SVM Detection for Superposed Pulse Amplitude Modulation in Visible Light Communications

Transcription

1 SVM Detection for Superposed Pulse Amplitude Modulation in Visible Light Communications Youli Yuan 1, Min Zhang 1*, Pengfei Luo, Zabih Ghassemlooy, Danshi Wang 1, Xiongyan Tang 1, Dahai Han 1 1 State Key Laboratory of Information Photonics and Optical Communications, Beijing University of Posts and Telecommunications, Beijing , China 2 Research Department of HiSilicon, Huawei Technologies Co., Ltd, Beijing , P. R. China 3 Optical Communications Research Group, NCRLab, Faculty of Engineering and Environment, Northumbria University, Newcastle upon Tyne NE1 8ST, U.K. *mzhang@bupt.edu.cn Abstract A support vector machine (SVM)-based data detection for 8-superposed pulse amplitude modulation in visible light communication is proposed and experimentally demonstrated. In this work, the SVM detector contains three binary classifiers with different classification strategies. And the separating hyperplane of each SVM is constructed by training data. The experiment results show that the SVM detection offers 35% higher data rates when compared with the traditional direct decision method. Keywords support vector machine; superposed pulse amplitude modulation; visible light communiction; direct decision. I. INTRODUCTION With the increasing popularity of smart mobile devices, the wireless data traffic is growing exponentially [1]. This has resulted in shortage of the radio frequency (RF) spectrum and made the RF spectrum become a precious commodity. Since the RF spectrum is limited and expensive due to the license fees, the optical wireless communications (OWC) has evolved to become a high-capacity complementary technology to the RF communications, which can alleviate the spectrum shortage. In addition, the RF wireless technologies are not allowed to be used in electromagnetic sensitive areas such as hospitals, aircraft cabins, mines and others because of the electromagnetic interference [2]. As a new lighting technology, the white light-emitting diodes (LEDs) are being installed at a fast rate globally, thus creating a golden opportunity for full scale practical realization of visible light communications (VLC) (or LiFi) technology. In VLC systems, the information signal is used to intensity modulate (IM) LEDs. On-off-keying (OOK) is one of the most widely used modulation format, which is relatively easily to implement in hardware [3]. However, the transmission data rate R d is limited by the low bandwidth of white LEDs, which is only several MHz due to the slow time constant of the yellow phosphor [4]. Hence, R d of conventional OOK could only reach several Mbps. Consequently, several studies have been performed to increase R d in VLC systems. Firstly, quadrature amplitude modulation (QAM) with orthogonal frequency division multiplexing (OFDM) including directcurrent-biased optical OFDM (DCO-OFDM) and asymmetrically clipped optical OFDM (ACO-OFDM) are widely used in VLC systems [5]. But OFDM requires that the transmitter (Tx) (including LED and the driven circuit) has a wide dynamic range and reasonably good linearity characteristics. Secondly, OFDM based on employing a discrete power level stepping technology was proposed in [6] to address the shortcoming of conventional optical Txs. The core of this technology is the digital to analog conversion (DAC) implemented in the optical domain [7-8]. However, these modulation formats rely on complex digital signal processing (i.e., fast Fourier transform (FFT) and inverse fast Fourier transform (IFFT)). This leads to increase complexity for both Tx and receiver (Rx). To avoid such complexity, spatial modulation techniques were proposed with higher spectral efficiency as in [9-15]. Superposed pulse amplitude modulation (SPAM), which is low-cost, insensitive to nonlinearity of LEDs, was proposed to reduce the complexity of the transceiver and enhance the frequency efficiency [12]. In SPAM, an N parallel NRZ-OOK signals with different amplitudes are used for IM of N-LED, whose optical power are then superposed in a free space channel to produce a 2 N -level signal. Following this principle, the Tx can provide a larger dynamic range than traditional one of LED system. Machine learning has been widely employed to solve problems in different areas such as big data processing, personalized recommendation, pattern recognition, data mining and artificial intelligence, etc. [16]. What s more, this technique has been used in VLC systems to reduce the effect of fluorescent light interference, inter-symbol interference (ISI) and artificial light interference [17-19]. In VLC systems, the background noise, electrical noise from transceiver and bandwidth limitation of LED contribute to the distortion of the received waveform. Amplitude modulation schemes such as pulse amplitude modulation (PAM) are especially sensitive to the signal distortion, which leads to poor performance in IM/direct decision (DD) based systems. In this paper, we introduce a support vector machine (SVM) algorithm in 8-SPAM VLC system to improve system performance. After being trained with a training data, SVM is able to learn the properties of received signal and generate an improved decision hyperplane for more precise data detection. In order to test the bit-error rate (BER) performance of a SVMbased 8-SPAM VLC system, an experimental test-bed has been built. The results show that the SVM detection scheme can achieve improved performance when compared with the traditional DD method /16/$ IEEE

