and its frequency range from this gure. However, if we extend our simulated data sequence to a data sequence of length = 8192 points with the chirp signal starting at the 2048th point with length 4096 points, a better time-frequency plane result is

Total length:2048,Chirp start at:512,Chirp length:1024,SNR=3dB


70

Total length:2048,Chirp start at:512,Chirp length:1024,SNR=3db

10

60

20

50

30

40

30

40
20

50
10

60 10 20 30 40 50 Image Result after applying MULW & Local Index Max 60


0 0 10 20 30 40 50 Result After Applying Local Maximum Index Method 60 70

(a)

(b)

Total length:2048,Chirp start at:512,Chirp length:1024,SNR=6dB


70

Total length:2048,Chirp start at:512,Chirp length:1024,SNR=6dB

10

60

20

50

30

40

30

40
20

50
10

60 10 20 30 40 50 Image Result after applying MULW & Local Index Max 60


0 0 10 20 30 40 50 Result after applying MULW & Local Index Max 60 70

(c)

(d)

Figure 10: The image time-frequency plane and local maximum index gure after applying L=2 level MWFT and local maximum scheme, where (a) and (b) are for with SNR = ?3 dB, and (c) and (d) are with SNR = ?6 dB. obtained as shown in Figure 11(d). We can nd the location and the frequency range of the chirp signal clearly. Even though we may misdetect some pixels in the time-frequency plane, but most pixels are correctly detected. That means it reduces the rate of misdetecting when increasing the total number of points of signal.

4.3 GENERAL NARROW-BAND SIGNAL DETECTION

The propose algorithm of using multi-window Fourier transform and the local maximum indexing provides a general procedure to detect the signal. They are not only applicable to the chirp signal but also any narrow-banded signal whose energy is concentrated in a certain narrow frequency component varying with the time index. We show in Figure 12 a signal which contains one cosine wave, one sine wave with di erent frequency components and a chirp signal. Figure 13 shows the time-frequency plane result of applying MWFT and the local maximum index to this signal buried in white noise with SNR = 3 dB and Figure 14 shows the corresponding loca maximum index gure.

4.4 DISCUSSION

Based on the above experimental results, we can summarize our observations as follows.

Total length:2048,Chirp start at:512,Chirp length:1024,SNR=6dB


140

Total length:2048,Chirp start at:512,Chirp length:1024,SNR=6dB

20

120

40

100

60

80

60

80
40

100
20

120 10 20 30 40 50 Image Result after applying 3 level MULW & Local Index Max 60
0 0 10 20 30 40 50 60 Result after applying 3 level MULW & Local Index Max 70

(a)

(b)

Total length:2048,Chirp start at:512,Chirp length:1024,SNR=9dB

Total length:8192,Chirp start at:2048,Chirp length:4096,SNR=9dB

10

10

20

20

30

30

40

40

50

50

60 10 20 30 40 50 Image Result after applying MULW & Local Index Max 60

60 50 100 150 200 Image Result after applying MULW & Local Index Max 250

(c)

(d)

Figure 11: Figure (a) and (b) show the results of applying 3 level MWFT and local maximum scheme. It performs better than the results of Fig. 10 (c) and (d), for which only the 2-level MWFT is applied only. Figure (c) shows the result of applying 2 level MWFT and local maximum index to a chirp signal with length=1024 points and SNR=-9 dB. Figure (d) shows the result of applying 2 level MWFT and local maximum index to a chirp signal with length=4096 points and SNR = -9 dB. By comparing (c) and (d), we can see that the longer the signal length is, the better the performance is. 1. Our proposed signal detection algorithm is simple to apply. Unlike the Gabor transform, it does not create the coe cient dictionary which requires a lot of computations. 2. It has better results in detecting the signal location and frequency range when compared with wavelet shrinkage and matching pursuit. 3. For xed values of Ns and L, the larger the signal length is, the better the performance is. 4. For xed values of Ns and signal length, the larger the level number of Multi-Window Fourier Transform is, the better the result is. However, a larger level number requires more computation. 5. The local maximum index gure can be applied to signal matching. We can decide whether any two signals corrupted by a noisy environment are identical by comparing their local maximum index gures. To reduce the e ects of some misdetecting points in the local maximum index gure, we suggest to perform the following preprocessing procedure:

1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1 0

500

1000

1500

2000

2500

Figure 12: The original multi-frequency signal.


10

20

30

40

50

60 10 20 30 40 50 60

Figure 13: The time-frequency plane after applying 2 level MWFT and the local maximum index scheme with SNR = 3 dB. For kth step, k = 1; 2; 3; : : :; tTP , compute the average frequency value of kth and its following ?1 1 points. That is MEANk = k+ =k `( ) where `( ) means the result of the time-frequency plane after applying MWFT and the local maximum index method. Compute current MEANk+1. If the di erence (k) between MEANk and MEANk+1 is larger than than D(k) which is de ned as D(k) = 100 MEANk , and all of (k ? 1), (k ? 2), : : :, (k ? ) are smaller than their corresponding D(i), i = k ? 1; : : :; k ? , change the value of `(k + ) to MEANk , otherwise, no change. Note: is the total pixel number in the time domain, where NT is the total number of the data sequence. in the average post-processing window size which is less than Ns . (k) = kMEANk ? MEANk+1k. is a constant less than Ns . is a constant less than 100.

tT = 2

NT Ns

70

60

50

40

30

20

10

0 0

10

20

30

40

50

60

70

Figure 14: The local maximum index gure after applying 2-level MWFT and the local maximum index scheme with SNR = 3 dB.

5 CONCLUSION

By using both MWFT and the local maximum index scheme, the chirp signal buried in a noisy environment can be detected and located correctly in the image time-frequency plane. This approach requires less computation and performs better than the wavelet shrinkage and matching pursuit when SNR is below 0 dB. Moreover, the idea of noise removal in multi-window Fourier transform provides an excellent signal enhancement technique which is superior to other standard methods.

6 ACKNOWLEDGMENT 7 REFERENCES

This work was supported by the National Science Foundation Presidential Faculty Fellow (PFF) Award ASC-9350309. 1] R.R. Co man and D.L. Donoho, \Translation-Invariant De-Noising," Yale University and Stanford
University.

2] D.L. Donoho, G.K. Johnstone and D. Picard, "Wavelet Shrinkage: Asymptopia," J. Roy. Statist. Soc., B 57 2, pp. 301{369. 3] J.B. Buckheit and D.L. Donoho, "WaveLab and Reproducible Research," Stanford University. 4] S.G. Mallat and Z. Zhang, \Matching Pursuits with Time-Frequency Dictionaries," IEEE Trans. Signal Processing, Vol. 41, No. 12, Dec. 1993, pp. 3397{3415. 5] S. Qian and D. Chen, \Discrete Gabor Transform," IEEE Trans. Signal Processing, Vol. 41, No. 7, July. 1993. pp. 2429{2438. 6] V.C. Chen, \Reconstruction of Inverse Synthetic Aperture Radar Image Adaptive Time-Frequency Wavelet Transform," SPIE, Vol. 2491, pp. 373{386. 7] Leon Cohen, \Time-Frequency Distributions - A Review," Proceedings of IEEE, Vol. 77, No. 7, July. 1989. pp. 941{949. 8] L.C. Lin, T.J. Lee and C.-C. Jay Kuo, \A Wavelet Based Extrapolation Technique for High Resolution Beamforming," SIPI, University of Southern California

Anda mungkin juga menyukai