A high-order modulation constellation design method for hybrid automatic repeat request systems

Abstract

The present invention discloses a high-order modulation constellation design method suitable for a hybrid automatic repeat request system. The neural network cost function in the design method adopts a cross entropy function; the training method adopts a back propagation algorithm; the high-order modulation signal generated after the binary bit stream is mapped through the constellation is used as the input of the neural network encoder, and the output of the neural network encoder represents the transmission signal corresponding to the input high-order modulation signal; the real part and the imaginary part of the received signal after passing through the Gaussian white noise channel are used as the input of the neural network decoder, and the output result of the neural network decoder represents the probability value of each bit in the high-order modulation signal being 1; the activation functions of the hidden layer and the output layer of the neural network respectively adopt conventional Relu functions and Linear functions; the present invention not only outperforms the traditional square high-order modulation constellation in the process of single transmission, but also can obtain greater performance gain through retransmission merging of multiple transmissions.

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless communications, and in particular to a method for designing a high-order modulation constellation in a hybrid automatic repeat request system based on a deep learning network. Background Art

[0002] In mobile communication systems, forward error correction (FEC) technology and automatic repeat request (ARQ) technology are usually used to ensure the reliability of data transmission. FEC technology is to protect the transmitted data in groups, and improve the reliability of the system at the cost of introducing redundant bits. This technology can keep the system throughput constant. The use of ARQ technology can greatly improve the reliability of the system, but in the case of poor channel conditions, ARQ technology will increase the number of retransmissions of the system, resulting in a decrease in the system throughput. Hybrid automatic repeat request (HARQ) is a core key technology in the eMBB scenario and is expected to continue to play an important role in the next generation of mobile communication systems. HARQ technology combines the advantages of ARQ technology and FEC technology, and improves the decoding capability of the receiving end at the cost of increasing less redundancy. Chase combining (CC) is a type of HARQ. Its characteristic is that the data packet during retransmission is exactly the same as the data packet transmitted for the first time. Each transmitted data packet contains information bits and redundant bits, which means that each transmitted data packet can be decoded separately. This solution has been written into the IEEE 802.16e protocol as a standard.

[0003] In the Additive White Gaussian Noise (AWGN) channel, in order to maximize the mutual information between the channel input and output, the channel input signal needs to satisfy the Gaussian distribution. Forney et al. proposed that when the modulation order approaches infinity, compared with the uniformly distributed QAM signal, the input signal that satisfies the Gaussian distribution can obtain a performance gain of up to 1.53dB, and this part of the gain is defined as the shaping gain. The modulation signal used in the 5G standard protocol does not satisfy the Gaussian distribution, but satisfies the uniformly distributed QAM signal. On the one hand, the use of QAM signals is simple in the engineering implementation process and there is a simplified algorithm for demodulation. On the other hand, high-order modulation signals that satisfy the Gaussian distribution are not easy to design. For high-order QAM modulation signals, each bit on the constellation symbol that satisfies the Gray mapping has different reliability. When the CC scheme is adopted, the reliability gap between different bits on the symbol will become larger, which will cause a partial loss of decoding performance. C. Wengerter et al. proposed a HARQ technology based on the Constellation Rearrangement (CoRe) scheme, which changes the constellation mapping method during each retransmission to make the reliability of each bit on the superimposed constellation symbol consistent. The disadvantage of this method is that the CoRe scheme often requires a specific number of transmissions to achieve the goal of consistent bit reliability, that is, the CoRe scheme with a limited number of transmissions cannot make the reliability of each bit on the symbol consistent. Summary of the invention

[0004] The present invention aims to provide a high-order modulation constellation design method applicable to a hybrid automatic repeat request system to solve the technical problem of different reliability gaps between bits on constellation symbols in the retransmission system.

