Integrated Multifrequency Recognition and ... - IEEE Xplore

Integrated Multifrequency Recognition and Downconversion Based on Photonics-Assisted Compressive Sampling Volume 4, Number 3, June 2012 Li Yan Yitang Dai Kun Xu Jian Wu Yan Li Yuefeng Ji Jintong Lin

DOI: 10.1109/JPHOT.2012.2195482 1943-0655/$31.00 ©2012 IEEE

IEEE Photonics Journal

Recognition and Downconversion Based on CS

Integrated Multifrequency Recognition and Downconversion Based on Photonics-Assisted Compressive Sampling Li Yan, Yitang Dai, Kun Xu, Jian Wu, Yan Li, Yuefeng Ji, and Jintong Lin State Key Laboratory of Information Photonics and Optical Communications, Beijing University of Posts and Telecommunications, Beijing 100876, China DOI: 10.1109/JPHOT.2012.2195482 1943-0655/$31.00 Ó2012 IEEE

Manuscript received January 31, 2012; revised April 5, 2012; accepted April 11, 2012. Date of publication April 19, 2012; date of current version April 24, 2012. This work was supported in part by National 973 Program (2012CB315705), 863 Program (2011AA010306), National Science Foundation of China Programs (61107058 and 61120106001), and the Beijing Excellent Doctoral Thesis Project under Grant YB20101001301. Corresponding author: K. Xu (e-mail: [email protected]).

Abstract: On the basis of photonics-assisted compressive sampling (CS), we demonstrate for the first time an integrated system with the capacity of multifrequency recognition and downconversion of intercepted radio frequency (RF) component. Using a single analogto-digital converter (ADC) with sub-Nyquist analog bandwidth of 826.75 MHz, a multicomponent signal with unknown frequencies ranging from 15 to 20 GHz is precisely located with maximum detection error of 100 kHz. The system’s ability of photonic downconversion of identified RF component is characterized in terms of spur-free dynamic range (SFDR) as well. Index Terms: Microwave photonics signal processing, fiber optics systems.

1. Introduction Identification of frequencies of an intercepted microwave signal from a radar or communication system is of critical importance in the field of electronic warfare. Over a broad frequency range, it is highly desired to provide the radio frequency (RF) spectrum information of intercepted signal with multifrequency resolving, high accuracy, and real-time operation. Moreover, the downconversion to intermediate frequency (IF) of the interested RF component therein for further digitalizing and processing is also in demand. Conventional RF receivers, however, suffer from high weight, limited bandwidth, vulnerability to electromagnetic interference, and reduced dynamic range due to the nonlinearity of electronic components. As a result, photonics-assisted multifrequency RF recognizing systems as well as all optical RF downconversion techniques have been proposed to overcome these limitations. Multifrequency resolving has been demonstrated by channelizing: A broadband RF signal is modulated on one or multiple optical carriers and then demultiplexed by an optical channelizer, which could be a diffraction grating [1], a parallel filter bank [2], or two comb filters with different free spectral ranges [3]. However, an optoelectronic converter array is always a must, and the detection uncertainty is only around the gigahertz range. Though much higher accuracy (tens of MHz) was reported in other schemes [4], [5], time-consuming scanning is required. Besides, the desired extraction of the identified RF component is not integrated in those schemes, resulting in extra cost of devices as well as volume and power consumption. Note that in a radar system, especially in military circumstance, the RF signal under detection is generally spectrally sparse, which contains multiple (assume K ) RF components covering a broadband but possesses little energy outside these K discrete frequencies [6]. The compressive

Vol. 4, No. 3, June 2012

Page 664



Fig. 1. (a) System configuration of the proposed scheme. (b) Principle of compressing and encoding of a signal. B[ represents convolution.

