1. Field of the Invention
The present invention relates to ultra-wideband signal processing, and particularly to a peak detection method using blind source separation.
2. Description of the Related Art
Efficient utilization of radio spectrum has gained recent attention. It has been observed that utilization of spectrum by licensed wireless systems, for instance TV broadcasting, is quite low. Transition from voice only data services to multimedia services requires high data rates. Current static frequency allocation schemes cannot cope forever with increasing data rates. Some frequency bands are overcrowded, and some are barely used. A spectrum occupancy measurement project concluded that the average spectrum occupancy over multiple locations is 5.2%, with a maximum of 13.1%.
Cognitive radio (CR) seems a tempting solution to resolve the perceived bandwidth scarcity versus under-utilization dilemma. CR uses opportunistic usage of bands that are not crowded by licensed users. They use spectrum sensing to sense the frequency bands that are unoccupied by licensed users and transmit on these bands to avoid harmful interference to licensed users.
CRs front end architecture is dependent on spectrum utilization. For spectrum utilization under 20%, a wideband architecture for the CR front end is suggested. The observed spectrum consists of numerous frequency bands. Power spectral density (PSD) within each frequency band is smooth. Transition of PSD from one band to another band is considered as irregularities in PSD. Such irregularities can be studied using wavelet transforms, which are capable of characterizing local regularity of a signal.
Applying a wavelet transform on an incoming signal results in peaks at locations where signal PSD is irregular. Irregularity could be a jump or a fall in PSD, depicting the change of frequency band. Jump depicts that the next user has higher PSD than the current one, whereas fall shows that the next user has lower PSD than the current one, or it could also be a vacant band. When the incoming signal is noisy, these peaks are accompanied by noisy peaks. In some known methods, multiscale wavelet products are used to extract true peak information. This technique requires multiplication of various wavelet transform gradients (for the same signal).
As a result of this, true peaks will be enhanced, whereas random noisy peaks will be suppressed. However this technique requires a priori knowledge regarding the total number of occupied bands in a spectrum at a given instance. This information is normally unknown to CR.
Thus, a peak detection method using blind source separation solving the aforementioned problems is desired.
The peak detection method using blind source separation extracts true peaks from noisy peaks in a more robust way that does not require any a priori information. Information regarding true peak location is obtained by thresholding the output of a wavelet transform. The value of the threshold is dependent on noise variance. While noise variance is normally unknown, the method implements a blind source separation technique to calculate the noise variance. The blind source separation technique does not require information of the incoming signal or the channel noise, and hence is suitable for CR peak detection.
These and other features of the present invention will become readily apparent upon further review of the following specification and drawings.
Similar reference characters denote corresponding features consistently throughout the attached drawings.
At the outset, it should be understood by one of ordinary skill in the art that embodiments of the present method can comprise software or firmware code executing on a computer, a microcontroller, a microprocessor, or a DSP processor; state machines implemented in application specific or programmable logic; or numerous other forms without departing from the spirit and scope of the method described herein. The present method can be provided as a computer program, which includes a non-transitory machine-readable medium having stored thereon instructions that can be used to program a computer (or other electronic devices) to perform a process according to the method. The machine-readable medium can include, but is not limited to, floppy diskettes, optical disks, CD-ROMs, and magneto-optical disks, ROMs, RAMs, EPROMs, EEPROMs, magnetic or optical cards, flash memory, or other type of media or machine-readable medium suitable for storing electronic instructions.
The peak detection method using blind source separation extracts true peaks from noisy peaks in a more robust way that does not require any a priori information. Performing spectrum sensing using a wavelet edge detection technique provides edges (peaks) that contain information regarding the start and end locations of a frequency band. In the presence of noise, there is a mixture of true peaks and noisy peaks. Knowledge of noise variance is required to extract true peaks efficiently from the mixture. Information regarding true peak location is obtained by thresholding the output of the wavelet transform. The value of the threshold is dependent on noise variance. Noise variance is also normally unknown. Here, a blind source separation technique is implemented to calculate the noise variance. Blind source separation does not require information of the incoming signal or the channel noise, and hence is suitable for CR peak detection.
The PSD (power spectral density) of an incoming signal is flat within each band, and transition occurs at the beginning of a new band. Plot 100 of
The continuous wavelet transform of an incoming signal is given as follows:
W
s
S
r(f)=Sr(f)*φS(f), (1)
where φs(f) is the dilated wavelet smoothing function and “*” defines the convolution operator. The variable ‘s’ depicts the dilation factor of the wavelet smoothing function, and it takes values in terms of power of 2. A common example of a wavelet smoothing function is Gaussian function. For detection of edges, the first derivative of the wavelet transform can be used, which is given as:
Local maxima of the first derivative provide information of edges, which corresponds to the start and end locations of a frequency band. We can take a second derivative of equation (2), and we can detect these edges. But with the second derivative, we have to look for zero crossings, not the local maxima. Once the frequency boundaries, i.e., {fn}n=0N−1, are detected, then the next step is to calculate the PSD within each band and decide about the presence or absence of a primary user. Calculation of the PSD is given as follows:
Blind source separation has found very useful applications in the area of signal processing and neural networks. Blind source separation does not require knowledge of the channel and the transmitted signal. In fact, its goal is to recover the unobserved signals, i.e., ‘source signals’, from a set of observed signals. The term ‘Blind’ refers to the fact that the source signals are not observed, and the fact that there is no a priori knowledge available about the mixing system.
