Deep learning detection method for SCMA-VLC system based on codebook prior knowledge

Abstract

The present invention discloses a deep learning detection method for an SCMA-VLC system based on codebook prior knowledge, which relates to the field of visible light communication technology, including: the user bit information at the transmitting end is SCMA-encoded to obtain a superimposed codeword, the SCMA superimposed codeword is converted into a non-negative real number signal under a DC bias optical orthogonal frequency division multiplexing modulation scheme, and an LED is driven to send an optical signal; the optical signal received at the receiving end is collected by a photodiode and converted into an electrical signal, and then DCO-OFDM demodulation is performed, and finally a deep learning detection network based on codebook prior knowledge is used to process the received signal to realize user bit information recovery. The present invention adopts an integrated noise reduction and detection network to enhance the ability to extract network signal features, thereby reducing the impact of LED nonlinear distortion and system noise on multi-user detection performance, so as to solve the problem that the traditional MPA detection algorithm has limited ability to resist signal nonlinear distortion, resulting in reduced multi-user detection performance of the system under large nonlinear distortion.

Description

Technical Field

[0001] The present invention relates to the technical field of visible light communication, and in particular to a deep learning detection method for an SCMA-VLC system based on codebook priori knowledge. Background Art

[0002] With the rapid growth of wireless data traffic, the shortage of radio frequency spectrum resources is facing extremely high communication pressure. Visible Light Communication (VLC) has attracted extensive research attention due to its advantages such as abundant unlicensed bandwidth, no electromagnetic interference, low energy consumption and high confidentiality, and has gradually become an effective supplementary solution for wireless radio frequency communication. The VLC system uses light emitting diodes (LEDs) as transmitters, and its narrow modulation bandwidth greatly limits the transmission capacity of the system. In response to this, non-orthogonal multiple access (NOMA) technology has been introduced into the VLC system, which improves the transmission capacity and spectrum utilization of the system by allowing multiple users to share the same resource blocks.

[0003] Sparse Code Multiple Access (SCMA) technology is a code domain NOMA technology that combines modulation and spread spectrum. It provides coding gain and shaping gain for the system through a multi-dimensional codebook, and improves the system's bit error rate performance under the same user and resource block conditions. Applying SCMA technology to VLC systems can effectively utilize the limited modulation bandwidth of LEDs to achieve high spectrum efficiency and large-capacity transmission.

[0004] One of the main challenges in the SCMA-VLC system is the design and optimization of the multi-user detection algorithm. The traditional message passing algorithm (MPA) is highly complex and difficult to implement in hardware. At the same time, the nonlinear characteristics of LED electro-optical conversion introduce nonlinear signal distortion in the system, and the MPA detection algorithm has limited ability to resist nonlinear signal distortion, resulting in reduced multi-user detection performance of the receiver. Summary of the invention

[0005] The purpose of the present invention is to provide a deep learning detection method for an SCMA-VLC system based on codebook prior knowledge, propose an integrated noise reduction and detection network, enhance the ability to extract network signal features, thereby reducing the impact of LED nonlinear distortion and system noise on multi-user detection performance, so as to solve the problem that the traditional MPA detection algorithm has limited ability to resist signal nonlinear distortion, resulting in reduced multi-user detection performance of the system under large nonlinear distortion.

[0006] To achieve the above object, the present invention provides the following technical solution: a deep learning detection method for SCMA-VLC system based on codebook prior knowledge, the deep learning detection method comprising:

[0007] The user bit information at the transmitting end is encoded by SCMA to obtain a superimposed codeword, which is converted into a non-negative real signal under the DC bias optical orthogonal frequency division multiplexing (DCO-OFDM) modulation scheme and drives the LED to send an optical signal;

[0008] The optical signal received at the receiving end is collected by the photodiode and converted into an electrical signal, which is then demodulated by DCO-OFDM. Finally, a deep learning detection network based on codebook prior knowledge is used to process the received signal to recover the user bit information.