sampling (CS) theory has been proposed as a promising methodology for reconstructing sparse signals with very few measurements [7]–[9]. Benefiting from the nature of photonics, the optical CS has also been proposed recently for larger bandwidth and superior size, weight, and power (SWaP) [6], [10]. However, the effective implementation, especially for photonics-assisted, multifrequency, broadband RF signals sensing, is still highly concerned. In [10], even though signals below 5 GHz were compressively digitalized, 40 optical paths, each of which contained a mode-locked laser (MLL), a modulator, an encoding sequence, a photodetector (PD), and an analog-to-digital converter (ADC) were required. Such complicated scheme was only demonstrated by numerical examples. In [6], one optical path was employed, but only single tone, which had to be the harmonics of ADC sampling rate, was faithfully reconstructed. Additionally, due to the conversion of noise from the whole band to frequency bins containing information, the extracted signal suffered deteriorated signal-to-noise ratio (SNR) and spurious tones in both schemes. In this paper, we propose a novel photonics-assisted CS system integrated with downconversion to achieve the aimed capacity of broadband, high accuracy, multifrequency resolving, as well as high-dynamic-range extraction of identified component. Experimentally, single and coupled RF frequencies ranging from 15 GHz to 20 GHz are distinguished and recognized through a single ADC with sub-Nyquist analog bandwidth of 826.75 MHz. The maximum detection error is 100 kHz. Under the same setup, the spur-free dynamic range (SFDR) of the downconversion can reach 93.3 dB Hz2=3 . Precisely resolving of as many as ten frequencies is demonstrated numerically. The principle and the essentials of system design are discussed.

2. System Design and Principle The proposed system integrated with RF frequency resolving and extraction is shown in Fig. 1(a). The sampling source for CS contains a short optical pulse train with repetition rate of frep , which is modulated by an N-bit pseudorandom bit sequence (PRBS). The RF signal x ðt Þ, which covers a known bandwidth of BRF but consists of multiple components centered around unknown frequencies, is modulated onto the PRBS train by a 1 2 Mach–Zehnder modulator (MZM). The MZM is biased so that one output arm works the same as a regular quadrature-biased 1 1 MZM, while the other arm outputs the complementary optical power. The time delays from the outputs of MZM to the inputs of balanced PD (BPD) are equal, and therefore, the output of the BPD is iðt Þ / pðt Þ sin x ðt Þ pðtÞ x ðt Þ (1) V V where pðt Þ is the optical power of the PRBS-modulated pulse train, and V is the half-wave voltage. After the low-pass filter (LPF), only baseband with bandwidth of BADC is digitalized by an ADC. Finally, the original RF spectral information is recovered in digital signal processor (DSP) unit periodically at a predefined data acquisition time.

Vol. 4, No. 3, June 2012

Page 665



The principle of multifrequency resolving is shown in Fig. 1(b). The pulse train is modulated by the PRBS pattern, resulting in a set of PRBS frequency lines equally spaced by frep =N. The PRBS lines have different complex amplitudes, which also cover a broadband. After mixing at the 1 2 MZM, each component of the RF signal is downconverted and multicast to a set of base-band lines (compression). The amplitudes of such base-band lines follow the pattern of a part of the original PRBS lines while their frequencies are corresponding to the particular RF component (encoding). Consequently, on one hand one could obtain the broadband RF spectral information by a muchnarrower-band ADC thanks to the spectrum compression. On the other hand, the original spectrum can also be recovered from the digitalized data based on the encoded patterns. This process is mathematically described in [8], and the recovery is theoretically supported by CS [7], [11]. In order to obtain the easier implementation, better frequency resolving performance, and integrated higher-dynamic-range information extraction, three improvements have been made compared to the previous photonics-based CS schemes. First, Fig. 1(a) shows a feasible scheme rather than [10] and offers the advantage of acquiring multiple and nonharmonic tones over [6]. While other designs [8], [10] focus on compressing the incoming signals with different encoding sequences, our scheme exploits the capacity of a single encoding sequence with only one optical sampling channel. In this case, the observation matrix in the reconstruction algorithm (which is defined in [8]) is a Toeplitz matrix, whose performance has been proved to guarantee the signal recovery [11]. Extensive researches [7], [8] indicate that stable recovery of incoming signal requires approximately (2) BADC ¼ O K logð2NBRF =Kfrep Þ frep =N where Oðx Þ means the same order of magnitude of x . Equation (2) shows that BADC grows logarithmically, rather than linearly, with BRF , which implies the possibility of sub-Nyquist sampling in our scheme. Performance according to the key parameters should be concerned for the proposed system design. On one hand, Fig. 1 shows the compression of broadband RF signal requires a set of weights covering also a broadband. The PRBS encoding is chosen since it provides flat Fourier coefficients (weights), which guarantee a uniform SNR and recovery probability [10]. Note that such weights are symmetrical around harmonics of frep =2 and repeat every frep . Therefore, two components will not be clearly distinguished after compression if their interval is any harmonics of frep =2 because they will have the same pattern. So, the RF bandwidth is limited by BRF G frep =2:

(3)