2 (a) Fig. 1. A waveform emitted by three LEDs with different amplitudes. The structure of this paper is as follows. The operation of each technique is detailed in Section II. The experiment setup of the SVM-based detection for 8-SPAM is described in Section III. The experiment result and discussion are presented in Section IV. Finally, Section V concludes this paper. II. OPERATION PRINCIPLE A. Concept of SPAM The basic concept of SPAM is to use multiple LEDs to realize DAC in the optical domain [7-8] - multiple optical pulses transmitted in parallel from different spatial positions in a LED array, which eventually produces multi-amplitude light levels to realize PAM in the optical domain. In this scheme, a parallel binary sequences control multiple LEDs (i.e., on or off) with each LED transmitting at a specific intensity (i.e., optical power) level. With the superposition of light intensities in the free space channel, a multi-level light signal is generated. Assuming I is the maximum light intensity among all independent LEDs, in order to generate an equally spaced PAM signal, the corresponding LED light intensities are set as: I 1 = I / 2 N - 1 ), I 2 = I / 2 N - 2 ),..., I k = I / 2 N - k ), and I N = I, where I k is the k th LED s output optical power, and N is the number of LEDs. Therefore, the equally spaced PAM signal could be generated by combining different optical power levels in the free space channel, which can also be considered as DAC of signal. Finally, the 2 N -level optical PAM signal can be detected at the Rx using only a single photo-detector (PD). A standard PAM demodulator can then be employed for demapping the received multi-amplitude signal back to the transmitted serial data stream. Therefore, by employing N-LED and a single PD, the SPAM scheme could archive N times that of the one-led and one-pd OOK system. Unlike OFDM or MIMO scheme which have high spectral efficiency, the SPAM scheme is a relatively simple scheme not requiring complex data processing and DAC. It is known that the power versus current characteristics of LEDs are linear over a certain range, beyond which the characteristics becomes non-linear. In this paper, we generated three OOK signals for IM of 3 LEDs with different intensities to realize the 8-SPAM scheme. Fig. 1 shows an example of (b) Fig SPAM modulation scheme: (a) three waveforms that drive three LEDs, and (b) the corresponding combined waveforms and signal mapping strategy. Fig. 3. SVM schematic diagram: (a) for linearly separable data, (b) for linearly inseparable data, and (c) using kernel function to realize linear classfiction for linearly inseparable data. a normalized amplitude waveform for LEDs A, B and C having the same frequency. To achieve 8-level of light intensities, the normalized amplitude of LEDs A, B, and C are set to 0.25, 0.5 and 1.0, respectively. Fig. 2(a) depicts the waveforms that drive the three LEDs, whereas Fig. 2(b) illustrates the corresponding combined 8- SPAM waveform and the signal mapping strategy. It is notable that the latter is the same as the traditional 8-PAM signal, therefore, an 8-PAM demodulator can be adopted in the proposed 8-SPAM to recover the information.

3 Fig. 5. Block diagram of experimental setup of the proposed system. Fig. 4. Classification strategy based on SVM for 8-SPAM. B. SVM for 8-SPAM Detection SVM is a well-known statistical classification and regression technique, which has been used to solve a range of engineering problems, notably in statistical signal processing, pattern recognition, image analysis, and digital communication systems. In the latter application SVM has been used as a new method for channel equalization [20] and for combating ISI and the co-channel interference [21]. A SVM is fundamentally a two-class classifier, which could generate boundary to classify two segments of data. Based on the statistical theory, SVM aims to find an optimal (maximize) margin, which is defined by the smallest distance between the decision boundary and any point of the samples, as is shown in Fig. 3(a). A subset of the data points which determine the location of the boundary are known as the support vectors. The SVM is very effective in solving linear classification problems. In practice, however, the data are usually linearly inseparable, see Fig. 3(b). Hence, the use of the kernel trick [16], which maps linearly inseparable data from lower-dimensional feature space to higherdimensional feature space. In the higher-dimension of the data, which in the original dimension is nonlinearly separable, is linearly separable, as depicted in Fig. 3(c). Through this way, we can solve this linear classify problem in higher-dimension, which is equal to handling the nonlinear problem in the original dimension space. As stated SVM is a two-class classifier as shown in Fig. 3. However, in many situations there is the requirement for classifying signals into