Visible Light Communication Non-Orthogonal Multiple Access System Based on Neural Network Precoding

Abstract

The present invention belongs to the technical field of visible light communication, and specifically relates to a visible light communication non-orthogonal multiple access system based on neural network precoding. The system of the present invention mainly includes a feedback neural network (NN) at the transmitting end, a modulo encoder, and a super-Nyquist (FTN) multi-band carrierless amplitude-phase (CAP) modulator; a waveform-symbol demodulation equalization neural network, a modulo decoder and other modules at the receiving end. The present invention introduces crosstalk in the time domain through active FTN filtering to compress the signal bandwidth, and through the preprocessing at the transmitting end and the postprocessing at the receiving end, reduces the crosstalk problem caused by the compression of the time-domain symbol interval of the FTN multi-band CAP signal, and improves the spectral efficiency of the visible light communication system. Most of the complexity of this method is concentrated at the transmitting end, and the receiving end only requires an independent feedforward network, meeting the high confidentiality and low complexity requirements of the downlink multiple access user terminal.

Description

Technical Field

[0001] The present invention belongs to the technical field of visible light communication, and particularly relates to a non-orthogonal multiple access system based on neural network precoding applied to a visible light communication system. Background Art

[0002] With the large-scale deployment of 5G, the next-generation (6G) mobile communication has become a hot issue of common concern in the academic and industrial fields. At present, the radio frequency-based wireless communication network has faced congestion and spectrum exhaustion problems, and is insufficient to meet the requirements of high capacity and wide coverage. Visible Light Communication (VLC) uses the visible band of 380nm - 780nm and uses LEDs as light sources, which can provide integrated lighting and communication at the same time, and has the advantages of low cost, high capacity, high security, environmental friendliness, electromagnetic radiation resistance, etc. This new communication mode will promote the integrated development and technological progress of the next-generation lighting and access network, and has become the focus and high point of domestic and international competition. From the national strategic level and the potential market scale, the development of visible light communication is of great significance.

[0003] However, the implementation of a high-speed visible light communication system is limited by the optoelectronic characteristics of devices, and the signal modulation bandwidth limitation of commercial white LEDs is the most important factor restricting system performance. In order to make full use of limited bandwidth resources, high spectral efficiency transmission is required. Carrierless Amplitude and Phase (CAP) modulation is a variant of Quadrature Amplitude Modulation (QAM). Different from QAM, CAP realizes the real-numberization of complex signals through a pair of orthogonal shaping filters. As a high spectral efficiency amplitude modulation, it has been widely studied in visible light communication systems.

[0004] In a multi-band CAP system, different user multi-access can be achieved by adjusting the center frequency of each sub-band. To achieve high spectral efficiency, "non-orthogonality" can be introduced in the time domain or frequency domain to compress the bandwidth occupied by the overall signal and realize super-Nyquist modulation. On the one hand, time-domain compression can achieve FTN by introducing inter-symbol interference (ISI) through FTN active filtering; on the other hand, frequency-domain compression can achieve FTN by introducing ICI by reducing the center frequency interval. However, the method of achieving FTN by introducing ICI requires joint decoding of each channel at the receiving end and is not suitable for high-secrecy scenarios. In addition, the ICI introduced by changing the center frequency is time-varying and difficult to process. Therefore, we will focus on discussing the non-orthogonal multiple access visible light communication system based on compressed FTN multi-band CAP in the time domain. Compared with the orthogonal multiple access visible light communication system based on Nyquist multi-band CAP, the FTN multi-band CAP system no longer satisfies the zero ISI condition after matched filtering. In addition, due to the limited number of taps of the shaping filter, the truncation effect in the time domain will introduce frequency-domain bandwidth broadening, making it inevitable to have ICI between sub-bands. After transmission through the visible light communication channel, it will also be affected by high-frequency fading and nonlinear effects. These interferences will limit the performance of the visible light communication system.