On the other hand, the power of PRBS lines decays with the increasing of RF frequency due to the nonzero optical pulse width, t . As a result, the proposed CS source gives an upper bound fmax for recoverable RF frequency as fmax 1=t :

(4)

Besides, according to the well-known property of Fourier transformation, the longer the data acquisition time, the smaller frequency grid one can get in DSP [5]. Therefore, the uncertainty of the detected frequencies is determined by the acquisition time as f 1=Tacquisition :

(5)

The second improvement is to employ a 1 2 MZM with BPD rather than the ordinary MZM in [6] and [10]. Mathematically, the output of BPD is PRBS-modulated short pulse train multiplied by RF signal, under the approximation of small-signal modulation on the 1 2 MZM. Such design provides not only a higher SNR due to balanced receiver, but also the capacity of stably detecting. Ordinary MZM cannot suppress the original PRBS frequency lines. Therefore, when the RF frequency is any harmonics of frep =N, the multicast and patterned lines overlap with the original PRBS lines, which will incur significant recovering instability. However, our design ensures a cancellation of the original

Vol. 4, No. 3, June 2012

Page 666



Fig. 2. Configuration of pulse generator. PS: phase shifter; LD: laser diode.

PRBS lines and keeps the modulation linear as well. This implementation makes the recognition of particular frequencies more likely. The third improvement is to integrate the capacity of downconverting identified signal. Though CS algorithm recognizes the frequency of the interested component, fs , precisely according to (5), the recovered signal usually has large noise and spurious tones (refer to the following experiment results as well as [6], [10]). Instead, we turn off the PRBS modulation in Fig. 1(a) and tune the repetition rate of the short pulse train to fLO (LO: local oscillator). Such pulse train is then employed directly for downconversion. From a spectral viewpoint, the pulse train is a broadband frequency comb with equal line interval of fLO . The desired downconversion occurs when the nth line of the frequency comb is aligned properly to the interested signal with fs n fLO ¼ fs fIF

(6)

where fIF is the predefined IF. Note that other lines should not overlap with the entire monitored RF band BRF to avoid interference from mixing between harmonics of fLO and any other components. As a result, the following equation should be maintained: fLO 9 BRF :

(7)

In our system of Fig. 1(a), the downconversion signal could be either output after LPF or digitalized by ADC for further analysis. Note that as long as (6) and (7) are satisfied, only one frequency line of the short pulse train, with frequency offset of nfLO , participates in the downconversion. As a result, our downconversion scheme is equivalent to that where a continuous wave LO is employed to coherently downconvert a particular RF component, with the assistance of filter [12], [13]. The nonlinearity of such downconversion link comes from the sinusoidal transfer function of MZM, according to (1), and has been well analyzed in [12], [13]. However, the integrated implementation offers the advantages of shared devices and reduced cost.

3. Experiment and Simulation A system design for the proposed integrated multifrequency recognition and downconversion is experimentally and numerically demonstrated. In our experiment, the pulse generator is constructed using a cascade of MZM and phase modulator (PM) [14], [15], as shown in Fig. 2. Both modulators are driven at frep ¼ 10 GHz. The wave shaper (Finisar wave shaper 4000s) is used to dechirp and compress the pulse to 15 ps. N is equal to 127 in the following PRBS modulation. The advantages of our pulse generator over MLL include reduced cost and easier repetition rate tuning for downconversion. Single or multiple RF tones serve as the signal under detection. For the availability of our equipment, the mixed signal is received by a 10-GHz BPD and a 10-GS/s oscilloscope (LeCroy WavePro 7400A), which is, however equivalent to an ADC with analog bandwidth of 826.75 MHz, since a numerical filter with above bandwidth is applied to the digitalized data. Every about 10 s, the acquired data are fed to the recovering process, which is currently an offline MATLAB program using gradient pursuit algorithm [16]. The real-time operation could be achieved by a DSP unit in practice. According to (3) and (4), our scheme is capable of detection and extraction unknown RF components within a given 5-GHz bandwidth at frequency as high as 60 GHz. Under the ADC

Vol. 4, No. 3, June 2012

Page 667



Fig. 3. Recovered spectra for one-tone signals. PSD: power spectral density.

Fig. 4. Recovered spectrum for two-tone signal where two frequencies are far separated. Insets: enlarged views.

