NOMA transceiver design method and system based on multi-user variational information bottleneck

Abstract

The invention discloses an NOMA transceiver design method and system based on a multi-user variational information bottleneck, and relates to the communication technology, and the method comprises the steps: enabling an NOMA end-to-end system to be equivalent to an auto-encoder network model, enabling the auto-encoder network model to comprise an encoder network and a decoder network, enabling the encoder network to be equivalent to a transmitting end of the NOMA system, enabling the decoder network to be equivalent to a receiving end of the NOMA system, and enabling the encoder network to be equivalent to a decoding end of the NOMA system; the decoder network is equivalent to a receiving end of the NOMA system; according to the constructed NOMA end-to-end system, constructing a cost function for training an auto-encoder network model based on an information bottleneck theory; and training the constructed auto-encoder network model to complete the design of the NOMA transceiver. According to the method, under a deep learning framework, the reliability and the overload rate of the system are considered, and the design requirements of different bandwidth requirements and different modulation orders of different users are considered.

Description

Technical Field

[0001] The present application relates to the field of communication technology, and in particular to a NOMA transceiver design method and system based on multi-user variational information bottleneck. Background Art

[0002] As an alternative air interface technology for large-scale access in the future 6G, NOMA has received widespread attention from the industry in recent years. Existing research mainly focuses on the codebook design of NOMA and the design of multi-user detection algorithms, with the goal of improving system performance, obtaining higher overload performance or reducing the average bit error rate. In general, current research is generally based on the idea of ​​divide and conquer - the codebook design and the receiver multi-user detection algorithm are designed and optimized separately. However, the divide and conquer optimization method can generally only obtain local optimal results. If the codebook design and the multi-user detection algorithm are jointly optimized, the optimization complexity is often too high and it is difficult to perform as a whole. Moreover, since the NOMA system allows users to use communication resources non-orthogonally, the superposition of signals also leads to a complex coupling relationship between the parameters to be optimized.

[0003] Existing research has adopted a joint optimization method for NOMA transceivers based on deep learning. The NOMA transceiver is jointly optimized using a deep learning autoencoder network. However, most of this type of work currently uses minimizing the mean square error function, minimizing the cross entropy, or minimizing the lower bound of evidence to optimize the network. The essence is to minimize the reconstruction error, that is, to improve the reliability of the system. In addition, it is usually assumed that users use the same order of modulation. For the NOMA system, in addition to improving reliability, it is also necessary to consider improving the overload rate, that is, carrying more users with as few communication resources as possible. In addition, the bandwidth requirements of different users in the network may also be different, so different orders of modulation are used. Summary of the invention

[0004] The embodiments of the present application provide a NOMA transceiver design method and system based on multi-user variational information bottleneck. Under the deep learning framework, it takes into account the reliability and overload rate of the system, and considers the different bandwidth requirements of different users and the design requirements of different modulation orders.

[0005] The embodiment of the present application provides a NOMA transceiver design method based on multi-user variational information bottleneck, including:

[0006] The NOMA end-to-end system is equivalent to an autoencoder network model, wherein the autoencoder network model includes an encoder network and a decoder network. The encoder network is equivalent to the transmitter of the NOMA system, realizes symbol-level processing, and maps the user's transmitted bits to K-dimensional codewords; the decoder network is equivalent to the receiver of the NOMA system, realizes multi-user detection, and restores the transmitted bits of each user;

[0007] According to the constructed NOMA end-to-end system, the cost function for training the autoencoder network model is constructed based on the information bottleneck theory;

[0008] The constructed autoencoder network model is trained to complete the NOMA transceiver design.

[0009] The present application also provides an embodiment of a NOMA transceiver design system based on a multi-user variational information bottleneck, including a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the steps of the aforementioned NOMA transceiver design method based on a multi-user variational information bottleneck are implemented.

[0010] The method of the embodiment of the present application takes into account the reliability and overload rate of the system under the deep learning framework, and considers the design requirements of different bandwidth requirements and modulation orders of different users.