Since the development of the blind source separation technique, many new algorithms have been formulated for various problems. Some of these techniques depend on exploiting the second-order statistics and stationary or non-stationary conditions of the received signal, while others need higher order statistics and some exploitations of the time-frequency diversities. All these algorithms obtain a cost function through some optimization process, which normally is computationally complex.
In a known blind source separation algorithm, the maximum signal-to-noise ratio (SNR) can be achieved when sources are separated completely. The cost function of this algorithm is based on the SNR definition. This algorithm achieves a low computational complexity solution based on an instantaneous mixing method. The assumption is that source signals come from different sources and could be considered as statistically independent. The received signal can be written as:
x
i(t)=Σj=1naijsj(t) (4)
where aij represents the instantaneous mixing matrix (i, j) element. In vector form, we can write (4) as:
x(t)=As(t), (5)
where x(t) is a vector of mixed signals.
BSS algorithms have information of mixed signals and the statistical independence property of the source signals. Assuming W is an un-mixing matrix for the aforementioned problem, the BSS problem can be stated as follows:
y(t)=Wx(t)=WAs(t), (6)
where y(t) is the estimate of the source signals, i.e., s(t). The difference between the original signal and the estimated signal is the noise signal. Thus, the SNR may be defined as:
Optimized processing of equation (7) results in an Eigenvalue problem. The resultant Eigenvalue matrix corresponds to the un-mixing matrix W. Once the un-mixing matrix is calculated, the source signals can be obtained using equation (6). The un-mixing matrix calculation is given as follows:
(xx)×W=(({circumflex over (x)}−x)({circumflex over (x)}−x)T)×W×D, (8)
where {circumflex over (x)} is the moving average estimate of x. For energy detection, the received signal can be written in terms of its sample covariance matrix, i.e.:
R
x(N)=Rs(N)+σ2I, (9)
where
are the received and transmitted signal sample covariance matrices, respectively. Also, σ2 is the noise variance.
Since we do not have information regarding the transmitted (or source) signal, we cannot calculate the transmitted signal sample covariance matrix. The blind source separation algorithm can calculate the un-mixing matrix for the received signal. Using un-mixing matrix and received signal, we can estimate the transmitted signal as shown in equation (6) and its corresponding sample covariance matrix. Noise variance can be calculated as:
In our case, signal x is the output of the wavelet edge detection technique. Noise variance is not sufficient to threshold one such signal. In order to calculate exact threshold values, we have to normalize noise variance with the sample mean of the received signal. Hence, the threshold value can be written as:
where r represents the received signal. Using T, we can threshold the output of the edge detection technique, and hence can calculate the frequency edge locations. The present peak detection method 200 is illustrated in
Here we assume that our wideband signal of interest lies in the range of [0, 1000]Δ Hz, where Δ is frequency resolution. We also assume that during the transmission there are total of N=11 bands in the wideband signal with frequency boundaries {fn}n=010=[0, 100, 119, 300, 319, 500, 519, 700, 719, 900, 919, 1000]. Out of these eleven bands, only five bands are carrying primary user transmission and the remaining six bands are available for secondary users, i.e., they are spectrum holes. In the simulation, we used a Gaussian wavelet for the edge detection technique.
We studied the effect of noise on spectrum sensing performance. We calculated the success ratio and the probability of detection for each SNR value over 1000 realizations. The success ratio is defined as the probability of accurately detecting the frequency boundaries (i.e., the start and end of a frequency band) using the thresholding method, as described earlier. The probability of detection is based on the PSD value calculation within each band. Plot 300 of
In the present method, we proposed calculation of noise variance for detecting true peaks using the blind source separation method. This noise variance information is useful when the output of the edge detection technique contains noisy peaks along with true peaks. Noisy peaks are suppressed by thresholding the signal. This process directly affects the probability of detection of a primary user when performing spectrum sensing. The present method gained 4 dB in term of success ratio, and 8 dB in the probability of detection compared to the multiscale wavelet products technique, hence allowing cognitive radio devices to work efficiently on low power in a wideband regime. The present method will capture the interest of telecommunication equipment vendors, mobile phone manufacturers, and research institutes who are interested in bringing new era of wireless communication devices.
It is to be understood that the present invention is not limited to the embodiments described above, but encompasses any and all embodiments within the scope of the following claims.