[0005] In order to solve the above technical problems, the specific technical solutions of the present invention are as follows:

[0006] A high-order modulation constellation design method for a hybrid automatic repeat request system is used for a binary bit stream X = {x1, x2, ..., x N}, specifically including the following steps:

[0007] Step 1, the transmitter obtains the number of bits m in a single M-order orthogonal amplitude modulation signal according to the modulation order M and m=log2M, and obtains the number of groups K according to K=N / m, and then proceeds to step 2; wherein N represents the number of bits in the binary bit stream X;

[0008] Step 2: The transmitter sequentially groups the bits in the binary bit stream X with the number of bits m in a single M-order orthogonal amplitude modulation signal to obtain K modulation orders of M-order orthogonal amplitude modulation signals S, where S = {s1, ..., s k , ..., s K},s k represents the kth M-order orthogonal amplitude modulation signal in the M-order orthogonal amplitude modulation signal S; wherein, if the last s in the sequence k If the number of bits in is less than m, 0 is added to the end until the number of bits is m.

[0009] Step 3: Construct two sets of deep neural networks DNN-Enc-A and DNN-Enc-B for encoders;

[0010] Step 4: Construct two decoder deep neural networks DNN-Dec-A and DNN-Dec-B with fully connected structures and the same parameters;

[0011] Step 5: Modulate the M-order quadrature amplitude modulation signal S = {s1, ..., s k , ..., s K} as the input of DNN-Enc-A and DNN-Enc-B; the output of DNN-Enc-A is signal V 1st = {v1, ..., v k , ..., v K},v k Indicates signal V 1st The kth M-order quadrature amplitude modulation signal in the DNN-Enc-B is composed of a real part and an imaginary part. 2nd = {v1, ..., v k , ..., v K}, the output of DNN-Enc-A is signal V 1st After power normalization, the signal is obtained The output of DNN-Enc-B is signal V 2nd After power normalization, the signal is obtained

[0012] Step 6: Transmit the signal and signal The signal N with superimposed Gaussian noise is obtained through the AWGN channel respectively. 1st and N 2nd ;

[0013] Step 7: Transform the signal N 1st and N 2ndAs the input of DNN-Dec-A and DNN-Dec-B respectively; the output of each node of the output layer of the decoder deep neural network is the probability that the corresponding bit in each M-order orthogonal amplitude modulation signal is 1, and the two decoder deep neural networks respectively obtain the output P 1st ={p1, ..., p k , ..., p K} and P 2nd ={p1, ..., p k , ..., p K};

[0014] Step 8: P 1st ={p1, ..., p k , ..., p K} and P 2nd ={p1, ..., p k , ..., p K} and the modulation order is M-order orthogonal amplitude modulation signal S to calculate the binary classification cross entropy loss Loss 1st and Loss 2nd , and loss 1st and loss 2nd Add up to get the total network loss Loss total ; Using the total network loss Loss total Using the gradient descent method, calculate the gradient of each node in each intermediate layer of the deep neural network, update the node parameters of each layer of DNN-Enc-A, DNN-Enc-B, DNN-Dec-A and DNN-Dec-B; repeat steps 5 to 8 until the total network loss Loss total The trained neural network is obtained after the numerical stability no longer has a downward trend;

[0015] Step 9: Generate an M-order quadrature amplitude modulation signal Q={q1, ..., q m , ..., q M}, take the signal Q as the input of the trained DNN-Enc-A and DNN-Enc-B, and get the output of DNN-Enc-A As the constellation of the first transmission process in the retransmission system, the output of DNN-Enc-B As the constellation for the second transmission in a retransmission system.

[0016] Furthermore, in step 3, both deep neural networks DNN-Enc-A and DNN-Enc-B adopt a fully connected structure with the same structural parameters, and are composed of 5 layers respectively, and their structures are as follows: input layer->first hidden layer->second hidden layer->third hidden layer->linear output layer;

[0017] The parameters of each layer of DNN-Enc-A and DNN-Enc-B are set as follows: the number of nodes in the input layer is set to the number of bits m, the number of nodes in the first hidden layer to the third hidden layer is set to 256, and the number of nodes in the linear output layer is set to 2.

[0018] Furthermore, in step 4, the decoder deep neural networks DNN-Dec-A and DNN-Dec-B are composed of 4 layers respectively, and their structures are as follows: input layer->first hidden layer->second hidden layer->linear output layer, and the parameters of each layer of DNN-Dec-A and DNN-Dec-B are set as follows: the number of nodes in the input layer is set to 2, the number of nodes in the first and second hidden layers is set to 256, and the number of nodes in the linear output layer is set to m.