[0011] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present application. Also, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0013] Figure 1 The basic process diagram of the NOMA transceiver design method of the embodiment of the present application is shown;

[0014] Figure 2 A schematic diagram of a NOMA system model of a NOMA transceiver design method according to an embodiment of the present application;

[0015] Figure 3 A schematic diagram of a NOMA communication system based on an autoencoder for a NOMA transceiver design method according to an embodiment of the present application;

[0016] Figure 4 For the BER performance comparison under the AWGN channel of the application example of the embodiment of the present application, J=6, K=4;

[0017] Figure 5 Comparison of BER performance at different overload rates in an AWGN channel of an application example of an embodiment of the present application;

[0018] Figure 6 BER performance comparison under Rayleigh channel for the application example of the embodiment of this application, J = 6, K = 4. Detailed implementation manners

[0019] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be fully conveyed to those skilled in the art.

[0020] The embodiment of this application provides a design method for a NOMA transceiver based on multi-user variational information bottleneck, as Figure 1 shown, including the following steps:

[0021] In step S101, the NOMA end-to-end system is equivalent to an autoencoder network model, where the autoencoder network model includes an encoder network and a decoder network. The encoder network is equivalent to the transmitting end of the NOMA system, realizing symbol-level processing and mapping the user's transmitted bits to a K-dimensional codeword; the decoder network is equivalent to the receiving end of the NOMA system, realizing multi-user detection and restoring the transmitted bits of each user. In some embodiments, as Figure 2 shown, the NOMA end-to-end system is equivalent to an autoencoder network model, where

[0022] the NOMA end-to-end system includes J users, and each user has a single antenna;

[0023] For any user j ∈ [1, 2, 3... J], with modulation order M j transmits information on K orthogonal resources, and the log2M j binary bits of user j are mapped to a K-dimensional symbol;

[0024] Let x j = [x j1 , x j2 ,...., x jK , in the NOMA system, K < J, and the signals transmitted by J users are non-orthogonally superimposed, which is expressed at the receiving end as:

[0025]

[0026] where h j = [h j1 , h j2 ,..., h jK] is the channel vector between user j and the receiver, n is additive white Gaussian noise, and the receiver recovers the information sent by each user through multi-user detection based on the received signal y;

[0027] For any user j in the system, the entire process from sending information to recovering information is expressed as:

[0028]

[0029] Among them, the function f j (b j ) represents the bit stream b of user j j Mapping to send symbol x j process, function Indicates that user j recovers the transmitted symbol x by receiving the signal j process.

[0030] In some embodiments, Figure 3 As shown in the figure, the NOMA end-to-end system based on the autoencoder model is divided into two parts, the encoder part and the decoder part, which specifically include:

[0031] Construct an autoencoder, which corresponds to the transmitter composed of each user in the NOMA end-to-end system, and the last layer of the encoder corresponds to the communication channel, and the distribution formed satisfies:

[0032]

[0033] In the formula represents the baseband codeword generation function for J users, F Θ (b j ) indicates that user j sends bit b j After bit mapping, a 2K-dimensional real-valued vector is output, where the K dimension is the I-path symbol and the remaining K dimensions are the Q-path symbols. Mean variance Gaussian distribution of;

[0034] The decoder corresponds to the multi-user detection of the receiver. When the receiving end obtains complete channel state information, the decoder network of each user at the receiving end outputs the posterior probability of each user sending a symbol that satisfies:

[0035]

[0036] Among them, b j Indicates the 01 bit corresponding to each user sending symbol, G Φ (y,H) ji represents the posterior probability of the i-th bit sent by user j, H represents the channel matrix, and g jIndicates that user j demaps the received signal into its transmitted bit b j The process of M j represents the modulation order of user j.

[0037] Based on the autoencoder model, a new cost function is constructed based on the information bottleneck theory to train the network, which is different from the cost functions based on mean square error, cross entropy, and minimization of the lower bound of evidence used in current research. The cost function based on the information bottleneck theory in this application takes into account the system overload rate while considering reliability.

[0038] In step S102, according to the constructed NOMA end-to-end system, a cost function for training the autoencoder network model is constructed based on the information bottleneck theory;

[0039] In step S103, the constructed autoencoder network model is trained to complete the NOMA transceiver design.