[0009] In the above multi-user detection method, the signal nonlinear distortion caused by the nonlinear characteristics of the LED is regarded as a special noise processing in the received signal, and a deep neural network is used to perform integrated noise reduction and multi-user detection on the received signal to reduce the interference of noise on signal detection and enhance the ability to extract network signal features, so as to achieve better multi-user detection performance under large nonlinear distortion.

[0010] Preferably, the deep learning detection network based on codebook prior knowledge is used to process the received signal, specifically:

[0011] A deep neural network is used to perform integrated processing of noise reduction and multi-user detection on the received signal to reduce the interference of noise on signal detection and enhance the ability to extract network signal features, resulting in better multi-user detection performance under large nonlinear distortion.

[0012] Preferably, the deep learning detection network based on codebook prior knowledge includes:

[0013] Multi-user detection is converted into a multi-output classification problem. The log2M-bit user input bit information bj constitutes M categories, and each output is used to judge the corresponding input bit information of a user. The proposed network is based on DNN structure and consists of a denoising subnetwork and J user subnetworks.

[0014] Preferably, the denoising subnetwork is used to reconstruct the original noise-free SCMA superimposed codewords in the received signal to reduce the interference of noise on subsequent multi-user detection; wherein, the denoising subnetwork includes 5 fully connected hidden layers, and a batch normalization operation and a ReLU activation function are added after each hidden fully connected layer.

[0015] Preferably, the user sub-network is used for detecting bit information sent by each user, and comprises 2 fully connected hidden layers and 1 fully connected output layer.

[0016] Preferably, the SCMA-VLC system is a SCMA-VLC system using a complex codebook.

[0017] Preferably, the system comprises J users and K resource blocks, and the used bandwidth is divided into N subcarriers;

[0018] At the transmitting end, each user is assigned a predefined codebook consisting of M K-dimensional sparse codewords. According to the codebook, the log2M-bit input binary information bj of each user is mapped into a K-dimensional complex codeword by the SCMA encoder. The encoder for user j can be defined as the following mapping rule:

[0019]

[0020] in, represents the user codebook, B and C represent the binary domain and the complex domain respectively; the coded J user codewords are added on K resource blocks to obtain the SCMA superimposed codeword

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] The present invention proposes a deep learning detection scheme for the SCMA-VLC system based on codebook prior knowledge. The nonlinear distortion of the signal caused by the nonlinear characteristics of the LED is treated as a special noise in the received signal, and the received signal is subjected to noise reduction and multi-user detection integrated operations to reduce the interference of noise on signal detection. Compared with the traditional message passing algorithm, the present invention exhibits better bit error rate performance in the case of strong nonlinear distortion, and can effectively combat the deterioration of detection performance caused by nonlinear distortion. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 Block diagram of the SCMA-VLC system using a deep learning detection network.

[0024] Figure 2 This is a model diagram of the SCMA-VLC system based on DCO-OFDM.

[0025] Figure 3 Schematic diagram of codeword mapping for 4×6 SCMA system.

[0026] Figure 4 This is the PU nonlinear characteristic curve of LED.

[0027] Figure 5 Integrated network diagram for noise reduction and detection.

[0028] Figure 6 This is a model diagram for comparing the performance of multi-user detection algorithms under different peak-to-peak conditions. DETAILED DESCRIPTION

[0029] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0030] like Figure 1 As shown, this implementation provides a deep learning detection method for the SCMA-VLC system based on codebook prior knowledge, and the deep learning detection method includes:

[0031] The user bit information at the transmitting end is encoded by SCMA to obtain a superimposed codeword, which is converted into a non-negative real signal under the DCO-OFDM modulation scheme and drives the LED to send an optical signal;

[0032] The optical signal received at the receiving end is collected by the photodiode and converted into an electrical signal, which is then demodulated by DCO-OFDM. Finally, a deep learning detection network based on codebook prior knowledge is used to process the received signal to recover the user bit information.

[0033] The embodiment of the present invention is based on a SCMA-VLC system using a complex codebook, such as Figure 2 As shown in Figure 2, the system includes J users and K resource blocks, and the bandwidth used is divided into N subcarriers.