[0005] The present invention proposes to use a pre-coding based on a feedback neural network at the transmitting end to pre-subtract the interference term at the transmitting end, and use a waveform-symbol direct mapping neural network at the receiving end to replace the traditional processes of equalization, matched filtering, and downsampling to mitigate the interference problem existing in the non-orthogonal multiple access visible light communication system of FTN multi-band CAP. By compressing the total signal bandwidth of multi-band CAP, high spectral efficiency transmission is achieved. And most of the complexity of this method is concentrated at the transmitting end, and the receiving end only needs an independent feed-forward network, meeting the high-secrecy and low-complexity requirements of the downlink multiple access user end. Summary of the Invention

[0006] The purpose of the present invention is to provide a visible light communication non-orthogonal multiple access system based on neural network pre-coding with high spectral efficiency; this system can further compress the signal bandwidth by reducing ISI and ICI interferences, and obtain the same or higher system transmission rate with less modulation bandwidth; at the same time, most of the complexity is concentrated at the transmitting end, and the receiving end only needs an independent feed-forward network, meeting the high-secrecy and low-complexity requirements of the downlink multiple access user end.

[0007] The non-orthogonal multiple access visible light communication system based on neural network pre-coding proposed by the present invention, as Figure 1 shown, includes two parts: a transmitting end and a receiving end; the transmitting end is composed of a feedback neural network, a modulo encoder, and a super-Nyquist multi-band CAP modulator connected in sequence; the receiving end is composed of a waveform-symbol demodulation and equalization neural network, a modulo decoder and other modules connected in sequence; where:

[0008] The feedback neural network module is used to perform precoding on the original symbol data to subtract the inter-symbol interference in advance;

[0009] The modulo encoder module is used to perform a modulo operation on the data preprocessed by the neural network to limit the distribution range of constellation points;

[0010] The super-Nyquist multi-band CAP modulator is used to modulate the precoded symbol data to generate a super-Nyquist CAP signal;

[0011] The waveform-symbol demodulation and equalization neural network module is used to demodulate and equalize the received waveform signal and directly map it to the symbol data with an extended constellation point distribution;

[0012] The modulo decoder module is used to perform a modulo process on the symbol data with an extended constellation point distribution to obtain the symbol data with a normal constellation point distribution for further bit error rate testing.

[0013] In the present invention, at the transmitting end, first, the feedback neural network module is used to perform precoding on the original symbol data to subtract the inter-symbol interference in advance; assuming that the number of subbands of the multi-band CAP signal is m, where the i-th subband signal is denoted as a i , and the precoded i-th subband signal is denoted as The input of the feedback neural network is a q×m vector expanded into a qm×1-dimensional vector, denoted as:

[0014]

[0015] where vec a,b (·) is a matrix reshaping operation, and the output is a qm×1-dimensional vector; q represents the memory length considered for inter-symbol interference.

[0016] In the present invention, the feedback neural network can use a complex neural network. Taking a neural network with one hidden layer and a complex ReLU (CReLU) as the non-linear activation function as an example, its output can be expressed as:

[0017] y(n) = w 2 ·CReLU(w 1 ·x(n) + b 1 ) + b 2 (2)

[0018] where w 1 , w 2 are weight matrices respectively, and b 1 , b 2They are bias term matrices respectively. The output is an m×1 vector. The structure of this neural network can also be extended to a linear structure neural network without hidden layers, a neural network with multiple hidden layers, and a neural network with other non-linear activation functions.

[0019] The output of the feedback neural network is added to the original symbol to obtain the preprocessed signal, which is used as the input of the modulo encoder module.

[0020] In the present invention, at the transmitting end, the modulo encoder module performs a modulo operation on the data preprocessed by the neural network to limit the distribution range of constellation points. Its output formula can be expressed as:

[0021]

[0022] Among them, Λ represents the modulo limit range. In the present invention, taking 16-QAM as an example, the modulo limit range is a square region where the real part and the imaginary part are limited between [-4, 4). The feedback neural network module and the modulo encoder module jointly complete the precoding step at the transmitting end. This structure of feedback structure plus modulo is based on the Tomlinson-Harashima precoding (THP) structure and jointly encodes multi-channel signals, so it is called a multi-input multi-output THP neural network (MIMO-THPNN).

[0023] In the present invention, at the transmitting end, the super-Nyquist multi-band CAP modulator module is used to modulate the precoded symbol data to generate a super-Nyquist CAP signal. First, for the input signal of this module, an upsampling operation is performed to obtain a symbol sequence b after R-fold upsampling i . Then, for each sub-band, a pair of orthogonal FTN shaping filters are used to convolve with the real part and the imaginary part of b i respectively, and the result after subtraction is the CAP signal of this sub-band. For different sub-bands, the center frequency interval of the shaping filter is the bandwidth of a single sub-band. Adding the CAP signals of each sub-band together, the FTN multi-band CAP signal s(n) is obtained, as shown in the following formula:

[0024]

[0025] Among them, τ is the compression coefficient, satisfying 0 < τ ≤ 1. The shaping filter is obtained by multiplying the root-raised cosine (SRRC) pulse by and respectively, where T s is the sampling symbol interval, is the center frequency of the i-th sub-band. Specifically, it can be expressed as:

[0026]

[0027]

[0028] where α is the roll-off factor.