[0019] A high-order modulation constellation design method applicable to a hybrid automatic repeat request system of the present invention has the following advantages:

[0020] The present invention realizes the design of high-order modulation constellation through deep autoencoder neural network to reduce the reliability gap between bits on constellation symbols in the retransmission system, and optimizes the constellation symbol position and mapping method during multiple transmissions, thereby improving the decoding performance of the retransmission system under limited times. Compared with the prior art, the constellation diagram generated in step 9 of the present invention is better than the square QAM constellation in decoding performance in a single transmission, and the decoding performance in the retransmission system under two retransmissions is significantly improved compared with the CoRe scheme. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 is a system structure flow chart of the present invention;

[0022] FIG. 2( a ) is a 16QAM constellation diagram for the first transmission process of a hybrid automatic repeat request system generated according to the first embodiment of the present invention;

[0023] FIG2( b ) is a 16QAM constellation diagram for the second transmission process of a hybrid automatic repeat request system generated according to the first embodiment of the present invention;

[0024] Figure 3 The bit error rate curve of the 802.11n standard LDPC code with a code rate of 1 / 2 and an information bit length of 972 transmitted using the 16QAM constellation diagram generated by the first embodiment of the present invention under AWGN channel conditions;

[0025] FIG4(a) is a 64QAM constellation diagram for the first transmission process of a hybrid automatic repeat request system generated by the second embodiment of the present invention;

[0026] FIG4( b ) is a 64QAM constellation diagram for the second transmission process of the hybrid automatic repeat request system generated by the second embodiment of the present invention;

[0027] Figure 5 The bit error rate curve of 802.11n standard LDPC code with a code rate of 1 / 2 and an information bit length of 972 transmitted using the 64QAM constellation diagram proposed in the second embodiment of the present invention under AWGN channel conditions is shown. DETAILED DESCRIPTION

[0028] In order to better understand the purpose, structure and function of the present invention, a high-order modulation constellation design method applicable to a hybrid automatic repeat request system of the present invention is further described in detail below with reference to the accompanying drawings.

[0029] A high-order modulation constellation design method for a hybrid automatic repeat request system is used for a binary bit stream X = {x1, x2, ..., x N}, realizing the transmission control from the transmitter to the receiver, which specifically includes the following steps:

[0030] Step 1: The transmitter obtains the number of bits m in a single M-order quadrature amplitude modulation (M-QAM) signal according to the modulation order M and m=log2M, and obtains the number of groups K according to K=N / m, and then proceeds to step 2; wherein N represents the number of bits in the binary bit stream X;

[0031] Step 2: The transmitter sequentially groups the bits in the binary bit stream X with the number of bits m in a single M-order orthogonal amplitude modulation signal to obtain K modulation orders of M-order orthogonal amplitude modulation signals S, where S = {s1, ..., s k , ..., s K},s k represents the kth M-order orthogonal amplitude modulation signal in the M-order orthogonal amplitude modulation signal S; wherein, if the number of bits in the last M-order orthogonal amplitude modulation signal in the sequence is less than m, 0 is added to the end until the number of bits is satisfied;

[0032] Step 3: Construct two sets of deep neural networks (DNNs) DNN-Enc-A and DNN-Enc-B for the encoder; both sets of deep neural networks for the encoder adopt a fully connected structure with the same structural parameters, and are composed of 5 layers respectively, and their structures are as follows: input layer->first hidden layer->second hidden layer->third hidden layer->linear output layer;

[0033] The parameters of each layer of DNN-Enc-A and DNN-Enc-B are set as follows: the number of nodes in the input layer is set to m, the number of nodes in the first hidden layer to the third hidden layer is set to 256, and the number of nodes in the linear output layer is set to 2;

