The surge of visible-light communication (VLC) derives from the ever-increasing demand of wireless network capacity and the limited spectral resources in the radiofrequency domain
Opto-Electronic Advances, Volume. 3, Issue 8, 200009-1(2020)
Demonstration of a low-complexity memory-polynomial-aided neural network equalizer for CAP visible-light communication with superluminescent diode
Visible-light communication (VLC) stands as a promising component of the future communication network by providing high-capacity, low-latency, and high-security wireless communication. Superluminescent diode (SLD) is proposed as a new light emitter in the VLC system due to its properties of droop-free emission, high optical power density, and low speckle-noise. In this paper, we analyze a VLC system based on SLD, demonstrating effective implementation of carrierless amplitude and phase modulation (CAP). We create a low-complexity memory-polynomial-aided neural network (MPANN) to replace the traditional finite impulse response (FIR) post-equalization filters of CAP, leading to significant mitigation of the linear and nonlinear distortion of the VLC channel. The MPANN shows a gain in Q factor of up to 2.7 dB higher than other equalizers, and more than four times lower complexity than a standard deep neural network (DNN), hence, the proposed MPANN opens a pathway for the next generation of robust and efficient neural network equalizers in VLC. We experimentally demonstrate a proof-of-concept 2.95-Gbit/s transmission using MPANN-aided CAP with 16-quadrature amplitude modulation (16-QAM) through a 30-cm channel based on the 442-nm blue SLD emitter.
Introduction
The surge of visible-light communication (VLC) derives from the ever-increasing demand of wireless network capacity and the limited spectral resources in the radiofrequency domain
In VLC systems, both the light emitters and detectors
The necessity for VLC transmission or applications under a high-attenuation channel inevitably requires a high signal amplitude, i.e., more than the 0.5 V used in previous demonstrations
CAP as a single-carrier modulation scheme has been proposed in VLC depicting a lower complexity and an improved PAPR when compared to orthogonal frequency division multiplexing (OFDM) schemes (i.e. DMT)
Here, we design and propose an innovative memory-polynomial-aided neural network (MPANN) for the novel high-speed SLD-based VLC system with CAP modulation. We implement the MPANN as a post-equalization filter and study the necessity and efficiency of the MPANN equalizer by providing a comprehensive comparison of both the equalization performance and computation complexity between MPANN and other traditional CAP equalization schemes, including the least mean square (LMS) equalizer, the 2nd-order Volterra series (VOLT2) equalizer, the digital pre-distortion (DPD), and a standard DNN. As compared to the LMS equalizer, the MPANN provides up to 2.7 dB gain of Q factor, and as compared to the DNN, the MPANN shows equivalent Q factor with nearly a fifth of the spatial complexity. These results demonstrate the superior equalization performance and the practical low complexity of the developed MPANN equalizer. A proof-of-concept VLC transmission through a 30-cm free-space link based on a 442-nm blue SLD emitter achieves 2.95-Gbit/s using the MPANN-aided CAP modulation. The demonstration suggests that the MPANN equalizer is a promising option for a nonlinear VLC system, especially SLD-based VLC systems.
Principle
Blue superluminescent diode
The SLD was fabricated by etching a c-plane GaN-based commercial laser epitaxial structure, designed with a facet tilting of 12°, and a ridge waveguide with a length of ~1 mm and a width of ~15 μm is similar to reported elsewhere
Figure 1.SLD device and electro-optical characteristics.(
The electroluminescence spectra were measured with an optical spectrum analyzer (Yokogawa AQ6373B) at injection currents from 100 mA to 1000 mA as seen in
CAP and MPANN equalizer
For traditional CAP, the finite impulse response (FIR) filter is usually used in the receiver side to mitigate the linear damage, such as the signal attenuation and the ISI from the channel and imperfect devices. However, a long filter length is usually needed for the FIR filter
Most of these advanced equalization schemes obtain the optimal equalization performance at the cost of computational complexity, hence, it is important to consider the trade-off between the performance and the complexity of the filters when a practical implementation is intended. For example, the DPD is believed to be an effective and popular method which simply uses the memory polynomial to shrink the equalization's complexity
By combining the theory of the DPD and MLP, a novel MPANN equalizer was designed on a three-layers structure, including a memory-polynomial layer (MP layer), a first hidden layer, and an output layer. Its structure is presented in
Figure 2.(
where m and r represent the linear memory depth and nonlinear memory depth of the channel, respectively, which can be further studied in Ref.
In order to apply MPANN at the receiver side, the first task is to estimate the optimal network weights
where L stands for the number of layers of MPANN, d and q represent the node number of the l-th layer and the (l+1)-th layer. Thus, the output of (l+1)-th layer
where f(x) is the rectified linear unit (ReLU) used as the activation function in MPANN, which has been verified to be better than other activation functions in the gradient propagation and computation
where Y is the value of the estimated symbol for the received signal X outputted from MPANN. The Z is the correct label value of the estimated symbol Y. The B is the batch size. The Yi and Zi stand for the i-th estimated symbol and its label in one batch. Finally, the BP algorithm is utilized to carry out the training process to reduce the loss function by updating the network weights. Validation set and test set are also adopted to prevent the over-fitting problem and evaluate the performance of MPANN, respectively. During the training and test process of MPANN, the order of all samples in every batch is randomized to avoid the memory effect for PRBS sequence
In order to verify the quality of the MPANN in the SLD-based VLC system, it should be fairly compared with LMS, VOLT2, DPD and DNN technology.