[0029] For Nyquist multi-band CAP signals, the bandwidth of a single sub-band signal after CAP modulation is B(1 + α), where B = 1 / RT s , and the total bandwidth of the multi-band CAP signal is mB(1 + α);

[0030] For FTN multi-band CAP signals, considering the same symbol sampling rate, the bandwidth of a single sub-band signal after CAP modulation is τB(1 + α), and the total bandwidth of the FTN multi-band CAP signal is mτB(1 + α). It can be seen that the bandwidth of the FTN multi-band CAP signal is compressed to τ times the original.

[0031] In the implementation simulation example, the modulation order designed by this system is: 16QAM. In the implementation simulation example, m = 3, R = 12, that is, three sub-band CAP signals are used for illustration, and the upsampling factor is 12. Without loss of generality, it can be extended to multi-sub-band CAP signals.

[0032] After that, the generated FTN multi-band CAP signal is subjected to digital-to-analog conversion (DAC), transmitted through the visible light communication channel, and after being sampled by analog-to-digital conversion (ADC) at the receiving end, it is sent to subsequent digital signal processing.

[0033] In the present invention, at the receiving end, first, the waveform-symbol demodulation and equalization neural network module is used to demodulate and equalize the received waveform signal, and directly map it to the symbol data with extended constellation point distribution. For Nyquist multi-band CAP signals, after the matched filtering at the receiving end, the zero inter-symbol interference (ISI) condition is satisfied. However, for FTN multi-band CAP signals, there is still ISI after matched filtering. And after the transmission of the FTN multi-band CAP signal through the visible light channel, there is not only the ISI introduced by FTN filtering, but also the influence of channel high-frequency power fading and nonlinear effects. Therefore, the present invention proposes to use a waveform-level to symbol-level feed-forward neural network (W2S-FFNN) to replace the equalization, matched filtering, and downsampling steps of traditional multi-band CAP, which can reduce the residual crosstalk and the computational complexity of digital signal processing at the receiving end. The input of the W2S-FFNN is the received waveform signal:

[0034] x W2S (n) = [r(R(n - k)), r(R(n - k)+1),..., r(Rn),..., r(R(n + k))] T , (7).

[0035] In the present invention, the W2S-FFNN can use a plurality of neural networks. Taking a neural network with one hidden layer and a complex rectified linear unit (CReLU) as the non-linear activation function as an example, its output is the symbol sequence of the estimated sub-band, which can be expressed as:

[0036] y W2S (n) = w 2 ·CReLU(w 1 ·x W2S (n) + b 1 ) + b 2 (8)

[0037] where w 1 and w 2 are weight matrices respectively, and b 1 and b 2 are bias term matrices respectively; the output is an m×1 vector. The structure of this neural network can also be extended to a linear structure neural network without a hidden layer, a neural network with multiple hidden layers, and a neural network with other non-linear activation functions (n).

[0038] In the present invention, at the transmitter, the modulo decoder module performs modulo processing on the symbol data with an extended constellation point distribution to obtain symbol data with a normal constellation point distribution for further bit error rate testing. After the symbol sequence after THP precoding passes through the channel and is equalized by the W2S-FFNN, a constellation point distribution extended in the I-Q plane is obtained. Therefore, it is necessary to use the same modulo operator as that at the transmitter to process the symbol data with an extended constellation point distribution, so as to obtain symbol data with a normal constellation point distribution, and then use it for further bit error rate testing.

[0039] Compared with the current Nyquist multi-band CAP visible light communication system, the high spectral efficiency visible light communication non-orthogonal multiple access system based on neural network precoding proposed by the present invention compresses the signal bandwidth by introducing ISI from the time domain through FTN filtering; reduces crosstalk through neural network-based precoding at the transmitter and waveform-symbol demodulation and equalization neural network at the receiver, improves the signal bandwidth compression ratio, and obtains the same or higher system transmission rate with less modulation bandwidth; at the same time, most of the complexity is concentrated at the transmitter, and the receiver only needs an independent feedforward network, meeting the high confidentiality and low complexity requirements of the downlink multiple access user side.