[0034] Step 4: Construct two decoder deep neural networks DNN-Dec-A and DNN-Dec-B with fully connected structures and the same parameters. They are composed of 4 layers respectively, and their structures are as follows: input layer->first hidden layer->second hidden layer->linear output layer. Set the parameters of each layer of DNN-Dec-A and DNN-Dec-B as follows: set the number of nodes in the input layer to 2, the number of nodes in the first and second hidden layers to 256, and the number of nodes in the linear output layer to m;

[0035] Step 5: Modulate the M-order quadrature amplitude modulation signal S = {s1, ..., s k , ..., s K} as the input of DNN-Enc-A and DNN-Enc-B; the output of DNN-Enc-A is signal V 1st = {v1, ..., v k , ..., v K},v k Indicates signal V 1st The kth M-order quadrature amplitude modulation signal in the DNN-Enc-B is composed of a real part and an imaginary part. 2nd = {v1, ..., v k , ..., v K}, the output of DNN-Enc-A is signal V 1st After power normalization, the signal is obtained The output of DNN-Enc-B is signal V 2nd After power normalization, the signal is obtained

[0036] Step 6: Transmit the signal and signal The signal N with superimposed Gaussian noise is obtained through the AWGN channel respectively. 1st and N 2nd ;

[0037] Step 7: Transform the signal N 1st and N 2nd As the input of DNN-Dec-A and DNN-Dec-B respectively; the output of each node of the output layer of the decoder deep neural network is the probability that the corresponding bit in each M-order orthogonal amplitude modulation signal is 1, and the two decoder deep neural networks respectively obtain the output P 1st ={p1, ..., p k , ..., p K} and P 2nd ={p1, ..., p k , ..., p K};

[0038] Step 8: P 1st ={p1, ..., P k , ..., p K} and P 2nd ={p1, ..., p k , ..., p K} and the modulation order is M-order orthogonal amplitude modulation signal S to calculate the binary classification cross entropy loss Loss 1st and Loss 2nd , and loss 1st and loss 2nd Add up to get the total network loss Loss total ; Using the total network loss Loss total Using the gradient descent method, calculate the gradient of each node in each intermediate layer of the deep neural network, update the node parameters of each layer of DNN-Enc-A, DNN-Enc-B, DNN-Dec-A and DNN-Dec-B; repeat steps 5 to 8 until the total network loss Loss total The trained neural network is obtained after the numerical stability no longer has a downward trend;

[0039] Step 9: Generate an M-order quadrature amplitude modulation signal Q={q1, ..., q m , ..., q M}, take the signal Q as the input of the trained DNN-Enc-A and DNN-Enc-B, and get the output of DNN-Enc-A As the constellation of the first transmission process in the retransmission system, the output of DNN-Enc-B As the constellation for the second transmission in a retransmission system.

[0040] Embodiment 1:

[0041] The high-order modulation constellation design method designed by the present invention for a hybrid automatic repeat request system is used to design a constellation diagram with a modulation order of 16QAM. It is necessary to build a communication system simulation platform in the presence of AWGN noise interference during the transmission process, wherein the main physical layer parameters of the communication system simulation platform are shown in Table 1 below.

[0042]

[0043] Table 1

[0044] like Figure 1As shown, the high-order modulation constellation design method for a hybrid automatic repeat request system designed by the present invention is applied to generate a binary bit stream X with a length of 64000, that is, N=64000; the system modulation order M is set to 16, and the number of bits m=4 and the number of groups K=16000 in a single QAM signal are obtained; the transmitter applies a constellation mapping method to sequentially group each bit in the binary bit stream X to obtain 16000 QAM signals to form a 16-QAM signal S; the signal S is used as the input of DNN-Enc-A and DNN-Enc-B, and the output signals P of DNN-Dec-A and DNN-Dec-B are obtained after the signal is processed by the system. 1st and P 2nd ; P 1st and P 2nd Calculate the cross entropy with the 16-QAM signal S to get the total network loss Loss total Using the total loss Loss total The gradient descent method is used to calculate the gradient of each node in each intermediate layer of the deep neural network and update the node parameters of each layer of DNN-Enc-A, DNN-Enc-B, DNN-Dec-A and DNN-Dec-B. The number of training times is set to 600,000 times; after the network training is completed, a 16-QAM signal Q = {q1, ..., q m , ..., q 16}, take the signal Q as the input of the trained DNN-Enc-A and DNN-Enc-B, and get the output of DNN-Enc-A As the constellation of the first transmission process in the retransmission system, the output of DNN-Enc-B As the constellation for the second transmission in the retransmission system. The constellation diagram obtained for the first transmission process is shown in Figure 2(a), and the constellation diagram for the second transmission process is shown in Figure 2(b). The specific constellation point data results of the obtained constellation diagram are shown in Table 2.