Experimental setup
The experimental setup and the schematic of the MPANN-aided CAP modulation SLD-VLC system are shown in
Figure 3.VLC experimental setup.(
At the digital signal processing (DSP) part, the MPANN serves as the first-stage equalizer to mitigate both the linear and nonlinear damage from the devices and channel. To highlight the role of the MPANN as a robust and relatively practical equalizer, the LMS, VOLT2, DNN and DPD equalizers are also set as the first-stage equalizers. After the equalization, the real and imaginary parts of the output signal are respectively filtered by two matching filters mfI and mfQ, given in Ref.
Results and discussion
This section details the results of the SLD-based CAP VLC link and the fair comparison of MPANN with other equalization schemes based on their equalization performance and computation complexity.
First, we tested the BER performance under the different injection currents and the various peak-to-peak voltages (Vpp) simultaneously, as seen in
Figure 4.(
Even though the bias point has been set up correctly, the nonlinear effect induced by the exponential behavior of the SLD's L-I-V relationship (
In order to verify the MPANN as an effective first-stage nonlinear equalizer, three traditional linear and nonlinear equalization schemes, including LMS, Volterra-series and DPD technology, were investigated in an evaluation benchmark.
Figure 5.BER performance when using(
Next, we try to find an optimal structure of DNN and MPANN to achieve a balance between complexity and performance. That is to say, all hyperparameters including the number of nodes in every layer, the number of the hidden layers in DNN and MPANN, and the linear and nonlinear memory depths in MPANN need to be discussed. The simultaneous iteration for all hyperparameters is time-costly while it can find the optimal results. Thus, the hyperparameters are optimized in the control variate technique, which means that one hyperparameters is investigated while the other hyperparameters are fixed. The hyperparameters that have been optimized maintain the optimal value. The hyperparameters that are waiting in line maintain a high value to avoid becoming the limited factor to the equalization performance of the MPANN. According to the results of Ref.
Given the BER performance in
Figure 6.BER performance when varying(
Figure 7.(
Simultaneously, it is important to compare the computation complexity of all the equalization schemes used under their optimal structure. Usually, computation complexity is divided into time complexity and spatial complexity. Time complexity is difficult to be analyzed fairly for each equalization scheme, because it is not only related to the number of weights, but also is related to the optimization algorithm in the training process of every equalization scheme. However, the spatial complexity represents the required memory space for running each equalization scheme, which is a fair figure of merit to evaluate the complexity of equalization schemes for practical implementation and can be simply represented by the required number of weights to be updated in the training process. The structure and spatial complexity comparison of the investigated equalization schemes are all summarized in
|
We measured the achievable transmission data rate of CAP-16-QAM when MPANN and other equalization schemes serve as the first-stage equalizer. The results are shown in
Even though the LMS, DPD, and VOLT2 also improve the constellation diagram (
Figure 8.The constellation diagram of CAP-16-QAM at 2.4 Gbit/s using the setup in
DNN needs a large number of layers and nodes as the consuming source to refine the optimal network weights from the coarse samples in a low data dimension. To eliminate complexity concerns, MPANN takes advantage of the prior knowledge of the nonlinear effect model to simplify its network structure. This approach lifts the original samples to a higher data dimension by memory polynomial expansion, helping the MPANN to find the optimal network weights easily and efficiently. We observe that the MPANN's small-size structure can achieve a performance equivalent to the DNN using nearly a fifth of the DNN's spatial complexity. To further explain why the MPANN outperforms other equalizers, two points can be mentioned: First, the neural structure and the advanced back-propagation algorithm have a more powerful ability to deal with classification and regression problems, on which the signal prediction is just a related application. Second, the use of prior knowledge from the nonlinearity model helps to reduce the number of layers and nodes needed for MPANN, resulting in a simplified structure and lower complexity as compared to DNN.
As a result, given the balance between performance and complexity, we propose MPANN as an efficient post-equalizer in the SLD-based VLC system. As seen in
Conclusions
The evolution of VLC as an attractive tool for the next generation of communications is based on its capability to provide multi-Gbit/s data rates while using unlicensed and interference-free channels, complementing the limited spectral resources of the conventional radio frequency regime. In this paper, we demonstrate a multi-Gbit/s VLC link using a novel MPANN-aided CAP modulation and a GaN-based SLD. We analyzed the communication performance of SLD-based CAP modulation using the MPANN as a post-equalizer. We compared the MPANN with equivalent LMS, 2nd-order Volterra-series, DPD and MLP (DNN) equalizers to measure the advantages and disadvantages of the MPANN in terms of data rate, correction of the linear and nonlinear distortion, and complexity. MPANN achieves Q factor gain of up to 2.7 dB higher than the LMS equalizer and complexity over 4 times less than that of DNN at a similar data rate. Such facts prove MPANN a low-complexity and practical equalizer for SLD-based VLC system. Using a 442-nm SLD emitter and given the superior performance of MPANN-aided CAP to resist the linear and nonlinear distortion, we successfully achieved up to 2.95 Gbit/s data rate in a VLC link through 30 cm free-space for a proof-of-concept demonstration.