Fig. 5. Recovered spectrum for two-tone signal where two frequencies are very close. Inset: enlarged view.

analog bandwidth of 826.75 MHz, the number of detectable sparse components is around 10, according to (2). In this paper, multifrequency recognition within 15 to 20 GHz is verified by the following examples. First, single-tone RF signals are sent into the proposed system to demonstrate the capacity of sensing single frequency. Frequencies equal to harmonics of frep =N are also tested. The recovered RF spectra, corresponding to different trials, are plotted in Fig. 3. Our experiment shows that unknown single frequencies, in the range of 15 GHz to 20 GHz, can be recovered precisely: The recovered (original) frequencies are 15.74813 GHz (15.74803 GHz, the 200th harmonics), 16.77175 GHz (16.77165 GHz, the 213th harmonics), 17.30002 GHz (17.30000 GHz), and 18.81899 GHz (18.81900 GHz, the 239th harmonics), respectively. The maximum detection error is 100 kHz. Second, the multicomponent resolving is shown by inputting two coupled RF tones. When the RF frequencies are 15.03000 GHz and 19.97000 GHz, which are located at the two sides of the given 5-GHz band, the recovered spectrum is shown in Fig. 4. The sensed two frequencies are 15.03006 GHz and 19.96998 GHz, with a maximum error of 6 kHz. Even if the two frequencies

Vol. 4, No. 3, June 2012

Page 668



Fig. 6. Simulation result of recovered spectrum for a ten-tone RF signal.

Fig. 7. IF signal and IMD3 powers versus the input RF power. The measured noise floor is 143 dB. Inset: the measured power spectrum of down-converted signal with input frequencies of 19.589 GHz and 19.59091 GHz, respectively.

are very close, i.e., 19.58900 GHz and 19.59091 GHz in another experiment, they are also reconstructed and distinguished clearly, as shown in Fig. 5. The sensed two frequencies are 19. 58899 GHz and 19.59099 GHz, with a maximum error of 80 kHz. Due to the limitation of our equipment, we show the capacity of recognizing more frequencies by a numerical example. In simulation, the parameters are the same as those in experiment, except that the input RF signal consists of ten tones randomly located between 15 GHz and 20 GHz. The reconstructed spectrum is shown in Fig. 6. Ten frequencies are recognized simultaneously with negligible frequency error. Note that spurious tones in the experiments (see Figs. 3–5) are owing to the downconversion of noise from the entire spectrum to baseband [10]. Since the noise is ignored in simulation, spurious tones disappear in Fig. 6. The maximum frequency detection error in experiment is 100 kHz, which is consistent with the calculation of (5) under the 10-s acquisition time. As a result, the setup imperfections, including the timing jitter of the sampling pulses (about 1 ps, which induces dithering of the PRBS lines in the frequency domain) and the time delay difference of the BPD (less than 10 ps), show no significant degeneration on the estimation accuracy. Besides, V of the 1 2 MZM (EOSpace) is 6.5 V, and the power of each RF tone is 10 dBm in our experiment. Input with low RF power will reduce the SNR of digitalized data, while very high signal power will introduce large nonlinearity as well as spurious tones, according to (1); either case will decrease the estimation accuracy. The final experiment is performed to illustrate the capacity of downconverting identified RF component. As shown in Fig. 5, the RF signal has already been recognized around 19.590 GHz by CS. Assume fIF ¼ 150 MHz. On the basis of (6) and (7), we choose n ¼ 2 and tune the repetition

Vol. 4, No. 3, June 2012

Page 669



rate of our pulse generator to fLO ¼ 9:72 GHz. The spectrum of downconverted two-tone signal with 3rd order intermodulated distortion (IMD3) is shown in Fig. 7. The SFDR after downconversion is also measured, which reaches 93.3 dB Hz2=3 at 1 Hz noise bandwidth. In our experiment, the SFDR is mainly limited by the relatively high noise floor (143 dB), which is induced by the optical amplifier in the fiber link. In addition, the dispersed fLO power is also attributable: The short pulse train contains many frequency lines, while only one of them contributes to the downconversion. The SFDR could be improved by higher optical power source, low noise optical amplifier, as well as BPD that tolerates higher power.

4. Conclusion On the basis of CS, we demonstrated for the first time a system capable of precise multi-RFfrequency resolving as well as high-dynamic-range downconversion. With an example of RF range from 15 to 20 GHz, our system estimated frequencies by ADC with sub-Nyquist bandwidth of only 826.75 MHz. The maximum error is 100 kHz. Such design was expected to distinguish as many as ten RF components within the 5-GHz band. By employing an optical short pulse generator with tunable repetition rate, our system showed the integrated capacity of signal extraction for further digitalizing or processing. SFDR of 93.3 dB Hz2=3 was obtained in the prototype experiment.