[0040] It can be seen from the visible light communication non-orthogonal multiple access system based on neural network precoding provided by the present invention described above that the present invention combines the machine learning method with the THP structure, and proposes neural network-based precoding at the transmitter and waveform-symbol demodulation and equalization neural network at the receiver to reduce crosstalk, having the following advantages:

[0041] (1) For the first time, the neural network is combined with the THP precoding structure of MIMO and applied to FTN multi-band CAP. By significantly improving the tolerance of FTN signals to crosstalk, the same or higher system transmission rate can be obtained with less modulation bandwidth.

[0042] (2) By using a neural network with direct waveform-symbol mapping at the receiving end to replace the traditional post-equalization, matched filtering, and downsampling processes, the tolerance of FTN signals to crosstalk is significantly improved, further enhancing the system capacity.

[0043] (3) For the first time, the precoding based on neural network at the transmitting end and the demodulation and equalization neural network of waveform-symbol at the receiving end to reduce crosstalk are applied to the visible light communication system, and simulations are carried out, theoretically proving a visible light communication system with higher spectral efficiency.

[0044] (4) The proposed visible light communication non-orthogonal multiple access system concentrates most of the complexity at the transmitting end, and only an independent feedforward network is required at the receiving end, meeting the high confidentiality and low complexity requirements of the downlink multiple access user side.

[0045] The present invention is applicable to the field of short-distance, high-speed, and high spectral efficiency visible light communication. Combining machine learning algorithms, it provides a new non-orthogonal multiple access solution, concentrating most of the complexity at the transmitting end, meeting the high confidentiality and low complexity requirements of the downlink multiple access user side. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a block diagram of the visible light communication non-orthogonal multiple access system based on neural network precoding of the present invention.

[0047] Figure 2 It is a spectrogram obtained by FFT of the multi-band CAP of the present invention. Among them, (a) is the spectrum obtained by FFT of the Nyquist multi-band CAP signal, and (b) is the spectrum obtained by FFT of the FTN multi-band CAP signal (τ = 0.8).

[0048] Figure 3 It is a study on the bit error performance of the crosstalk cancellation method based on neural network precoding of the present invention in the FTN multi-band CAP VLC system with the change of the compression factor.

[0049] Figure 4 It is a study on the effective mutual information performance of the crosstalk cancellation method based on neural network precoding of the present invention in the FTN multi-band CAP VLC system with the change of the compression factor.

[0050] Figure 5 It is a study on the effective mutual information performance of the crosstalk cancellation method based on neural network precoding of the present invention in the non-linear FTN multi-band CAP VLC system with the change of the signal amplitude.

[0051] Reference numerals in the figure: 101 is the feedback neural network module at the transmitting end, 102 is the modulo encoder module, 103 is the super-Nyquist multi-band CAP modulator module, 104 is the waveform-symbol demodulation and equalization neural network module at the receiving end, and 105 is the modulo decoder module. Detailed implementation manners

[0052] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will elaborate on the various implementation manners of the present invention with reference to the accompanying drawings. However, those of ordinary skill in the art can understand that in the various implementation manners of the present invention, many technical details are provided to help readers better understand the present application. However, even without these technical details and various changes and modifications based on the following implementation manners, the technical solutions claimed in the various claims of the present application can still be implemented.

[0053] The visible light communication non-orthogonal multiple access system based on neural network precoding provided by the present invention can further compress the signal bandwidth by reducing crosstalk, and obtain the same or higher system transmission rate with less modulation bandwidth; at the same time, most of the complexity is concentrated at the transmitting end, and the receiving end only needs an independent feedforward network, meeting the high confidentiality and low complexity requirements of the downlink multiple access user side.

[0054] The system proposed by the present invention is as Figure 1 shown, and is composed of a feedback neural network, a modulo encoder, and a super-Nyquist multi-band CAP modulator at the transmitting end; a visible light channel; a waveform-symbol demodulation and equalization neural network and a modulo decoder at the receiving end, and other modules.