Acknowledgements
This work was supported in part by the National Key Research, Development Program of China (2017YFB0403603), and the NSFC project (No. 61925104). JAHL, YM, TKN and BSO gratefully acknowledge the financial support from King Abdullah University of Science and Technology (KAUST) through BAS/1/1614-01-01, REP/1/2878-01-01, GEN/1/6607-01-01, and KCR/1/2081-01-01. This publication is partially supported by the King Abdullah University of Science and Technology (KAUST) Office of Sponsored Research (OSR) under Award No. OSR-CRG2017-3417. JAHL further acknowledge access to the KAUST Nanofabrication Core Lab for the fabrication of devices.
Competing interests
The authors declare no competing financial interests.
[1] [1] 11th IEEE International Symposium on Personal Indoor and Mobile Radio Communications 1325-1329 (IEEE, 2000); http://doi.org/10.1109/PIMRC.2000.881634.
[9] C H Cheng, C C Shen, H Y Kao, D H Hsieh, H Y Wang et al. 850/940-nm VCSEL for optical communication and 3D sensing. Opto-Electron Adv, 1, 180005(2018).
[13] [13] Optical Fiber Communication Conference (OFC) M3I.5 (OSA, 2019); http://doi.org/10.1364/OFC.2019.M3I.5.
[14] [14] Wei L Y, Chow C W, Liu Y, Yeh C H. Multi-Gbit/s phosphor-based white-light and blue-filter-free visible light communication and lighting system with practical transmission distance. Opt Express28, 7375-7381 (2020).
[20] [20] Proceedings of the SPIE 10483, Optical Coherence Tomography and Coherence Domain Optical Methods in Biomedicine XXⅡ, 104832T (SPIE, 2018); http://doi.org/10.1117/12.2288246.
[22] [22] Proceedings of the SPIE Digital Optical Technologies 2019 110620F (SPIE, 2019); http://doi.org/10.1117/12.2527626.
[25] C Shen, J A Holguin-Lerma, A A Alatawi, P Zou, N Chi et al. Group-Ⅲ-nitride superluminescent diodes for solid-state lighting and high-speed visible light communications. IEEE J Sel Top Quantum Electron, 25, 2000110(2019).
[26] [26] Proceedings of the SPIE 11307, Broadband Access Communication Technologies XIV 113070H (SPIE, 2020); http://doi.org/10.1117/12.2543983.
[28] [28] Wu F M, Lin C T, Wei C C, Chen C W, Chen Z Y et al. Performance comparison of OFDM signal and CAP signal over high capacity RGB-LED-based WDM visible light communication. IEEE Photonics J5, 7901507 (2013).
[29] [29] 2019 IEEE/CIC International Conference on Communications in China (ICCC) 173-176 (IEEE, 2019); http://doi.org/10.1109/ICCChina.2019.8855926.
[30] [30] Rodes R, Wieckowski M, Pham T T, Jensen J B, Turkiewicz J et al. Carrierless amplitude phase modulation of VCSEL with 4 bit/s/Hz spectral efficiency for use in WDM-PON. Opt Express19, 26551-26556 (2011).
[33] [33] Zhang J. Memory-polynomial digital pre-distortion for linearity improvement of directly-modulated multi-IF-over-Fiber LTE mobile fronthaul. In Optical Fiber Communications Conference (OFC), Tu2B.3 (OSA, 2016).
[35] [35] Ramachandran P, Zoph B, Le Q V. Searching for activation functions. arXiv: 1710.05941 (2017).
[36] [36] 2018 European Conference on Optical Communication (ECOC) 1-3 (IEEE, 2018); http://doi.org/10.1109/ECOC.2018.8535327.
[38] [38] Proceedings of the 20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining 661-670 (ACM Press, 2014); http://doi.org/10.1145/2623330.2623612.
[41] [41] Haigh P A, Chvojka P, Zvánovec S, Ghassemlooy Z, Darwazeh I. Analysis of Nyquist pulse shapes for carrierless amplitude and phase modulation in visible light communications. J Light Technol36, 5023-5029 (2018).
[45] [45] Fehenberger T, Hanik N. Mutual information as a figure of merit for optical fiber systems. arXiv: 1405.2029 (2014).
Get Citation
Copy Citation Text
Fangchen Hu, Jorge A. Holguin-Lerma, Yuan Mao, Peng Zou, Chao Shen, Tien Khee Ng, Boon S. Ooi, Nan Chi. Demonstration of a low-complexity memory-polynomial-aided neural network equalizer for CAP visible-light communication with superluminescent diode[J]. Opto-Electronic Advances, 2020, 3(8): 200009-1
Received: Apr. 6, 2020
Accepted: May. 17, 2020
Published Online: Jan. 7, 2021
The Author Email: Ooi Boon S. (nanchi@fudan.edu.cn)