[0034] At the transmitting end, each user is assigned a predefined codebook consisting of M K-dimensional sparse codewords. According to the codebook, the log2M-bit input binary information bj of each user is mapped into a K-dimensional complex codeword by the SCMA encoder. The encoder for user j can be defined as the following mapping rule:

[0035]

[0036] in, represents the user codebook, B and C represent the binary domain and the complex domain respectively; the coded J user codewords are added on K resource blocks to obtain the SCMA superimposed codeword

[0037] like Figure 3 A schematic diagram of the mapping and superposition process of the SCMA encoder with J=6, K=4 and M=4 is given. In the figure, the rows of the user codebook correspond to resource blocks, the white part indicates that the user does not occupy the resource block, and the other color parts represent different codewords transmitted on the resource block. The six users map the 2-bit information to the corresponding codewords in the codebook according to their specific codebooks, and perform superposition transmission on the corresponding resource blocks.

[0038] The VLC system is based on intensity modulation and direct detection, which requires the transmitted signal to be a non-negative, real signal. In the SCMA-VLC system using complex codewords, the complex codewords can be converted into non-negative real signals for transmission based on the DCO-OFDM modulation scheme through Hermitian symmetry and DC bias.

[0039] First, the superimposed codewords are converted from K-way parallel to N / 2-1-way parallel, and the following mapping operations are performed:

[0040] X mapping =[0,X(1),X(2),…,X(N / 2-1),0,X * (N / 2-1),…,X * (2), X * (1)] T

[0041] Among them, X * (n) represents the complex conjugate of X(n). The mapped signal X mapping Satisfies Hermitian symmetry, that is, X mapping (n) = X * mapping (Nn). After the Inverse Fast Fourier Transform (IFFT) module, the frequency domain mapping signal X mapping Converted into a real-number time-domain OFDM signal x. After the time-domain signal x is converted from parallel to serial (P / S), the LED is driven under a DC bias, and the modulated optical signal output by the LED can be expressed as:

[0042] x LED =f(x+xDC)

[0043] Where xDC is the DC bias and f(·) represents the nonlinear conversion function of the LED from voltage to optical power.

[0044] The embodiment of the present invention experimentally measures the optical power of LED under different input voltage conditions, uses the linear equation P=aU-b to fit the linear part of the LED's electro-optical conversion, and uses the Rapps model to describe the nonlinear characteristics of the LED. The LED optical power-voltage (PU) conversion curve can be expressed as:

[0045]

[0046] Where U is the voltage loaded on the LED, P is the LED output light power, a=66.8 and b=224.8 are the slope and intercept of the linear part of the electro-optical conversion, b / a is the starting voltage, k=3 is used to characterize the smoothness of the transition from the linear region to the saturation region of the curve, P max=3.8mW is the maximum output optical power, c = 1.6mW is the starting output optical power of the LED approximate linear range, Figure 4 The PU nonlinear curve of LED is shown. It can be seen from the figure that the modulation voltage of LED has a linear range. Beyond this range, the modulation signal will be clipped, resulting in nonlinear distortion. This phenomenon is more serious in the SCMA-VLC system based on DCO-OFDM with relatively high peak-to-average power.

[0047] The optical signal output by the LED is transmitted through the VLC channel, collected by the PD at the receiving end and converted into an electrical signal. After removing the DC bias, the signal received by user j can be expressed as:

[0048] y j =γh j f(x)+nj

[0049] Among them, γ represents the responsiveness of PD, n j ~N(0,σ' 2 ) represents additive Gaussian white noise, h j represents the channel gain between LED and user j. The received signal y j After serial-to-parallel conversion (S / P), fast Fourier transform (FFT) and Hermitian symmetric demapping, it is restored to a complex signal. Finally, the complex signal is converted into K parallel signals. A deep learning detection method of SCMA-VLC system based on codebook prior knowledge is used to recover user bit information.

[0050] The embodiment of the present invention regards the signal nonlinear distortion caused by the nonlinear characteristics of LED as a special noise processing in the received signal, and proposes a deep learning detection network based on codebook prior knowledge, which makes full use of the advantages of deep neural networks (DNN) in processing multidimensional signals with nonlinear characteristics, and performs integrated processing of noise reduction and multi-user detection on the received signal to reduce the interference of noise on signal detection, so as to have better multi-user detection performance under the condition of large nonlinear distortion.