References [1] W. Wang, R. L. Davis, T. J. Jung, R. Lodenkamper, L. J. Lembo, J. C. Brock, and M. C. Wu, BCharacterization of a coherent optical RF channelizer based on a diffraction grating,[ IEEE Trans. Microw. Theory Tech., vol. 49, no. 10, pp. 1996–2001, Oct. 2001. [2] D. B. Hunter, L. G. Edvell, and M. A. Englund, BWideband microwave photonic channelised receiver,[ in Proc. Int. Topical Meeting Microw. Photon., Oct. 2005, pp. 249–252. [3] X. Zou, W. Pan, B. Luo, and L. Yan, BPhotonic approach for multiple-frequency-component measurement using spectrally sliced incoherent source,[ Opt. Lett.35, no. 3, pp. 438–440, Feb. 2010. [4] Y. Wang, H. Chi, X. Zhang, S. Zheng, and X. Jin, BPhotonic approach for microwave spectral analysis based on Fourier cosine transform,[ Opt. Lett., vol. 36, no. 19, pp. 3897–3899, Oct. 2011. [5] B. Vidal, T. Mengual, and J. Marti, BPhotonic technique for the measurement of frequency and power of multiple microwave signals,[ IEEE Trans. Microw. Theory Tech., vol. 58, no. 11, pp. 3103–3108, Nov. 2010. [6] J. M. Nichols and F. Bucholtz, BBeating Nyquist with light: A compressively sampled photonic link,[ Opt. Exp., vol. 19, no. 8, pp. 7339–7348, Apr. 2011. [7] E. Candes, J. Romberg, and T. Tao, BRobust uncertainty principles: Exact signal reconstruction from highly incomplete frequency information,[ IEEE Trans. Inf. Theory, vol. 52, no. 2, pp. 489–509, Feb. 2006. [8] M. Mishali and Y. C. Eldar, BFrom theory to practice: Sub-Nyquist sampling of sparse wideband analog signals,[ IEEE J. Sel. Topics Signal Process., vol. 4, no. 2, pp. 375–391, Apr. 2010. [9] J. A. Tropp, J. N. Laska, M. F. Duarte, J. K. Romberg, and R. G. Baraniuk, BBeyond Nyquist: Efficient sampling of sparse bandlimited signals,[ IEEE Trans. Inf. Theory, vol. 56, no. 1, pp. 520–544, Jan. 2010. [10] H. Nan, Y. Gu, and H. Zhang, BOptical analog-to-digital conversion system based on compressive sampling,[ IEEE Photon. Technol. Lett., vol. 23, no. 2, pp. 67–69, Jan. 2011. [11] J. Haupt, W. U. Bajwa, G. Raz, and R. Nowak, BToeplitz compressed sensing matrices with applications to sparse channel estimation,[ IEEE Trans. Inf. Theory, vol. 56, no. 11, pp. 5862–5875, Nov. 2010. [12] A. Agarwal, T. Banwell, P. Toliver, and T. K. Woodward, BPredistortion compensation of nonlinearities in channelized RF photonic links using a dual-port optical modulator,[ IEEE Photon. Technol. Lett., vol. 23, no. 1, pp. 24–26, Jan. 2011. [13] A. Agarwal, T. Banwell, and T. K. Woodward, BOptically filtered microwave photonic links for RF signal processing applications,[ J. Lightwave Technol., vol. 29, no. 16, pp. 2394–2401, Aug. 2011. [14] T. Yamamoto, T. Komukai, K. Suzuki, and A. Takada, BMulticarrier light source with flattened spectrum using phase modulators and dispersion medium,[ J. Lightwave Technol., vol. 27, no. 19, pp. 4297–4305, Oct. 2009. [15] R. Wu, V. R. Supradeepa, C. M. Long, D. E. Leaird, and A. M. Weiner, BGeneration of very flat optical frequency combs from continuous-wave lasers using cascaded intensity and phase modulators driven by tailored radio frequency waveforms,[ Opt. Lett., vol. 35, no. 19, pp. 3234–3236, Oct. 2010. [16] T. Blumensath and M. E. Davies, BGradient pursuits,[ IEEE Trans. Inf. Theory, vol. 56, no. 6, pp. 2370–2382, Jun. 2008.

Vol. 4, No. 3, June 2012

Page 670