[0040] The method of this application constructs a new cost function based on the information bottleneck theory. From the perspective of information theory, the joint design process of NOMA transceivers is to find the best function pair. The signal reconstruction error is as small as possible under certain constraints, making the mutual information of the bit stream information sent by all users and the received signal as small as possible. This optimization goal conforms to the original definition of the information bottleneck theory. represents the function for each user to map 01 bits to baseband symbols, Indicates the function of each user demapping the received signal into the transmitted bit. According to the physical meaning of the NOMA system, in some embodiments, the cost function of constructing the autoencoder network model based on the information bottleneck theory includes:

[0041] The original definition based on the information bottleneck theory is described as a multi-user mutual information expression, which solves the following optimization problem:

[0042]

[0043] In formula (3) J represents the K-dimensional codeword reconstructed by user j through the decoder network, y represents the received signal obtained at the receiving end after the codewords sent by each user are superimposed through the channel, and b j (j=1,2,3...J) represents the transmitted bits of user j; I(y; x1,x2,...,x J ) represents the mutual information between the received signal and the reconstructed symbol, I(b1,b2,...,b J; y) represents the mutual information between the transmitted bit stream and the received signal, and β is the Lagrange multiplier, whose size determines the importance of information compression rate in joint optimization. The larger β is, the higher the required information compression rate is, and the higher the overload rate of the NOMA system is.

[0044] The multi-user mutual information expression is used as Figure 3 The cost function of the autoencoder is used to jointly optimize the NOMA system transceiver.

[0045] It is difficult to directly solve the optimization problem formula (3) analytically. In some embodiments, the joint optimization of the NOMA system transceiver includes:

[0046] The lower bound of the multi-user mutual information expression is constructed based on variational inference theory:

[0047]

[0048] The expression on the right side of the inequality sign has an analytical expression. By using the expression on the right side of the maximization to replace the maximization formula (3), it can be further expressed as:

[0049]

[0050] Where z nk for The kth item of ;

[0051] In Rayleigh channel:

[0052]

[0053] In some embodiments, training the constructed autoencoder network model includes:

[0054] The number of users in the specified network is J. Each user selects different modulation orders according to different service requirements. The modulation order of user j is M. j , the number of resources used in the network is K;

[0055] The autoencoder maps the bits sent by the user to the codeword. The codeword generation of each user is realized by K DNNs, corresponding to the baseband symbols of the codeword on K resources. The encoder consists of JK sub-networks in total. In the specific example, the input width is log2M j , there are 3 hidden layers with a width of 64 in the middle, and the output layer has a width of 2, which correspond to the real and imaginary parts of the codeword sent by the user in one of the K resources.

[0056] The activation function of the hidden layer is the ReLU function. The real and imaginary parts of the JK sub-networks are linearly superimposed to obtain a transmission layer with a width of 2K. K of them correspond to the real part of the superimposed codeword, and the other K correspond to the imaginary part of the superimposed codeword. The decoder network contains 3 hidden layers, each with a width of 64, and the activation function is also the ReLU function. The width of the output layer is The activation function of the output layer is the Sigmoid function, and the function value range is in the interval (0,1), which indicates the probability that the bit sent by the user is 1.

[0057] Let the total number of modulation symbols contained in the training set be N, and let M1>M2>L>M J , corresponding to the proportion of each modulation order α1>α2>L>α J , and α1+α2+L+α J =1, each time from α1N, α2N...α J A training sample is formed by extracting one symbol from each of N, and a complete sample set is formed by traversing all cases. Each training sample is a vector composed of 01 bits. The tth sample in the data set can be expressed as For any of b j Contains log2M j Bit.

[0058] The autoencoder is trained round by round in an end-to-end manner. In each round of training, I training samples are randomly selected from the dataset B as a small batch. Send to the encoder network to generate superimposed signal;

[0059] After the superimposed signal passes through the channel, it is input into the decoder network at the receiving end, and the decoder calculates the posterior probability of each user sending a bit and outputs it.

[0060] The actual output and expected output are used as a measure to calculate the average loss function over the entire training data set. or

[0061] Calculate the gradient using stochastic gradient descent and And use the Adam optimizer to update Θ and Φ to perform training.