[0055] The crosstalk reduction method based on neural network precoding and the waveform-symbol borrowing equalization neural network at the receiving end of the above system specifically includes the following steps:

[0056] Step 101: Use the feedback neural network module at the transmitting end to precode the original symbol data and subtract the inter-symbol crosstalk in advance. Assume that the number of subbands of the multi-band CAP signal is m. Among them, let the i-th subband signal be expressed as a i , and the i-th subband signal after precoding be expressed as The input of the feedback neural network is a q×m vector expanded into a qm×1-dimensional vector, which can be expressed as:

[0057]

[0058] where, vec a,b(·) is a matrix reshaping operation, and the output is a matrix with dimensions a×b. In the present invention, a complex neural network is used. Taking a neural network with one hidden layer and a complex ReLU (CReLU) as the non-linear activation function as an example, its output can be expressed as:

[0059] y(n) = w 2 ·CReLU(w 1 ·x(n) + b 1 ) + b 2

[0060] Where, w 1 , w 2 are weight matrices respectively, and b 1 , b 2 are bias term matrices respectively. The output is an m×1 vector. The structure of this neural network can also be extended to a linear structure neural network without a hidden layer, a neural network with multiple hidden layers, and a neural network with other non-linear activation functions. The output of the feedback neural network is added to the original symbol to obtain a preprocessed signal, which is used as the input of the modulo encoder module.

[0061] Step 102: Pass through the modulo encoder module, which is used to perform a modulo operation on the data preprocessed by the neural network to limit the constellation point distribution range. Its output formula can be expressed as:

[0062]

[0063] Where, Λ represents the modulo limit range. In the present invention, taking 16-QAM as an example, the modulo limit range is a square region where the real part and the imaginary part are limited between [-4, 4). The feedback neural network module and the modulo encoder module jointly complete the precoding step at the transmitting end. This structure of feedback structure plus modulo is based on the Tomlinson-Harashima precoding (THP) structure and jointly encodes multi-channel signals, so it is called a multi-input multi-output THP neural network (MIMO-THPNN).

[0064] Step 103: Pass through the super-Nyquist multi-band CAP modulator, which is used to modulate the precoded symbol data to generate a super-Nyquist CAP signal. First, for the input signal of this module, an upsampling operation is performed to obtain a symbol sequence b i after R-fold upsampling. Then, for each sub-band, a pair of orthogonal FTN shaping filters are used to convolve with the real part and the imaginary part of b i respectively, and the results are subtracted to obtain the CAP signal of this sub-band. For different sub-bands, the center frequency interval of the shaping filter is the bandwidth of a single sub-band. Adding the CAP signals of each sub-band together, the FTN multi-band CAP signal s(n) is obtained, as shown in the following formula:

[0065]

[0066] Among them, τ is the compression coefficient, satisfying 0 < τ ≤ 1. The shaping filter is obtained by multiplying the root-raised cosine (SRRC) pulse by and respectively, where T s is the sampling symbol interval, is the center frequency of the i-th subband. Specifically, it can be expressed as:

[0067]

[0068]

[0069] Among them, α is the roll-off factor. For Nyquist multi-band CAP signals, the bandwidth of a single subband signal after CAP modulation is B(1 + α), where B = 1 / RT s , and the total bandwidth of the multi-band CAP signal is mB(1 + α); for FTN multi-band CAP signals, considering the same symbol sampling rate, the bandwidth of a single subband signal after CAP modulation is τB(1 + α), and the total bandwidth of the FTN multi-band CAP signal is mτB(1 + α). It can be seen that the bandwidth of the FTN multi-band CAP signal is compressed to τ times the original.

[0070] Step 104: Use the waveform-symbol demodulation and equalization neural network module at the receiving end to demodulate and equalize the received waveform signal and directly map it to the symbol data of the constellation point extended distribution. The present invention proposes to use a waveform-level to symbol-level feedforward neural network (W2S-FFNN) to replace the equalization, matched filtering, and downsampling steps of traditional multi-band CAP, which can reduce residual crosstalk and the computational complexity of digital signal processing at the receiving end. The input of the W2S-FFNN is the received waveform signal:

[0071] x W2S (n) = [r(R(n - k)), r(R(n - k)+1),..., r(Rn),..., r(R(n + k))] T .

[0072] In the present invention, taking a complex neural network with one hidden layer and a complex ReLU (CReLU) as the non-linear activation function as an example, its output is the estimated symbol sequence of the subband, which can be expressed as:

[0073] y W2S (n) = w 2 ·CReLU(w 1 ·x W2S (n)+b 1 )+b 2

[0074] Among them, w 1 and w 2 are weight matrices respectively, and b 1 and b 2 are bias term matrices respectively. The output is an m×1 vector. The structure of this neural network can also be extended to a linear structure neural network without hidden layers, a neural network with multiple hidden layers, and a neural network with other non-linear activation functions.

