Compared with multicolor-chip integrated white LEDs, phosphor-based white LEDs are more attractive for daily illumination due to lower cost and complexity, and thus they are preferable for future commercial use of visible light communication (VLC) systems. However, the application of phosphorescent white LEDs has a lower data rate than multicolor-chip integrated LEDs because of severe nonlinear impairments and limited bandwidth caused by the slowresponding phosphor. In this paper, for the fifirst time we propose to employ phosphorescent white LEDs based on silicon substrate with adaptive bit-loading discrete multitone (DMT) modulation and a memoryless polynomial based nonlinear equalizer to achieve a high-speed VLC system. We also present a comprehensive comparison among nonlinear equalizers based on the Volterra series model, memory polynomial model, memoryless polynomial model and deep neural network (DNN) with experimental results utilizing a silicon substrate phosphorescent white LED, and provide detailed suggestions on how to choose the most suitable nonlinear mitigation scheme considering difffferent practical conditions and the tradeoffff between complexity and performance. Beyond 3.00 Gb/s DMT VLC transmission over 1-m indoor free space is successfully demonstrated with bit error rate (BER) under the 7% forward error correction (FEC) limit of 3.8×10−3 . As far as we know, this is the highest data rate ever reported for VLC systems based on a single high-power phosphorescent white LED.
In recent years, white LED based VLC has become an attractive method for next generation wireless communication. Compared with traditional wireless communication, VLC has several advantages such as high security, low cost, immunity to electromagnetic interference and license-free. There are two kinds of commonly used white LEDs, phosphor-based LEDs and multicolor-chip integrated LEDs, such as red-green-blue (RGB) and red-green-blue-yellow (RGBY) LEDs. Phosphorescent white LEDs are more promising for general illumination due to lower cost and complexity. But the bandwidth of a phosphorescent white LED is only few tens of MHz even with a blue fifilter because of the slow response of the phosphors, and therefore the achieved data rates applying this kind of LED are lower than those applying multicolor-chip integrated LEDs. In a hardware pre-equalizer is proposed to extend the modulation bandwidth of phosphorescent white LEDs. Furthermore, many advanced modulation formats, such as multi-level carrier-less amplitude and phase modulation (CAP), pulse amplitude modulation (PAM), and orthogonal frequency division multiplexing (OFDM), have been applied to improve spectrum effiffifficiency and achieve high-speed VLC systems based on phosphorescent white LEDs. Apart from linear distortions caused by limited modulation bandwidth, phosphorescent white LEDs suffffer from intrinsic nonlinearity due to the nonlinear voltage-current and current-light intensity relationships, which is the major source of the nonlinearity in VLC systems. And the nonlinear effffect becomes particularly detrimental when high-order spectrally effiffifficient modulation formats are applied to achieve high-speed transmission.
In this paper, for the fifirst time we propose to employ a vertical structure phosphorescent white LEDs based on silicon substrate with adaptive bit-loading DMT modulation and a memoryless polynomial based nonlinear equalizer to achieve a high-speed VLC system. This kind of LED has larger bandwidth. It employs hemispherical and pyramidal pattern surface textured GaN to enhance light extraction effiffifficiency, Ag reflflector to improve single side luminescence, and complementary electrode to reduce light absorption. We also present a comprehensive comparison among nonlinear equalizers based on the Volterra series model, memory polynomial model, memoryless polynomial model and DNN with experimental results, and provide advices on how to choose the most suitable scheme considering difffferent conditions and the tradeoffff between complexity and performance. Data rates of beyond 3.00 Gb/s over 1-m indoor free space VLC transmission are successfully demonstrated with BER under the 7% FEC limit of 3.8×10−3 . As far as we know, this is the highest data rate ever reported for VLC systems based on a single high-power phosphorescent white LED.
Fig1
The DNN is tested with difffferent node number of each layer to achieve a compromise between complexity and performance, and the results will be addressed in the following part 3.2. Thus, the node numbers of the layers are 29, 32, 16 and 1 respectively. The received signal is divided into three sets including the training set, the validation set and the test set with the ratios of 30%, 20% and 50%, respectively. The batch size is 256 and the epoch is 10. The details of the training parameters are listed in Table 1.
The spatial complexity of DNN, second-order Volterra, memory polynomial and memoryless polynomial is compared in Table 2. The detailed corresponding spatial complexity based on parameters set for difffferent equalizers which will be addressed in the part 3.2 is 1456, 50, 22, and 2 using DNN, second-order Volterra, second-order memory polynomial and second-order memoryless polynomial with N = 14 and M = 8. The DNN based equalizer is an effiffifficient scheme to mitigate nonlinearity in high-speed VLC systems, but it is always questioned by its high complexity and diffiffifficulty to be used in practical environments. There is still a long way to go before the actual commercial use of DNN based equalizers. And thus, the other conventional nonlinear equalizers based on Volterra, memory polynomial and memoryless polynomial with lower complexity may be a good substitute for DNN based equalizers at now.
To generate the adaptive bit-loading DMT signal, a binary phase shift keying (BPSK) DMT signal is fifirstly transmitted to test the signal-to-noise ratio (SNR) distribution of each subcarrier. The bit number for each subcarrier is then estimated based on the SNR results. Then, M-QAM symbol mapping for each subcarrier is carried out according to the bit allocation scheme. 128 subcarriers are employed to load the M-QAM symbols, and zeros are inserted into the signal. The signal combined with its conjugation is then upsampled by a factor of 4. After 256-point inverse fast Fourier transform (IFFT), a real-valued signal is generated. An 8-sample cyclic prefifix (CP) is used to mitigate intersymbol interference (ISI). During the offlfflffline processing, the nonlinear equalizer is applied at fifirst to mitigate nonlinearity. Channel estimation and zero-forcing post-equalization is also utilized in the following procedures. At last, the original bit sequence is recovered after M-QAM demapping.