[0051] The deep learning detection network based on codebook prior knowledge has the following structure: Figure 5 As shown in the figure. The network converts multi-user detection into a multi-output classification problem, with log2M bits of user input information b j M categories are formed, and each output is used to judge the corresponding input bit information of a user. The proposed network is based on DNN and consists of a denoising subnetwork and J user subnetworks.

[0052] The denoising subnetwork is used to reconstruct the original noise-free SCMA superposition codeword in the received signal to reduce the interference of noise on subsequent multi-user detection. The reconstructed SCMA superposition codeword is represented by The denoising subnetwork contains 5 fully connected hidden layers, and a batch normalization (BN) operation and a ReLU activation function are added after each hidden fully connected layer. Since neural networks can only process real numbers, Figure 3 The complex signal output by the FFT module at the receiving end The real and imaginary parts of are connected as the input of the denoising sub-network, that is, The input dimension is 2K. The output of the denoising subnetwork is represented as x d =d(x in θ d ), which represents the reconstruction signal The connection of the real and imaginary parts θ d Represents the parameters of the denoising subnetwork. During the training process, the mean square error (MSE) is used as the training loss function of the denoising subnetwork to estimate the reconstruction loss of the denoising subnetwork in order to update the weights and biases and reduce the loss of the next iteration. The mean square error between the reconstructed signal and the noise-free input SCMA superimposed codeword is expressed as:

[0053]

[0054] The user sub-network is used to detect the bit information sent by each user. It contains 2 fully connected hidden layers and 1 fully connected output layer. The output of the j-th user sub-network is represented as x u,j =u(x d θ u,j );θ u,j Represents the parameters of the user sub-network. The activation function of the fully connected output layer of each user sub-network is the Softmax function, which represents

[0055]

[0056] in, Input vector x to the Softmax function of the j-th user sub-network u,j The mth element of Output vector p for the Softmax function j The mth element of . It can be seen that and but It can be regarded as the probability of the mth bit information combination of user j, by selecting the maximum The corresponding bit information combination of is used to detect the input data of user j. In order to enable the proposed network to accurately detect user input information, the cross entropy loss function is minimized as the objective function of network parameter optimization. The cross entropy loss function of a single user is expressed as:

[0057]

[0058] Among them, l j is the bit information b input by user j j Tags, is the mth element of lj. The joint loss function of the multi-output classification problem is usually the weighted sum of the loss functions of the individual classification problems. In order to ensure user fairness in the multi-user detection process, the weight of each user's cross entropy loss function is set to 1, and the loss function of the detection network of all users is expressed as:

[0059]

[0060] Where P = {p1,···,p J}, l={l1,···l J}.

[0061] In order to realize the integrated processing of received signal noise reduction and multi-user detection, the present invention simultaneously trains the noise reduction sub-network and the user sub-network, and trains the θ d ,θ u,1 ,···θ u,J The parameters are optimized, and the overall loss function can be expressed as:

[0062] L=L d +L u

[0063] This application considers a SCMA-VLC system under the conditions of J=6, K=4 and M=4. The number of nodes corresponding to the 5 fully connected hidden layers of the denoising subnetwork is {128, 64, 64, 32, 8}, and the number of nodes corresponding to the 2 fully connected hidden layers and 1 fully connected output layer of the user subnetwork is {16, 16} and 4, respectively.

[0064] The data used for network training is generated by MATLAB. The user input binary data is generated by random functions and used as the label of the data set. The user bit data is encoded at the transmitter of the SCMA-VLC system, Hermitian symmetry, IFFT, channel transmission, FFT at the receiver, and Hermitian symmetry removal to obtain the signal as input data. The training data contains 2×10 6The samples are divided into training set and validation set in a ratio of 4:1. The data set samples are generated under the conditions of noise power of 0.01mW and peak-to-peak modulation signal of 30mV to 150mV, so that the network can fully learn effective information to compensate for the nonlinear distortion of the signal and detect multi-user information. During the network training process, the Adaptive Moment Estimation (ADAM) optimizer is used to optimize the overall network loss function. The initial learning rate is set to 0.001, and the learning rate is gradually reduced according to the cosine annealing learning rate adjustment strategy. The batch size is 2000. The trained detection network parameters are saved, and the test data is input for network performance testing.