[0045] Compared with the existing constellation design transmission algorithm in the retransmission system, the superiority of the constellation diagram designed by the neural network is evaluated from the perspective of the system bit error rate (BER), such as Figure 3 As shown, the horizontal axis represents SNR. The constellation diagram designed by the neural network not only has better single transmission performance than the existing square QAM constellation, but also performs significantly better than the CoRe scheme in the two-transmission retransmission system.

[0046]

[0047] Table 2

[0048] Second embodiment:

[0049] The high-order modulation constellation design method designed by the present invention for the hybrid automatic repeat request system is used to design a constellation diagram with a modulation order of 64QAM. It is necessary to build a communication system simulation platform in the presence of AWGN noise interference during the transmission process, wherein the main physical layer parameters of the communication system simulation platform are shown in Table 3 below.

[0050]

[0051]

[0052] Table 3

[0053] like Figure 1 As shown, the high-order modulation constellation design method for a hybrid automatic repeat request system designed by the present invention is applied to generate a binary bit stream X with a length of 60,000, that is, N=60,000; the system modulation order M is set to 64, and the number of bits m=6 and the number of groups K=10,000 in a single QAM signal are obtained; the transmitter applies a constellation mapping method to sequentially group each bit in the binary bit stream X to obtain 10,000 QAM signals to form a 64-QAM signal S; the signal S is used as the input of DNN-Enc-A and DNN-Enc-B, and the output signals P of DNN-Dec-A and DNN-Dec-B are obtained after the signal is processed by the system. 1st and P 2nd ; P 1st and P 2nd Calculate the cross entropy with the 16-QAM signal S to get the total network loss Loss total Using the total loss Loss total The gradient descent method is used to calculate the gradient of each node in each intermediate layer of the deep neural network and update the node parameters of each layer of DNN-Enc-A, DNN-Enc-B, DNN-Dec-A and DNN-Dec-B. The number of training times is set to 600,000 times; after the network training is completed, a 64-QAM signal Q = {q1, ..., q m , ..., q 64}, take the signal Q as the input of the trained DNN-Enc-A and DNN-Enc-B, and get the output of DNN-Enc-A As the constellation of the first transmission process in the retransmission system, the output of DNN-Enc-B As the constellation for the second transmission in the retransmission system. The constellation diagram for the first transmission process is shown in Figure 4(a), and the constellation diagram for the second transmission process is shown in Figure 4(b). The specific constellation point data results of the obtained constellation diagram are shown in Table 4.

[0054] Compared with the existing constellation design transmission algorithm in the retransmission system, the superiority of the constellation diagram designed by the neural network is evaluated from the perspective of the system bit error rate (BER), such as Figure 5 As shown, the horizontal axis represents SNR. The constellation diagram designed by the neural network not only has better single transmission performance than the existing square QAM constellation, but also performs significantly better than the CoRe scheme in the two-transmission retransmission system.

[0055]

[0056]

[0057] Table 4

[0058] It is to be understood that the present invention is described by some embodiments, and it is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the scope of protection of the present invention.