[0075] Step 105: A modulo decoder module is used to perform modulo processing on the symbol data with an extended constellation point distribution to obtain symbol data with a normal constellation point distribution for further bit error rate testing. After the symbol sequence after THP precoding passes through the channel and W2S-FFNN equalization, a constellation point distribution extended in the I-Q plane will be obtained. Therefore, it is necessary to use the same modulo operator as at the transmitter end to process the symbol data with an extended constellation point distribution, so as to obtain symbol data with a normal constellation point distribution.

[0076] So far, the visible light communication non-orthogonal multiple access system based on neural network precoding ends.

[0077] Next, the structure and simulation verification results of the visible light communication non-orthogonal multiple access system based on neural network precoding proposed by the present invention will be introduced. The specific steps are as follows:

[0078] Figure 1 is the block diagram of the visible light communication non-orthogonal multiple access system based on neural network precoding of the present invention. It includes a feedback neural network module 101 at the transmitter end, a modulo encoder module 102, an ultra-Nyquist multi-band CAP modulator module 103, a waveform-symbol demodulation and equalization neural network module 104 at the receiver end, and a modulo decoder module 105. In this example, an exponential function is used to simulate the high-frequency fading effect H of the visible light channel; secondly, additive white Gaussian noise (AWGN) with a signal-to-noise ratio of SNR is introduced. In a non-linear visible light channel, the non-linearity is simulated through the following function:

[0079]

[0080] Figure 2 is the spectrogram obtained by performing FFT on the multi-band CAP of the present invention. Among them, (a) is the spectrum obtained by performing FFT on the Nyquist multi-band CAP signal, and (b) is the spectrum obtained by performing FFT on the FTN multi-band CAP signal (τ = 0.8). It can be intuitively seen that the visible light communication non-orthogonal multiple access system with high spectral efficiency based on neural network precoding proposed by the present invention realizes bandwidth reduction under the condition that the order of the transmitted signal remains unchanged, so as to achieve high spectral efficiency.

[0081] Figure 3Research curve of the bit error performance of the neural network precoding crosstalk cancellation method of the present invention varying with the compression factor in the FTN multi-band CAP VLC system. The simulation channel is a linear visible light communication channel. Here, the bit error performance is the average bit error rate (BER) of three sub-bands. Number 31 represents the curve of the bit error performance of the proposed neural network precoding (MIMO-THPNN) combined with the waveform-symbol feedforward neural network (W2S-FFNN) algorithm at the receiver varying with the compression factor; number 32 represents the curve of the bit error performance of only using the waveform-symbol decision feedback neural network (W2S-DFNN) algorithm at the receiver varying with the compression factor; number 33 represents the curve of the bit error performance of the symbol-symbol feedforward neural network (S2S-FFNN) algorithm at the receiver after using traditional matched filtering and downsampling varying with the compression factor. It can be seen that as the compression factor decreases, the compression degree of the modulation bandwidth increases, and the BER of all three algorithms increases accordingly, which is caused by the increased crosstalk introduced by FTN. Among them, the performance of the W2S-DFNN algorithm is better than that of the traditional S2S-FFNN algorithm when the compression factor is greater than 0.55. However, due to the problems of decision error transmission and accumulation in the decision feedback structure, when the compression factor is less than 0.55, the BER is higher than that of the traditional algorithm. The algorithm used in the present invention can avoid error transmission through precoding. Therefore, the bit error performance of the algorithm used in the present invention is higher than that of the above two algorithms.

[0082] Figure 4 Research curve of the effective mutual information performance of the neural network precoding crosstalk cancellation method of the present invention varying with the compression ratio in the FTN multi-band CAP VLC system. The simulation channel is a linear visible light communication channel. The effective mutual information (EMI) here is calculated by the following formula:

[0083]