multiple classes. Thus, the need for a multi-class SVM classifier, which can be realized by combining multiple two-class SVMs. In this work, 3-ary SVM is adopted as an 8-SPAM demodulator. In this approach, different amplitudes of 8-SPAM signals can be considered as eight different classes of the received data where a better decision boundary can be found using SVM than the traditional DD based PAM. Note that in Fig. 4 each symbol bit is labeled as +1 or -1 by a binary SVM. For example, if a bit sequence of 110 is correctly detected by all three SVMs, then the least significant bit 0, the middle bit 1, and the most significant bit 1 are labeled as -1-1 and +1 by SVMs 1, 2 and 3, respectively. Since all the permutation of the label created by SVMs are unique, the combination of three SVMs could eventually perform as an 8-SPAM decoder. Fig. 6. Experiment setup for the SVM 8-SPAM VLC system. TABLE I. Parameter Test distance Test bit rate Traning data set Half-angle of conser cup Signal amplitude for each LED The bandwith of PD Peak Response Diameter of the condenser lens Focus length of the condenser lens KEY PARAMETERS OF EXPERIMENT Value 1.6 m Mbps 6000 symbols mv pp, 1 V pp, 2 V pp 150 MHz 0.45 A/W (750nm) 100 mm 132 mm SVM library LibSVM [22] III. EXPERIMENTAL SETUP Fig. 5 shows the system block diagram of the proposed SVM based 8-SPAM VLC system, whereas Fig. 6 illustrates the experimental setup. In our prototype, we produced the parallel data stream (i.e., SPAM) in Matlab and use an arbitrary waveform generator (AWG) (TTi TGA12104) to generated the electrical version of SPAM for IM of three multi-chip white LED chips (Cree MC-E). The 3 db bandwidth of each LED chip is ~2 MHz. At the Rx, a condenser lens, a PD (THORLABS PDA 10A-EC) was used to focus light and convert the focused optical signal into an electric signal, respectively. A digital oscilloscope (Agilent DSO9104H) was employed to capture the signal for further off-line signal

4 Fig. 7. BER as a function of transmission rates for SVMs detection and direct decision. processing in the Matlab domain. A 3-ary SVM model containing three SVMs was used for detection of the captured signal. Since SVM can acquire signal characteristics from the training data for classification, then no prior knowledge of the channel state information (CSI) is necessary. Finally, based on the output labels provided for the 3-ary SVM, mapping of the data to corresponding symbols was carried out in Matlab domain, as in Fig.. All the key system parameters are listed in Table I. IV. RESULT AND DISCCUSSION Note that following completion of the training process, the received data was decoded using SVMs. The measured BER vs. the data rate performance for the SVM and DD based systems are shown in Fig. 7. For each data rate, the total number of recorded 8-SPAM symbols were ~ Therefore, the total number of raw binary bits was ~ It can be inferred from Fig. 7 that the BER performance deteriorate with increasing data rate. However, compared to the DD method, the SVM detection display superior BER performance. For example, at a BER the maximum data rates are ~5.1 Mbps and ~6.1 Mbps for DD and SVM detections, respectively. At the FEC limit of the SVM detection offer a maximum data of 8.1 Mbps over a transmission span of 1.6 m compared to the 5.8 Mbps of DD detection. This is due to the enhanced decision boundary based on the specific distribution of received signal, which was obtained during the training process, could achieve more accurate classification. Note that also shown are the eyediagrams. Finally, Fig. 8 shows an example of the detection process based on using SVM to decode 8-SPAM signal at a date rate of 6 Mbps. The horizontal and vertical axis represents time and signal amplitude, respectively. In each subplot, the received signal corresponding labels of +1 and -1 are marked by Fig ary SVM detection examples for 8-SPAM signal with signal calssification by: (a) SVM1, (b) SVM2, and (3) SVM3. green triangles and red circles, respectively. As is depicted in Fig. 8, due to the limited LED bandwidth, back ground noise and electrical noise, the constellations are fluctuating around the ideal constellation point. It is clear that SVM can obtain the characteristics from the training data and be able to adjust the decision boundary. By doing so, SVM can mitigate the impact of fluctuations, thus improving the BER performance compared to the traditional DD method. V. CONCLUSION In this paper, we presented an experimental VLC system using SPAM modulation to realize data transmission for an indoor environment. The SVM scheme, which makes accurate decision in symbol detection, was employed to obtain the optimum decision boundary for decoding the SPAM signal. A commercial multi-chip was used to generate the 8-SPAM signal and a PD was used as a Rx. We showed that the transmission data rate is three times that of a single LED and a single PD based OOK system. The experimental results indicated a maximum data rate of 8 Mbps at a BER of can be archived using the SVM scheme over a communication range of 1.6 m. This is an increase of ~35% compared with the DD scheme. Note that higher order SPAM signals with higher data rates can be realized by using more LEDs.