[0062] The specific training methods are as follows:

[0063]

[0064] The embodiment of the present application further proposes a simulation case of a NOMA transceiver design method based on a multi-user variational information bottleneck:

[0065] The simulation scenario in this example is a typical NOMA uplink access scenario, where J users share K resources. The modulation order of each user is There is no relative mobility between the base station and the user. The encoder network contains J DNNs, and the details of the network structure are given in the following table. When simulating the Rayleigh channel, the fading coefficient of each user channel follows a complex Gaussian distribution. It is also assumed that the base station has performed uplink channel estimation before multi-user detection and obtained accurate uplink channel state information between each user and the base station. The network is trained with accurate channel state information, as shown in the following table:

[0066]

[0067]

[0068] The network training parameter settings are shown in the following table:

[0069]

[0070] Figure 4 The BER performance of different algorithms in AWGN channels is compared. The modulation orders include 2nd, 4th and 8th order modulation. In this simulation, 6 users are still assumed to share 4 resources. Obviously, several other algorithms based on deep learning are better than traditional design methods in bit error rate performance. In particular, the joint design based on information bottleneck theory has a performance gain of about 2dB compared with the traditional design method at a BER of 10-3. Its BER performance is the best among all the algorithms involved in the comparison.

[0071] Figure 5 The BER performance comparison of different optimization methods under different overload rates is further given. For the case of an overload rate of 200%, the number of users J=6 and the number of resources K=3 are set. For the case of an overload rate of 300%, the number of users J=12 and the number of resources K=4 are set, and the Lagrange multiplier β is adjusted to β=1. The modulation order is still 2nd order, 4th order and 8th order modulation. From the performance comparison in the figure, it can be seen that the three joint design methods based on deep learning are superior to the traditional design methods in different overload rate configurations, reflecting the advantages of joint design over single optimization. In addition, by adjusting the Lagrange multiplier β, the autoencoder can achieve a better adaptation to different overload rate requirements. When the overload rate is large, there are more users on the same resource, and the distance between the constellation points is closer without changing the transmission power, resulting in a decrease in BER performance. At this time, by reducing the β value, the constellation points can be expanded from the center to the outside, and a larger distance between the constellation points can be obtained, but at the cost of increasing power. Therefore, the joint design method based on multi-user variational information bottleneck in this application has the best performance among all the algorithms involved in the comparison.

[0072] Figure 6 The bit error performance of the joint design based on variational information bottleneck and several other benchmark algorithms under Rayleigh channel is compared. Among all the algorithms involved in the comparison, the joint design algorithm based on variational information bottleneck of the present application has the best bit error performance.

[0073] This application proposes a joint design method for NOMA transceivers based on multi-user variational information bottleneck. This joint synthesis method takes into account the reliability and overload rate of the system under the deep learning framework. At the same time, it also considers the design requirements of different bandwidth requirements and modulation orders of different users.

[0074] An embodiment of the present application also proposes a NOMA transceiver design system based on a multi-user variational information bottleneck, including a processor and a memory, wherein a computer program is stored on the memory, and when the computer program is executed by the processor, the steps of the NOMA transceiver design method based on a multi-user variational information bottleneck as described above are implemented.

[0075] It should be noted that in the various embodiments of the present application, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0076] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0077] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for a terminal (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in each embodiment of the present application.

[0078] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present application, ordinary technicians in this field can also make many forms without departing from the purpose of the present application and the scope of protection of the claims, all of which are within the protection of the present application.

Claims

1. A NOMA transceiver design method based on multi-user variational information bottleneck, characterized in that: include: The NOMA end-to-end system is equivalent to an autoencoder network model, wherein the autoencoder network model includes an encoder network and a decoder network. The encoder network is equivalent to the transmitter of the NOMA system, realizes symbol-level processing, and maps the user's transmitted bits to K-dimensional codewords; the decoder network is equivalent to the receiver of the NOMA system, realizes multi-user detection, and restores the transmitted bits of each user; According to the constructed NOMA end-to-end system, the cost function for training the autoencoder network model is constructed based on the information bottleneck theory; The constructed autoencoder network model is trained to complete the NOMA transceiver design.

2. The NOMA transceiver design method based on multi-user variational information bottleneck as claimed in claim 1, characterized in that: The NOMA end-to-end system is equivalent to an autoencoder network model where: The NOMA end-to-end system includes J users, each of which has a single antenna; For any user j∈[1,2,3...J], with modulation order M j Sending information on K orthogonal resources, user j's log2M j Binary bits are mapped to a K-dimensional symbol; Let x j = [x j1 , x j2 ,...., x jK , in the NOMA system, K < J, and the signals sent by J users are non - orthogonally superimposed, which is represented at the receiver as: where h j =[h j1 ,h j2 ,...,h jK ] is the channel vector between user j and the receiver, n is additive white Gaussian noise, and the receiver recovers the information sent by each user through multi-user detection based on the received signal y; For any user j in the system, the entire process from sending information to recovering information is expressed as: Among them, the function f j (b j ) represents the bit stream b of user j j Mapping to send symbol x j process, function Indicates that user j recovers the transmitted symbol x by receiving the signal j process.