[0084] Where represents the mutual information between the estimated symbol sequence and the transmitted symbol sequence of the i-th subband, τ is the compression coefficient, and m is the number of subbands. No. 41 represents the curve of the achievable spectral efficiency performance of the proposed MIMO-THPNN+W2S-FFNN algorithm as the compression coefficient changes; No. 42 represents the curve of the achievable spectral efficiency performance of the W2S-DFNN algorithm as the compression coefficient changes; No. 43 represents the curve of the error performance of the S2S-FFNN algorithm as the compression coefficient changes after the receiving end uses traditional matched filtering and downsampling. It can be seen that as the compression coefficient decreases, the compression degree of the modulation bandwidth increases, and the EMI of the three algorithms all show a trend of linear rise and then decline. In other words, there is an inflection point with the highest EMI in the curve. This is because there is a certain trade-off between spectral efficiency and crosstalk. The stronger the anti-crosstalk ability of the algorithm, the greater the compression degree that can be supported, the smaller the compression rate corresponding to the inflection point, and the greater the EMI. Among them, the performance of the W2S-DFNN algorithm is better than the traditional S2S-FFNN algorithm when the compression coefficient is greater than 0.55. When the compression coefficient is less than 0.55, the EMI is lower than the traditional algorithm. The EMI performance of the algorithm used in the present invention is higher than the above two algorithms. Figure 3 This is consistent with the conclusions in .

[0085] Figure 5 This is a research curve diagram of the effective mutual information performance of the neural network precoding crosstalk elimination method in the nonlinear FTN multi-band CAP VLC system as the signal amplitude changes. The simulation channel No. 51 is an AWGN channel, and the simulation channels No. 52 and 53 are nonlinear visible light communication channels. It can be seen that under the AWGN channel, as the signal amplitude (Vpp) increases, the signal-to-noise ratio improves, and the EMI also increases. For the nonlinear VLC channel, as Vpp increases, the nonlinear effect on the signal also increases. Therefore, there is a trade-off between the signal-to-noise ratio and nonlinearity. No. 52 is the effective mutual information curve using the linear MIMO-THPNN+W2S-FFNN algorithm under the nonlinear VLC channel as the Vpp changes; No. 53 is the effective mutual information curve using the nonlinear MIMO-THPNN+W2S-FFNN algorithm under the nonlinear VLC channel as the Vpp changes. Nonlinearity is introduced here by adding hidden layers and nonlinear activation functions. In this example, the number of hidden layers is 1, the number of hidden layer nodes is 31, and the nonlinear activation function is a complex ReLU. The results show that under nonlinear VLC channels, the performance of nonlinear methods is better than that of linear methods, and as Vpp increases, the improvement of nonlinear methods over linear methods becomes more obvious.

[0086] In summary, through simulation comparisons from multiple dimensions, the simulation results of this embodiment show that, compared with the existing crosstalk cancellation methods, the proposed crosstalk cancellation method based on neural network precoding can achieve higher spectral efficiency in the FTN multi-band CAP VLC system, and can achieve the same or higher transmission rate with less bandwidth. Most of the complexity of the proposed visible light communication non-orthogonal multiple access system based on neural network precoding is concentrated at the transmitter, and the receiver only requires an independent feedforward network, which meets the requirements of high confidentiality and low complexity for the downlink multiple access user terminal.

[0087] In this embodiment, the division of each step is only for clear description. When implemented, they can be combined into one step or some steps can be split into multiple steps. As long as the same logical relationship is included, they are all within the protection scope of this patent.

[0088] Those of ordinary skill in the art can understand that the above embodiments are specific embodiments for implementing the present invention. In actual applications, various changes can be made in form and details without departing from the spirit and scope of the present invention.