5 ACKNOWLEDGMENT This work was supported by NSFC Project No , Doctoral Scientific Fund of MOE of China (No ) and the Royal Society Newton International Exchanges between U.K. and China under Grant NI REFERENCES [1] S. Dimitrov and H. Haas, Principles of LED Light Communications: Towards Networked Li-Fi: Cambridge University Press, [2] Z. Ghassemlooy, W. Popoola, and S. Rajbhandari, Optical wireless communications: system and channel modelling with Matlab : CRC Press, [3] H. Elgala, R. Mesleh, and H. Haas, "Indoor optical wireless communication: potential and state-of-the-art," Communications Magazine, IEEE, vol. 49, pp , [4] H. Li, X. Chen, B. Huang, D. Tang, and H. Chen, "High bandwidth visible light communications based on a post-equalization circuit," Photonics Technology Letters, IEEE, vol. 26, pp , [5] O. González, R. Pérez-Jiménez, S. Rodriguez, J. Rabadán, and A. Ayala, "OFDM over indoor wireless optical channel," in Optoelectronics, IEE Proceedings-, 2005, pp [6] T. Fath, C. Heller, and H. Haas, "Optical wireless transmitter employing discrete power level stepping," Journal of Lightwave Technology, vol. 31, pp , [7] J. Armstrong, "Optical domain digital-to-analog converter for visible light communications using LED arrays [Invited]," Photonics Research, vol. 1, pp , [8] J. Yew, S. D. Dissanayake, and J. Armstrong, "Performance of an experimental optical DAC used in a visible light communication system," in Globecom Workshops (GC Wkshps), 2013 IEEE, 2013, pp [9] T. Fath, H. Haas, M. D. Renzo, and R. Mesleh, "Spatial modulation applied to optical wireless communications in indoor LOS environments," in Global Telecommunications Conference (GLOBECOM 2011), 2011 IEEE, 2011, pp [10] W. O. Popoola and H. Haas, "Demonstration of the merit and limitation of generalised space shift keying for indoor visible light communications," Lightwave Technology, Journal of, vol. 32, pp , [11] W. O. Popoola, E. Poves, and H. Haas, "Spatial pulse position modulation for optical communications," Lightwave Technology, Journal of, vol. 30, pp , [12] J. Li, Z. Huang, R. Zhang, F. Zeng, M. Jiang, and Y. Ji, "Superposed pulse amplitude modulation for visible light communication," Optics express, vol. 21, pp , [13] L. Song, P. Luo, M. Zhang, Z. Ghassemlooy, D. Han, and H. L. Minh, "Undersampled Digital PAM Subcarrier Modulation for Optical Camera Communications," in Asia Communications and Photonics Conference 2015, Hong Kong, 2015, p. AM1E.2. [14] S. H. Yang, D. H. Kwon, S. J. Kim, and S. K. Han, "Overcoming bandwidth limitation of light emitting diode in visible light communication using differential pulse amplitude modulation," Optical Engineering, vol. 54, pp , [15] N.Ishikawa, S.Sugiura, "Maximizing Constrained Capacity of Power- Imbalanced Optical Wireless MIMO Communications Using Spatial Modulation," Lightwave Technology, Journal of, vol.33, pp , [16] C. M. Bishop, Pattern recognition and machine learning: springer, [17] S. Rajbhandari, Z. Ghassemlooy, and M. Angelova, "Effective denoising and adaptive equalization of indoor optical wireless channel with artificial light using the discrete wavelet transform and artificial neural network," Journal of Lightwave technology, vol. 27, pp , [18] S. Rajbhandari, P. A. Haigh, Z. Ghassemlooy, and W. Popoola, "Wavelet-neural network VLC receiver in the presence of artificial light interference," Photonics Technology Letters, IEEE, vol. 25, pp , [19] P. A. Haigh, Z. Ghassemlooy, S. Rajbhandari, I. Papakonstantinou, and W. Popoola, "Visible light communications: 170 Mb/s using an artificial neural network equalizer in a low bandwidth white light configuration," Lightwave Technology, Journal of, vol. 32, pp , [20] D. Sebald and J. Bucklew, Support Vector Machine Techniques for Non Linear Equalization, IEEE Trans. Sig. Proc., vol. 48, 2000, pp [21] J. P. Dubois, and M. Abdul-Latif, Improved M-ary Signal Detection using Support Vector Machine Classifiers, PROCEEDINGS OF WORLD ACADEMY OF SCIENCE, ENGINEERING AND TECHNOLOGY, Vol. 7, pp , Aug [22] C.-C. Chang and C.-J. Lin, "LIBSVM: A library for support vector machines," ACM Transactions on Intelligent Systems and Technology (TIST), vol. 2, p. 27, 2011.