3. The NOMA transceiver design method based on multi-user variational information bottleneck as claimed in claim 1, characterized in that: Building a NOMA end-to-end system based on the autoencoder model includes: Construct an autoencoder, which corresponds to the transmitter composed of each user in the NOMA end-to-end system, and the last layer of the encoder corresponds to the communication channel, and the distribution formed satisfies: In the formula represents the baseband codeword generation function for J users, F Θ (b j ) indicates that user j sends bit b j After bit mapping, a 2K-dimensional real-valued vector is output, where the K dimension is the I-path symbol and the remaining K dimensions are the Q-path symbols. Mean variance Gaussian distribution of When the receiving end obtains complete channel state information, the decoder network of each user at the receiving end outputs the posterior probability of each user sending a symbol that satisfies: Among them, b j Indicates the 01 bit corresponding to each user sending symbol, G Φ (y,H) ji represents the posterior probability of the i-th bit sent by user j, H represents the channel matrix, and g j Indicates that user j demaps the received signal into its transmitted bit b j The process of M j represents the modulation order of user j.

4. The NOMA transceiver design method based on multi-user variational information bottleneck as claimed in claim 3, characterized in that: The cost function of constructing the autoencoder network model based on the information bottleneck theory includes: The original definition based on information bottleneck theory is described as the multi-user mutual information expression: Among them, x j (j=1,2,...,J) represents the K-dimensional codeword reconstructed by user j through the decoder network, y represents the received signal obtained at the receiving end after the codewords sent by each user are superimposed through the channel, and b j (j=1,2,3...J) represents the transmitted bits of user j; I(y; x1,x2,...,x J ) represents the mutual information between the received signal and the reconstructed symbol, I(b1,b2,...,b J ; y) represents the mutual information between the transmitted bit stream and the received signal, and β is the Lagrange multiplier; The multi-user mutual information expression is used as the cost function of the autoencoder to jointly optimize the NOMA system transceiver.

5. The NOMA transceiver design method based on multi-user variational information bottleneck as claimed in claim 4, characterized in that: Joint optimization of NOMA system transceivers includes: The lower bound of the multi-user mutual information expression is constructed based on variational inference theory: Using the maximum right-hand expression instead of the maximum upper bound, it can be further expressed as: Where z nk for The kth item of ; In Rayleigh channel:

6. The NOMA transceiver design method based on multi-user variational information bottleneck as claimed in claim 5, characterized in that: Training the constructed autoencoder network model includes: The number of users in the specified network is J. Each user selects different modulation orders according to different service requirements. The modulation order of user j is M. j , the number of resources used in the network is K; The autoencoder maps the bits sent by the user to the codewords. The codeword generation for each user is realized by K DNNs, corresponding to the baseband symbols of the codeword on K resources. The encoder consists of JK sub-networks in total. Let the total number of modulation symbols contained in the training set be N, and let M1>M2>L>M J , corresponding to the proportion of each modulation order α1>α2>L>α J , and α1+α2+L+α J =1, each time from α1N, α2N...α J One symbol is extracted from each of N to form a training sample, and all cases are traversed to form a complete sample set; The autoencoder is trained round by round in an end-to-end manner. In each round of training, I training samples are randomly selected from the dataset B as a small batch. Send to the encoder network to generate superimposed signal; After the superimposed signal passes through the channel, it is input into the decoder network at the receiving end, and the decoder calculates the posterior probability of each user sending a bit and outputs it. The actual output and expected output are used as a measure to calculate the average loss function over the entire training data set. or Where Θ is the parameter of the encoder network, and Φ is the parameter of the decoder network; Calculate the gradient using stochastic gradient descent and And use the Adam optimizer to update Θ and Φ to perform training.

7. A NOMA transceiver design system based on multi-user variational information bottleneck, characterized in that: It includes a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the steps of the NOMA transceiver design method based on multi-user variational information bottleneck as described in any one of claims 1 to 6 are implemented.

Images