Claims

1. A visible light communication non-orthogonal multiple access system based on neural network precoding, characterized in that, it includes two parts: a transmitter and a receiver; the transmitter is composed of a feedback neural network, a modulo encoder, and an ultra-Nyquist multi-band CAP modulator connected in sequence; the receiver is composed of a waveform-symbol demodulation equalization neural network and a modulo decoder connected in sequence; where: The feedback neural network module is used to precode the original symbol data and subtract the inter-symbol interference in advance; The modulo encoder module is used to perform modulo operation on the data preprocessed by the neural network to limit the distribution range of constellation points; The ultra-Nyquist multi-band CAP modulator is used to modulate the precoded symbol data to generate an ultra-Nyquist CAP signal; The waveform-symbol demodulation equalization neural network module is used to demodulate and equalize the received waveform signal and directly map it to the symbol data with extended constellation point distribution; The modulo decoder module is used to perform modulo operation on the symbol data with extended constellation point distribution to obtain the symbol data with normal constellation point distribution for further bit error rate testing; At the transmitter, first, through the feedback neural network module, the original symbol data is precoded to subtract the inter-symbol interference in advance; specifically: Assume that the number of sub - bands of the multi - band CAP signal is m. Among them, let the i - th sub - band signal be denoted as a i , and the i - th sub - band signal after precoding is denoted as The input of the feedback neural network is a q×m vector expanded into a qm×1 - dimensional vector, which is expressed as: Among them, vec a,b (·) is a matrix reshaping operation, and the output is a qm×1 dimensional vector; q represents the memory length considered for inter-symbol interference; The feedback neural network uses a complex neural network with one hidden layer, and the non-linear activation function is complex ReLU (CReLU), and its output is expressed as: y(n) = w 2 ·C ReLU(w 1 ·x(n) + b 1 ) + b 2 ,(2) where, w 1 , w 2 are weight matrices respectively, and b 1 , b 2 are bias matrices respectively; the output is a vector of m×1; The output of the feedback neural network is added to the original symbol to obtain the preprocessed signal, which is used as the input of the modulo encoder module; At the transmitter, the modulo encoder module performs modulo operation on the data preprocessed by the neural network to limit the distribution range of constellation points; its output formula is expressed as: where, Λ represents the limit range of modulo; the feedback neural network module and the modulo encoder module jointly complete the precoding at the transmitter. This structure of feedback plus modulo uses the Tomlinson-Harashima precoding (THP) structure and jointly encodes multi-channel signals, which is called multi-input multi-output THP neural network (MIMO-THPNN); At the receiver, the waveform-to-symbol feedforward neural network (W2S-FFNN) of the waveform-symbol demodulation equalization neural network module is used to replace the equalization, matched filtering, and downsampling steps of the traditional multi-band CAP; the input of the W2S-FFNN is the received waveform signal: x W2S x(n) = [r(R(n - k)), r(R(n - k)+1),..., r(Rn),..., r(R(n + k))] T , (7) The W2S-FFNN uses a complex neural network with one hidden layer, and the non-linear activation function is complex ReLU (CReLU), and its output is the estimated symbol sequence of the sub-band, expressed as: y W2S (n) = w 2 ·CReLU(w 1 ·x W2S (n) + b 1 ) + b 2 , (8) where, w 1 , w 2 are weight matrices respectively, and b 1 , b 2 are bias matrices respectively; the output is a vector of m×1.

2. The visible light communication non-orthogonal multiple access system according to claim 1, characterized in that, At the transmitter, the ultra-Nyquist multi-band CAP modulator module is used to modulate the precoded symbol data to generate an ultra-Nyquist CAP signal; the specific process is as follows: First, perform an upsampling operation on the input signal of this module to obtain the symbol sequence b after upsampling by a factor of R i ; Then, for each sub-band, a pair of orthogonal FTN shaping filters are used to convolve with the real part and the imaginary part of b i respectively, and the sub-band CAP signal is obtained by subtracting the results; for different sub-bands, the center frequency interval of the shaping filter is the bandwidth of a single sub-band; adding each sub-band CAP signal together, the FTN multi-band CAP signal s(n) is obtained, as shown in the following formula: where τ is the compression coefficient, satisfying 0 < τ ≤ 1; the shaping filter is obtained by multiplying the root-raised cosine (SRRC) pulse by and respectively, where T s is the sampling symbol interval, is the center frequency of the i-th sub-band; specifically expressed as: where, α is the roll-off factor; For a Nyquist multi-band CAP signal, the bandwidth of a single sub-band signal after CAP modulation is B(1 + α), where B = 1 / RT s , and the total bandwidth of the multi-band CAP signal is mB(1 + α); For the FTN multi-band CAP signal, considering the case where the symbol sampling rates are the same, the bandwidth of a single sub-band signal after CAP modulation is τB(1 + α), and the total bandwidth of the FTN multi-band CAP signal is mτB(1 + α); thus, it can be seen that the bandwidth of the FTN multi-band CAP signal is compressed to τ times the original bandwidth; After that, the generated FTN multi-band CAP signal undergoes digital-to-analog conversion (DAC), is transmitted through the visible light communication channel, and after being sampled by analog-to-digital conversion (ADC) at the receiving end, is sent to subsequent digital signal processing.

3. The visible light communication non-orthogonal multiple access system according to claim 1, characterized in that, at the receiving end, the modulo decoder module processes the symbol data with an extended distribution of constellation points using the same modulo operator as that at the transmitting end.

Images