Claims

1. A method for designing a high-order modulation constellation for a hybrid automatic repeat request system, characterized in that: For a binary bit stream X = {x1, x2, ..., x N }, specifically including the following steps: Step 1, the transmitter obtains the number of bits m in a single M-order orthogonal amplitude modulation signal according to the modulation order M and m=log2M, and obtains the number of groups K according to K=N / m, and then proceeds to step 2; wherein N represents the number of bits in the binary bit stream X; Step 2: The transmitter groups the bits in the binary bit stream X in sequence according to the number of bits m in a single M-order orthogonal amplitude modulation signal to obtain K orthogonal amplitude modulation signals S with a modulation order of M, where S = {s1, ..., s k ,...,s K },s k represents the kth M-order orthogonal amplitude modulation signal in the M-order orthogonal amplitude modulation signal S; wherein, if the last s in the sequence k If the number of bits in is less than m, 0 is added to the end until the number of bits is m. Step 3: Construct two sets of deep neural networks DNN-Enc-A and DNN-Enc-B for encoders; Step 4: Construct two decoder deep neural networks DNN-Dec-A and DNN-Dec-B with fully connected structures and the same parameters; Step 5: Modulate the M-order quadrature amplitude modulation signal S = {s1, ..., s k ,...,s K } as the input of DNN-Enc-A and DNN-Enc-B; the output of DNN-Enc-A is signal V 1st ={v1,...,v k ,...,v K },v k Indicates signal V 1st The kth M-order quadrature amplitude modulation signal in the DNN-Enc-B is composed of a real part and an imaginary part. 2nd ={v1,...,v k ,...,v K }, the output of DNN-Enc-A is signal V 1st After power normalization, the signal is obtained The output of DNN-Enc-B is signal V 2nd After power normalization, the signal is obtained Step 6: Transmit the signal and signal The signal N with superimposed Gaussian noise is obtained through the AWGN channel respectively. 1st and N 2nd ; Step 7: Transform the signal N 1st and N 2nd As the input of DNN-Dec-A and DNN-Dec-B respectively; the output of each node of the output layer of the decoder deep neural network is the probability that the corresponding bit in each M-order orthogonal amplitude modulation signal is 1, and the two decoder deep neural networks respectively obtain the output P 1st ={p1,...,p k ,...,p K } and P 2nd ={p1,...,p k ,...,p K }; Step 8: P 1st ={p1,...,p k ,...,p K } and P 2nd ={p1,...,p k ,...,p K } respectively calculate the binary classification cross entropy loss Loss with the modulation order M-order orthogonal amplitude modulation signal S 1st and Loss 2nd , and loss 1st and loss 2nd Add up to get the total network loss Loss total ; Using the total network loss Loss total Using the gradient descent method, calculate the gradient of each node in each intermediate layer of the deep neural network, update the node parameters of each layer of DNN-Enc-A, DNN-Enc-B, DNN-Dec-A and DNN-Dec-B; repeat steps 5 to 8 until the total network loss Loss total The trained neural network is obtained after the numerical stability no longer has a downward trend; Step 9: Generate an M-order quadrature amplitude modulation signal Q={q1,...,q m ,...,q M }, take the signal Q as the input of the trained DNN-Enc-A and DNN-Enc-B, and get the output of DNN-Enc-A As the constellation of the first transmission process in the retransmission system, the output of DNN-Enc-B As the constellation for the second transmission in a retransmission system.

2. The method for designing a high-order modulation constellation for a hybrid automatic repeat request system according to claim 1, characterized in that: In step 3, both deep neural networks DNN-Enc-A and DNN-Enc-B adopt a fully connected structure with the same structural parameters. They are composed of 5 layers respectively, and their structures are as follows: input layer->first hidden layer->second hidden layer->third hidden layer->linear output layer; The parameters of each layer of DNN-Enc-A and DNN-Enc-B are set as follows: the number of nodes in the input layer is set to the number of bits m, the number of nodes in the first hidden layer to the third hidden layer is set to 256, and the number of nodes in the linear output layer is set to 2.

3. The method for designing a high-order modulation constellation for a hybrid automatic repeat request system according to claim 2, characterized in that: In step 4, the decoder deep neural networks DNN-Dec-A and DNN-Dec-B are composed of 4 layers respectively, and their structures are as follows: input layer -> first hidden layer -> second hidden layer -> linear output layer. The parameters of each layer of DNN-Dec-A and DNN-Dec-B are set as follows: the number of nodes in the input layer is set to 2, the number of nodes in the first and second hidden layers is set to 256, and the number of nodes in the linear output layer is set to m.

Images