[0065] Figure 6 Shows different V PP Under the same conditions, the Bit Error Ratio (BER) performance of the traditional MPA detection algorithm and the deep learning detection method based on codebook prior knowledge is compared, where the bias voltage VDC is set to 3.39V and the power of the added Gaussian white noise is 0.02mW and 0.01mW respectively. As can be seen from the figure, due to the nonlinear characteristics of the LED, as the modulation signal V PP As V PP When V is low, the dynamic range of the modulation signal is within the linear region of the LED. PP As V increases, the signal-to-noise ratio of the received signal gradually increases, and the average BER of the user decreases. When the dynamic range of the modulated signal enters the nonlinear region, the two ends of the signal are truncated to produce nonlinear distortion. At this time, the BER performance decreases with V PP When 0.02mW of noise is added, MPA and the proposed deep learning detection network based on codebook prior knowledge achieve the minimum BER performance at 80mV and 90mV respectively; when 0.01mW of noise is added, the minimum BER performance is achieved at 70mV and 80mV respectively. As can be seen from the figure, in the linear region of the LED, the deep learning detection network based on codebook prior knowledge can basically maintain the near-optimal detection performance of MPA. In the nonlinear region of the LED, the deep learning detection network based on codebook prior knowledge can effectively improve the performance deterioration caused by nonlinear distortion and expand the dynamic range of the LED modulation voltage.

[0066] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0067] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A deep learning detection method for SCMA-VLC system based on codebook prior knowledge, characterized in that: The deep learning detection method includes: The user bit information at the transmitting end is encoded by SCMA to obtain a superimposed codeword, which is converted into a non-negative real signal under the DC bias optical orthogonal frequency division multiplexing modulation scheme and drives the LED to send an optical signal; The optical signal received at the receiving end is collected by a photodiode and converted into an electrical signal, which is then demodulated using DC biased optical orthogonal frequency division multiplexing. Finally, a deep learning detection network based on codebook prior knowledge is used to process the received signal to recover the user bit information. The deep learning detection network based on codebook prior knowledge is used to process the received signal, specifically: A deep learning detection network is used to perform integrated processing of noise reduction and multi-user detection on the received signal to reduce the interference of noise on signal detection and enhance the ability to extract network signal features, so as to achieve better multi-user detection performance under nonlinear distortion. The deep learning detection network based on codebook prior knowledge includes: The multi-user detection is converted into a multi-output classification problem. The log2M-bit user input bit information bj constitutes M categories, and each output is used to judge the corresponding input bit information of a user. The proposed network is based on the deep learning detection network, which consists of a denoising sub-network and J user sub-networks. The denoising subnetwork is used to reconstruct the original noise-free SCMA superposition codewords in the received signal to reduce the interference of noise on subsequent multi-user detection; wherein the denoising subnetwork includes 5 fully connected hidden layers, and a batch normalization operation and a ReLU activation function are added after each fully connected hidden layer; The user sub-network is used for detecting bit information sent by each user, and comprises two fully connected hidden layers and one fully connected output layer.

2. According to the SCMA-VLC system deep learning detection method based on codebook prior knowledge according to claim 1, the SCMA-VLC system is a SCMA-VLC system using a complex codebook.

3. The SCMA-VLC system deep learning detection method based on codebook prior knowledge according to claim 2, characterized in that: The system contains J users and K resource blocks, and the used bandwidth is divided into N subcarriers; At the transmitting end, each user is assigned a predefined codebook consisting of M K-dimensional sparse codewords, and the log2M-bit input b of each user in the codebook is j The SCMA encoder is mapped into a K-dimensional complex codeword The encoder for user j can be defined as the following mapping rule: in, represents the user codebook, b j Represents binary information, B and C represent the binary domain and the complex domain respectively; the coded J user codewords are added on K resource blocks to obtain the SCMA superimposed codeword

Images